(Quantitative Geology and Geostatistics 19) J. Jaime Gómez-Hernández, Javier Rodrigo-Ilarri, María Elena Rodrigo-Clavero, Eduardo Cassiraga, José Antonio Vargas-Guzmán (eds.) - Geostatistics Valencia .pdf
(Quantitative Geology and Geostatistics 19) J. Jaime Gómez-Hernández, Javier Rodrigo-Ilarri, María Elena Rodrigo-Clavero, Eduardo Cassiraga, José Antonio Vargas-Guzmán (eds.) - Geostatistics Valencia .pdf
(Quantitative Geology and Geostatistics 19) J. Jaime Gómez-Hernández, Javier Rodrigo-Ilarri, María Elena Rodrigo-Clavero, Eduardo Cassiraga, José Antonio Vargas-Guzmán (eds.) - Geostatistics Valencia .pdf
J. Jaime Gómez-Hernández
Javier Rodrigo-Ilarri
María Elena Rodrigo-Clavero
Eduardo Cassiraga
José Antonio Vargas-Guzmán Editors
Geostatistics
Valencia
2016
Quantitative Geology and Geostatistics
Volume 19
More information about this series at http://www.springer.com/series/6466
J. Jaime Gómez-Hernández • Javier Rodrigo-Ilarri
Marı́a Elena Rodrigo-Clavero • Eduardo Cassiraga
José Antonio Vargas-Guzmán
Editors
v
Acknowledgments
The 10th International Geostatistical Congress would have not been possible without
the financial support of a number of institutions and companies. The Organizing
Committee is very thankful to the Universitat Politècnica de València; the Polytechnic
City of Innovation; the Valencian Regional Ministry of Education, Research, Culture
and Sport (ref. AORG/2016/050); the International Association of Mathematical Geol-
ogy; and Geovariances. Iberia and Geoconnexion are also thanked for their support.
Last, but certainly not least, special thanks are due to the Saudi Arabian Oil
Company (SAUDI ARAMCO) for undertaking the exclusive sponsorhip of this
Geostatistics Valencia 2016 proceedings book by Springer.
vii
Organizing Committee
ix
Scientific Committee
xi
xii Scientific Committee
Part II Theory
Functional Decomposition Kriging for Embedding Stochastic
Anisotropy Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
J.A. Vargas-Guzmán and B. Vargas-Murillo
Can Measurement Errors Be Characterized from Replicates? . . . . . . . 45
Chantal de Fouquet
Modelling Asymmetrical Facies Successions Using Pluri-Gaussian
Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
Thomas Le Blévec, Olivier Dubrule, Cédric M. John, and Gary J. Hampson
Considerations for the Use of Sequential Sampling Techniques . . . . . . . 77
J. Leguijt
A New High-Order, Nonstationary, and Transformation
Invariant Spatial Simulation Approach . . . . . . . . . . . . . . . . . . . . . . . . . 93
Amir Abbas Haji Abolhassani, Roussos Dimitrakopoulos,
and Frank P. Ferrie
A Truly Multivariate Normal Score Transform Based
on Lagrangian Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
Ute Mueller, K. Gerald van den Boogaart, and Raimon Tolosana-Delgado
xiii
xiv Contents
Part V Hydro(geo)logy
Building Piezometric Maps: Contribution of Geostatistical Tools . . . . . 761
B. Bourgine, M. Saltel, N. Pedron, and E. Lavie
A Gradient-Based Blocking Markov Chain Monte Carlo Method
for Stochastic Inverse Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 777
Jianlin Fu, J. Jaime Gómez-Hernández, and Song Du
Geostatistical Modelling and Simulation Scenarios as Optimizing
Tools for Curtain Grouting Design and Construction
at a Dam Foundation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 789
V. Gavinhos and J. Carvalho
xviii Contents
W. Assibey-Bonsu
Abstract This paper provides a write-up of the first Professor Danie Krige memo-
rial lecture in 2014, which was organised by the University of the Witwatersrand in
collaboration with the Southern African Institute of Mining and Metallurgy
(SAIMM) and the Geostatistical Association of Southern Africa, where his wife
Mrs Ansie Krige, the SAIMM and Professor R.C.A. Minnitt also spoke. The
memorial lecture was presented by his previous PhD graduate student, Dr Winfred
Assibey-Bonsu.
During that inaugural memorial lecture, the SAIMM highlighted three activities
that the institute would hold going forward, so as to remember this great
South African mining pioneer:
• The publication of a Danie Krige Commemorative Volume of the SAIMM
Journal
• An annual Danie Krige Memorial Lecture to be facilitated by the School of
Mining Engineering of the University of the Witwatersrand
• The annual award of a Danie Krige medal
What follows is both a tribute to his work and a testimony to the great man’s
deep personal integrity, belief in family, humility and faith in Christ, all of which
led him to become not only a giant in the South African mining industry but indeed
worldwide.
The paper was first published by the Southern African Institute of Mining and Metallurgy during
the Danie Krige Geostatistical Conference, Johannesburg, 2015.
W. Assibey-Bonsu (*)
Gold Fields Limited, Perth Office, PO Box 628, West Perth 6872, Australia
e-mail: [email protected]
1 Introduction
It has been said that ‘we make a living by what we receive, but we make a life by
what we give’. Professor Krige epitomised this in both thought and deed, where he
showed that true success in life does not revolve around material accomplishments
accrued as an individual but is defined by that which one does and leaves for others.
It has been the author’s privilege to have had an association with Professor Krige
for over 20 years, both initially as a student during a doctorate thesis at the
University of the Witwatersrand and later with him as mentor, counsellor and
‘father figure’ for the period that followed.
This paper will cover the two aspects that defined Professor Krige, firstly with
respect to his personal life and career, including the achievements of both, whilst
the second part will briefly touch on his immense contribution to industry and the
world for over half a century, through his pioneering work in ore evaluation,
economics and of course geostatistics. Indeed, his passing was recorded in
Wikipedia under notable persons, a distinction he shared with renowned persons
such as Margaret Thatcher.
This memorial lecture would be incomplete without firstly throwing light on some
of the things Professor Krige held very dear in his life, taken from his interview in
2012 with Professor R.C.A. Minnitt of the University of the Witwatersrand.
Professor Danie Krige’s First Memorial Lecture: The Basic Tenets of. . . 5
Fig. 2 A 1930 photograph of Professor Krige aged 11 ( front middle) with family
Professor Krige was born in Bothaville in the Free State and was the youngest of
nine children born to a pastor.
Figure 2 shows a 1930 photograph of Professor Krige aged 11 with family.
Professor Krige was a devout Christian, who always emphasised that what made
a difference in his life was his belief in Jesus Christ. He also acknowledged that he
had been the recipient of gifts of grace from the Creator – ‘grace given to him’ –
drawing attention to six specific areas, in which he could identify the grace of the
Almighty at work in his life and career:
The first gift of grace:
It was a tribute to his parents for the practical application of a godly lifestyle, the
establishment of a firm foundation and a life philosophy that was modelled by
them in every area of life. An example being that even with the limited resources
at their disposal, they ensured that seven of the nine siblings received a tertiary
education.
The second gift of grace:
The second of the gifts of grace that he acknowledged was the support he had
received from his two spouses. He was happily married for 45 years to his first
wife (until her death) and thereafter for 20 years to Ansie.
The third gift of grace:
The third gift of grace was the way in which his career developed and the various
changes in direction that it took, as his research unfolded.
6 W. Assibey-Bonsu
Professor Krige matriculated from Monument High School at the age of 15 and, in
1938 at the age of 19, graduated as a Mining Engineer from the University of the
Witwatersrand. It was clear early on that he was destined for great achievements.
The two photos in Fig. 6 show the difference between the robe of a university
graduate and typical clothes of an underground miner. It provides a perfect illus-
tration of Professor Krige’s values regarding theoretical developments aimed at
solving practical problems.
Figure 6 shows Professor Krige as a university graduate and also learning the
trade, 1939.
2.2.1 Career
Professor Krige worked with Anglo Transvaal on a number of gold mines in the
Witwatersrand until 1943 and thereafter joined the Government Mining
Professor Danie Krige’s First Memorial Lecture: The Basic Tenets of. . . 7
Fig. 5 Professor Krige celebrating his 90th birthday with wife Ansie
Fig. 6 Shows Professor Krige as a university graduate and also learning the trade
Professor Danie Krige’s First Memorial Lecture: The Basic Tenets of. . . 9
As a professional engineer, Professor Krige served for many years on the mining
committee of the Engineering Council of South Africa and on the Council of the
SAIMM; he was a co-founder of the International Association of Mathematical
Geology, Geostatistical Association of Southern Africa, Geostatistical Association
of Australia and the Statistical Association of South Africa.
He also served as a director of several companies, on the subcommittee of the
South African Prime Minister’s Economic Advisory Council during 1967/1968, as
well as on various committees of the South African Chamber of Mines. He was a
member of the SAMREC Working Committee for The South African Code for
Reporting of Exploration Assets, Mineral Resources and Mineral Reserves
(SAMREC Code) as first published in 2000.
Amongst all of this, he still managed to find time to (i) design the state aid
formula, which assisted a large number of gold mines to survive the period of low
10 W. Assibey-Bonsu
gold prices, (ii) establish the original South African uranium contracts and, (iii) in
1955 and in Afrikaans, published probably one of the first papers on risk analysis
for new mining investment. He also gave major inputs in the fields of financial
analysis and taxation.
Professor Krige was especially committed to the Application of Computers and
Operations Research in the Mineral Industry (APCOM). He was South Africa’s
representative on the International APCOM Council from its inception, served as
the Chairman of the International APCOM Council and was the first member,
outside of the United States, to be elected to this position. He initiated and was
directly involved with all arrangements for the symposia held in South Africa in
1972, 1987 and 2003 and is believed to have attended all APCOM symposia until he
was almost 90 years old. In 2003, 2 weeks after a major operation, he managed to
convince his medical doctors to allow him to attend the 2003 APCOM in Cape
Town, South Africa, where he was a keynote speaker and also presented two other
papers.
During his time as a Professor of Mineral Economics at the University of the
Witwatersrand, he was responsible for postgraduate courses in geostatistics and
mineral economics and supervised many masters and doctoral theses. Both whilst
there and afterwards, he presented courses in geostatistics and/or lectured at local
South African and also international universities in Australia, Germany, Taiwan,
Chile, Russia and China, to name but a few. He also still found the time to provide
valuable consultancy work both locally and internationally and also participated in
and contributed to many international congresses all over the world.
Over his life time, he was the recipient of numerous awards locally and interna-
tionally, too many to mention all. His academic achievements and awards included:
– DSc (Eng.) 1963, University of the Witwatersrand
– DIng (HC) 1981, Honorary Degree, University of Pretoria
– Honorary Doctorate Degree from Moscow State Mining University
– Honorary Doctorate Degree from University of South Africa (UNISA)
– Order of Meritorious Service Class 1, Gold, by South African State President
– The highest award of the SAIMM, the Brigadier Stokes Award, 1984
– Many other merit awards from SAIMM including two gold medals in 1966 and
1980 and two silver medals in 1979 and 1993
– International Association of Mathematical geology – William Krumbein medal,
1984
– One of the highest awards from the American Society of Mining Engineers – the
Daniel Jackling Award
– Several awards from APCOM International Council, including the Distin-
guished Achievement Award, 1989
Professor Danie Krige’s First Memorial Lecture: The Basic Tenets of. . . 11
Although impossible to provide a thorough list in this paper, the author will try to
detail at least some of the many principles that Professor Krige brought forth over
half a century.
Acquision, mine development or expansion cost Acquision, mine development or Expansion cost
(US$ billions) (US$ billions)
8.5
38.1
27.0
2.5 2.3
1.1
Rio Tinto Billiton Gold Fields Ltd Gold Fields Ltd Newmont Barrick
(2007) (2014) (F2007) (2011) (2011) (2012)
strategic objective is to explore, acquire, develop and ultimately mine them, but one
critical risk exists in the uncertainty of the estimation of resources/reserves. If, after
intensive capital investments, it is subsequently found that the expected mineral
resources and mineral reserves were inefficiently estimated or valued, billions of
dollars may be lost. Professor Krige’s pioneering and research work provides
technical solutions to mitigate these technical and financial risks when evaluating
these fundamental assets.
Professor Krige emphasised the critical importance of data integrity as the lifeblood
of mineral resource and reserve evaluation. This includes data validation and
authorisation, use of standards and blanks with approved laboratories and also
database safety and security, which are all critical requirements of the Sarbanes-
Oxley Act of 2002 (SOX) that is necessary for compliance with the New York
Stock Exchange regulations.
In the field of mineral resource and reserve evaluation, geology and geostatistics are
two inseparable sides of the same coin. As highlighted above, on the one side,
geology concentrates on the physical features of the ore body, such as structures,
source, deposition and type of mineralisation. Geostatistics is the other side of the
coin and provides mathematical, statistical and geostatistical models for the
14 W. Assibey-Bonsu
Geostatistics as such did not really originate until the basic concept of ore grades as
a spatial variable, with a spatial structure, was introduced in 1951/1952 by Profes-
sor Krige.
This arose firstly in his endeavour to explain the experience seen on the
South African gold mines for many decades, where ore reserve block estimates
consistently showed significant undervaluation in the lower-grade categories and
overvaluation for estimates in the higher-grade categories, during subsequent
mining, i.e. what is now known as conditional biases, and illustrated below in the
form of a simple diagram (Fig. 10). His pioneering work provided the geostatistical
explanation of conditional biases, as unavoidable errors resulting from the use of
limited data on the periphery of blocks, which were used in evaluating ore reserve
blocks. He proposed and implemented corrective measures to eliminate these
significant conditional biases. The regression corrections were applied routinely
to block estimates on several mines in the early 1950s and represented the actual
birth of kriging. The regressed estimate was, in effect, a weighted average of the
peripheral estimate and the global mean of the mine section; it was the first
application of kriging. It could be called ‘simple elementary kriging’, being based
on the spatial correlation between the peripheral values and the actual grades of the
Professor Danie Krige’s First Memorial Lecture: The Basic Tenets of. . . 15
ore inside the blocks and giving proper weight to the data outside the block
periphery via the mean. In this way, the spatial concept and kriging were intro-
duced. The concept of ‘support’ is very basic to geostatistics and was first covered
by Ross (1950) and further developed by Krige (1951), including Krige’s variance
size of area relationship.
Figure 10 illustrates conditional biases.
Professor Krige’s pioneering work in the early 1950s aroused interest worldwide,
particularly in France where, under Professor Allais, Professor Krige’s papers were
republished in French (Krige 1955). One of Professor Allais students, later to
become world renowned as Professor Matheron, started the development of the
16 W. Assibey-Bonsu
theory of regionalised variables. Matheron also then proposed the use of the
variogram to define the spatial structure. This model is an extension and refinement
of the concept covered by De Wijs (1951/1953). Professor Krige’s regressed
estimates were then still called ‘weighted moving averages’ until Matheron’s
insistence in the middle 1960s on the term ‘kriging’ in recognition of Professor
Krige’s pioneering work.
During 1963 to 1966 (Krige and Ueckermann 1963, 1966), the spatial patterns
were defined in far more detail. These studies covered the spatial correlations
between individual ‘point’ sample values, as well as those between regularised
data blocks. The corresponding correlograms or covariograms were used on a
simple kriging basis for block evaluations. Kriging on a routine basis for ore reserve
evaluation was, therefore, already in use on some Anglovaal gold mines more than
50 years ago.
It is instructive to observe that on the South African gold mines, the improvement in
the standard of block evaluations due to the elimination of conditional biases
accounts for some 70 % of the total level of improvement achievable today, using
the most sophisticated geostatistical techniques. It is for this reason that Professor
Krige placed so much emphasis on the ‘proper’ implementation of the methods to
mitigate conditional biases. Thus, it is critical that the elimination of conditional
biases is not only the major contributor to the reduction of uncertainty in assessing
the mineral resources of mining companies but also an integral and fundamental
part of any kriging and mineral resource and reserve assessment process.
The elimination of these biases is basic to ore evaluation and all geostatistical
procedures. David, in his 1977 popular geostatistics book (‘Geostatistical Ore
Reserve Estimation’), emphasised that the elimination of these biases is basic to
ore evaluation and all geostatistical procedures. As per David, conditional unbi-
asedness is ‘the key point of Krige’s 1951 paper, one of the key points of his 1976
paper but even then, still appeared as a revelation to many people’.
Any increase in knowledge and available data relevant to any uncertainty being
studied will reduce the level of uncertainty, provided the knowledge is applied
properly. Knowledge will never be perfect and data never complete, and therefore
Professor Danie Krige’s First Memorial Lecture: The Basic Tenets of. . . 17
The graphs and tables that follow, some also taken from Professor Krige’s historical
and practical work, illustrate the effect and outcomes of conditional biases.
Figure 11 illustrates feasibility block estimates versus final production blast-hole
averages of an aluminium deposit, showing no correlation between the feasibility
block estimates and that observed during production, as demonstrated by the
regression trend, which could lead to significant risk in invested capital. Figure 12
illustrates similar conditional biases problems, demonstrates why they are impor-
tant and shows how they result in misclassification of ore blocks, which lead to
levels of profit well below what can be achieved. Figure 13 demonstrates the
improved estimates of Fig. 12 analysis that can be achieved through using ‘proper’
kriging with an adequate search.
More recent practical examples of conditional biases are included in the tables
and graphs that follow (Tables 1 and 2 and Fig. 14). Table 1, an open-pit historical
mined-out case study, demonstrates that even the latest sophisticated geostatistics
method used to estimate recoverable resources can suffer from inherent conditional
biases. Table 2 shows the effect of conditional biases over time, from a historically
mined-out case study, with consistently large negative percentage errors for tonnes
and positive ones for grade, over various time periods and cut-offs. Figure 14
illustrates the financial impact the errors would have, over the respective cut-offs
and time periods.
Mineral resource estimation for a new or an existing mine covers two major stages:
– At the initial or first stage, the data is limited and is obtained either from a broad
drill hole grid or from an initial main development grid.
18 W. Assibey-Bonsu
– During the second or final stage, more data becomes available from grade
control or from stope faces and auxiliary developments.
Apart from providing a basis for short- and longer-term mine planning and
viability studies, evaluations are frequently required to provide resource and
reserve classification figures (measured/indicated/inferred and proven/probable)
and to substantiate a major capital investment and/or the raising of loans. At both
stages of evaluation, the evaluation technique should ensure minimum error
Professor Danie Krige’s First Memorial Lecture: The Basic Tenets of. . . 19
where:
BV ¼ block variance (i.e. the variance of actual block values, calculated from a
variogram).
KV ¼ kriging variance (i.e. the error variance of respective block estimates).
For perfect evaluations, KV ¼ 0, the dispersion variance (DV) of the estimates
(calculated from the observed kriged model) ¼ BV, and then:
ðBV 0Þ
Efficiency ¼ ¼ 100% ð2Þ
BV
where only a global estimate of all blocks is practical, all blocks will be valued at
the global mean, i.e.:
However, with conditional biases present, this relationship does not hold, and
then because of insufficient smoothing:
and:
In order to avoid this unacceptable negative efficiency for block estimates, the
following critical control limit test is proposed for the regression slope to test for
conditional biases (Assibey-Bonsu and Muller 2014):
The regression slope can be written as:
where: ξ ¼ þ1 or 1 depending on the sign of LM (ie, the kriging system set-up) and:
LM ¼ respective Lagrange multiplier for ordinary kriging, and BV and KV are as
defined above.
Where only a global estimate of all blocks is practical, all blocks will be valued
at the global or subdomain mean, i.e. KV ¼ BV and Efficiency ¼ 0.
Substituting KV ¼ BV into Eq. 7:
Thus, a regression slope of less than 0.5 will always lead to a negative block
efficiency estimate (i.e. worthless kriged estimates). This highlights the danger of
accepting block estimates that have a slope of regression less than 0.5.
The critical regression slope limit of 0.5 should only be used to identify blocks
that will result with negative kriging efficiencies. Ideal slopes of regression should
be greater than 0.95, as proposed by Krige (Krige 1996).
An extensive study of some 70 cases by Professor Krige covering a wide range
of spatial and data patterns used indicated a correlation between kriging efficiency
and the regression slope (actuals on estimates) of 87.5 % (Krige 1996). Thus, the
slope (or the extent of conditional biases present) effectively incorporates all the
major factors affecting the efficiency of block evaluations.
22 W. Assibey-Bonsu
where σy and σx are the respective standard deviations of actuals and estimates,
respectively, and r is the correlation coefficient.
If the slope is to be unity (i.e. slope ¼ ~1) for unbiased block estimates and (r) is
less than unity, because estimates are never perfect, then:
σ y =σ x > 1 ð10Þ
i.e. the standard deviation (or variance) of the actual or real block values must be
larger than that of the estimated block values. The gap between these two variances
(the smoothing effect) can therefore only be reduced by increasing the correlation
(r) between block estimates and actual values, i.e. by improving the efficiency of
the estimation technique or by providing more data. No mathematical manoeuvring
can achieve this objective.
Various post-processing techniques are available to remove smoothing effects
(e.g. Assibey-Bonsu and Krige 1999; Journel et al. 2000) and should be applied
only to block estimates that are conditionally unbiased.
4 Conclusion
The industry seems to be going backwards in certain areas, with a widely spread
misunderstanding of the causes and consequences of conditional biases. The fol-
lowing are some of the possible causes:
– In certain universities, as well as training provided elsewhere in the industry,
geostatistics is taught using commercially available computer programmes, with
the emphasis being how to use the programmes.
– Unfortunately, this is what many mining companies expect: graduates or prac-
titioners who are good at operating programmes (‘black-box approach’). This
does not give much time to teaching the fundamentals of geostatistics and the
consequences of misusing the technology.
– What complicates matters is that, certain universities rarely have large databases
to demonstrate the strengths and weaknesses of various methods in different
environments and research is by its own nature geared towards only develop-
ment of theoretical geostatistics, often based on strong stationarity assumptions.
. . .after half a century of phenomenal developments in geostatistics, conditional biases
which gave birth to this subject, are still encountered in practical applications. . . the main
concern is that this record will be tarnished by the all too ready acceptance (in certain cases)
of estimates, which are still conditionally biased. For the future, I would like to see
geostatistics continue to grow from strength to strength with new models, techniques and
24 W. Assibey-Bonsu
applications, but where these are all validated properly by way of (follow-up) checks to
confirm the absence of biases and the practical advantages to be gained when they are
applied in practice. . . . Professor Danie Krige.
Professor D. Krige was indeed a pioneering giant, which the South African
mining industry was blessed to have had for his immense contributions. He always
gave willingly and unselfishly, with the rewards not gold, platinum and diamonds
but the tools for others to utilise in finding and evaluating mineral resources, so as to
achieve a positive financial return whilst minimising the associated risk. He took
the industry far along the road, but the journey is not over, and it now remains the
responsibility of those that follow, to both adhere to his principles and indeed
continue to build thereon, to ensure his legacy lives on.
Figure 15 shows Professor Danie Gerhardus Krige, who indeed made the world a
better place.
Acknowledgements The author wishes to thank Gold Fields Limited for the support and time it
has allowed in collating and presenting this paper.
Bibliography
Annual Report (2007) Gold fields limited 2007 annual report. p. 18. 30th June 2007
Assibey-Bonsu W, Krige DG (1999) Use of direct and indirect distributions of selective mining
units for estimation of recoverable resource/reserves for new mining projects. APCOM’99
International Symposium, Colorado School of Mines, Golden, October 1999
Assibey-Bonsu W, Muller C (2014) Limitations in accepting localised conditioning recoverable
resource estimates for medium term, long-term, and feasibility-stage mining projects, partic-
ularly for sections of an ore deposit. J South Afr Inst Min Metall 114:619–624
Bouw B (2011) Newmont to acquire Fronteer Gold in $2.3-billion deal. The Globe and Mail. 3rd
Feb 2011
Professor Danie Krige’s First Memorial Lecture: The Basic Tenets of. . . 25
David M (1977) Geostatistical ore reserve estimation. Elsevier Scientific Publishing Co,
Amsterdam
De Wijs HJ (1951/1953) Statistical ore distribution. Geologie en Mijnbouw. November 1951,
January 1953
Hill L (2013) Barrick to suspend work on Pascua-Lama to conserve cash. Bloomberg. 31 Oct 2013
Jamasmie C (2014) Argentina lobbies to overturn Barrick’s Pascua Lama freeze in Chile. www.
mining.com. 15th Feb 2013.
Journel AG, Kyriadkidis PC, Mao S (2000) Correcting the smoothing effect of estimators: a
spectral postprocessor. Math Geol 32(7):787–813, October 2000
Krige DG (1951) A statistical approach to some basic mine valuation problems on the Witwa-
tersrand. J Chem Metall Min Soc S Afr 52:119–139
Krige DG (1952) A statistical analysis of some of the borehole values in the Orange Free State
goldfield. J Chem Metall Min Soc S Afr 53:47–64
Krige DG (1955) Travaux de M.D.G.KRIGE sur l’evaluation des gisements dans les mines d’or
sudafricaines. Annales des Mines. December 1955
Krige DG (1960) On the departure of ore value distributions from the lognormal model in
South African gold mines. J S Afr Inst Min Metall 61:231–244
Krige DG (1966) Two-dimensional weighted moving average trend surfaces for ore valuations.
Proc. Symp. on Math. Stats. and Comp. Appls. In Ore Valn. pp 13–38. S Afr Inst Min Metall
Krige DG (1978) Lognormal-de Wijsian geostatistics for ore valuation. S Afr Inst Min Metall.
Monograph No. 1. 1978
Krige DG (1996) A practical analysis of the effects of spatial structure and data available and used,
on conditional biases in ordinary kriging. 5th Int. Geostatistics Congress, Wollongong. 1996
Krige DG, Ueckermann HJ (1963) Value contours and improved regression techniques for ore
reserve valuations. JSAfr Inst Min Metall 63:429–452
McCrae MA (2013) The worst mining deal ever: Rio Tinto buying Alcan for US$ 38.1 billion.
www.mining.com. 15th Feb 2013
Minnitt RCA (2012) Thoughts from an interview Professor D. Krige gave to Professor
R.C.A. Minnitt during 2012. Personal correspondence
News.com.au (2012) BHP Billiton’s decision won’t be covered by tax hike says SA Premier Jay
Weatherill. www.news.com.au/finance/business/. 23rd Aug 2012
Ross FWJ (1950) The development of some practical applications of statistical value distribution
theory for the Witwatersrand auriferous deposits. MSc (Eng.) thesis, University of the Witwa-
tersrand, Johannesburg
Sichel HS (1947) An experimental and theoretical investigation of bias error in mine sampling
with special reference to narrow gold reefs. Trans Inst Min Metall Lond 56:403–473
Sichel HS (1952) New methods in the statistical evaluation of mine sampling data. Bull Inst Min
Metall Lond 61:261–288
Sichel HS, Kleingeld WJ, Assibey-Bonsu W (1992) A comparative study of the frequency
distribution models for use in ore valuation. J S Afr Inst Min Metall. April. pp 91–99. 1992
Trustcott SJ (1929) The computation of the probable value of ore reserves from assay results.
Trans Inst Min Metall Lond 39:482–496
Watermeyer GA (1919) Application of the theory of probability in the determination of ore
reserves. J Chem Metall Min Soc S Afr 19:97–107
Part II
Theory
Functional Decomposition Kriging
for Embedding Stochastic Anisotropy
Simulations
1 Introduction
performed by taking the domain to a higher dimensional isotropic space, where the
covariance should yield positive definite matrices, by avoiding non-Euclidean
space (Curriero 2005). Although large computers may accelerate the
multidimensional scaling, the approach requires the input of deterministic ortho-
rhombic variogram anisotropy orientations, which must be known a priori at every
location.
Anisotropy of covariances is predictable using inverse methods. The anisotropy
field can be modified by correlating errors in the expected response attribute with
sensitive modifications in the predicted anisotropy parameters. An example is that
errors in predicted versus measured rock proportions in new drilled wells are used
to enhance knowledge about anisotropy directions (Dossary and Vargas-Guzmán
2011). Reservoir models are reported to be made better by changing the orientation
of geostatistical anisotropy (Caeiro et al. 2015).
Predictions of anisotropy tensors are known in computational fluid dynamics
(CFD), but the results do not match field samples in detail (e.g., storm patterns,
turbulence forecast, and sedimentary dynamics). Turbulent 3D flow models with
physically realistic streamlines are unavailable (Ladyzhenskaya 2003). Therefore,
forecasting components of non-orthogonal anisotropy tensors, with geostatistics
embedding variable anisotropy at non-sampled locations, is a necessary alternative
to vorticity prediction. Additionally, it is essential to embed the uncertainty of
anisotropy in geostatistical predictions, considering the observed parameters for the
directional covariances are spatially variable, uncertain, and usually continuous.
In this study, the spatially stationary single anisotropy constraint is circumvented
by developing a truly functional analysis of the kriging algorithm. The approach is
achieved via functional Hilbert space decompositions of the random field that yield
continuous field analytic functions. Vector components of tensors with uncertain
anisotropy are also predicted as non-orthogonal members. They enable stochastic
parameters for input covariance functions.
2 Theory
A set of random variables fZðxi Þg, each at a given physical location, xi , conforms a
spatial random field, Z ðxÞ, which is defined in the probability space fB; Ω; pg, where
p is the probability measure. The sample space Ω and Borel algebra B conform a
measurable space. Each random variable is a vector Z i in L2 Hilbert space. The
random variables Zi¼q and Z i¼r are related pairwise by inner products Zq ∙ Z r in L2 .
A stationary and deterministic covariance function of a lag distance c ð hÞ
(in general, a stationary co-cumulant function of higher-order κ h1 , . . . hq ) does
not exploit the possibility of spatially variable anisotropy parameters, as observed
in natural phenomena (Vargas-Guzmán 2011). This is due to the variable
32 J.A. Vargas-Guzmán and B. Vargas-Murillo
(e.g., samples and simulated values). This is similar to the Hahn-Banach extension
theorem.
Projection of a vector Ri¼α onto another neighboring vector Ri¼q is an application
of the orthogonal projection theorem. Furthermore, a projection in Hilbert space is
applied to comfortably minimize the residual norm. For example, k R2jα1 ðxα2 Þ k¼
k R1 ðxα2 Þ k wα1 , α2 k R1 ðxα1 Þ k, where R1 ðxα2 Þ is the sampled random variable
from unconditional random field R1 ðxÞ at location xα2 , and wα1 , α2 is the cosine
between the vectors corresponding to variables R1 ðxα1 Þ and R1 ðxα2 Þ. The projections
of R2jα1 ðxα2 Þ on the physical domain generate a new random field, termedR2 ðxÞ,
which avoids any redundancy from a prior field analytic function based on R1 ðxα1 Þ.
In general, a stochastic realization of the functional f αn ðxÞ must optimize
k Rnjα... ðxÞ k.
Remark 1 The projection theorem is generalized as a consecutive process (Fig. 1),
where the component norm residual is from a domino effect, simplified in terms of
uncorrelated components
34 J.A. Vargas-Guzmán and B. Vargas-Murillo
k Rnjα1 αn1 ðxαn Þ k¼k Rn ðxαn Þ k wα1 , αn k R1 ðxα1 Þ k þwα2 , αn k R2jα1 ðxα1 Þ k
ð2Þ
þ wαn1 , αn k Rn1jα1 αn2 ðxαn1 Þ k
0 1
ϑ1 ðxÞ
^z ðxÞ ¼ R_ α1 ðxα1 Þ R_ α n jαn1
ðxαn Þ @ ⋮ A ð3Þ
ϑn ðxÞ
^ i ðxÞ is continuous in the spatial domain x. Note that R_ α jα ðxαi Þ is a new sample
R i i1
residual datum after prior projections have been removed from the datum.
If instead of R_ αi jαi1 ðxαi Þ, the second-order covariance is used in Eq. 5, the result
would be the variance of estimates as a continuous spatial field analytic function
pertaining to one residual datum.
The weights are computed from the standardization of an anchored linear
innovation function, f ̇ijα̌ ðxÞ, as ϑi ðxÞ, for i ¼ 1 . . . n, which represents a
i
nonstationary weight function, and can also be interpreted as a local hat function.
Equation 5 benefits from the fact that the cosine of the angle between vectors in L2 ,
is a bounded function, ϑi ðxÞ, in the spatial domain, x, which has a Fourier transform
but is not stationary. This equation upgrades the estimation approach to an entirely
“functional” method based on large yet simple analytic expressions for weight
functions and estimates.
If the data values are replaced by simulated residuals (i.e., white noise scaled by
the residual norm), then the approach becomes a functional simulation of orthog-
onal nonstationary parametric field analytic functions.
The FDK approach follows the consecutive projection process described in the
previous section. The first functional, anchored at the first sample location, f ̇1jα̌ ðxÞ,
1
is expressed on the physical space or domain x. The subindex 1jðαˇ1 Þ indicates that it
is the first component f 1 functional, and the distances are offsets, e.g.,jx xα̌ j. The
j
first datum is projected onto the entire field as an analytic function. Computing a
residual datum at the second sampling location removes R b 1 ðxÞ from the estimation,
_ _ b
which yields R α2 jα1 ðxα2 Þ ¼ R α2 ðxα2 Þ R 1 ðxα2 Þ: Projecting the new nonredundant
datum R_ α2 jα1 ðxα2 Þ onto the entire field requires an updated weight function. When a
second sample is collected, the Hilbert space L2 would give
̇
f 2jðxjα̌ Þ ðxÞ : Rα2 jα1 ! R2j . A straightforward way to compute this function, as a
2
stepwise optimization, is to anchor a function, f ̇2jðxjα̌ Þ ðxÞ. At this stage, f ̇1jðxjα̌ Þ ðxÞ
2 2
is defined as a hat function at location α2 , which may not be stationary in relation to
f ̇1jðxjα̌ Þ ðxÞ, centered at the first sampled location; however a symmetry condition
1
exists, f ̇1jðxjα̌ Þ ðxα2 Þ ¼ f ̇1jðxjα̌ Þ ðxα1 Þ. Consequently, following Eq. 4, the component
1 2
Functional Decomposition Kriging for Embedding Stochastic Anisotropy Simulations 37
^ ðxÞ ¼ Ṙ
estimate function is R 2 α2 jα1 ðxα2 Þϑ2 ðxÞ. A third sample is assimilated in the
same fashion. Hence, the innovative functional is generalized as
Equation 6 enables all weight functions, ϑn ðxÞ, and field analytic functions R^ ðxÞ,
i
estimated or simulated, to be stored in analytical form, provided that all the required
priors have been previously stored. Computer saving entire analytical models is
analogous to storing the algorithm parameters, but alternatively, a numerical
solution can also be stored at ultrahigh resolution, and it can contain upscaled
values. Nevertheless, updating models is easier if analytic weights are stored. This
analytic type of FDK and functional decomposition simulation (FDS) model can be
automatically updated by inputting new sample coordinates and a new realization
of stochastic covariance parameters. The developer has to decide if the computer
resources required are available to store all the produced field analytic functions or
alternatively target numerical realizations.
A specific analytical expression for the functional requires analytic consistency.
For example, using the exponential model for the covariance, the first function
serves as input in all consecutive functions. It is straightforward to replace second
stage component functionals to obtain third stage functionals and so forth.
FDK process is handled with a program loop in the implementation algorithm.
Note that instead of inputting range parameters a1 ; a2 , and an into the equations, the
distances can be stretched or scaled in rotated coordinates, during numerical display
generation. The transformed coordinates have to remove the anisotropy, as in the
classic orthorhombic case, though separately for each field analytic function. The
monoclinic and triclinic deformations have been handled with Cartesian equivalent
coordinates.
FDK, as introduced for the first time here, is the result of Riesz mapping on
continuous space and/or time after functionals pertaining a random field in normed
Hilbert space are analytically decomposed. FDK generates an explicit spatial
equation for a random field in an nth dimensional domain, which can be inferred
as a complete field analytic function. FDK enables the continuity of the field along
all spatial dimensions including time, but is quantized by the number of discrete
samples and simulated values introduced in the prediction. FDK avoids the use of
matrix algebra or systems of equations; thus, the number of samples that could be
potentially used is infinite, and there are no restrictions regarding their physical
proximity or resolution (i.e., a continuous resolution is theoretically claimed).
Practicality may require a representative elementary volume.
38 J.A. Vargas-Guzmán and B. Vargas-Murillo
An FDK model provides parameters for each conditional probability density func-
tion (PDF) at every spatial location in the continuum spatial domain. Functional
decomposition simulation (FDS) is the continued FDK using simulated input data.
FDS delivers the anisotropy tensor random fields and the scalar members of any
anisotropic scalar, vector, or tensor property. The random field models are used as
input to client algorithms (e.g., nonlinear transformations, flow equations, or
project optimization). There are two ways to conceptualize stochastic models:
1. A stochastic model can be made of PDFs that carry conditional parameters for
each location and attribute. In this case, the client algorithms use FDK predicted
moments as input to deliver response moments. That is a genuine parametric
Functional Decomposition Kriging for Embedding Stochastic Anisotropy Simulations 39
Fig. 3 Functional decomposition indicator kriging (FDIK) model for river channels in Lena
Delta, based on 890 predicted field analytic functions
before updating the conditional functional and performing each Hilbert space
projection. The approach must rotate and stretch the system of coordinates for
each weight function (i.e., to assimilate a new sample). Consequently, the spatial
model is updated with each new component. Note that due to the prior covariance
between two samples is in the initial system of coordinates, it could become
inconsistent, overestimated, or underestimated as the tensor principal components
rotate. Then, the ranges of the covariance must be made properly consistent to
match the required zero covariance with prior samples. This aspect is illustrated in
Fig. 4.
The idea of decomposing a random field can be intuitively extended to each of the
members of the anisotropy tensor, as seen in Fig. 1. Subsequently, the tensor data
representing anisotropy parameter random variables are continuously embedded
Functional Decomposition Kriging for Embedding Stochastic Anisotropy Simulations 41
Fig. 4 Successive anisotropy. (a) Consistent classic. (b) Arbitrary. (c) Ranges consistent with
rotation. (d) FDK anisotropic example for three indicator samples
into the functional for generating a random field pertaining to each attribute of
anisotropy itself, at non-sampled locations. The process creates stochastic func-
tionals that are multivariate because they contain as many parameters as delivered
by the anisotropy tensor at each physical location. Furthermore, the functionals for
simulated data are stochastic because they use covariance parameters from the
simulated anisotropy tensors. Thus FDS is a very powerful approach because it
allows for spatially variable stochastic covariance parameters simulated in the form
of anisotropy tensor random fields.
Remark 5 As aforementioned, the anisotropy can be converted into vectors
describing surfaces or manifolds. In the simple case of a 2D topology, the anisot-
ropy streamlines are converted to normal vectors. These vectors represent the
gradient vector of the attribute. Therefore, one can consider that a new norm
becomes the sum of the norm for the attribute plus the norm for the gradient, this
is, kZðxÞ k þ k∇Z ðxÞk, a Sobolev space. One finding is that FDK with stochastic
anisotropy parameters in second-order covariances is a functional embedding that
can be used to model a field pertaining to a Sobolev space constructed from L2
spaces.
Case in point, uncertain anisotropy tensor members are estimated and simulated
using local orientations and ratios from anisotropy data, consecutively for each
vector or non-orthogonal component. Non-orthorhombic vectors are modeled in the
same fashion as non-orthogonal components (Vargas-Guzmán 2003). The differ-
ence is that simulated anisotropy values are introduced into the stochastic func-
tionals during simulation of each non-orthogonal scalar component.
Structural uncertainty is anisotropic and is represented with random variables in a
tensor random field. In 3D, the three vectors in the tensor require nine cosine directors
42 J.A. Vargas-Guzmán and B. Vargas-Murillo
Fig. 5 Functional decomposition kriging (FDK) model of anisotropy tensors, from 270 stream-
located samples from Lena Delta. Stream vector is a principal tensor component (the color blue in
the background represents streams and water)
and three modules. In the orthorhombic case, with one vector on the vertical modeling
plane, only three cosine directors and two anisotropy ratios are sufficient to handle the
symmetry; the procedure is simplified further for 2D. Hence, each component can be
predicted and simulated in the same way as non-orthogonal attributes with conditional
components (Vargas-Guzmán 2003). However, FDK must be used to account for a
circumvented embedding of anisotropy through covariance functions. For composi-
tional reasons, it is convenient to handle cosine directors as components of unit
vectors. It is not a good practice to estimate angles; to assure resulting vectors have
unit modules and correct orientations, direction cosines must be estimated or simu-
lated. A separate model of scalar anisotropy magnitudes, conditional to the orienta-
tions, is made using measured data and expert inputs or analogs. Modeling of the
non-orthogonal scalar components must be performed with caution, as the anisotropy
range ratios and principal components of anisotropy may conform non-orthorhombic
tensor and components could hold third-order correlations.
A simplified example of FDK in 2D is presented in Fig. 5, where the vector
component of the tensor of anisotropy represents averaged pseudo-streamlines. The
Functional Decomposition Kriging for Embedding Stochastic Anisotropy Simulations 43
Bibliography
Boisvert J, Manchuk J, Deutsch C (2009) Kriging in the presence of locally varying anisotropy
using non-euclidean distances. Math Geosci 41:585–601
Caeiro M, Soares A, Demyanov V (2015) Geostatistical modelling of complex deltaic reservoirs
integrating production data through optimized history matching. Society of Petroleum Engi-
neers, SPE 177813-MS
Curriero F (2005) On the use of non-euclidean distance measures in geostatistics. Math Geol
38:907–926
Dossary AA, Vargas-Guzmán JA (2011) Spatial modeling of complex sandstone bodies to
maximize reservoir contact for wells drilled in clastic formations. GeoArabia J Middle East
Pet Geosci 16:182
Giraldo R, Delicado P, Mateu J (2011) Ordinary kriging for function-valued spatial data. Environ
Ecol Stat 18:411–426
Gutjahr A, Bullard B, Hatch S (1997) General joint conditional simulation using a fast fourier
transform method. Math Geol 29:361–389
Horta A, Caeiro M, Nunes R, Soares A (2010) Simulation of continuous variables at meander
structures: application to contaminated sediments of a lagoon. In: Atkinson P, Lloyd C (eds)
GeoENV VII – Geostatistics for environmental applications. Springer, London, pp 161–172
Ladyzhenskaya O (2003) Sixth problem of the millennium: Navier–Stokes equations, existence
and smoothness. Uspekhi Mat Nauk 58:251
Riesz F (1907) Sur une espèce de géométrie analytique des systèmes de fonctions sommables. C R
Acad Sci 144:1409–1411
Soares A (1990) Geostatistical estimation of orebody geometry: morphology kriging. Math Geol
22:787–802
Stroet C, Snepvangers J (2005) Mapping curvilinear structures with local anisotropy kriging. Math
Geol 37:635–649
van den Boogaart K, Schaeben H (2002) Kriging of regionalized directions, axes and orientations
I. Directions and axes. Math Geol 34:479–503
Vargas-Guzmán JA (2003) Conditional components for simulation of vector random fields. Stoch
Env Res Risk A 17:260–271
Vargas-Guzmán JA (2004) Geostatistics for power models of Gaussian fields. Math Geol
36:307–322
Vargas-Guzmán JA (2011) The kappa model of probability and higher-order rock sequences.
Comput Geosci 15:661–671
Xu W (1996) Conditional curvilinear stochastic simulation using pixel-based algorithms. Math
Geol 28:937–949
Yao T, Calvert C, Jones T, Foreman L, Bishop G (2007) Conditioning geological models to local
continuity azimuth in spectral simulation. Math Geol 39:349–354
Can Measurement Errors Be Characterized
from Replicates?
Chantal de Fouquet
Abstract Sample measurements (of grade, depth, etc.) are almost inevitably
affected by errors. Several error models were studied in the literature. But the
interest of replicates for selecting the error model received limited attention. If
measurement errors are supposed to be additive, homoscedastic, without correlation
between them, and spatially not correlated with the exact values, the variances of
the measurement errors are computable from the sample, simple, and cross-
variograms of replicate data sets, even if the variogram of the exact value is pepitic
(Aldworth W, Spatial prediction, spatial sampling, and measurement error. Retro-
spective Theses and Dissertations. Paper 11842. Iowa State University Digital
Repository @ Iowa State University, 1998; Faucheux et al. Characterisation of a
hydrocarbon polluted soil by an intensive multi-scale sampling. Geostats 2008,
proceedings of the 8th international geostatistics congress, 1–5 Dec. 2008, Santi-
ago, Chile. Ortiz J-M, Emery X (eds) for an example, 2008). But what about the
other cases? When the error is additive, its correlation with the exact value can
remain undetectable. The variance of the measurement errors is thus not always
computable. It’s the same for an error of multiplicative type. Except in some special
cases, keeping the different measurement values rather than their average improves
the precision of the estimation.
1 Introduction
C. de Fouquet (*)
Centre de géosciences, Ecole des mines de Paris, Mines ParisTech, PSL. 35, rue Saint-Honoré,
77305 Fontainebleau, France
e-mail: [email protected]
Models of measurement errors are proposed in the literature and used in the
estimation for a long time. In an instructive paper, Bourgault (1994) examines
several models of additive or multiplicative “noise” which take heteroscedasticity
(the amplitude of the error variance increases with the measured values) and
conditional bias into account. A case study based on a simulated image contami-
nated by noise shows that the performances of factorial kriging to reproduce the
spatial patterns of high signal values fall when the noise is correlated with the
signal. However, the paper does not examine how to select the error model from
the data. The literature remains poor on this practical matter.
In the literature, measurement errors are almost always supposed to be not
correlated with the exact values. Is this usual assumption made for lack of a better
option? To answer this question, three error models are revisited. Their properties
are examined so as to show if the error models can be characterized from
replicated data.
Kriging in the presence of replicates is then briefly examined in a practical
context: how to best use all the data for the estimation? Should replicates be
averaged or introduced individually in the kriging system? It is well known
(Rivoirard et al. 1993) that a model for measurement errors makes the kriging
matrix regular (not singular) when several data are available at the same data points.
On a toy example, the precision of kriging with replicates is compared to that with
their average.
In the following, the exact value is designated by “grade” and the measurement
values of replicates as the “observations.”
a b
BPE Vertical pas=1m 30000 12
250000 10
Variogram : BPEcd
Variogram : A2CPG
200000
20000 17 12
23 14 ab
150000
17 14
a&b
100000 m(ab,cd)36 10000
ab & cd 20
50000 20
1
01
0
0 1 2 3 0 5 10 15 20 25
Distance (m) Distance (m)
Fig. 1 Variogram of the average on two replicate sets (upper curve) and their cross-variogram
(lower curve) for two measurement types and two sampling grids on the same polluted site.
Hydrocarbon grades. (a) Vertical variogram of pyrogram measurements and (b) horizontal
variogram of gas chromatography (Extracted from Faucheux and de Fouquet (2009))
Let us consider a site sampled by soil cores. Each homogenized core constitutes
a sample, of which subsamples are extracted and analyzed. The grade at core
support is modeled by a random function Y(x), supposed to be stationary of order
two (in order to simplify the presentation) with mean (expectation) m and covari-
ance C(h). The grade variogram is denoted γ(h). Z1(x) and Z2(x) represent two
observation sets.
where the random variable Ri(x) denotes the error of the measurement set i at
sample point x and Ti(x) the associated standardized variable. bi is supposed to be
constant for each set and positive.
In practice, the error expectation μi and the coefficients bi (and thus the error
variances) are unknown.
Let us first suppose that measurement errors R1 and R2 are mutually not
correlated and spatially not correlated with grade Y. The variance of the observa-
tions of the set i is
48 C. de Fouquet
The variances of the measurement errors σ 2i are obtained from the difference
between the variance of each observation set and their covariance:
The simple variogram of the observations per set and their cross-variogram are
When the variogram of the observations is pepitic at the scale of the sampled
domain, the previous calculation allows splitting the nugget effect between the
measurement error part σ 2i and the microstructural part, given by the cross-
variogram.
In this model, the average of two observations per data point
1
ZA ðxÞ ¼ ðZ1 ðxÞ þ Z 2 ðxÞÞ ð7Þ
2
b21 þ b22
γ A ð hÞ ¼ þ γ ð hÞ ð8Þ
4
is pepitic:
The variogram analysis shows that this measurement error model is not suited if
the simple variograms of the observations sets Zi or of their average ZA are not
parallel to the cross-variogram γ 12(h) or if the variogram of the difference is not
pepitic.
Rivoirard (1983) has shown that the sample variance is equal to the average of
the variogram cloud (origin included) and thus to the average of variogram points at
all distances, weighted by their pair numbers. Thus, the “statistical” calculation of
the measurement error variance (from sample variance and covariance, Eq. 4) is
equivalent to its “geostatistical” calculation from weighted variogram and cross-
variogram points (Eq. 6). In practice, these two results can slightly differ because
only the first lags of the sample variograms and cross-variogram are computed.
with the same hypotheses on the Ti(x) as previously. The link between grade and
observation
is described by the model with orthogonal residuals (Rivoirard 1994; Chilès and
Delfiner 2012). The special case ai ¼ 0 corresponds to the absence of spatial
correlation between measurement error and grade.
In practice, the error expectation μi and the coefficients ai and bi are unknown.
The covariance between measurement error and grade at same point is
Cov ðRi ðxÞ, Y ðxÞÞ ¼ ai Cð0Þ. Their correlation coefficient (sgn denoting the sign)
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
b2
ρ Y , Ri ¼ sgnðai Þ= 1 þ 2 i ð13Þ
ai Cð0Þ
The expectation and the variance of the observations of set i are, respectively,
EðZ i ðxÞÞ ¼ μi þ m and
is
Var ZL ðxÞ ¼ ð‘1 ð1 þ a1 Þ þ ‘2 ð1 þ a2 ÞÞ2 Cð0Þ þ ‘21 b21 þ ‘22 b22 ð17Þ
When measurement errors and grade are correlated (with ai 6¼ 1), the compo-
nent of the variogram of observations linked to the spatial variability of grades is no
longer equal to the variogram of grades but is proportional to it. In practice, the
proportionality factor is unknown.
The simple variograms of observations and their cross-variogram are no longer
parallel (unless a1 ¼ a2) but are proportional, up to a nugget component on the
simple variograms.
The variogram of the linear combination ZL(x) of observations is
b21 þ b22 ð ð 1 þ a1 Þ þ ð 1 þ a2 Þ Þ 2
γ A ð hÞ ¼ þ γ ð hÞ ð21Þ
4 4
that is exactly the same form as in the absence of correlation (a ¼ 0) but with grade
variogram ð1 þ aÞ2 γ ðhÞ (Eq. 5). These two models are compatible with Fig. 1a.
Remark: when the error is additive and correlated with grade, all variances,
covariances, and simple and cross-variograms of the observations or of any of their
linear combination (Eqs. 14, 15 and 17, 18, 19, 20, 21, 22, and 23) are identical to
those obtained with the covariance C0 (h) and the coefficients ai 0 such that
0 1 þ ai
C0 ðhÞ ¼ k2 CðhÞðk > 0Þand 1 þ a i ¼ ð24Þ
k
Thus, the grade covariance C(h) and the coefficients 1+ai are determined only up
to the factor k2 and 1/k, respectively, but they are not computable individually. As a
consequence, an additional hypothesis on the grade covariance is necessary to write
the kriging system for estimating the grade.
In practice, the coefficients υi and si are unknown, as well as the expectation and
the variance of the random variables Ui(x) and Ti(x) and those of the random
function Y(x).
The expectations of the observation sets are EðZ i ðxÞÞ ¼ miU þ νi m, and their
variances are
Var Z i ðxÞ ¼ Var U i ðxÞ þ Cð0Þ ν2i þ Cð0Þ þ m2 s2i ð26Þ
When the two observation sets have same expectation and same variance, let us
put vi ¼ v and s2 ¼ si 2 . With the additional hypothesis of no additive component
Ui(x) on the error, we obtain
and
Cð0Þ þ m2 s2 ¼ Var Zi ðxÞ Covð Z 1 ðxÞ, Z 2 ðxÞÞ ð29Þ
These expressions are analogous to those obtained in the case of an additive error
correlated with grade: Eqs. 18 and 19 have the same form as Eq. 30. These two
models are compatible with Fig. 1b.
The two cases are impossible to be distinguished in practice: if present, the
structured component is proportional to the simple and cross-variograms. The
difference is that in Eq. 12, the coefficient of proportionality is determinist, while
it is random in Eq. 25.
Can Measurement Errors Be Characterized from Replicates? 53
If the two measurement sets have same mean, the simplest is to assume that the
mean error of each set is zero. If the data sets have different means, additional
hypotheses are needed, for writing the non-biased condition for the estimation.
From now on, the error is supposed to be additive. As previously mentioned,
when the two simple variograms of the observations sets and their cross-variogram
are parallel, it is not possible to determine if the error is spatially correlated with
grade or not.
When the two simple variograms and the cross-variogram present a proportional
(but not parallel) structured component, Eqs. 18 and 19 are a special case of the
linear model of coregionalization, in which the correlation between the components
of Z1(x) and Z2(x) proportional to the grade variogram is maximal (equal to 1).
Indeed, if we note ðh > 0Þγ i ðhÞ ¼ C0i þ C1i γ ðhÞ and γ 12 ðhÞ ¼ C112 γ ðhÞ, then
qffiffiffiffiffiffiffiffiffiffiffiffi
C112 ¼ C11 C12 .
The previous relation can be written differently. Let us denote
1 þ a2
R¼ ð32Þ
1 þ a1
the square root of the ratio of the sills of the structured component (up to the same
factor k which disappears from the ratio). Equations 18 and 19 give
1
γ 1 ðhÞ ¼ b21 þ γ ðhÞ and γ 2 ðhÞ ¼ b22 þ Rγ 12 ðhÞ ð33Þ
R 12
On a practical point of view, it has been known for a long time that the absence
of nugget effect on the cross-variogram of the observations indicates its absence on
the grade variogram. The coefficients b2i can then be derived from the nugget
component of the simple variogram of the observations (Eq. 18). When the grade
variogram presents a microstructural nugget component, this component is present
on the cross-variogram too. The coefficient R can be computed from the structured
part of the simple and cross-variograms and then applied (Eq. 33) in order to
determine the microstructural part of the nugget component on the simple
variograms. The difference from the nugget component of each simple variogram
determines the coefficients b2i .
In practice, when the simple variograms of the observations are parallel and
parallel to their cross-variogram, the simplest model consists in assuming the error
of additive type and without correlation with grade, in order to determine the grade
variogram and the pseudo-cross-variogram between grade and observations. In
particular, this is the case if the simple and cross-variograms of the observations
are pepitic. However, this model does not necessary describe at best the reality.
54 C. de Fouquet
When the structured components of the simple and cross-variograms are pro-
portional, additional hypotheses are needed in order to write the kriging system, as
already mentioned.
The coefficients ai and, up to a factor, the grade variogram being unknown, the
pseudo-cross-variogram between observations and grade
1
gi ðhÞ ¼ EðZi ðxÞ Y ðx þ hÞÞ2
2 ð34Þ
1
¼ μ2i þ b2i þ a2i Cð0Þ þ ð1 þ ai Þγ ðhÞ
2
In the presence of replicates, a measurement error model makes the kriging matrix
regular (i.e., nonsingular). This property is known for quite a long time (Rivoirard
et al. 1993, cited in Aburto 2012; Bourgault 1994) and used, for example, in order to
map a variable surveyed on profiles (geophysics, measurements from ships) on
which the different values do not coincide at their intersections. An efficient model
consists in assuming an acquisition error constant by profile and independent
between profiles.
To precise the influence of the model of measurement error on the estimation, let
us examine the elementary case of the point estimation from only two observations
at the same data point. The errors are assumed to be additive, spatially not
correlated with grade, with mean equal to 0, and with respective variance σ 2i .
The (pseudo)-variogram between observations and that between observation and
grade are
1 1
g12 ð0Þ ¼ EðZ1 ðxÞ Z 2 ðxÞÞ2 gi ðhÞ ¼ EðZi ðxÞ Y ðx þ hÞÞ2
2 and 2 ð35Þ
σ 21 þ σ 22 σ2
¼ ¼ i þ γ ð hÞ
2 2
According to Eq. 5, the point kriging system of Y at x+h from two replicates at
point x is
Can Measurement Errors Be Characterized from Replicates? 55
8
> σ 2 þ σ 22 σ2
< ð1 λÞ 1 þ μ ¼ 1 þ γ ð hÞ
2 2 ð36Þ
: λ σ 1 þ σ 2 þ μ ¼ σ 2 þ γ ð hÞ
> 2 2 2
2 2
The weights of the data with respective measurement error variance σ 21 and σ 22
are λ ¼ 1σ2 and 1 λ ¼ 1σ2 : they depend on the ratio of the measurement error
1þ 1 1þ 2
σ2 σ2
2 1
1
σ 2K ðhÞ ¼ 2γ ðhÞ þ ð37Þ
1
σ 21
þ σ12
2
1 1
Y K ðx þ hÞ ¼ σ 21
Z 1 ðxÞ þ σ2
Z 2 ðxÞ ð38Þ
1þ σ 22
1 þ σ22
1
can be written
1 1
Y K ðx þ hÞ ¼ Y ðxÞ þ σ 21
R1 ðxÞ þ σ2
R 2 ðxÞ ð39Þ
1þ σ 22
1 þ σ22
1
In this model, if the two variances of measurement errors are equal, the kriging
weights are equal, assuming the variogram is the same: it is equivalent to make the
estimation with all the replicates or with their average. But if the measurement error
variances are different, the average of replicate values is generally not optimal.
Indeed the estimation variance of Y(x+h) from the average of the two replicates is
σ 21 þ σ 22
σ2 ðZA ðxÞ, Y ðx þ hÞÞ ¼ þ 2γ ðhÞ ð40Þ
4
The precision gain of kriging in regard to the estimation from the data average is
ðσ21 σ22 Þ
2
4ðσ 1 þσ2 Þ
2 2 which is obviously null if the measurement errors have same variance. At
fixed σ 1 þ σ22 , the precision gain increases with the deviation between σ 21 and σ22 .
2
56 C. de Fouquet
4 Conclusion
Replicates are useful but they do not allow to fully characterizing the measurement
error.
First of all, the usual hypothesis that the error is not spatially correlated with the
grade appears to be somewhat conventional. When the two observation sets have
the same variogram, this hypothesis is the simplest one, but a correlation between
additive error and grade is also possible. Secondly, when the structured component
appears to be proportional on the simple variograms and on the cross-variogram,
the grade variogram is proportional to the cross-variogram, up to an unknown
factor.
If the variance of measurement errors varies according to the set, taking the
average of the replicate values is not optimal for the estimation. Realistic modeling
of measurement errors allows improving the precision of the estimation, using all
replicates.
In the presence of replicates, it is thus better to record all the data with the
measurement characteristics (laboratory, device, date, etc.) rather than only their
average per data point. Indeed, when they are numerous enough, the replicates can
allow specifying a model of measurement errors, even if some indeterminations
remain. Of course, if available, the precision given by the laboratory or the device
characteristics should be considered for building the error model.
Acknowledgment The author thanks cordially Hélène Beucher and the reviewers for their
attentive reviews and suggestions for improving the paper.
Bibliography
Chilès J-P, Delfiner P (2012) Geostatistics: modeling spatial uncertainty, 2nd edn. Wiley,
New York
Faucheux C, de Fouquet C (2009) Interprétation géostatistique des analyses CPG et Pollut-Eval du
site 2. PRECODD LOQUAS, 9ème rapport d’avancement (D20). R2009-345CFCF. Rapport
d’étude. 72p
Faucheux C, Lefebvre E, de Fouquet C, Benoit Y, Fricaudet B, Carpentier C, Gourry J-C (2008)
Characterisation of a hydrocarbon polluted soil by an intensive multi-scale sampling. Geostats
2008, proceedings of the 8th international geostatistics congress, 1–5 Dec 2008, Santiago,
Chile. Ortiz J-M, Emery X (eds)
Rivoirard J (1983) Remarques pratiques a propos des variances et du variogramme. Note de cours
C-75. Ecole des mines de Paris, Fontainebleau
Rivoirard J (1994) Introduction to disjunctive kriging and non linear geostatistics. Oxford Uni-
versity Press, Oxford
Rivoirard J, Renard D, Léger M (1993) Quantification de l’incertitude sur les profondeurs estimées
par sismique et par puits (rapport FSH). Ecole des mines de Paris, Fontainebleau, 14p
Modelling Asymmetrical Facies Successions
Using Pluri-Gaussian Simulations
1 Introduction
While it is easy to constrain the models with the proportion and autocovariance
of each facies (Alabert 1989; Armstrong et al. 2011), it is more complex to model
the cross-indicator covariances between facies. For instance, SIS (sequence indi-
cator simulation) by modelling every facies independently (Alabert 1989) does not
reproduce cross-covariances between different facies, possibly resulting in
non-realistic geological models.
With the aim of modelling spatial relationships between different facies, Carle
and Fogg (1996) constrain cross-covariances using the parameters of a continuous-
time Markov chain. An important outcome of their method is the possibility to
model spatial asymmetry between the indicator variables. The probability of facies
A to be on top of facies B can be different from that of facies A being under B. Such
asymmetrical vertical stacking patterns of facies are common in the stratigraphic
record as sedimentological processes tend to create and preserve shallowing-
upward facies successions which are asymmetric (Burgess et al. 2001; Grotzinger
1986; Strasser 1988; Tucker 1985). However, the model used by Carle and Fogg
(1996) is memoryless and so prevents from using a hole-effect covariance and
reproducing cyclicity, which is another common feature of vertical facies succes-
sions (Burgess et al. 2001; Fischer 1964; Goldhammer et al. 1990; Grotzinger 1986;
Masetti et al. 1991). Another approach uses non-parametric indicator variograms
for bivariate probabilities to simulate facies with asymmetrical patterns (Allard
et al. 2011; D’Or et al. 2008). The approach presented in the current paper aims to
use parametric auto- and cross-covariance models that are “realizable”, that is
associated with valid random set models (Chilès and Delfiner 2012).
Pluri-Gaussian simulations (PGS) can handle facies interactions thanks to the
use of underlying continuous gaussian variables and truncation rules defining facies
ordering and geometries (Armstrong et al. 2011). Moreover, by construction, the
PGS formalism leads to a general cross-covariance model between facies that is
realizable (Chilès and Delfiner 2012). Developing a flexible multivariate gaussian
framework allows to increase the range of facies patterns. For instance, the original
linear model of co-regionalization (Wackernagel 2013), applied to the underlying
gaussian functions, provides flexibility in the resulting facies thicknesses and
distributions. However, the cross-correlations between the underlying gaussian
functions are symmetrical and so are the facies relations.
To overcome this limitation, some authors have proposed to use spatial shifts to
transform the cross-covariances between gaussian functions (Apanasovich and
Genton 2010; Li and Zhang 2011; Oliver 2003). Armstrong et al. (2011) proposed
to use a similar approach when defining the linear model of co-regionalization of
the underlying gaussian variables. Although it is natural to expect that an asym-
metrical cross-correlation between the gaussian functions should lead to asymmet-
rical relations between facies, this approach has not yet, to our knowledge, been
fully developed and tested. Moreover, the relation between the spatial shift, the
correlation and the facies asymmetry has not been studied explicitly.
In this article, we expand on the previous work described above to demonstrate
that a spatial shift applied to the underlying gaussian functions can be used to create
asymmetries in the vertical stacking of facies. The sensitivity of vertical facies
Modelling Asymmetrical Facies Successions Using Pluri-Gaussian Simulations 61
2 Methodology
We focus here on a simple example with three facies. The truncation rule that
defines the contacts between facies and their proportion, relative to their area, can
be drawn as follows (Fig. 1):
If I1, I2 and I3 are the indicators of the three facies, the truncation rule defines
them as follows for every location x on a vertical section:
1, Z 1 ðxÞ < t1
I 1 ðxÞ ¼ ð1Þ
0, else
1, Z 1 ðxÞ > t1 , Z2 ðxÞ > t2
I 2 ðxÞ ¼ ð2Þ
0, else
1, Z 1 ðxÞ > t1 , Z2 ðxÞ < t2
I 3 ðxÞ ¼ ð3Þ
0, else
When the indicator of a facies equals 1, the corresponding facies is present at the
location x. The marginal gaussian cumulative function G applied to each gaussian
function Z1 and Z2 allows to have a truncation rule on which the area of a facies
equals its proportion. However, if there is a correlation between the two functions, it
affects the proportion as the points tend to be located along the transformation of
the correlation line ρ (Fig. 1) which is plotted in the axes (G(Z1),G(Z2)) and thus has
for equation
Y ¼ G ρ*G1 ½X ð4Þ
G(t1)
1.0
Facies 2
0.8
Facies 1
G(Z2) 0.6
0.4
0.2 G(t2)
Facies 3
0.0
Fig. 1 Truncation rule defining three facies with two gaussian random functions Z1 and Z2. t1 and
t2 are the truncations associated with each gaussian functions and G is the gaussian cumulative
function. The red curve is defined by Eq. 4, with the correlation ρ ¼ 0.7. One thousand random
generations with a correlation ρ ¼ 0.7 are performed thanks to the R package MASS (Venables and
Ripley 2002) and displayed
rule could be obtained by applying the bi-normal gaussian cumulative function with
correlation ρ on Z1 and Z2, but its analytical expression is not known.
The truncation rule does not contain spatial information and so does not control
asymmetries. As the aim of this study is to model asymmetrical relations, the
transition probability from one facies i to another j should be different in opposite
directions h and h:
P I i ðxÞ ¼ 1, I j ðx þ hÞ ¼ 1
tij ðhÞ ¼ 6¼ tij ðhÞ ð5Þ
P½ I i ð x Þ ¼ 1
Understanding the link between the facies transition probabilities and the parame-
ters of the underlying bi-gaussian function would help in inferring a pluri-Gaussian
model resulting in the correct asymmetrical transition probabilities. Armstrong
et al. (2011) show that the covariance of the facies indicator can be expressed as
a multivariable integral of the underlying bi-gaussian density. For instance, the
non-centred cross-covariance, between facies 2 and 3, C23(h), is defined as
Modelling Asymmetrical Facies Successions Using Pluri-Gaussian Simulations 63
C23 ðhÞ ¼ P½Z 1 ðxÞ > t1 , Z 2 ðxÞ > t2 , Z 1 ðx þ hÞ > t1 , Z 2 ðx þ hÞ < t2 ð7Þ
This is the joint probability of four gaussian events with their dependence
described by the correlation matrix:
0 1
1 ρ ρZ1 ðhÞ ρZ1 Z2 ðhÞ
X B ρ 1 ρZ1 Z2 ðhÞ ρZ2 ðhÞ C
ð hÞ ¼ B
@ ρZ ðhÞ
C
A ð8Þ
1
ρZ1 Z2 ðhÞ 1 ρ
ρZ1 Z2 ðhÞ ρZ 2 ð h Þ ρ 1
As we work with three facies (Fig. 1), the covariance between facies 1 and facies
2 is expressed by a triple integral, while a double integral defines the
autocovariance of facies 1.
The spatial shift, a, is the distance at which the correlation between the two
gaussian functions Z1 and Z2 is maximal, and ρ is the correlation between the two
simulated gaussian functions Z1 and Z2 at the same location. We can directly deduce
from the square root term in Eq. 10 the condition of validity of the model:
64 T. Le Blévec et al.
This condition originally results from the fact that the variance of the gaussian
functions Z1 and Z2 is one. It is now possible to relate the covariances ρZ1 and ρZ2 of
the gaussian fields Z1 and Z2 to the covariances of Y1 and Y2:
8 " #
< ρZ 1 ð hÞ ¼ ρY 1 ð hÞ ρ2
ρ 2
ρY 1 ð hÞ þ 1 ρY 2 ð hÞ ð12Þ
: ρZ 2 ð hÞ ¼ 2 ρY 1 ð aÞ 2
ρY 1 ð aÞ
3 Results
In this section, we study the indicator transiograms derived from the shifted linear
model of co-regionalization applied with PGS and with the truncation rule in Fig. 1.
We first express the analytical expressions for a special case and then develop a
sensitivity study in the general case thanks to numerical gaussian integrations.
Gaussian variograms for the gaussian functions are used in order to have a linear
behaviour at the origin on the indicator transiograms.
With the truncation rule used in Fig. 1, the transition probability between facies
1 and 2 can be written as a triple integral. Its analytical expression, developed in the
appendix (Eqs. 25 and 26), is the following:
1 1 ρρZ1 ðjh þ ajÞ
t12 ðhÞ ¼ þ arccos þ arccos ρZ1 ðhÞ þ arcsinðρÞ ð15Þ
4 2π ρZ 1 ð aÞ
Therefore, the shift a and the correlation ρ must be non-zero to bring asymmetry
(Fig. 2). We can also deduce the relation:
which means that changing the sign of the shift allows the asymmetry between the
two facies to be switched.
We can see that if the correlation and shift are positive, and the transition
probability tends towards a facies with low proportions, the curve has a very high
concavity with a maximum before the range (Fig. 2, right). If the correlation is
negative and the transition probability tends towards a facies with high proportion,
the curve has an inflexion point (Fig. 2, left). In the opposite direction, the
behaviour is always different, highlighting the asymmetry. If there is no shift,
there is no asymmetry (Fig. 2).
66 T. Le Blévec et al.
Fig. 2 Influence of a positive shift on the transition probabilities from facies 1 to facies 2 with
different values of the proportion P2 of facies 2. The coefficient ρ is either 0.8 or 0.8. The
gaussian function has a gaussian variogram with range 8 (practical range ¼ 13.85) and the shift is
3. The upward and downward transitions are deduced from Eq. 15, such as the dotted line obtained
with a shift equal to 0, and the black tangents are obtained from Eqs. 17 to 18
The frequency of contacts between two given facies can be derived from the
derivative of the cross-transition probability at the origin, which is the rate of
transition from one facies to the other per unit length. We can express the rate of
transition upward T12+ and downward T12 in the case of a gaussian variogram by
differentiating Eq. 15:
"pffiffiffi #
1 2 2 ρ
Tþ 0
12 ða; ρ; a1 Þ ¼ limh!0 t12 ðhÞ ¼ þ a pffiffiffiffiffiffiffiffiffiffiffiffiffi ð17Þ
2π a1 a21 1 ρ2
"pffiffiffi #
0 1 2 2ρ
T 12 ða; ρ; a1 Þ ¼ limh!0 t12 ðhÞ ¼ a pffiffiffiffiffiffiffiffiffiffiffiffiffi ð18Þ
2π a1 a21 1 ρ2
From these equations, it is clear that if the correlation ρ and the shift a are
positive, the probability of having facies 2 on top of facies 1 is higher than of having
facies 1 on top of facies 2. It can be interesting to see for which shift the transition
rate is maximal; let’s take
Modelling Asymmetrical Facies Successions Using Pluri-Gaussian Simulations 67
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2ð 1 ρ2 Þ
alim ¼ a1 ð19Þ
2ρ
With this shift, facies 1 cannot make a transition to facies 2 going downwards as
the transition rate is 0. For the upward transition, it can be noticed that the
expression of the transition rate is the inverse of the mean length of facies
1 (Lantuéjoul 2002). This implies that the upward transition rate from facies 1 to
facies 3 is zero with the closing relations of the transition rate matrix Q:
0 1
1=L1 1=L1 0
Q¼@ 0 1=L2 1=L2 A ð21Þ
1=L3 0 1=L3
with Li as the mean lengths of the different facies. Therefore, this shift gives the
maximum of asymmetry and allows to build perfect geologic asymmetrical
sequences. However, the shift is also bounded by Eq. 11, and consequently
Eqs. 19, 20 and 21 are not possible. As the transition rates increase linearly with
the shift, the maximum of asymmetry is obtained for the higher shift which is the
following according to Eq. 11:
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
amax ¼ a1 logðρÞ ð22Þ
It can be noted that the expressions of alim and amax converge to each other when
ρ tends to one. Thus, for a correlation that tends to one, amax gives upward and
downward transition rates that tend, respectively, to 1/Li and 0, allowing to create
perfect asymmetrical sequences (Eq. 21). This limit case can also be obtained by
simulating only one gaussian function and use the shifted equivalent as the second
gaussian function.
The expressions of the multi-gaussian integrals have allowed asymmetries for a
truncation rule with cut-off at 0 to be analytically expressed. Lantuéjoul (2002)
gives a solution for a general truncation rule when the correlation tends to 1. This
might allow development of more general expressions with thresholds.
The gaussian integral cannot be computed analytically in the general case with
cut-offs different from 0. However, it can be computed numerically (Genz 1992)
68 T. Le Blévec et al.
Fig. 3 Comparison of the impact of the correlation and the shift on the transition probability from
facies 2 to facies 1 upwards. The step for the black curves is 0.1 for the correlation (left) and 0.3 for
the shift step (right). The range of the first gaussian variogram is 8, the proportion of facies 1 is 0.3
and facies 2 is 0.4
using a code available on R (Genz et al. 2009; Renard et al. 2015). Consequently,
we have access to all the transition probabilities, and the correlation ρ can be
changed while keeping the proportions constant which is not possible analytically.
This is carried out by minimizing an objective function quantifying the differ-
ence between the targeted and simulated proportions computed with the gaussian
numerical integral (Genz 1992). It can also be done with a maximum likelihood
estimation of the target proportions by generating random correlated gaussian
values. Understanding the impacts of the correlation and the shift at constant
proportions is important for manually fitting transition probabilities (Fig. 3).
We can see in Fig. 3 that both the correlation and the shift have an impact on the
tangent at the origin which provides a flexibility to match the asymmetry between
facies contacts. The asymmetrical limit behaviour alim (Eq. 19) seems to have been
reached with ρ ¼ 0.8 and a ¼ 3 as the transition rate is close to 0 for these values.
The two parameters also affect differently the curvature of the transition probability
increasing the flexibility of the method.
Modelling Asymmetrical Facies Successions Using Pluri-Gaussian Simulations 69
4 Case Study
This section presents a case study for illustrating the method described earlier with
three facies and the truncation rule of Fig. 1. The geostatistical package RGeostats
is used for the simulation (Renard et al. 2015). The transiograms are studied for two
facies as the relation with the third can be automatically deduced from them.
The transition probabilities of Fig. 5 were derived from the data shown in Fig. 4.
They can be fitted with the shifted linear model of co-regionalization manually
through a trial-and-error process, by maximum likelihood estimation or by mini-
mizing an objective function. In a more general context, a manual procedure is
preferred as transiogram modelling is a step where geological conceptual knowl-
edge can be incorporated. Therefore, we choose to fit manually the transition
probabilities of Fig. 5.
As seen in Fig. 5, the model fitted by trial and error honours the tangent at the
origin of the transition probabilities. This means that the transition rates are well
constrained. Moreover, a possible hole-effect is observed in the experimental
transition probabilities due to a low variance in the facies thicknesses. This effect
cannot be modelled with the current model, but a hole-effect variogram on the
gaussian function should be able to model it (Dubrule 2016).
70 T. Le Blévec et al.
Fig. 4 Comparison between the vertical section of the Latemar section reported by Egenhoff et al.
(1999) and simulations with asymmetrical pluri-Gaussian simulations. The parameters for the
simulation are the same as described in Fig. 5
Fig. 5 Match between experimental transition probabilities (red) observed in Fig. 4 and the model
(blue). Facies 1 is subtidal and facies 2 intertidal. The parameters used for the model are 0.9 for the
range of the first gaussian, 0.52 for the range of the second gaussian, 0.13 for the shift and 0.8 for
the correlation
Modelling Asymmetrical Facies Successions Using Pluri-Gaussian Simulations 71
Fig. 6 Comparison between transition probabilities model (blue) and simulated (grey) and mean
of the simulated (red) on 50 simulations of the Latemar section presented in Fig. 4. Facies 1 is
subtidal, and facies 2 is intertidal
We build two gaussian fields, and then we apply the transformations described in
Eq. 10 on three simulations (see Fig. 4). The asymmetry is still preserved in the
simulations, with supratidal facies always on top of the intertidal facies and the
intertidal facies on top of the subtidal facies. However, the limit shift alim (Eq. 19)
has not been reached as the probability of having subtidal on top of supratidal is not
1, which is also observed on the data. To go further in the simulation analysis, the
experimental transiograms are computed on 50 simulated sections and compared to
the model variogram (Fig. 6).
72 T. Le Blévec et al.
This Monte Carlo study shows that the simulated transition probabilities seem to
match the model well at the origin and for other distances as the mean transiogram
of the simulations matches with the transiogram model (Fig. 6).
This study has shown that the shifted linear model of co-regionalization seems well-
suited to model facies transitions asymmetries using PGS. For the case of modelling
three facies, the first two gaussian variograms allow to define facies mean thick-
nesses, while the shift and the correlation determine the asymmetrical patterns.
Therefore, every transition rate of the transiogram matrix can be inferred indepen-
dently making the method very flexible. Moreover, we saw analytically and numer-
ically that the maximum rate of transitions could be reached asymptotically, which
allows to build perfect asymmetrical sequences.
More precisely, the gaussian integral allows to fix the transition rates as with a
Markov process (Carle and Fogg 1996). However, if the number of facies is
increased, it would be more difficult to respect the different asymmetries, and
manual fitting of the different transition probabilities would be more complex.
Automatic procedures such as maximum likelihood estimations might address
that issue.
The advantage of PGS over continuous-time Markov chains is that it provides a
framework in which the resulting indicator variograms are automatically valid but
also quite flexible. Beyond just transition rates, the parametrical covariances can
lead to linear or fractal behaviour of the indicator variogram at the origin (Chilès
and Delfiner 2012; Dubrule 2016). Other models than the linear model of
co-regionalization would allow to select different behaviours for every facies. For
instance, the multivariate Matern model would allow cross-transition probabilities
to have different smoothness parameters for every facies (Gneiting et al. 2012), and
the spatial shift could be applied to it (Li and Zhang 2011) which would also result
in facies asymmetries. This is currently investigated by the authors.
Acknowledgements The authors would like to thank the Earth Science and Engineering Depart-
ment of Imperial College for a PhD studentship grant for T. Le Blévec and Total for funding
O. Dubrule professorship at Imperial College.
location x and facies 2 at location x+h correspond to one variate being negative and
two positive. The indicator covariance C12(h) quantifies the probability of the
intersection of these three events. The correlation matrix between the three gaussian
variates is the following:
0 1
ρρZ1 ðjh þ ajÞ
1 ρZ 1 ð h Þ
X B ρZ1 ðaÞ C
B C
ð hÞ ¼ B
B ρZ1 ðhÞ 1 ρ C
C ð23Þ
@ ρρZ ðjh þ ajÞ A
1
ρ 1
ρZ1 ðaÞ
C12 ðhÞ
Z 0 Z þ1 Z 0 Z þ1
1
¼ 1 3*0:5 þ gρ Z ðhÞ ðu; vÞdudv þ gρρZ1 ðjhþajÞ ðu; vÞdudv
2 1 0 1
1 0 ρZ ð a Þ
1
Z Z !
þ1 þ1
þ gρ ðu; vÞdudv ð25Þ
0 0
Sheppard (1899) gives then the solution of the double integral that allows to obtain
the final expression of the transition probability between facies 1 and 2 (Eq. 15):
Z þ1 Z þ1 Z þ1 Z þ1
1 1 1
gρ ðu; vÞdudv ¼ gρ ðu; vÞdudv ¼ þ arcsinðρÞ
0 0 2 0 0 4 2π
ð26Þ
74 T. Le Blévec et al.
Bibliography
Alabert F (1989) Non‐Gaussian data expansion in the earth sciences. Terra Nova 1(2):123–134
Allard D, D’Or D, Froidevaux R (2011) An efficient maximum entropy approach for categorical
variable prediction. Eur J Soil Sci 62(3):381–393
Apanasovich TV, Genton MG (2010) Cross-covariance functions for multivariate random fields
based on latent dimensions. Biometrika 97(1):15–30
Armstrong M, Galli A, Beucher H, Loc’h G, Renard D, Doligez B, . . . Geffroy F (2011)
Plurigaussian simulations in geosciences. Springer Science & Business Media, New York
Burgess P, Wright V, Emery D (2001) Numerical forward modelling of peritidal carbonate
parasequence development: implications for outcrop interpretation. Basin Res 13(1):1–16
Carle SF, Fogg GE (1996) Transition probability-based indicator geostatistics. Math Geol 28
(4):453–476
Catuneanu O, Galloway WE, Kendall CGSC, Miall AD, Posamentier HW, Strasser A, Tucker ME
(2011) Sequence stratigraphy: methodology and nomenclature. Newslett Stratigr 44
(3):173–245
Chilès J-P, Delfiner P (2012) Geostatistics: modeling spatial uncertainty, vol 713. Wiley, Hoboken
D’Or D, Allard D, Biver P, Froidevaux R, Walgenwitz A (2008) Simulating categorical random
fields using the multinomial regression approach. Paper presented at the Geostats 2008—
proceedings of the eighth international geostatistics congress
Dubrule O (2016) Indicator variogram models – do we have much choice?. Manuscript submitted
for publication
Egenhoff SO, Peterhänsel A, Bechstädt T, Zühlke R, Gr€otsch J (1999) Facies architecture of an
isolated carbonate platform: tracing the cycles of the Latemar (Middle Triassic, northern Italy).
Sedimentology 46(5):893–912
Fischer AG (1964) The Lofer cyclothems of the alpine Triassic. Princeton University, Princeton
Genz A (1992) Numerical computation of multivariate normal probabilities. J Comput Graph Stat
1(2):141–149
Genz A, Bretz F, Miwa T, Mi X, Leisch F, Scheipl F, Hothorn T (2009) mvtnorm: multivariate
normal and t distributions. R package version 0.9-8. URL http://CRAN.R-project.org/
package¼mvtnorm
Gneiting T, Kleiber W, Schlather M (2012) Matérn cross-covariance functions for multivariate
random fields. J Am Stat Assoc 105(491):1167–1177
Goldhammer R, Dunn P, Hardie L (1990) Depositional cycles, composite sea-level changes, cycle
stacking patterns, and the hierarchy of stratigraphic forcing: examples from Alpine Triassic
platform carbonates. Geol Soc Am Bull 102(5):535–562
Grotzinger JP (1986) Cyclicity and paleoenvironmental dynamics, Rocknest platform, northwest
Canada. Geol Soc Am Bull 97(10):1208–1231
Kendall M, Stuart A, Ord J (1994) Vol. 1: Distribution theory. Arnold, London
Lantuéjoul C (2002) Geostatistical simulation: models and algorithms. Springer Science &
Business Media, Berlin
Li B, Zhang H (2011) An approach to modeling asymmetric multivariate spatial covariance
structures. J Multivar Anal 102(10):1445–1453
Masetti D, Neri C, Bosellini A (1991) Deep-water asymmetric cycles and progradation of
carbonate platforms governed by high-frequency eustatic oscillations (Triassic of the Dolo-
mites, Italy). Geology 19(4):336–339
Oliver DS (2003) Gaussian cosimulation: modelling of the cross-covariance. Math Geol 35
(6):681–698
Renard D, Bez N, Desassis N, Beucher H, Ors F, Laporte F (2015) RGeostats: the geostatistical
package (Version 11.0.1). Retrieved from http://cg.ensmp.fr/rgeostats
Sena CM, John CM (2013) Impact of dynamic sedimentation on facies heterogeneities in Lower
Cretaceous peritidal deposits of central east Oman. Sedimentology 60(5):1156–1183
Modelling Asymmetrical Facies Successions Using Pluri-Gaussian Simulations 75
Sheppard W (1899) On the application of the theory of error to cases of normal distribution and
normal correlation. Philos Trans R Soc London Ser A Containing Pap Math Phys Charact
192:101–531
Strasser A (1988) Shallowing‐upward sequences in Purbeckian peritidal carbonates (lowermost
Cretaceous, Swiss and French Jura Mountains). Sedimentology 35(3):369–383
Tucker M (1985) Shallow-marine carbonate facies and facies models. Geol Soc London Spec Publ
18(1):147–169
Venables WN, Ripley BD (2002) Modern applied statistics with S, 4th edn. Springer, New York
Wackernagel H (2013) Multivariate geostatistics: an introduction with applications. Springer
Science & Business Media, Berlin
Considerations for the Use of Sequential
Sampling Techniques
J. Leguijt
1 Introduction
J. Leguijt (*)
Shell International, Kessler Park 1, 2288 GS Rijswijk, The Netherlands
e-mail: [email protected]
1. The algorithm is used to generate geostatistical models that are not constrained
by measurements.
2. The algorithm is used to generate models that are only constrained by direct
perfect measurements of the properties that are modelled.
3. The algorithm is used in a situation where the measurements are contaminated
with noise.
First, an analysis of the sequential sampling method will be presented, using a
model that only consists of two locations. Second, a more realistic example will be
presented using a grid of 101 101 locations.
An alternative for the sequential sampling method will be presented. This is able
to properly sample the posterior distribution that arises from the assimilation of
measurements that are contaminated with noise.
pð v 2 j v 1 ; d 2 ; d 1 Þ ¼ p ð v 2 j v 1 ; d 2 Þ ð5Þ
When the value of v1 is known, a measurement, d1, thereof, does not add any extra
information and can be skipped from the list of conditions. It would be convenient if
the decomposition in Eq. 6 was valid, since then, Eq. 7 would be valid:
d 1 ¼ v1 þ n1 and
ð8Þ
d 2 ¼ v 2 þ n2
It is assumed that the noise at location 1 is independent of the noise at location 2 and
that the values of v1 and v2 are dependent on each other.
The pdfs that describe the probabilistic behaviour of the noise realisations are
Gaussian distributions with an expectation value of 0.0 and a standard deviation of
2.0. The value of d1 is 10.0 and the value of d2 is 10.0.
The pdf that describes the probabilistic behaviour of v1 and v2 is a two-variable
Gaussian distribution. The parameters of which are specified in Table 1.
The prior pdf is updated with the measurements using Bayes’ rule, resulting in a
posterior distribution. This is a Gaussian distribution from which the expectation
values and the covariance matrix can be computed exactly. Samples from this
posterior distribution are drawn in three different ways, as shown in Fig. 3:
1. Samples are drawn directly from the two-dimensional Gaussian distribution,
using the Cholesky decomposition of the covariance matrix. This method is very
efficient and is described extensively in literature. These samples are indicated
with the label “Posterior 1”.
2. Samples are drawn using sequential sampling, starting with location 2. These
samples are indicated with the label “Posterior 2”.
3. Samples are drawn using sequential sampling, starting with location 1. These
samples are indicated with the label “Posterior 3”.
Considerations for the Use of Sequential Sampling Techniques 81
Samples drawn from the prior pdf are indicated with the label “Prior”. The
picture on the left-hand side of Fig. 3 shows the actual samples. The right-hand
picture shows both the mean values of the samples that are obtained with each
algorithm and the ellipsoids that are derived from the covariance matrices of each
algorithm. The size of the ellipsoids is determined by the eigenvalues of the
covariance matrices and the orientation by the eigenvectors. This means that the
contours are one standard deviation away from the expectation value, in the
direction of the eigenvectors.
Comparison of the samples from the “Prior” group with samples from the
“Posterior” group clearly shows how strongly the samples are constrained by the
measurements. Comparison of the samples from the “Posterior 1” group with
samples from the “Posterior 2” group and the “Posterior 3” group shows the
deterioration caused by the sequential sampling algorithm. The ellipsoids
corresponding with the “Posterior 2” group and the “Posterior 3” group are shifted
away from the ellipsoid of the “Posterior 1” group and they are also larger. This
82 J. Leguijt
means that both the expectation values and the covariance matrices of the samples
that are obtained with the sequential sampling algorithm are incorrect. This first,
somewhat fictional example, has been specifically designed to show the problems
with sequential sampling with a very simple model.
Example 2
The second example is a more realistic scenario. It consists of a model that is
defined on a grid with 101 grid points in both x and y directions. The grid spacing
between the grid points is 50 m, in both directions. The correlation between the
variables at the grid nodes is defined with an isotropic exponential variogram with a
range of 2000 m, a sill of 2.0 and a nugget of 0.0. The quality of the samples is
inspected with χ 2, ψ 2 and φ2 statistics, each of which is described in the Appendix.
These statistics are different from those described by Leuangthong et al. (2004).
The χ 2, ψ 2 and φ2 statistics have the advantage that the statistical quality of the
whole ensemble of samples can be examined.
First, the quality of the sequential sampling algorithm has been tested by
sampling an ensemble of 100 realisations from the prior. Each location was
constrained by a maximum of 314 nearest neighbours. This is a rather large number
of neighbours, but it has been chosen such that the ensemble of samples that is
generated with the sequential sampling algorithm has the correct χ 2, ψ 2 and φ2
statistics. A map of typical realisation from this algorithm is shown in Fig. 4.
A variogram was estimated by using ten realisations, and a comparison was
made with the variogram that was used to define the prior pdf, in Fig. 5. There is
little discrepancy between the two variograms, which is one indication that the
algorithm performs well.
From the ensemble of 100 realisations, which are generated from the prior, the
average values of the diagnostic statistics, χ 2, ψ 2 and ϕ2, are computed and
compared with their expectation values. The results are listed in Table 2 and
show that the average values are within one standard deviation of the expectations.
This means that the sequential sampling algorithm does an excellent job when it is
used to sample the prior.
When the sequential sampling algorithm is used to sample a posterior pdf, the
results are not particularly encouraging. The observations that are used to constrain
the geostatistical model have been generated by sampling a prior pdf that has also
been defined with an isotropic, exponential variogram, with a range of 2000 m, a sill
of 2.0 and a nugget of 0.0. An observation is generated for each grid location, as
typical for seismic data. It is assumed that the observations are contaminated with
additive noise, with a standard deviation of 1.0, and that the noise is not correlated
between grid locations. These synthetic observations are shown in Fig. 6.
Realisations of the posterior distribution have been generated with the sequential
sampling method and with a method that directly draws samples from the posterior.
An example of a realisation from the sequential sampling algorithm is shown in
Fig. 7.
The imprint of the observations is clearly visible in this realisation, but a more
proper way to judge the validity of the realisations that are obtained from the
sequential sampling algorithm is to compare them with realisation that is generated
Considerations for the Use of Sequential Sampling Techniques 83
Fig. 4 Grid with prior sample for realisation 0. The x and y axes are in metres
Fig. 5 Estimated variogram, prior, variogram of the prior; estimated, variogram estimated from
ten realisations
84 J. Leguijt
Fig. 6 Observations generated with an exponential variogram, with a range of 2000 m. X and
Y axes are in metres
with an algorithm that correctly generates samples from the posterior. This algo-
rithm is based on the Cholesky decomposition of the matrix of the true posterior
distribution. This will be denoted by direct sampling and should not be confused
with direct sequential sampling. With the chosen grid, the size of the covariance
matrix was such that this was still possible.
First, the mean of the realisations obtained with sequential sampling method,
shown in Fig. 8, is compared with the actual posterior expectation values, shown in
Fig. 9.
A map of the differences between the means of the realisations obtained with the
sequential sampling method and the actual posterior expectation values is shown in
Fig. 10. These differences are significant when they are compared with the values of
the observations and the standard deviation of the prior. The mean and the variance
of all the realisations that are generated with the direct sampling method and the
sequential sampling method are listed in Table 3.
Considerations for the Use of Sequential Sampling Techniques 85
Fig. 7 Realisation from the posterior, generated with the sequential sampling method. X and
Y axes are in metres
This shows that both the mean and the variance differ considerably between the
two methods.
The prior variogram, the variogram estimated from the realisations that use the
direct sampling method and the variogram estimated from the realisations that use
the sequential sampling method are shown in Fig. 11. There is a remarkable
difference between the latter two variograms, especially at larger distances. Note
that the variogram that has been estimated from the realisations of the direct
sampling method does not coincide with the variogram of the prior, although
both the measurement and the prior are made with the same variogram. This
stems from the fact that the covariance matrix of the prior and the posterior is
different. A variogram that would have been estimated from the expectation value
of the posterior would coincide with the prior variogram in this situation, but a
realisation of the posterior also has a random component that is determined by the
covariance matrix of the posterior. A more objective method to inspect the validity
of a sampling method is to look at the values of the χ 2, the ψ 2 and the ϕ2 statistics.
Table 4 shows that their average values are far beyond the statistically feasible
range.
The difference between the expected diagnostic statistics and the actual values
of those differs by more than 100 standard deviations. This demonstrates that the
sequential sampling used in practice does not properly sample the posterior pdf.
86 J. Leguijt
Seismically constraint stochastic models with lateral constraints suffer the same
affliction. Seismic measurements are available at each location, but they are
contaminated with noise. For this reason, a different method has been developed
for seismically constrained stochastic modelling. In this method, each location is
revisited many times, after an initial assignment of values to the model parameters
at each location. The initial values can be obtained from the prior expectation
values from a sample of the prior pdf.
Every time that a location is revisited, a local posterior pdf is computed using a
local prior pdf that is constrained by the neighbourhood of that location. Subse-
quently a sample is generated from this local posterior pdf. This effectively creates
a Metropolis algorithm that only creates local updates. The convergence speed of
this algorithm is dependent on the type of the variogram that is used to construct the
prior pdf.
3 Conclusion
locations into account. The sequential sampling method does not correctly sample a
posterior that results from taking inaccurate measurements into account from all of
the locations that are considered in the model.
Fig. 10 Mean of realisations minus posterior expectation. X and Y axes are in metres
χ 2 Statistics
The χ 2 test has been designed to test whether a realisation has been drawn from a
specific normal distribution. Consider a multidimensional normal distribution p(x)
with an expectation value of <x> and a covariance matrix Cx as given by
1 1 T 1
pðxÞ ¼ D 1 exp ðx hxiÞ Cx ðx hxiÞ ; ð9Þ
ð2π Þ 2 jCx j2 2
where D is the dimension of vector x. Assume that the ensemble that has to be tested
consists of N realisations and is denoted by {x1, x1, . . ., xN}. The χ 2 value of
realisation i is given by
Fig. 11 Estimated variograms, prior, variogram of the prior; direct, variogram estimated from
samples that are directly generated from the true posterior; sequential, variogram estimated from
samples that are generated with sequential Gaussian simulation
2
χ2i
¼ D and ð11Þ
Var χ i ¼ 2D
1X N
χ2 ¼ χ2 ð12Þ
N i¼1 i
Since all the χ 2i values are independent, the expectation value and the variance of
the average are equal to
90 J. Leguijt
D E
χ 2 ¼ D and
n o 2D ð13Þ
Var χ 2 ¼
N
When the χ 2i values and χ 2 deviate too much from the expected value, this indicates
that the vectors xi have not been properly sampled from p(x). This may have several
reasons. The samples may be generated with a bias, or the “shape” of the ensemble
does not match the “shape” that is described by the covariance matrix. To make a
distinction between these possibilities, two other statistics, ψ 2 and φ2, have been
defined.
ψ 2 Statistics
where
1X N
x ¼ xi ð15Þ
N i¼1
The expectation value and the variance and the standard deviation of ψ 2i are
N1
ψ 2i ¼ D and
N ð16Þ
2
N1
Var ψ i ¼ 2D
N
1X N
ψ2 ¼ ψ2 ð17Þ
N i¼1 i
φ2 Statistics
The expectation value, the variance and the standard deviation of φ2 are
D
φ2 ¼ and
N ð20Þ
2
2D
Var φ ¼ 2
N
χ 2 ¼ ψ 2 þ φ2 ð21Þ
Bibliography
Deutsch CV (2002) Geostatistical reservoir modeling. Oxford University Press, New York
Gelderblom P, Leguijt J (2010) Geological constraints in model-based seismic inversion, SEG
Denver 2010, 2010-2825 SEG Conference Paper
Leuangthong O, McLennan JA, Deutsch CV (2004) Minimum acceptance criteria for
geostatistical realizations. Nat Resour Res 13(3):131–141
A New High-Order, Nonstationary, and
Transformation Invariant Spatial Simulation
Approach
1 Introduction
Since the early 1990s (Guardiano and Srivastava 1993), several new approaches to
geostatistical simulation have been developed to move this area of research beyond
the second-order or two-point methods and their limits. These methods, well
developed to date, are placed under the term multipoint statistics (MPS). The
basic idea of MPS approaches is that the two-point statistical tools (variogram,
covariance, correlogram) of a given attribute of interest are replaced by a so-called
training image (TI). The TI is then used as a source to provide multiple point
statistics and spatial relations that are used along with the hard data to generate
simulated realizations for the attributes of interest. The first implemented
multipoint method is SNESIM (Strebelle 2002) and it is TI driven; thus, similar
to all conventional MPS methods, the simulated realizations reproduce the high-
order spatial relations in the TI. As a result, in applications with a dense set of hard
data, the complex spatial relations in the data are overridden by those in the TI and
do not assist with the application of MPS methods to applications with relatively
dense datasets.
Several MPS methods are well known to date; examples are discussed next.
FILTERSIM (Zhang et al. 2006) is based on the classification of both data and TIs
using linear filters; it is efficient and also sensitive to the shape and size of the spatial
template, and the number and form of the filters are employed. The direct sampling
method (Mariethoz et al. 2010), in the other hand, does not produce a pattern
database from the TIs like FIMTERSIM. Instead, in a multigrid simulation setup,
first coarser grid nodes are simulated, a template is chosen about a simulation point,
and the data event is extracted. This data event is then compared with the data event
of a randomly chosen TI pattern, in an L2-norm distance basis. If the distance is less
than a threshold, then the pattern is pasted onto the simulation grid; otherwise,
another TI pattern is randomly selected and compared, and so on. The direct
sampling simulation method is fast and effective in simulating based on sparse
data with a given TI set and is also TI driven. Mariethoz et al. (2010) show the
influence of data statistics on realizations generated from this method as a dataset
increases. Other pattern-based MPS methods include the one suggested by Arpat and
Caers (2007); first, a pattern database is generated by sliding a fixed template over
the TIs. Then, the data event on the grid used for the simulation and at each location
is compared with the data event of the pattern database, and the one with the least
L2-norm distance is chosen and pasted on the grid nodes involved. Abdollahifard
and Faez (2012) first cluster the pattern database generated from the TI using a
Bayesian framework. Then each cluster is modeled by a set of simple linear features
and the extraction of features for each incomplete pattern on the simulated grid
follows. Honarkhah and Caers (2010), instead of building a raw pattern database,
classify TIs using some fixed simple features and compare them, using the L2-norm,
to the same features extracted from a point of the grid being simulated. The most
similar pattern is pasted onto the grid nodes, until all of them are visited.
A New High-Order, Nonstationary, and Transformation Invariant Spatial. . . 95
The following sections present the proposed simulation approach; then the
results of initial tests using known datasets follow. Conclusions and future work
complete the presentation.
2.1 Overview
The goal of the proposed method is to simulate a random field Z(x), in a sequential
multigrid process, given a grid with nodes x, a training image (TI),
zðYÞ ¼ fzðy1 Þ, . . . , zðyM Þg; with nodes Y ¼ fy1 ; . . . ; yM g in the training image,
and a sparse set of N hard data zðxi Þ, i 2 ½1; . . . ; N on a regular grid xi , i 2 ½1; . . . ; N .
The hierarchy of the sequential simulation is illustrated in Fig. 1. The blue nodes
represent the hard data (a). In the first sequence, each red node is conditioned on the
four closest blue nodes and simulated. The lines represent the conditions in (b). This
continues sequentially until all the nodes in the grid are simulated (Fig. 1c–e).
Unicolor lines in each figure represent the spatial templates, connecting the condi-
tioning data and a single node to be simulated. The size of the templates reduces
after each sequence to maintain the same number of conditioning nodes. This is the
natural representation of the multigrid approach for the simulation.
At each sequence, the path is chosen randomly and saved into a vector containing
the indices of the visiting nodes. Each successive random variable Z(x) at node x is
conditioned to n-nearest neighbors, selected from the set of previously simulated
nodes and the hard data {x1, . . ., xn} (Goovaerts 1998). A template is formed
spatially by connecting each conditioning data, xi, to node x, presented by a lag
vector Lx ¼ fh1 ; . . . ; hn g ¼ fx1 x, . . . , xn xg. Consequently, the neighbor-
hood of x is denoted by Nx ¼ fx þ h1 , . . . , x þ hn g and the data event is denoted
by dNx ¼ fzðx1 Þ, . . . , zðxn Þg. The goal is to estimate and draw a sample from the
probability, PðZðxÞjdNx , zðYÞÞ, of each successive Z(x) in the next visiting node on
the path given its data event dNx and the TI, z(Y). This probability is intractable for
continuous variables. Alternatively, a model with a set of parameters, θ 2 Θ, may
be chosen, to represent this probability independent from the data event and TI. The
parameter θ is optimized to express the data event and TI. Thus, this probability can
be decomposed using the Bayes product and sum rule:
A New High-Order, Nonstationary, and Transformation Invariant Spatial. . . 97
Fig. 1 The hierarchy of the sequential multigrid simulation for n ¼ 5, the connecting lines present
the conditioning dependencies. (a) The hard data. (b) Red nodes are simulated conditioned on their
four nearest neighbors. (c) Black nodes are simulated conditioned on both blue and red nodes from
previous sequence. (d) Oranges are simulated conditioned on previous simulated nodes and hard
data. (e) Greens are simulated in the last sequence. At each sequence the resolution of the grid
doubles
Z
PðZ ðxÞjdNx , zðYÞÞ ¼ PðZðxÞjθÞPðθjdNx , zðYÞÞdθ: ð1Þ
θ2Θ
Estimation of the integrand further simplifies Eq. 1. Figure 2 represents the term
PðθjdNx , zðYÞÞ as a function of θ. The contribution of this function is negligible
except for a narrow band near an optimal value for the parameter’s maximum
likelihood estimate, θMLE.
Consequently, Eq. 1 could be estimated as in Eq. 2:
constant
zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{
PðZðxÞjdNx , zðYÞÞ PðZðxÞjθMLE Þ PðθMLE jdNx , zðYÞÞΔθ
1
¼ PðZ ðxÞjθMLE Þ: ð2Þ
CMLE
In Eq. 2, CMLE can Zbe regarded as the normalization factor to produce a valid
probability, CMLE ¼ PðzjθMLE Þdz.
z2Z
98 A.A.H. Abolhassani et al.
Fig. 2 Presentation of
PðθjdNx , zðY ÞÞ as a function
of θ. In practice this
probability is negligible
except at a narrow Δθ band
near θMLE
argmax
θMLE ¼ PðθjdNx , zðYÞÞ: ð3Þ
θ2Θ
uniform prior
zffl}|ffl{
argmax PðzðYÞjθ, dNx Þ PðθÞ
θMLE ¼ : ð4Þ
θ2Θ PðzðYÞÞ
|fflfflfflffl{zfflfflfflffl}
independant from θ
One assumes a uniform prior in parameters space Θ and note that the marginal
probability in TI, P(z(Y)), is independent from θ. Hence, Eq. 4 becomes
argmax
θMLE ¼ PðzðYÞjθ, dNx Þ: ð5Þ
θ2Θ
Each node in TI is only conditioned on its neighbors and the parameters set θ.
Hence, the joint distribution in Eq. 5 can be decomposed further.
argmax Y M
θMLE ¼ P zðyi Þjθ, dNx , dNyi : ð6Þ
θ2Θ i¼1
A New High-Order, Nonstationary, and Transformation Invariant Spatial. . . 99
∂ X M
log P z ð y Þjθ, d N , d N ¼ 0: ð8Þ
∂θ i¼1 i x y i
The exponential family is used herein to model the likelihood function in Eq. 8 with
the parameter set θ ¼ {θ1, θ2}.
!
1 1 2
2 ð z ð yi Þ θ 1 Þ
P zðyi Þjθ, dNx , dNyi ¼ exp ω dNx ; dNyi ; σ 0 : ð9Þ
c 2 θ2
ω dNx ; dNyi ; σ 20 is introduced as the similarity measure (SM) of the data event dNx
and dNyi. It ensures that the TI patterns with similar data events contribute more
toward building the likelihood function in Eq. 9. The SM is defined as
0 T 1
1
2D dN x ; dN y i D dN x ; dN y i C
B
ω dNx ; dNyi ; σ 20 ¼ exp@ A: ð10Þ
Σ dNx ; σ 20
where
D(d Nx, dNyi) is introduced as the high-order statistics disparity vector.
Σ dNx ; σ 20 is the covariance matrix of the disparity vector and is calculated
using the calculus of variations.
2010; Mariethoz et al. 2010, and Mustapha et al. 2013). When considering two
sets of data events dNx ¼ fx1 ; . . . ; xn g, dNy ¼ fy1 ; . . . ; yn g of order n, to develop
an isotropic L2-norm disparity measure, one must compare all possible ordering
of these two data events, which results in n n! number of operations. For n ¼ 5
the number of operations is 600. This is computationally expensive and can only
operate on small-size TIs. The following method is employed to reduce the
computing time.
First, Vieta’s formula (Funkhouser 1930) is used to calculate the coefficients of
two polynomials ps(X) and pt(Y) with the roots equal to the data events dNx and dNy,
respectively.
8 ps ðXÞ ¼ Xn þ s1 Xn1 þ . . . þ sn :
>
> s1 ¼ x1 þ x2 þ . . . þ xn
<
s2 ¼ x1 ðx2 þ x3 þ . . . þ xn Þ þ x2 ðx3 þ . . . þ xn Þ þ xn1 xn ð11Þ
:
>
> :
:
sn ¼ x1 x2 . . . xn
8 pt ðY Þ ¼ Y n þ t1 Y n1 þ . . . þ sn :
>
> t 1 ¼ y1 þ y2 þ . . . þ yn
<
t2 ¼ y1 ðy2 þ y3 þ . . . þ yn Þ þ y2 ðy3 þ . . . þ yn Þ þ yn1 yn ð12Þ
:
>
> :
:
t n ¼ y1 y2 . . . yn
These could be regarded as two mappings dNx ! s ¼ {s1,. . .sn} and dNy ! t ¼
{t1,. . .tn}. The advantage of these mappings is that they are invariant to the ordering
of the domain. This invariance results from the coefficients of a polynomial being
invariant to the order of the roots in a set due to Eqs. 11 and 12. s and t are in a
particular form of high-order moments:
X
sm ¼ k2K
N ðkÞmðuÞ ðkÞ,
1 XNðkÞ Y ð13Þ
mðuÞ ðkÞ ¼ xj l2k xjþl ;
N ðk Þ j¼1
X
tm ¼ k2K
N ðkÞmðuÞ ðkÞ,
1 XNðkÞ Y ð14Þ
mðuÞ ðkÞ ¼ yj l2k yjþl ;
N ðk Þ j¼1
where, k ¼ {k1,. . .,ku1} and N(k) is the support for estimating the moment. The
high-order statistics disparity vector is defined as follows:
D d N x ; d N yi ¼ s T t T : ð15Þ
It is worth mentioning that the number of operations for this new disparity measure
is reduced dramatically to n 2n1 per simulation node versus n n! for the
L2-norm. For n ¼ 5, op(L2 norm) ¼ 600 and op(L2 ord) ¼ 80.
A New High-Order, Nonstationary, and Transformation Invariant Spatial. . . 101
The dataset used in this section is the Stanford V Reservoir dataset (Mao and
Journel 1999). This exhaustive dataset consists of a 3D grid with porosity values.
The grid consists of 130 100 30 nodes, i.e., X Y Z. Here this dataset is
cropped into a grid of 100 100 30 in order to perform some linear transforma-
tion on the data, e.g., rotation. For each simulation, a layer, Z2 {1,. . .,30}, is
selected as the ground truth of the simulation, referred to as the original image.
This image is then down-sampled to produce the hard dataset, containing N points.
All layers except layer Z are considered as TI for each simulation. The results of the
simulation produced by HOSTISIM are compared, with an order 5 template, with
the ones produced by FILTERSIM, with a search grid size 11 11 and inner patch
size 7 7. Note that the TIs provided for HOSTISIM are rotated 90 clockwise. The
simulations are generated for layers z ¼ {1,. . .,4}, from top to bottom in Figs. 3 and
4, first with N ¼ 625 number of hard data points, 6.25 % of the original data, in
Fig. 3, and second with N ¼ 169 number of hard data points, 1.69 % of the original
data, in Fig. 4. For each simulation on Figs. 3 and 4, from left to right, the original
image, HOSTISIM and FILTERSIM typical realizations are all presented. For each
set of results, the histogram of the original image and two simulations are also
plotted. As a robust quantitative comparison, for each simulation, ten realizations
are generated by HOSTISIM and FILTERSIM methods, and for each one the PSNR
and SSIM scores (Wang et al. 2004) are calculated and averaged for each method
over all ten realizations and provided in Figs. 3 and 4. For every single case,
HOSTISIM outperforms FILTERSIM, visually by better representing the channels
and low-contrast structures of the original exhaustive image, with higher PSNR and
SSIM scores, and by better matching the histogram. Tables 1 and 2 are presenting
the average PSNR and SSIM for HOSTISIM and FILTERSIM.
The current implementation of the method is in Matlab, using the GPU parallel
computation library. It has not been optimized nor developed in C, nor Python, as of
yet. On the other hand, FILTERSIM has been optimized and developed in Python
and is available in the SGEMS software platform. Despite this disparity in optimi-
zation, we ran some tests to compare them as is. The system used for the tests was a
Unix OS server with eight cores Xeon CPU, 3.500 GHz with 8 MB cache size and
64 GB DDR4 memory and Nvidia Tesla k40c GPU with 12 GB DDR5 memory
with 2880 cores.
For each method, two sets of tests were performed. Each set of tests consisted of
generating 10 separate simulations and averaging the computing time for each
simulation. The hard data used for all cases were 12 12 real-valued data on a
regular grid. The goal was to generate a realization on a 100 100 SG, for each
102 A.A.H. Abolhassani et al.
Fig. 3 The results generated given 625 hard data (6.25 %). From top to bottom, layers Z ¼ 1,. . .4.
From left to right, original image, HOSTISIM and FILTERSIM simulations. On the bottom the
histograms of each set are presented (Red: Exhaustive, dark blue: HOSTISIM and light blue:
FILTERSIM)
A New High-Order, Nonstationary, and Transformation Invariant Spatial. . . 103
Fig. 4 The results generated given 169 hard data (1.69 %). From top to bottom, layers Z ¼ 1,. . .4.
From left to right, original image, HOSTISIM and FILTERSIM simulations. On the bottom the
histograms of each set are presented (Red: Exhaustive, dark blue: HOSTISIM and light blue:
FILTERSIM)
104 A.A.H. Abolhassani et al.
Table 1 Comparing average PSNR and SSIM for HOSTISIM and FILTERSIM methods with
N ¼ 625 number of hard data points (6.25 %)
PSNR SSIM
Z HOSTISIM FILTERSIM HOSTISIM FILTERSIM
1 24.94 23.10 0.61 0.50
2 24.31 22.77 0.60 0.47
3 23.84 22.42 0.57 0.44
4 23.10 21.29 0.56 0.42
Table 2 Comparing average PSNR and SSIM for HOSTISIM and FILTERSIM methods with
N ¼ 169 number of hard data points (1.69 %)
Z PSNR SSIM
Z HOSTISIM FILTERSIM HOSTISIM FILTERSIM
1 21.23 20.98 0.47 0.43
2 21.49 20.65 0.47 0.40
3 20.87 20.20 0.42 0.35
4 20.17 18.76 0.40 0.30
For more and higher-resolution results, please visit the link bellow: http://cim.mcgill.ca/~amir/
HOSTISIM.html
Table 3 Comparing average computation time for HOSTISIM and FILTERSIM methods
FILTERSIM HOSTISIM ***
Test#1 8s 3s +9 s GPU initialization in Matlab
Test#2 134 s 42 s +12 s GPU initialization in Matlab
*** Almost constant, overhead GPU initialization time in Matlab
case. Each TI was a 100 100 real-valued image. In the first test, only 1 TI was
used and in the second test 29 TIs were used. The average computing times are
presented in Table 3.
5 Conclusions
Acknowledgment Funding was provided by the Natural Sciences and Engineering Research
Council of Canada (NSERC) Discovery Grant 239019 and mining industry partners of the
COSMO Laboratory (AngloGold Ashanti, Barrick Gold, BHP Billiton, De Beers Canada, Kinross
Gold, Newmont Mining, and Vale) and the Group for Research in Decision Analysis (GERAD).
Thanks are given to Prasun Lala for his assistance.
Bibliography
Abdollahifard MJ, Faez K (2012) Stochastic simulation of patterns using Bayesian pattern
modeling. Comput Geosci 17(1):99–116
Arpat GB, Caers J (2007) Conditional simulation with patterns. Math Geol 39(2):177–203
Chatterjee S, Dimitrakopoulos R (2012) Multi-scale stochastic simulation with wavelet-based
approach. Comput Geosci 45(3):177–189
Dimitrakopoulos R, Mustapha H, Gloaguen E (2010) High-order statistics of spatial random fields:
exploring spatial cumulants for modeling complex non-Gaussian and non-linear phenomena.
Math Geosci 42(1):65–99
Funkhouser HG (1930) A short account of the history of symmetric functions of roots of equations.
Am Math Mon 37(7):357–365
Goovaerts P (1998) Geostatistics for natural resources evaluation. Cambridge University Press,
Cambridge
Guardiano FB, Srivastava RM (1993) Multivariate geostatistics: beyond bivariate moments.
Geosatistics Tróia’92, Springer, pp 133–144
Honarkhah M, Caers J (2010) Stochastic simulation of patterns using distance-based pattern
modeling. Math Geosci 42(5):487–517
Mao S, Journel AG (1999) Generation of a reference petrophysical and seismic 3D data set: the
Stanford V reservoir. Report—Stanford Center for Reservoir Forecasting Annual Meeting,
Stanford. URL: http://ekofisk.stanford . edu/SCRF.htmlSS
Mariethoz G, Renard P, Straubhaar J (2010) The direct sampling method to perform multiple-point
geostatistical simulations. Water Resour Res 46(11):W11536. doi: 10.1029/2008WR007621
Minniakhmetov I, Dimitrakopoulos R (2016a) Joint high-order simulation of spatially correlated
variables using high-order spatial statistics. Math Geosci, in press
Minniakhmetov I, Dimitrakopoulos R (2016b) A high-order, data-driven framework for joint
simulation of categorical variables. GEOSTAT2016, in this volume
Mustapha H, Dimitrakopoulos R (2010a) A new approach for geological pattern recognition using
high-order spatial cumulants. Comput Geosci 36(3):313–334
Mustapha H, Dimitrakopoulos R (2010b) High-order stochastic simulation of complex spatially
distributed natural phenomena. Math Geosci 42(5):457–485
Mustapha H, Dimitrakopoulos R (2011) HOSIM: a high-order stochastic simulation algorithm for
generating three-dimensional complex geological patterns. Comput Geosci 37(9):1242–1253
Mustapha H, Chatterjee S, Dimitrakopoulos R (2013) CDFSIM: efficient stochastic simulation
through decomposition of cumulative distribution functions of transformed spatial patterns.
Math Geosci 46(1):95–123
106 A.A.H. Abolhassani et al.
Stien M, Kolbjørnsen O (2011) Facies modeling using a Markov mesh model specification. Math
Geosci 43(6):611–624
Strebelle S (2002) Conditional simulation of complex geological structures using multiple-point
statistics. Math Geol 34(1):1–21
Wang Z, Bovik AC, Sheikh HR, Simoncelli EP (2004) Image quality assessment: from error
visibility to structural similarity. IEEE Trans Image Process 13(4):600–612
Zhang T, Switzer P, Journel A (2006) Filter-based classification of training image patterns for
spatial simulation. Math Geol 38(1):63–80
A Truly Multivariate Normal Score
Transform Based on Lagrangian Flow
1 Introduction
Many of the standard geostatistical simulation algorithms for continuous data are
based on the assumption of Gaussianity. However, this assumption is often vio-
lated, and so prior to simulation, a transformation is required. Some notable
exceptions are direct sequential simulation and multiple indicator simulation. In
U. Mueller (*)
School of Science, Edith Cowan University, 270 Joondalup Drive, Joondalup, WA 6027,
Australia
e-mail: [email protected]
K.G. van den Boogaart • R. Tolosana-Delgado
Helmholtz Zentrum Dresden-Rossendorf, Helmholtz Institute for Resources Technology,
Chemnitzerstrasse 40, 09599 Freiberg, Saxony, Germany
e-mail: [email protected]; [email protected]
where zi denotes the ith element that was analyzed and uα a sample location within
the study region A. Data of this nature cannot be modeled “raw,” and even before a
transformation to Gaussian space is performed, a log-ratio transform ought to be
performed to remove the positivity requirement as well as the constant sum
constraint (Aitchison 1986). Several transformations are available for that purpose,
one of which is the alr transformation, given as
xi ðuα Þ ¼ alrD zi ðuα Þ ¼ ln zi ðuα Þ=zD ðuα Þ ð2Þ
for i ¼ 1,. . .,D. The choice of the divisor is arbitrary, so that there are in fact
D distinct alr transformations, given a regionalized variable of D components. To
ensure independence of simulation results from the choice of log-ratio transforma-
tion, the transformation to normality should be affine equivariant, that is,
A Truly Multivariate Normal Score Transform Based on Lagrangian Flow 109
The spreads of all kernels evolve simultaneously according to Eq. 4. For each
location, time-dependent scores are defined relative to a decomposition M ¼ RRT of
the variance covariance matrix of the input data:
1 1
sα ðz; tÞ ¼ R ðz zα ðtÞÞ ð5Þ
σ ðt Þ
The local position X of a mass point of the αth kernel, which is at time t at location
z as a function of time τ, is given by
The average velocity of the total mass at z is defined as the weighted average:
Xn
vðz; tÞ ¼ α¼1
ωα ðz; tÞvα ðz; tÞ ð8Þ
where ωα(z,t) are weights derived from the time-dependent density of the αth kernel
at this location and time. The weights are functions of the Mahalanobis distance
between the vector z and the kernel center zα(t):
exp ksα ðz; tÞk2
ωα ðz; tÞ ¼ X n ð9Þ
2
α¼1
exp ks α ð z; tÞk
∂
p ðtÞ ¼ vðpz ðtÞ, tÞ, pz ð0Þ ¼ z; ð10Þ
∂t z
From the construction it is clear that the flow anamorphosis depends on the choice
of suitable values for σ 0 and σ 1. The initial spread σ 0 controls the deformation of the
underlying space: the greater the value, the weaker the deformation. It thus exer-
cises the greatest influence over the shapes of the final marginal distributions and
multivariate normality of the normal scores. The ranges of the marginal increase
with decreasing σ 0. Multivariate normality tests performed on test data indicate that
the choice of σ 0 depends both on the number of variables and the sample size. In the
case of an initial distribution as shown in Fig. 1, the output distribution is bivariate
normal for values of σ 0 0.2. This was ascertained for 100 replicates generated
from this model (Fig. 2). Here an energy test was applied to determine multivariate
normality (Szekely and Rizzo 2013).
Note that the p-value of a test under the null hypothesis follows a uniform
distribution on the interval [0, 1]. Thus, the target behavior should be represented
by a broad, centered boxplot covering the whole range and not a constant value of
1.0 (which would imply overfitting).
Fig. 1 Trajectories of points under flow deformation: scatter diagrams of raw data (left) and
transformed data (right) with marginal distributions are superimposed in the bottom
112 U. Mueller et al.
Fig. 2 Boxplots of p-values of normality tests (left) and ranges (right) based on different values
of σ 0
The value of σ 1 does not impact on multivariate normality but does influence the
ranges of the transformed distributions. A simplistic choice is σ 1 ¼ 1, resulting in
marginal distributions with standard deviation less than unity. The relative change
in the ranges of the marginal distributions when σ 1 ¼ f(σ 0) for some suitably chosen
positive function f is f ðσ 0 Þ 1 compared to those with σ 1 ¼ 1. These properties are
illustrated in Table 1 for the case of the data shown in Fig. 1 with σ 0 ¼ 0.8, 0.5, and
pffiffiffiffiffiffiffiffiffiffiffiffiffi
0.2 and σ 1 ¼ 1, σ 1 ¼ 1 þ σ 0 , σ 1 ¼ 1 þ σ 20 , respectively.
Here the application of the flow anamorphosis is illustrated with two examples from
Western Australian mining contexts. Both data come from a mining bench, so can
be deemed two-dimensional. The first set is derived from an iron ore mine, the
second from a manganese mine. The data configurations are shown in Fig. 3. For
both sets locations were transformed to maintain confidentiality.
This data set consists of 400 samples, and the variables of interest are Fe, Al2O3,
and SiO2. Since the data are compositional, but the three elements considered only
form a sub-composition, a filler variable was introduced to achieve closure. The
A Truly Multivariate Normal Score Transform Based on Lagrangian Flow 113
Table 1 p-values for multivariate normality tests (Korkmaz et al. 2014) applied to the transforms
of the data shown, ranges of the variables and relative change in range compared to σ 1 ¼ 1 for
selected values of σ 0 and σ 1
Statistic 0.80 0.50 0.20 0.80 0.50 0.20 0.80 0.50 0.20
pffiffiffiffiffiffiffiffiffiffiffiffiffi
σ1 ¼ 1 σ1 ¼ 1 þ σ0 σ 1 ¼ 1 þ σ 20
Mardia 0.00 0.00 0.10 0.00 0.00 0.09 0.00 0.00 0.10
Henze-Zirkler 0.00 0.02 0.90 0.00 0.02 0.89 0.00 0.02 0.90
Royston 0.00 0.03 0.64 0.00 0.03 0.64 0.00 0.03 0.64
Energy 0.00 0.01 0.90 0.00 0.01 0.90 0.00 0.02 0.92
R(V1) 3.26 3.69 4.28 5.86 5.53 5.14 4.17 4.12 4.37
R(V2) 3.13 3.48 3.98 5.63 5.21 4.77 4.00 3.89 4.06
Rel. change – – – 0.80 0.50 0.20 0.28 0.12 0.02
0
25 75 125
0 0 0
alrAl2O3
alrSiO2
alrSiO2
-3 -2.5 -2.5
-6 -5 -5
0.4 0.7 1 0.4 0.7 1 -6 -3 0
alrFe alrFe alrAl2O3
4 4 4
2 2 2
FAV2
FAV3
FAV3
0 0 0
-2 -2 -2
-4 -4 -4
-4 -2 0 2 4 -4 -2 0 2 4 -4 -2 0 2 4
FAV1 FAV1 FAV2
Fig. 4 Scatter diagrams of alr-transformed variables (top) and normal scores (bottom)
This data set consists of four variables, Mn3O4, Fe2O3, Al2O3, and SiO2, with
Mn3O4 as the commodity of interest. The sample size is 176. As in the case of
A Truly Multivariate Normal Score Transform Based on Lagrangian Flow 115
the previous data, the data will be treated as compositional. This required inclusion
of a filler variable followed by a transformation to log-ratios relative to that variable
(Table 4).
For this data set, a much stronger deformation is required to transform the data to
multivariate normality. For σ 0 ¼ 0.02 and σ 1 ¼ 1.02, the Henze-Zirkler and energy
tests support a conclusion of multivariate normality, but the Mardia and Royston
tests do not: the relevant p-values are 0.34, 0.005, 0.48, 0.003, and 0.05 for the
Mardia skewness, Mardia kurtosis, Henze-Zirkler, Royston, and energy tests. This
issue can be overcome by replacing a single transform by a pair of transforms,
requiring weaker deformations: the data are first transformed via a flow
anamorphosis with parameters σ 0 ¼ 0.2 and σ 1 ¼ 1.2, and then an anamorphosis
with parameters σ 0 ¼ 0.2 and σ 1 ¼ 1.2 is applied to the output of the first transfor-
mation. With this approach, all but the Mardia tests support a conclusion of
multivariate normality.
Summary statistics in Table 5 show that means are close to 0, and ranges of the
transformed data lie between 85 % and 95 % of those obtained from quantile
matching. Given that in the case of the flow anamorphosis, it is the underlying
space that is transformed, these ranges are satisfactory. The scatter diagrams in
Fig. 5 show that it is in essence the third variable of the single-step transform which
leads to a rejection of multivariate normality, and from correlation tests, it follows
that the two-step transformed data are statistically independent.
However, in this case, the spatial decorrelation is not as satisfactory. For the
two-step anamorphosis, the mean decorrelation efficiency is 0.99, but there is
evidence of remnant spatial correlation for distances of up to 15 m, which is clearly
116 U. Mueller et al.
alrFe2O3
5
FA1sV2
FA2sV2
2 2
0 0 0
-2 -2
-5
-6 -4 -2 0 -2 0 2 -2 0 2
alrAl2O3 FA1sV1 FA2sV1
alrMn3O4
FA2sV3
FA1sV3
2 2 2
0 0 0
-2 -2 -2
-4
-6 -4 -2 0 -2 0 2 -2 0 2
alrAl O FA1sV1 FA V
2 3 2s 1
5
FA1sV4
alrSiO2
FA2sV4
2 2
0 0 0
-2 -2
-5
-6 -4 -2 0 -2 0 2 -2 0 2
alrAl2O3 FA1sV1 FA2sV1
alrMn3O4
FA1sV3
2 2 FA2sV3 2
0 0 0
-2 -2 -2
-4
-5 0 5 -2 0 2 -2 0 2
alrFe2O3 FA1sV2 FA2sV2
5
FA1sV4
alrSiO2
FA2sV4
2 2
0 0 0
-2 -2
-5
-5 0 5 -2 0 2 -2 0 2
alrFe2O3 FA1sV2 FA2sV2
5
FA1sV4
FA2sV4
2 2
alrSiO2
0 0 0
-2 -2
-5
-4 -2 0 2 -2 0 2 -2 0 2
alrMn3O4 FA1sV3 FA2sV3
Fig. 5 Scatter diagrams of alr-transformed variables (left column), single-step transformed scores
(center), and two-step transformed scores (right column)
A Truly Multivariate Normal Score Transform Based on Lagrangian Flow 117
Fig. 6 Deviation from diagonality (left) and decorrelation efficiency (right) as functions of lag
spacing
visible in the plot of decorrelation efficiency against lag spacing in Fig. 6. This is
also evident from the mean deviation from diagonality whose value is 0.28.
Figure 6 also shows the corresponding plots for the single-stage anamorphosis
and the two-step anamorphosis followed by an application of MAF decomposition
(Bandarian and Mueller 2008) with lag value chosen to be 10 m. These indicate that
the overall best decorrelation is achieved by the single-step flow anamorphosis
although MAF achieves improvements in the mean deviation from diagonality
compared to the two-step transformation.
5 Conclusion
strongly dependent on the features of the input data. In the case of the iron ore data,
the sample data appeared to be reasonably simple with relatively high correlation
between the input variables, while the manganese data set showed much more
complex relationships. This impacts on the strength of deformation required to
successfully transform them to multivariate normality. A two-step approach might
provide a feasible alternative in such a scenario. The resulting data were statistically
independent, but in the case of the manganese data, a slight spatial correlation
remained and a postprocessing via MAF did not remove it.
Acknowledgments The authors acknowledge financial support through the ECU CES Travel
Scheme 2014, the DAAD-UA Grant CodaBlockKriging, and the Perth Convention Bureau Aspire
Professional Development Grant 2016. Clint Ward, Cliffs Resources, and Matt Cobb, Consoli-
dated Minerals, are thanked for provision of the data on which the example studies are based.
Bibliography
Aitchison J (1986) The statistical analysis of compositional data. Chapman & Hall, London
Bandarian EM, Mueller U (2008) Reformulation of MAF as a generalised eigenvalue problem. In
Ortiz JM, Emery X, Geostats2008, pp 1173–1178. Santiago.
Barnett RM, Manchuk JG, Deutsch CV (2014) Projection pursuit multivariate transform. Math
Geosci 46(2):337–360
Filzmoser P, Hron K (2008) Outlier detection for compositional data using robust methods. Math
Geosci 40:233–248
Korkmaz S, Goksuluk D, Zararsiz G (2014) MVN: an R package for assessing multivariate
normality. R J 151–162. Retrieved from http://journal.r-project.org/archive/2014-2/korkmaz-
goksuluk-zararsiz.pdf
Leuangthong O, Deutsch CV (2003) Stepwise conditional transformation for simulation of
multiple variables. Math Geol 35(2):155–173
Szekely GJ, Rizzo ML (2013) Energy statistics: a class of statistics based on distances. J Stat Plan
Infer 143(8):1249–1272
Tercan A (1999) The importance of orthogonalization algorithm in modeling conditional distri-
butions by orthogonal transformed indicator methods. Math Geol 31(2):155–173
van den Boogaart KG, Mueller U, Tolosana Delgado R (2015) An affine equivariant anamorphosis
for compositional data. In: Schaeben H, Tolosana Delgado R, van den Boogaart KG, van den
Boogaart R (eds) Proceedings of IAMG 2015, Freiberg (Saxony) Germany, September 5–13,
2015. IAMG, Freiberg, pp 1302–1311
van den Boogaart KG, Mueller U, Tolosana-Delgado R (2016) An affine equivariant multivariate
normal score transform for compositional data. Math Geosci. doi:10.1007/s11004-016-9645-y
Part III
Mining Engineering
Using Samples of Unequal Length
for Estimation
Marcel Antonio Arcari Bassani and Jo~ao Felipe Coimbra Leite Costa
1 Introduction
In various mineral deposits, it is common to have drill hole samples obtained from
different drilling campaigns. As each campaign has its own sampling protocol,
samples from different campaigns are usually taken with different nominal lengths
or supports. The construction of grade models would benefit if all available
information is used for estimation purpose. However, the difference in length
among samples must be considered during estimation.
One particular situation in which samples of different lengths occur is when the
mineralization consists of a thin seam or vein and the drill hole samples cross the
entire seam or vein. Each sample length corresponds to the seam thickness. In this
case, practitioners usually work with the variables accumulation (product of grade
and thickness) and thickness (Krige 1978; Bertoli et al. 2003; Marques et al. 2014).
The grade estimates are obtained by dividing the accumulation estimates by the
thickness estimates. This approach eliminates the vertical component (z), resulting
in a two-dimensional (2D) model. The problem with this method is that a 2D block
model may not be used directly for pit or stope optimization algorithms.
X
n
z*V ðuÞ ¼ λ i zðv i Þ ð1Þ
i¼1
where λi is the ordinary kriging weight associated to the datum z(vi). The kriging
weights are the solution of the ordinary kriging system.
Equation 2 defines the ordinary kriging system accounting for the support of the
data (Journel and Huijbregts 1978; Isaaks and Srivastava 1989; Goovaerts 1997):
Using Samples of Unequal Length for Estimation 123
8 n
> X
>
> vi ; vj þ μ ¼ Cðvi , V ðuÞÞ
λj C
>
<
j¼1
X
n i ¼ 1, . . . , n ð2Þ
>
>
>
> λj ¼ 1
:
j¼1
where μ is the Lagrange multiplier and C vi ; vj is the covariance block to block
between the block datum vi and the block datum vj. C vi ; vj is calculated as the
0 0 0
average of point covariances C(ui , uj ) defined between any discretizing point ui of
0
the block datum vi and any discretizing point uj of the block datum vj:
1 X Ni XNj 0 0
C vi ; vj ¼ C ui , uj ð3Þ
N i N j i¼1 j¼1
where Ni is the number of discretizing points of the block datum vi and Nj is the
ðvi , V ðuÞÞ in Eq. 2 is
number of discretizing points of the block datum vj. The term C
the covariance block to block (Eq. 3) between the block datum vi and the block to be
estimated V centred at location u.
3 Case Study
The dataset derives from a bauxite deposit located in the northern portion of the
Brazilian Amazon basin. The dataset contains 686 drill holes located on a relatively
regular grid of 200 200 m spacing along the east (X) and north (Y) directions. The
original Z coordinates were transformed into stratigraphic coordinates. The variable
of interest is the percentage of the total sample mass retained at the no. 14 sieve
aperture (REC14). As REC14 is an additive variable, similar to grades, the methods
presented here are also suitable for grades. In 343 out of the 686 drill holes, REC14
was sampled at the nominal length of 0.5 m (white points in Fig. 1). In the
remaining 343 drill holes (black points in Fig. 1), there is a single sample of
REC14 whose length corresponds to the ore thickness. The sample represents the
average value over the total ore thickness.
Figure 2a shows the histogram of REC14 weighted by the length of the samples
and summary statistics. The distribution is fairly symmetric around the mean with a
low coefficient of variation. As the drill holes are approximately regularly spaced,
these statistics are representative of the study area. Figure 2b shows the QQ plot
between the drill hole data (length weighted) and the declustered data (obtained
with a nearest neighbour estimate). The points in Fig. 2b are close to the line y ¼ x,
showing that the two distributions are similar, as expected. Figure 2c shows the
124 M.A.A. Bassani and J.F.C.L. Costa
histogram of the length of the samples. The length of the samples varies from 0.25
to 7.88 m. Since roughly 80 % of the samples are short, whose lengths are between
0.25 and 0.75 m, the geomodeler may feel tempted to retain only these short
samples for estimation. However, keeping only the short samples results in exces-
sive loss of information.
Only the samples with length between 0.25 and 0.5 m were used to calculate the
experimental variogram. As kriging with samples of different support needs a
variogram at a point scale to calculate the average covariances, long samples
were not used to calculate the experimental variogram.
Using Samples of Unequal Length for Estimation 125
Fig. 2 Histogram of REC14 (a), QQ plot between drill hole and declustered data (b), histogram of
the length of the samples (c)
3.3 Estimation
Ordinary kriging considering the different support in the data was used to estimate
REC14. The drill hole samples were discretized along the main direction of the
sample. The spacing of the discretization points corresponds to the length of the
small-scale data used to calculate the experimental variogram. These discretization
points of the samples were used to calculate the covariances block to block (Eq. 3)
between the samples. The estimation was performed at a block model with block
size of 50 50 0.5 m along X, Y and Z, respectively. The block discretization
was set to 5 5 1. The estimates were constrained to the blocks inside the
interpreted geological model.
126 M.A.A. Bassani and J.F.C.L. Costa
The grade model was checked with the following techniques: (1) visual inspection,
(2) swath plot and (3) cross-validation.
Visual inspection consists in comparing visually the grade model with the
samples. The grade model must be consistent with the data.
Swath plot consists in first defining a series of swaths or slices along the X, Y and
Z directions. Then, the average grade of the model and the declustered average
grade of the samples within the slices are compared. The samples were declustered
with a nearest neighbour estimate.
In the cross-validation, first one sample at a particular location is removed.
Second, the value is estimated at that location using the remaining samples. The
same estimation parameters used in the estimation of the block model were used in
the cross-validation. The estimation error (difference between the estimated and
true value) was calculated. Furthermore, the estimated and true values were com-
pared using a scatter plot.
Figure 3 shows a plan view of the block model together with the samples. The high-
grade areas of the block model are close to the high-grade samples, as expected.
The swath plots show that the block model reproduced the trend in the data along
the X, Y and Z directions (see Fig. 4a–c). The local mean of the block model is
similar to the local declustered mean for the three directions. In addition, the
estimates neither overestimate nor underestimate systematically the local mean
(see Fig. 4a–c; the local mean of the block model is not always either above or
below the local declustered mean).
Figure 5a shows the histogram of estimation errors weighted by the length. The
grade model is globally accurate, as the mean error is 0.14 %, which is close to zero.
It represents a relative difference of 0.21 % in comparison to the mean of the data,
which is 65.52 % of REC14 (see summary statistics in Fig. 2a). The scatter plot
between the estimated and true values (Fig. 5b) shows that the regression line
y ¼ ax + b (red line in Fig. 5b) is fairly close to the first bisector line. This diagram
indicates that the grade model does not suffer from a substantial conditional bias.
The results highlight the strength of kriging to deal with samples of different
support. Practitioners often overlook the fact that the kriging system handles data of
different support (Journel 1986). In this paper, we used kriging considering the
support of the data to estimate grades with samples of different lengths in a mineral
deposit.
Using Samples of Unequal Length for Estimation 127
Fig. 3 Block model and samples. The lines represent the blocks while the points represent the
samples
5 Conclusions
The study shows that kriging considering the support of the data is suitable to
estimate grades in a mineral deposit using drill hole samples of different lengths.
The methodology was applied to a bauxite deposit. The resultant grade model is
visually consistent with the data and reproduced the trend of the data. In addition,
the grade model is accurate and does not have significant conditional bias.
128 M.A.A. Bassani and J.F.C.L. Costa
Fig. 5 Histogram of estimation errors (a) and scatter plot between estimated and true values (b)
Acknowledgements The authors would like to thank CAPES and CNPq (research agencies in
Brazil) for financial support.
References
Bassani MAA, Machado PL, Costa JFCL, Rubio RJH (2014) Using production data to improve
grade estimation in underground mining. Appl Earth Sci (Trans Inst Miner Metall Sect B) 122
(4):243–248
Bertoli O, Vann J, Dunham S (2003) Two-dimensional geostatistical methods—theory, practice
and a case study from the 1A shoot nickel deposit, Leinster, Western Australia. In: Proceedings
of the 5th international mining geology conference. The Australian Institute of Mining and
Metallurgy, Melbourne, pp 189–195
Deutsch CV, Journel AG (1998) GSLIB: geostatistical software library and user’s guide. Oxford
University Press, New York
Deutsch CV, Srinivasan S, Mo Y (1996) Geostatistical reservoir modeling accounting for preci-
sion and scale of seismic data. SPE annual technical conference. Denver, Colorado, USA,
6–9 October 1996
Goovaerts P (1997) Geostatistics for natural resource evaluation. Oxford University Press, Oxford
Hansen TM, Mosegaard K (2008) Visim: sequential simulation for linear inverse problems.
Comput Geosci 34(1):53–76
Isaaks HE, Srivastava MR (1989) An introduction to applied geostatistics. Oxford University
Press, Oxford
Journel AG (1986) Geostatistics: models and tools for the earth sciences. Math Geol 18:119–140
Journel AG, Huijbregts CJ (1978) Mining geostatistics. Academic, New York
Krige DG (1978) Lognormal–de Wijsian geostatistics for ore evaluation. South African Institute of
Mining and Metallurgy Monograph Series, Johannesburg, 50 pp
Kyriakidis P (2004) A geostatistical framework for area-to-point spatial interpolation. Geogr Anal
36:259–289
Liu Y, Journel AG (2009) A package for geostatistical integration of coarse and fine scale data.
Comput Geosci 35(3):527–547
130 M.A.A. Bassani and J.F.C.L. Costa
Marques DM, Hundelshaussen RR, Costa JFC (2014) The effect of accumulation in 2D estimates
in phosphatic ore. Revista Escola de Minas 67(4):431–437
Pardo-Igúzquiza E, Chica-Olmo M, Atkinson PM (2006) Downscaling cokriging for image
sharpening. Remote Sens Environ 102(1–2):86–98
Poggio L, Gimona A (2013) Modelling high resolution RS data with the aid of coarse resolution
data and ancillary data. Int J Appl Earth Observ Geoinf 23:360–371
Tran TT, Deutsch CV, Xie Y (2001) Direct geostatistical simulation with multiscale well, seismic,
and production data, SPE annual technical conference and exhibition. New Orleans, LA,
30 September–3 October 2001
Yao T, Journel AG (2000) Integrating seismic attribute maps and well logs for porosity modeling
in a west Texas carbonate reservoir: addressing the scale and precision problem. J Pet Sci Eng
28(1–2):65–79
New Approach to Recoverable Resource
Modelling: The Multivariate Case at
Olympic Dam
C. Badenhorst
Olympic Dam, BHP Billiton, 5000 Adelaide, SA, Australia
e-mail: [email protected]
S. O’Connell
Manager Resource Department, Olympic Dam, BHP Billiton, 5000 Adelaide, SA, Australia
e-mail: [email protected]
M. Rossi (*)
GeoSystems International Inc., 33431 Boca Raton, FL, USA
e-mail: [email protected]
1 Introduction
The Olympic Dam is Australia’s largest underground sublevel open stoping mine
producing around 10 Mt of ore per annum. The Olympic Dam mine has been in
production since 1988 using a standard mining method of mechanised sublevel
longhole open stoping (SLOS), with cemented aggregate backfill. The processing
plant is a fully integrated circuit that consists of autogenous grinding mills, flotation
circuits to recover copper concentrate and tailings leach circuits to recover uranium.
The copper concentrate is treated in an on-site direct-to-blister-furnace (DBF) smelter,
whilst an onsite refinery produces copper cathode and recovers gold and silver.
The Olympic Dam is a very large iron oxide-hosted Cu-U-Au-Ag ore deposit.
The deposit is hosted entirely within the Olympic Dam Breccia Complex (ODBC)
and is unconformably overlain by approximately 300 m of unmineralised, flat-
lying, sedimentary rocks. The deposit was discovered in the late 1970s and the
geology has been studied and described by numerous authors (Roberts and Hudson
1983; Ehrig et al. 2012). A key feature of the deposit is a central core of haematite–
quartz breccia largely devoid of copper and uranium mineralisation. In general, the
host breccias are more haematite-rich towards the centre of the ODBC and more
granitic at the margins. Including the sulphide minerals, there are more than
100 identified ore and gangue minerals. The most common minerals include
haematite, quartz, sericite, feldspar, chlorite, barite, fluorite, siderite, pyrite, chal-
copyrite, bornite and chalcocite. The three primary uranium minerals account for
less than 0.1 % of the total rock mass and occur as uraninite, coffinite and
brannerite.
The bulk of economic mineralisation is associated with sulphide-bearing hae-
matite-rich breccias. The majority of copper mineralisation occurs as chalcopyrite,
bornite and chalcocite and dominantly manifests as binary pairings of
chalcopyrite bornite and bornite chalcocite. Uranium, gold, silver and copper
minerals are all correlated to a statistically significant degree.
New Approach to Recoverable Resource Modelling: The Multivariate Case at. . . 133
The business case for recoverable resource modelling is simple: the true monetary
value of the Olympic Dam is highly leveraged to grade. For example, depending on
the specific mine plan, for every 10 % increase in grade in the early part of the plan,
there is a 25 % increase in annual cash flow. Around two thirds of the life-of-mine is
based on widely spaced drilling and is therefore a smoothed and under-
representative view of the grade that will be realised when close-spaced drilling
and high-resolution estimation is undertaken. Modelling the correct distribution of
grade is worth billions of dollars to the value of the Olympic Dam.
In the past 40 or so years, there have been many attempts at producing effective
recoverable resource models with varying degrees of success; some relevant dis-
cussions and case studies include (Journel and Huijbregts 1978; Journel and
Kyriakidis 2004; Assibey-Bonsu and Krige 1999; Rossi and Parker 1993; Abzalov
2006; Krige and Assibey-Bonsu 1999; Rossi and Deutsch 2014 and Roth and
Deraisme 2000).
In an environment of falling commodity prices, with a focus on reducing costs and
improving financial metrics of cash flow, IRR, NPV and capital efficiency, an equally
important consideration to costs is improving the revenue. The old adage, “grade is
king”, is never more true in this environment. For deposits that are relatively high
grade, and where the spatial continuity of that grade is amenable to selective mining,
increasing the grade for the same tonnage delivered to the mill is very effective in
increasing the margin on each unit of metal produced. The mining cost per tonne of ore
may increase but the cost per pound of copper can greatly decrease.
A common strategy at the Olympic Dam has been to increase metal production
by increasing process plant ore tonnage throughput. Unless there is latent capacity
in the plant, there is a large capital requirement in order to accommodate the
increase in tonnes. Alternatively, increasing metal throughput in the operation
can be achieved by raising the grade of the ore feed from the mine, which is
generally much less capital intensive since the plant is large and complex.
The Olympic Dam deposit is relatively high grade with contiguous zones of
2–4 % copper grade. This continuity of high grade gives the operation an opportu-
nity in the planning and mining processes to increase the grade delivered to the
processing plant. Relatively close 20 m-spaced lines of Measured Resource drilling
are required to define these contiguous high-grade ore shoots. The Olympic Dam
Inferred Resource is defined by relatively wide-spaced drilling at 100–250 m
notional centres. Indicated Resource is defined by both wide-spaced surface and
wide-spaced underground drilling at 70–100 m centres. Comprehensive drill spac-
ing studies (see later section) have back-tested the effect of resampling Measured
Resource (20 m drill spacing) at Inferred and Indicated drill spacings in the deposit.
These studies have demonstrated that a traditional Ordinary Kriging (OK) linear
estimate using this wide-spaced drilling information is neither globally accurate in
terms of tonnes and grade nor locally spatially accurate in representing the orien-
tation of the high-grade zones which are to be mined (Fig. 1). This holds true for
134 C. Badenhorst et al.
Fig. 1 Representative cross section through the Olympic Dam showing differences in copper
grade, modelled using Inferred (250 m spaced) drill hole information and Measured (20 m spaced)
drill hole information. Each model depicted above has a unique geological interpretation that is a
function of the data spacing
The RRM is integral to the overall strategy to increase grade and value at the
Olympic Dam. This strategy is aimed at maximising the contribution of ore grade to
the value proposition for the operation. Higher ore grades have a direct impact on
the revenue stream and are very effective in lowering the unit cost of metal
production. This strategy is based on six key focus areas that all contribute to
improving the value contribution of ore grade:
1. Resource modelling. By using conditional simulation to estimate grade in model
areas supported by wide-spaced drilling, a higher grade from improved granu-
larity can be realised in the mine planning process.
2. Reserves. By using a more sophisticated automated stope generation technique,
a higher design grade can be realised through a more efficient capture of high-
grade resource.
3. Resource utilisation. By including all potential (pre-resource) mineralisation and
Inferred Resource in the mine plan, the full possible value of the underground
operation can be assessed, allowing a higher cut-off to be applied, and therefore
a higher grade to be realised.
4. Cut-off grade optimisation. Historically the operation has been valued on a fixed
cut-off grade. The future plan is to implement a variable cut-off strategy, where
the cut-off varies between stopes and over time. This change in stope design
practice will allow the optimisation process to add value by promoting grade in
time, beyond the level achieved using a fixed cut-off.
5. Rejection of subgrade material from the ore stream. Separate removal of below
cut-off material and low-grade development to the surface raises the average ore
grade and increases the proportion of high-grade stope material in the ore stream.
6. Stope sequencing and scheduling. Higher value mine plans can be achieved by
promoting higher-grade stopes forward in the schedule and deferring lower-
grade stopes in time.
Both the absolute and relative contribution of these areas to the value proposition
will depend on the context of the specific mine plan that is being evaluated. The key
point is that for the full value of the resource to be realised, all six areas are
necessary. No modelling technique alone can realise the full value.
4 Fundamental Challenges
The development of a recoverable resource model for the Olympic Dam has been a
significant challenge. Some of these challenges are purely technical in nature, but
there are also fundamental challenges that inhibit the acceptance of this modelling
technique.
136 C. Badenhorst et al.
Ultimately, if recoverable resource models are not used to generate mine plans,
then they are of little practical use. One of the critical factors to success is the
support of mine planning engineers in understanding the value these models hold
and to use them for reserving. Without this, it is just another low-value, esoteric
exercise. To extract the maximum value from the model, the cultural inertia
manifests as conservatism and fear of failure and must change, or the upside in
grade will never be realised.
Gaining understanding and support from decision makers in the business has
been one of the single greatest challenges. Concepts that are basic to
geostatisticians and resource geologists (e.g. averaging, grade above cut-off,
change of support and volume variance) are obscure to decision makers and are
treated with suspicion and even derision. The lack of knowledge regarding these
elementary concepts in the industry, its leaders and some of the broader consultancy
community to which these leaders defer, are real and significant barriers to the
success of this work.
5 Technical Challenges
With regard to a deposit as large as the Olympic Dam, there are several technical
hurdles to overcome in attempting to produce a recoverable resource model using
conditional simulation.
The model covers an area of 6 3 km and extends to a depth of 2 km. It is
comprised of 5 10 5 m block support, from estimates using a number of
techniques for different areas, elements and minerals in the deposit. All elements
and minerals in areas classified as Measured Resource are estimated using Ordinary
Kriging (OK). Within areas classified as Inferred and Indicated Resource, Cu,
U3O8, Au, Ag, S and SG are estimated using conditional simulation. The copper
mineralogy is stoichiometrically calculated from the simulated Cu and S estimates.
All other elements and minerals are estimated using OK.
The dimensions of the deposit present the foremost challenge. At a 2.5 m node
spacing, a single model covering the entire deposit comprising a minimum of
14 data variables would require ~1.5 billion nodes and constitute a model file
size in excess of 700GB. A single drill hole file used for this simulation would
comprise of ~950,000 2.5 m length samples. Working with these files is imprac-
tical, and there is no simulation software capable of handling such a large model
file. To deal with this challenge, the deposit is split into 16 individual sub-models
and corresponding drill hole files, based on gross geological differences and
also the ability of the simulation algorithm and software to perform the task in
a reasonable amount of time. Once these individual models are amalgamated
into one model and regularised, there are no boundary artefacts evident in the
combined model. The characteristics of these individual models are listed in
Table 1.
New Approach to Recoverable Resource Modelling: The Multivariate Case at. . . 137
Fig. 2 Scatter plots of Cu v S and Cu v U3O8 for the chalcopyrite bornite (220) domain and the
bornite chalcocite (230) domain for drilling information in Model 18_19 of the Olympic Dam
A significant amount of work has been undertaken over the last few years to
establish the foundations of the recoverable resource model at the Olympic Dam.
Significant improvements in deterministic modelling of the key sulphide domains,
and work on previous simulation models of the Olympic Dam, have also played a
prominent role in determining the most efficient path forward. The key attributes
required of techniques and software are the usability of models by mine planning,
the speed and stability of the algorithm and software, the honouring of multivariate
relationships between simulated variables and the ability to integrate several
models from several simultaneous sources.
In these authors’ experiences, conditional simulation is the preferred technique
for underground mine planning work rather than other techniques such as Uniform
Conditioning (UC) and Multiple Indicator Kriging (MIK), because they suffer from
the same smoothing effect evident in all forms of Kriging and also produce models
that are more suitable for input into open pit mine planning software.
Following extensive trials over several years, the resource team has settled on
sequential Gaussian simulation (Isaaks 1990) as the most appropriate algorithm to
use. Stepwise conditional transformations (Leuangthong and Deutsch 2003) are
applied to account for the correlation between metals. Other algorithms, including
co-simulation with Bayesian updating (Journel 1988; Rossi and Badenhorst 2010)
and projection pursuit multivariate transformation (Friedman and Tukey 1974;
Barnett et al. 2012), have been extensively trialled and rejected either because of
poor reproducibility of input statistics, difficulty in replicating multivariate rela-
tionships observed in the drilling information, poor spatial match to input data and
the inability to deal with large datasets or a combination of all of the above. SGS
point simulation rather than direct block simulation (DBS) is preferred since
validation against the input data of DBS shows issues caused by the proportional
effect present in the original data, which has been shown to introduce biases in the
final output (Leuangthong 2006).
Both commercially available and open source software have been used with
varying degrees of success. Significant issues which were encountered included
software errors introduced by inconsistently incorporating GSLib-based routines
(Deutsch and Journel 1998), very slow operation with large datasets, stability issues
and data corruption or problems with integration of models from multiple users
working simultaneously on different parts of the deposit. It is clear from almost a
decade of work on this topic at the Olympic Dam that it is unlikely that standard
commercially available software packages can be used to produce recoverable
resource estimates that meet the mine planning and the corporate governance
requirements. Therefore, a modified version of the GSLib programs has been
adopted to complete the work.
In order to overcome some of the known issues with SGS, including edge effects
and grade blowouts between drill holes and at the edges of data, as well as to
increase the program’s functionality, the original GSLib FORTRAN code was
New Approach to Recoverable Resource Modelling: The Multivariate Case at. . . 139
modified in-house. One of the key enhancements made to the code was the
implementation of multiple search passes with the ability to have different param-
eters for each pass. For this reason, the program has been named Dynamic Search
SGS or DS-SGS. Other enhancements include the addition of domain control, not
drawing from the global distribution, multiple coarse/fine grid redefined in terms of
user input of x, y, z grid spacing, independent soft nodes and hard data search with
assign to nodes selected and minimum number of soft nodes selection with assign to
nodes option.
7 Case Study
The aim of the case study was to examine the change in grade-tonnage information
for an existing well-drilled 20 m-spaced Measured Resource area (hereafter
referred to as “truth”), by estimating and simulating it using wider Indicated and
Inferred Resource drill spacings. The goal is to determine if the simulation param-
eters using the wide-spaced drilling yielded results that matched the 20 m-spaced
Measured Resource truth, thus providing a mechanism to calibrate and validate
simulation models in other areas of the deposit.
In order to do this, a conventional drill hole spacing-type approach was
followed, but with a few differences. The starting point was to take a vast area of
Measured Resource that has been drilled from underground on 20 m-spaced lines
and treat this as the “truth”. This area constitutes 1.3 billion tonnes of Mineral
Resource and 300 Mt of Proved Ore Reserve, of which 170 Mt has been mined over
the last 27 years (Fig. 3).
Fig. 3 Location of Measured Resource in relation to the life-of-mine stope set (Note the extensive
spatial coverage of Measured Resource across the deposit)
140 C. Badenhorst et al.
Table 2 Resource classification and drill spacing used at the Olympic Dam
Resource classification Drill hole spacing Model block size
Inferred Resource 250 m 120 100 5 m
Indicated Resource 100 m 60 50 5 m
Measured Resource 20 m 5 10 5 m
A fixed volume of 20 m-spaced Measured Resource was originally drilled from the
surface on 70–100 m centres and is equivalent to Indicated Resource spacing. These
surface holes were resampled at 250 m-spaced centres to approximate the equiva-
lent Inferred Resource drill spacing. This resampling was undertaken 25 times by
randomly selecting holes at 250 m centres in order to capture the range of possible
outcomes from variations in the drilling grid. This process approximates a drilling
program that starts at 250 m spacing and is then progressively infill drilled to 100 m
and 20 m spacings. The outcome is a comparison of the nearest neighbour-
declustered drill hole data statistics and grade-tonnage curves of the different grid
spacings of 20 m and 100 m and the 25 iterations of the 250 m spacing. The
differences in results were alarming (Fig. 4).
There is considerable difference among the twenty-five 250 m iterations with
23 (92 %) of them under-calling the actual tonnes and grade by a considerable
margin, whereas the 100 m-spaced drilling dataset is much closer to the 20 m-
spaced dataset. In classical geostatistics, this is termed the information effect or,
more precisely, in this case, the misinformation effect.
New Approach to Recoverable Resource Modelling: The Multivariate Case at. . . 141
Fig. 4 Declustered sample grade-tonnage curves within the Measured Resource volume. The
differences in information for the 100 m dataset and 250 m resampled datasets are clearly
demonstrated. Cueq ¼ Cu + (U3O8*2.44) + (Au*0.881) + (Ag*0.0048) (S* 0.191)
Each of the 250 m drilling datasets had its own geological model and domains and
was used to create 25 separate resource models that were estimated using Ordinary
Kriging into block sizes of 120 100 5m. The same process was applied using
the 100 m-spaced dataset to a model with unique geological and domain charac-
teristics and a 60 50 5m block size. The resource grade-tonnage results mir-
rored the differences observed in the drilling data noted previously. The
26 estimated models (25 Inferred and 1 Indicated) were then converted to reserves
by the mine planning engineers running each through an automated and
semiautomated stope design process.
The results from the reserve grade-tonnage curves mirrored that of the resource
models and also the underlying drill hole data. That is, if the drill hole dataset was
the lowest of the group on the grade-tonnage curve, then the corresponding resource
and reserve models were also the lowest (Fig. 5).
A sequential Gaussian simulation was then developed using 3 of the 25 Inferred
datasets by choosing a low, mid- and high iteration of drill hole data. An additional
simulation model was also generated using the single 100 m-spaced Indicated
dataset. The results mirrored those of the drill holes and resource and reserve
142 C. Badenhorst et al.
Fig. 5 Comparison between ore reserve grade-tonnage curves using Measured Resource and
resampled Indicated and Inferred Resource models. Cueq ¼ Cu + (U3O8*2.44) + (Au*0.881)
+ (Ag*0.0048) (S* 0.191)
models discussed earlier. The results are very clear; the underlying data exerts the
strongest control on whether or not the 20 m-spaced Measured Resource grade-
tonnage result can be achieved from wide-spaced Inferred or Indicated Resource
datasets. The conclusion is that all resource and reserve models, be they estimated
or simulated, are strongly anchored to the starting drill hole data. Locally, different
mine areas typically behaved differently from iteration to iteration. Most of the
25 Inferred Resource datasets were low with respect to the 20 m-spaced Measured
Resource “truth”, but a few were higher. This is a random feature and is attributed
to chance interactions with geological influences.
Fig. 6 Comparison between reserve stopes in the Measured Resource volume at the same cut-off,
using (a) Measured Resource model and (b) one of the 25 resampled Inferred Resource models
(Note how poorly the resource estimate based on Inferred Resource drill spacing estimates the
mineable tonnes for the same mining area)
The results clearly indicate that as the block size increases, the average grade
above the cut-off substantially decreases for the same tonnage of material. This
decrease in relative terms, which comes about solely from the change of support, is
an order of magnitude greater than the variability between individual simulation
realisations at the small block scale (Fig. 7, Table 3).
Thus, for recoverable resource modelling of this large deposit, the number of
realisations in a simulation is not critical. A single realisation, chosen at random
from a handful, is more than adequate to deal with the change of support problem.
However, to emphasise the difference between the realisations at the local scale, a
number of simulation models should be evaluated by the mine planning engineers.
Moreover, as the size and area of the simulation increase, there is no global low,
mid- or high realisation; this is a misnomer. One realisation that is lowest in one
particular area may be the highest in another. Globally, there is no difference
between realisations, so any realisation could be used to develop a mine plan as
well as the basis for Mineral Resource declaration purposes.
It should be noted that this local difference between simulation realisations does
not diminish the usefulness or applicability of the technique. Furthermore, the OK
version of the model is just as incorrect locally; it is just that it is almost always
overlooked. Inferred Resource is inferred because there is significant local uncer-
tainty, regardless of the modelling technique used. An estimation method sensu
146
Table 3 Change in average Cu reserve grade above a 1.5 % Cu cut-off for 100 simulation realisations at varying support sizes
Relative change in ave. grade
Block size (m) 5 10 5 60 50 5 120 100 5 120 100 5 to 60 50 5 to 120 100 5 to
5 10 5 5 10 5 60 50 5
Drill spacing 20 m 70–100 m 100–250 m
Average grade 2.65 % 1.86 % 1.66 % 60 % 43 % 12 %
Spread of ~1 % ~1 % ~1 %
realisations
C. Badenhorst et al.
New Approach to Recoverable Resource Modelling: The Multivariate Case at. . . 147
stricto does not significantly alter this fact. The only way to reduce the local
uncertainty is to gather more information through drilling.
The E-type estimate is often suggested as a suitable estimate for recoverable
resources. The E-type estimate of 5 10 5 blocks is very smooth and negates the
change of support correction that is sought (see Fig. 7). The variance of the E-type
under-calls the actual variance through the mere averaging process, and thus it is
inappropriate as a representation of the true grade-tonnage relationship.
The purpose of the recoverable resource modelling work is to produce a block
model suitable for life-of-mine underground mine planning and financial valuation,
public declaration of resource and reserves and drill targeting by mine geology. The
absolute accuracy of the spatial location of stopes in the RRM is not as important
for underground mine planning purposes as the representation of the spatial geom-
etry of ore and the grade architecture.
There is a strong and compelling business case for focussing on grade improve-
ments at the Olympic Dam operation. Two thirds of the life-of-mine reserves are
based on wide-spaced drilling, which under-calls the grade that will be realised
when close-spaced drilling and high-resolution estimation is undertaken. Modelling
the expected grade is worth billions of dollars to the operation.
Conditional simulation that exhaustively validate, both visually and statistically,
against the drill hole data is the only technique that:
• Avoids smoothing by reproducing the original data variance (high granularity
models)
• Estimates the tonnes and grade at the time of mining for any drill spacing
• Produces a better local spatial model that can be used for practical geological,
mine planning and financial valuation purposes
The recoverable resource model is the main element that underpins a six-point
strategy aimed at increasing the grade of the ore feed from the mine. The strategy
includes improvements to mine planning and mining practices which are all
required to realise the full impact of value improvement.
The development of a recoverable resource model has not been without chal-
lenges. The most significant is communicating the elementary resource concepts
and principles to decision makers whom lack the specialised technical skills
required to fully appreciate the importance of recoverable resource modelling.
This is possibly the single largest inhibitor to a successful recoverable resource
model.
Technical challenges for the Olympic Dam recoverable resource model are
mostly about the size of the orebody and the inability of available software to
address the large file sizes required. The only practical solution is to modify the
existing GSLib SGS routine and create a fit-for-purpose algorithm. Reproduction of
148 C. Badenhorst et al.
the correlations among multiple metals pushes the limits of application of the
Stepwise Conditioning transform method employed.
The application of a single realisation for recoverable resources is outside the
usual scope for conditional simulation; see, for example, Goovaerts (1997),
Dimitrakopoulos (1999), Krige et al. (2004) and Van Brunt and Rossi (1999). It
is therefore important to realise that in this case, a single realisation provides the
additional value that the operation requires and why the usual objections to using a
single realisation are not applicable.
The drill hole spacing case study demonstrably shows the impact of the
misinformation effect, a term coined to describe all the unknowns in a resource
estimate. All estimates, regardless of whether they are performed using traditional
linear or non-linear techniques, are highly leveraged to the starting dataset. No
amount of “alternate modelling”, multiple simulation realisations or range analysis
can fully describe the uncertainty inherent in the starting dataset.
Regardless of the drilling dataset, the best that can be done by the practitioner is
to honour the available information. This approach will not capture the range of
uncertainty, and thus the Olympic Dam Mine Planning Department will employ the
use of modifying factors to account for the misinformation effect.
The most common applications of conditional simulation require many
realisations to be evaluated. This is not the case for change of support modelling
at the Olympic Dam, and perhaps on many other large deposits, since the global
differences between realisations are an order of magnitude less than the difference
between change of support models. However, there can be significant local differ-
ences between realisations and to quantify this impact; a handful of realisations
should be given to mine planning for evaluation.
The internal company governance requirements have made it harder to produce
a recoverable resource model as the basis for a Mineral Resource declaration. It is
expected that the nontechnical hurdles that these requirements bring about may be
overcome as management is further educated in the value of using a conditional
simulation for resource estimation.
References
Abzalov MZ (2006) Localized uniform conditioning (LUC): a new approach for direct modeling
of small blocks. Math J 38:393–411
Assibey-Bonsu W, Krige DG (1999) Use of direct and indirect distributions of selective mining
units for estimation of recoverable resource/reserves for new mining projects. APCOM’99
Symposium. Colorado School of Mines, Golden, CO
New Approach to Recoverable Resource Modelling: The Multivariate Case at. . . 149
Barnett RM, Manchuk JG, Deutsch CV (2012) Projection pursuit multivariate transform. Centre
for Computational Geostatistics, Edmonton
Deutsch CV, Journel AG (1998) GSLIB: geostatistical software library and user’s guide. Oxford
University Press, New York
Dimitrakopoulos R (1999) Conditional simulations: tools for modeling uncertainty in open pit
optimization, Optimizing with Whittle. Whittle Programming Pty Ltd, Perth, pp 31–42
Ehrig K, McPhie J, Kamenetsky V (2012) Geology and mineralogical zonation of the Olympic
Dam iron oxide Cu-U-Au-Ag deposit, South Australia. Society of Economic Geologists,
Special Publication 16, pp 237–267
Friedman J, Tukey J (1974) A projection pursuit algorithm for exploratory data analysis. IEEE
C-23:881–890
Goovaerts P (1997) Geostatistics for natural resources evaluation. Oxford University Press,
New York
Isaaks E (1990) The application of Monte Carlo methods to the analysis of spatially correlated
data. Stanford University PhD thesis
Journel AG (1988) Fundamentals of geostatistics in five lessons. Stanford Center for Reservoir
Forecasting, Stanford
Journel AG, Huijbregts CJ (1978) Mining geostatistics. Academic, New York
Journel AG, Kyriakidis P (2004) Evaluation of mineral reserves: a simulation approach. Oxford
University Press, New York
Krige DG, Assibey-Bonsu W (1999) Practical problems in the estimation of recoverable reserves
when using simulation or block kriging techniques, Optimizing with Whittle. Whittle Pro-
gramming Pty Ltd., Perth
Krige DG, Assibey-Bonsu W, Tolmay L (2004) Post processing of SK estimators and simulations
for assessment of recoverable resources and reserves for South African gold mines,
Geostatistics Banff 2004. Springer, Banff, pp 375–386
Leuangthong O (2006) The promises and pitfalls of direct simulation. Center for Computational
Geostatistics/University of Alberta, Edmonton
Leuangthong O, Deutsch CV (2003) Stepwise conditional transformation for simulation of
multiple variables. Math Geol 35:155–173
Roberts DE, Hudson GR (1983) The Olympic Dam copper-uranium-gold deposit, Roxby Downs,
South Australia. Bull Soc Econ Geol 78:799–882
Rossi ME, Badenhorst C (2010) Collocated co-simulation with multivariate bayesian bpdating: a
case study on the Olympic Dam deposit. In: Castor R et al (eds) 4th international conference on
mining innovation, MINIM. Santiago, Chile, pp 385–394
Rossi ME, Deutsch CV (2014) Mineral resource estimation. Springer, Dordrecht
Rossi ME, Parker HM (1993) Estimating recoverable reserves: is it hopeless? Geostatistics for the
next century. McGill University, Montreal
Roth C, Deraisme J (2000) The information effect and estimating recoverable reserves. In: WJ
Kleingeld, DG Krige (eds) Proceedings of the sixth international geostatistics congress. Cape
Town, South Africa, pp 776–787
Van Brunt BH, Rossi ME (1999) Mine planning under uncertainty constraints, Optimizing with
Whittle. Whittle Programming Pty Ltd., Perth
Comparison of Two Multivariate Grade
Simulation Approaches on an Iron Oxide
Copper-Gold Deposit
Antonio Cortes
1 Introduction
A. Cortes (*)
AMEC Foster Wheeler plc, av Apoquindo 3846, Santiago, Chile
e-mail: [email protected]
The main mineralization control is lithology and was used to define the estima-
tion domains. Ordinary kriging was used to estimate the grades within interpreted
estimation domains. The estimated variables are total copper (CuT), acid-soluble
copper (CuS), iron (Fe), and gold (Au). These four variables were estimated
independently without considering the correlations among them. Mineral resources
were classified based on a drill hole spacing study that quantified relative confi-
dence intervals on production increment estimates versus drill hole spacing. Rea-
sonable prospects for economic extraction were obtained by restricting the
resources to within a reasonable open-pit shell. No particular work was done to
assess the geological and grade risk through conditional simulations.
This paper describes briefly both coregionalization and projection pursuit mul-
tivariate transform (PPMT) simulations and presents the results obtained from the
given data set including some validation focused on the correlations among the
variables. Correlations among simulated variables in the Gaussian space and raw
data are compared with the correlations of the input values. Finally, conclusions and
recommendations are presented.
The main objectives of the study are:
1. To describe briefly for both coregionalization and PPMT, the main practical
implementation steps on a multivariate database
2. To compare the two multivariate simulation approaches in terms of reproduction
of the correlations observed on the input data
3. Given the results, to discuss the usefulness of simulation in particular for this
project given the information available: heterotopic database, geological model-
ing and parameters used for a previous resource estimation, and some conceptual
information on the projected open pit
The paper does not take into account the geological uncertainty. Various
assumptions were made, and a deeper analysis and improvements are
recommended for future work that will contribute to the discussion that this
document aims to generate. The final comparisons in terms of reproduction of the
correlations (against the input data) are presented for an individual realization of
both approaches. However, consistency tests were performed for a set of
realizations.
2 Approaches Compared
The case study includes an initial approach using sequential Gaussian cosimulation
with a linear coregionalization model.
Comparison of Two Multivariate Grade Simulation Approaches on an Iron Oxide. . . 153
5 Deposit Outline
The deposit is an iron oxide copper-gold (IOCG) ore body located in the Atacama
Desert in the north of Chile. The deposit extends approximately 1.5 km N-S,
0.75 km E-W, and 0.45 km vertically. The CuT and Fe mineralization is mainly
154 A. Cortes
The database for the exercise includes 140 drill holes with assays values for CuT,
CuS, Fe, and Au. The database is heterotopic and has the maximum number of
samples for CuT and the lowest number of samples for Au. The exploratory data
analysis (EDA) envelope corresponds to the main mineralized domain which
corresponds to specularite breccia (BES) – usually hosted by the chalcopyrite
(Cpy) mineralization. Figure 1 shows this envelope. The sample values were
capped to limit the impact of outliers and then regularized (2 m composites).
7 Declustering
The declustered data were transformed to standard Gaussian values using the
modeled transform function, one variable at a time. This function makes it possible
to transform grades (Z) to Gaussian values (Y) and vice versa. Once the variables
were transformed, basic statistics and scatterplots were produced to evaluate the
correlation coefficients and the shape of the clouds of points. At first sight the
scatterplots between Gaussian variables look reasonably bi-Gaussian with an ellip-
tic shape and only some outlier points related to domaining. Figure 2 shows some
examples of scatterplots of the input transformed data.
Two theoretical checks of by-Gaussianity were performed: h-scatterplots
between normal pairs ½Y ðxÞ, Y ðx þ hÞ; and ratio of the square root of the semi-
variogram and the order 1 variogram (madogram). The h-scatterplots have shapes
Comparison of Two Multivariate Grade Simulation Approaches on an Iron Oxide. . . 155
Fig. 1 Orthogonal views of EDA envelope (BES domain) and CuT samples used in the study
with poor correlation (tending to circular shapes). The ratios are constant for all
distances as showed in Fig. 3 for CuT and CuS as examples.
9 Gaussian Simulation
The sequential Gaussian simulation was performed over a grid restricted by the
EDA domain, while search parameters for conditioning data were based on the
parameters used in the resource estimation of this domain (search ellipsoid up to
200 m and a maximum number of samples of 20). A set of 100 realizations were run
and then validations performed.
156 A. Cortes
cutc_gauss cutc_gauss
-3 -2 -1 0 1 2 3 4 5 -3 -2 -1 0 1 2 3 4 5
4 4 4 4
rho=0.693 rho=0.874
3 3 3 3
2 2
2 2
1 1
auc_gauss
auc_gauss
cus_gauss
cus_gauss
1 1
0 0
0 0
-1 -1
-1 -1
-2 -2
-2 -2
-3 -3
-3 -3 -4 -4
-3 -2 -1 0 1 2 3 4 5 -3 -2 -1 0 1 2 3 4 5
cutc_gauss cutc_gauss
cutc_gauss cus_gauss
-3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3
3 3 3 3
rho=0.678 rho=0.604
2 2 2 2
1 1 1 1
fec_gauss
fec_gauss
fec_gauss
fec_gauss
0 0 0 0
-1 -1 -1 -1
-2 -2 -2 -2
-3 -3 -3 -3
-3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3
cutc_gauss cus_gauss
Fig. 3 Sqrt of semi-variogram/madogram for CuT (left) and CuS (right) in Gaussian space
Comparison of Two Multivariate Grade Simulation Approaches on an Iron Oxide. . . 157
10 Variography
The single and cross-variograms of the Gaussian values were calculated and
modeled along the three main geological directions. The largest range (up to
200–250 m) and intermediate and shortest ranges were along the N-S, E-W, and
vertical directions, respectively.
Figure 4 shows the single and cross variogram models for the four variables.
Below the experimental curves, histograms of the number of pairs are presented.
Fig. 4 Experimental and modeled variograms of CuT, CuS, Fe, and Au Gaussian values
158 A. Cortes
Var_Omni_cut_gauss Var_Omni_cus_gauss
1.1 Samples variogram 1.1 Samples variogram
1 1
0.9 0.9
Gama (h)
Gama (h)
0.8 0.8
10 realizations
0.7 0.7
10 realizations
0.6 0.6
0.5 0.5
0 100 200 300 400 500 0 100 200 300 400 500
Fig. 5 Reproduction of CuT and CuS variogram models in the Gaussian space
11 Validation of Results
A set of validations in the Gaussian space were completed using the realizations.
Histograms and basic statistics of the simulated Gaussian values in general show
mean values very close to zero and variances slightly lower than 1.
Variograms of the simulated Gaussian values were also produced. As an exam-
ple, Fig. 5 shows the CuT and CuS (Gaussian values) experimental variograms;
brown denotes the variogram of the conditioning sample values, and blue denotes
the variogram for ten different realizations.
Figure 6 shows the means by slices (100 m along easting, 100 m along northing,
50 m along elevation) for samples, NN model, and the E-type (from the all
realizations) in the Gaussian space. The E-type has lower variance than the samples
and varies locally from the declustered NN grades. The simple kriging
(SK) algorithm used by the simulation could produce a portion of this due to the
attraction to the mean, especially where there are less data. An example is around
coordinate N7,079,700 where NN differs significantly from E-type; only 15 samples
were available in this 100 m slice.
To assess the obtained correlations and check them against those of the input data,
scatterplots were produced for a number of realizations. Individually they show a
cloud that is more homogeneous than of the input data as can be seen in the example
in Fig. 7, which shows the scatterplots of the Gaussian values of the fifth realization.
This cloud of points shows a good correspondence with the input data presented
earlier in Fig. 2. A comparison of the simulated correlations in the Gaussian space
and real data space with the projection pursuit approach is presented later.
Comparison of Two Multivariate Grade Simulation Approaches on an Iron Oxide. . . 159
0.50
0.50
0.50
0.50
0.00
0.00
0.00
0.00
Grade
Grade
-0.50
-0.50
-0.50
-0.50
384250 384500 384750 7079000 7079500 7080000 250 500
Easting Northing Elevation
cut_gauss nn sim_etype cut_cmp
1.00
1.00
1.00
Grade
Grade
0.00
0.00
0.00
0.00
-1.00
-1.00
-1.00
-1.00
Fig. 6 Swath plots of samples, NN, and E-type on the Gaussian values
The simulated Gaussian values were back-transformed to the original space and
basic statistics calculated (particularly the mean grade and the standard deviation).
Table 2 shows the main statistics from back-transformed simulated values. These
are comparable with the statistics from the original data (showed previously in
Table 1).
The database used for this exercise is heterotopic as showed in Table 1. This is an
issue for the projection pursuit multivariate transform (PPMT), which can only be
applied on homotopic databases. The exclusion of incomplete observations is not
desirable because it means loss of information and an unfair comparison against the
160 A. Cortes
simu_cut_gauss3 [00005]
2
simu_au_gauss3 [00005]
simu_au_gauss3 [00005]
2 2 2
1 1 1 1
0 0
0 0
-1 -1
-1 -1
-2 -2
-2 -2
-3 -3
-3 -3
-4 -4
-3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3
simu_cut_gauss3 [00005] simu_cut_gauss3 [00005]
simu_cut_gauss3[00005] simu_cus_gauss3[00005]
-3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3 4
4 4 4 4
rho=0.687 rho=0.577
3 3 3 3
simu_fe_gauss3 [00005]
simu_fe_gauss3 [00005]
simu_fe_gauss3 [00005]
simu_fe_gauss3 [00005]
2 2 2 2
1 1 1 1
0 0 0 0
-1 -1 -1 -1
-2 -2 -2 -2
-3 -3 -3 -3
-3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3 4
simu_cut_gauss3 [00005] simu_cus_gauss3 [00005]
Cut Cut
-3 -2 -1 0 1 2 3 4 5 -3 -2 -1 0 1 2 3 4 5
4 4 4 4
rho=0.670 rho=0.866
3 3 3
3
2 2
2 2
1 1
1 1
0 0
Cus
Cus
Au
Au
0 0
-1 -1
-1 -1
-2 -2
-2 -2
-3 -3
-3 -3 -4 -4
-3 -2 -1 0 1 2 3 4 5 -3 -2 -1 0 1 2 3 4 5
Cut Cut
Cut Cus
-3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3
3 rho=0.518 3 3 rho=0.445 3
2 2 2 2
1 1 1 1
Fe
Fe
Fe
Fe
0 0 0 0
-1 -1 -1 -1
-2 -2 -2 -2
-3 -3 -3 -3
-3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3
Cut Cus
Fig. 8 Scatterplots between different variables in Gaussian space (samples + imputed data)
14.2 PPMT
PPMT theory and its practical application (examples and software) are given in
publications referenced earlier. For this exercise, the methodology included the
162 A. Cortes
au_orig fe_orig
-2 -1 0 1 2 -2 -1 0 1 2
Type : Q-Q Plot Type : Q-Q Plot
2 2 2 2
1 1 1 1
au_imp
au_imp
fe_imp
fe_imp
0 0
0 0
-1 -1
-1 -1
-2 -2
-2 -2
-2 -1 0 1 2 -2 -1 0 1 2
au_orig fe_orig
Fig. 9 Q-Q Plots for Au (Left) and Fe (right) between original (horizontal axis) and imputed data
(vertical axis)
Var_Omni_cus_gauss Var_Omni_fe_gauss
1.3 1.3
1.1 Orginal values 1.1 Orginal values
Gama (h)
0.9 0.9
Gama (h)
0.7 0.7
0.5 0.5
10 orig+imput (3% of total) 10 orig+imput (45% of total)
0.3 0.3
0.1 0.1
0 100 200 300 400 500 0 100 200 300 400 500
Distance (m) Distance (m)
Var_Omni_Au_gauss
1.3
1.1 Orginal values
0.9
Gama (h)
0.7
0.5
10 orig+imput (32% of total)
0.3
0.1
0 100 200 300 400 500
Distance (m)
Fig. 10 Variography of original values (red) and original plus percentage of imputed values
(blue)
following steps: running the PPMT algorithm (ppmt.exe) whose input corresponds
to the input database (isotopic) and the output corresponds to the transformed
variables and a transformation table (binary) needed later when the back-
transformed algorithm is used. Transformed variables were then simulated by
sgsim (sgsim.exe). Finally, these simulated values are back-transformed using the
back-transformed PPMT algorithm (ppmt_b.exe).
Figure 11 shows the scatterplots between variables after 50 iterations of the
PPMT algorithm. The scatterplots show uncorrelated standard Gaussian variables.
Using this dataset with no correlation between the variables, one realization was
Comparison of Two Multivariate Grade Simulation Approaches on an Iron Oxide. . . 163
Fig. 11 Scatterplots of the PPMT transformed data (top left, CuT vs CuS; top right, CuT vs Fe;
below, CuT vs Au)
performed, and then the back correlation algorithm was used to introduce the
correlation between the simulated variables.
Once the back transformation PPMT_B is performed, the results (simulated values
in the real space) can be checked against the original database, focusing on the
examination of the correlations between variables in real space. To consider the
spatial cross-correlations after back-transforming, some cross-variograms for data
and simulated values were calculated. In general, both methods, linear coregiona-
lization and PPMT, reproduce the spatial cross-correlations observed in the data.
Figures 12, 13, and 14 show scatterplots for the different approaches for CuT and
CuS, CuT and Fe, and CuT and Au, respectively. It can be seen that for CuT-CuS,
differences are small and both methods reproduce the primary data correlation. The
main difference is in CuS grades, where there is a difference between the clouds of
points around 2 % CuT: the shape of the cloud of points seems better reproduced
using PPMT.
164 A. Cortes
cutc
0 1 2 3 4 5
3 rho=0.404 3
2 2
cus
cus
1 1
0 0
0 1 2 3 4 5
cutc
3 rho=0.447 3 3 rho=0.514 3
cus_simu_gauss [00005]
cus_simu_gauss [00005]
2 2 2 2
cus2
cus2
1 1 1 1
0 0 0 0
0 1 2 3 4 5 0 1 2 3 4 5
cut2 cut_simu_gauss [00005]
Fig. 12 CuT vs CuT scatterplot comparison (top, samples; below left, PPMT; below right,
cosimulation)
15 Conclusions
cutc
0 1 2 3 4 5
rho=0.662
60 60
50 50
40 40
fec
fec
30 30
20 20
10 10
0 0
0 1 2 3 4 5
cutc
50 50 50 50
fe_simu_gauss [00005]
fe_simu_gauss [00005]
40 40 40 40
fe
fe
30 30 30 30
20 20 20 20
10 10 10 10
0 0
0 1 2 3 4 5 0 1 2 3 4 5
cut cut_simu_gauss [00005]
Fig. 13 CuT vs Fe scatterplot comparison between (top, samples; below left, PPMT; below right,
cosimulation)
Acknowledgments Thanks to Georges Verly and Harry Parker for their valuable suggestions,
Amec Foster Wheeler for the time for this paper, Geovariances for providing me a temporary Isatis
license during the practical work, and CCG for providing the material and software to perform
application exercises on the server.
166 A. Cortes
cutc
0 1 2 3 4 5
auc
auc
0.5 0.5
0.0 0.0
0 1 2 3 4 5
cutc
au_simu_gauss [00005]
0.5 0.5
au
0.5 0.5
au
0 1 2 3 4 5 0 1 2 3 4 5
cut cut_simu_gauss [00005]
Fig. 14 CuT vs Au scatterplot comparison (top, samples; below left, PPMT; below right,
cosimulation)
References
Alabert FG (1987) Stochastic imaging of spatial distributions using hard and soft information. M.
Sc. Thesis. Stanford U., p 197
Barnett R, Deutsch C (2013) Imputation of geologic data. Paper 102. CCG Annual Report 15.
Barnett R, Deutsch C (2015) Guide to multivariate modeling with the PPMT. CCG, Guidebook,
Series Vol 20
Barnett R, Manchuk J, Deutsch C (2012) Projection pursuit multivariate transform. Paper 103.
CCG Annual Report 14
Barnett R, Manchuk J, Deutsch C (2013) Advances in the projection pursuit multivariate trans-
form. Paper 106. CCG Annual Report 15
Gomez-Hernandez J, Journel AG (1993) Joint sequential simulation of multigaussian fields. In:
Soares A (ed) Geostatistics Troia 92, part 1. Kluvert Academic Publishers, Dordrecht, pp 85–94
Isaaks E (1990) The application of Monte Carlo methods to the analysis of spatially correlated
data. Ph.D. Thesis. Standford U., p 213
Journel AG (1974) Simulations Conditionelles de Gisements Miniers – Théorie et Pratique.
Docteur Ingenieur Thesis. Nancy U., p 110
Journel AG, Huijbregts CJ (1978) Mining geostatistics. Academic, London, p 600, 2nd printing
revised
Verly G (1993) Sequential gaussian co-simulation: a simulation method integrating several types
of information. In: Soares A (ed) Geostatistics Troia 92, part 1. Kluver Academic Publishers,
Dordrecht, pp 543–554
Complexities in the Geostatistical Estimation
of Besshi-Type Mineral Deposits
on the Northwest of Pinar del Rı́o, Cuba
Abstract Geostatistics resources estimation of the copper (Cu) grade values was
carried out in the “Hierro Mantua” mineral deposit, which is located at the north-
west of Pinar del Rı́o province, Cuba. The geologic complexities in the region of the
deposit indicate the nonexistence of homogeneity in the Cu values. The structural
analysis showed a high asymmetric distribution in the variable studied. The non-
existence of normality was verified by different mean and median values, a coef-
ficient of variability greater than one, and the moving windows statistics of the
mean was different. Under the previous conditions, the data was log transformed to
assure the necessary stationarity in them and consequently to achieve an adequate
accuracy in the resources estimation, using a rational selective mining unit (SMU).
The log-transformed data revealed a homogeneous behavior in Cu grade values,
demonstrated by better results in basic and moving windows statistics.
Semivariograms showed defined structures with anisotropy in the 0 and 90
directions (considering 0 to the north). To apply the lognormal kriging estimation
is the main objective of this work, because of the complexities of the geology in the
studied area.
A. Dı́az-Carmona (*)
Department of Geology, University of Pinar del Rı́o, Pinar del Rı́o, Cuba
e-mail: [email protected]
J.Q. Cuador-Gil
Department of Physics, University of Pinar del Rı́o, Pinar del Rı́o, Cuba
e-mail: [email protected]
F. Giménez-Palomares
University Institute of Pure and Applied Mathematics, Polytechnic University of Valencia,
Valencia, Spain
e-mail: [email protected]
J.A. Monosoriu-Serra
Technical Superior School of the Design, Department of Physics Applied, Polytechnic
University of Valencia, Valencia, Spain
e-mail: [email protected]
1 Introduction
On the northwest of Cuba, at the province of “Pinar del Rı́o,” exists a group of
mineral deposits and other small accumulations that are classified as Besshi type.
The most important are “Hierro Mantua,” “Unión 1,” “Juan Manuel,” “Unión 2,”
and “Flanco Este.” Some of them are in the feasibility stage for future exploitation.
They associate to rocks of the Esperanza Formation (J3ox – K1ne), which is consti-
tuted by sequences of silts, carbonate, and volcanic rocks of basic composition
originated in a marginal basin. The rocks that prevail are calcareous, gritty,
argillites, concordant bodies, and diabases and gabbros sub-concordant representa-
tives of a submarine magmatism.
The mineral bodies form long lenses, which follow an NE direction, parallel to
the coast. The relief altitudes are between 60 and 75 m above sea level. These
bodies have many outcrops, concordant with the stockwork. Some secondary
enrichment processes have taken place in some of them, as a result of a process
of extreme oxidation, lixiviation, and concentration of cupper (Cu) in the primary
sulfurous mineralization of volcanogenic-sedimentary origin.
Historically, classic methods were used for resources estimation in those types of
mineral deposits. As it is known, these methods do not guarantee the necessary
precision, mainly because of the geological complexities in the region. In order to
achieve an appropriate accuracy of the resources estimation in “Hierro Mantua”
mineral deposit, two objectives are established: (1) to apply a methodology for the
rational SMU determination, which is used in quantifying the mineral resources in
the deposit studied, and (2) to apply a nonlinear geostatistics method, lognormal
kriging, for resources estimation because of the presence of Cu values asymmetric
distribution. Particularly, the occurrence of a few high Cu values in “Hierro
Mantua” mineral deposit makes it very difficult to handle data with lognormal
distribution. This mineral deposit was taken as an example in the present research.
Regarding lognormal kriging, there have been some publications related with
this estimator, classified as nonlinear. Some of the most important are Matheron
(1974), Marechal (1974), Rendu (1979), Journel (1980), Krige (1981), Dowd
(1982), Krige and Magri (1982), Thurston and Armstrong (1987), Armstrong and
Boufassa (1988), Rivoirard (1990), Marcotte and Groleau (1997), Lee and Ellis
(1997), Roth (1988), Clark (1998), Cressie (2006), Yamamoto (2007), Yamamoto
and Furuie (2010), and Paul and Cressie (2011).
Some ideas included in those works are “. . .the sensitivity of the experimental
semivariogram with regard to the few high values (Journel 1980).” High values
cause the deformity of the semivariograms that prevents the use of linear kriging.
According to Clark (1998), “. . .the potential problem in lognormal Kriging is in the
“back transformation” of the logarithmic estimates to the original sample values
scale, there appears to be some disagreement in the general geostatistical literature
as to how this back-transformation should be carried out.” In this sense, the
equation for back transformation of block kriging values is not easily found in the
literature. A review of lognormal estimators in an orderly way was presented in
Complexities in the Geostatistical Estimation of Besshi-Type Mineral. . . 169
Rivoirard (1990), but he concluded that the equations of back transformation for the
block estimation are not theoretically optimal. Roth (1998) concluded that “. . .the
practitioner must decide whether the lognormal Kriging estimator is suitable for
local estimator.” Cressie (2006) stated “. . .data that exhibit skewness may be
successfully modeled as coming from a lognormal spatial process.” In a practical
way, back transformation is achieved by the exponential of the kriging estimate
plus a non-bias term (Yamamoto 2007). Finally, in this research, Cu values are
estimated following ideas presented by Rivoirard (1990), mainly the back-
transformation equation, Eq. 6 on page 218. The precision of this method has
been proved by some authors in different mineral deposits and other studies when
skewness distributions have been found.
2 Data Description
The quality of the primary information is one of the most important elements for the
resources estimation using kriging. This assumption determines the quality of the
results (Armstrong and Carignan 1997). The “Hierro Mantua” deposit database
used in the present research work was validated in 2008 by the Coffey Mining
Company, for request of RSG Global to the “Cobre Mantua S.A. Company” (Arce-
Blanco and Santana-Okamoto 2008). An initial exploration of the samples was
performed, whose lengths vary between 0.05 and 4.65 m, with an average of
1.22 m. A regularization of the data support was needed; 1 m length composites
inside the mineral body were calculated. Cu values from 318 wells, regularly
distributed, in two 50 50 m and 100 100 m sampling networks were analyzed,
creating 5,192 composites of 1 m length.
Table 1 shows the descriptive statistics of Cu values. The Cu mean and median
values are very different and very small compared with the maximum value; the
coefficient of variation is greater than 1. So, the data distribution is asymmetric,
which can be seen in the histogram (Fig. 1). The Kolmogorov-Smirnov normality
test value is 0.27, which is greater than 0.05. The result shows the non-normal
behavior of the Cu values. For this reason, the linear geostatistics procedures cannot
be applied; therefore, it is more appropriate to use nonlinear geostatistics methods
(Journel and Huijbregts 1978; Isaaks and Srivastava 1989; Rivoirard 1990).
170 A. Dı́az-Carmona et al.
Histogram
3500
3000
Samples number
2500
2000
1500
1000
500
0
0.0 2.3 4.6 6.9 9.2 11.6 13.9 16.2 18.5 20.8 23.1 25.4 27.7 30.0 32.3
Cu%
Fig. 1 Histogram of Cu values
stationary, and the lognormal kriging can be applied (Journel and Huijbregts 1978;
Isaaks and Srivastava 1989; Cressie 1993).
The maximum separation distance between the points of the data is 650 m in
deposit strike direction, 230 m in the dip direction, and 60 m in the downhole
direction. The lag and directions used for the semivariogram calculation are
presented in Table 3. The variability was studied approximately up to half of the
maximum distance for each direction.
The experimental semivariogram and the fitted models are shown in Fig. 3. All
semivariograms were computed in Gemcom software. Geometric anisotropy can be
observed. The fit of the experimental semivariograms was performed by an imbri-
cated model, composed by two spherical models with a small nugget effect. The
analytical expression of the spatial variability model is as followed:
γ ðhÞ ¼ 0:1 þ 1:21 Sphð26, 15, 11Þ þ 1:11 Sphð114, 90, 40Þ ð1Þ
where
0.10 ¼ nugget effect
Sph ¼ spherical model
1.22 ¼ sill for the first spherical structure
(25, 20, 10) ¼ ranges for the first structure in direction: strike, dip, and downhole
1.19 ¼ sill for the second spherical structure
(110, 90, 30) ¼ ranges for the second structure in direction: strike, dip, and
downhole
172 A. Dı́az-Carmona et al.
Histogram
500
450
400
Samples number
350
300
250
200
150
100
50
0
–5.3 –4.7 –4.1 –3.5 –2.9 –2.4 –1.8 –1.2 –0.6 0.0 0.6 1.2 1.8 2.4 3.0
LnCu %
The spatial variability model obtained was validated using the cross validation
procedure. A high value was obtained for the correlation between estimated and
real values (Fig. 4); this figure shows a high concentration of values around the
slope 45 ; the coefficient of Pearson is 0.92. These results show that the variability
model obtained in the structural analysis describes adequately the spatial variability
and correlation of the lognormal transformed data in the mineralized phenomenon
under study.
Furthermore, an analysis of the estimated error was performed. 94.34 % of the
real data are included in the prediction interval Zi 2*σ (Alfaro-Sironvalle 2007)
considering σ as the standard deviation of the estimated error, and 87.11 % of the
real data are included in the prediction interval considering σ as the kriging standard
deviation (Szilágyi-kishné et al. 2003; De Oliveira et al. 1997).
Complexities in the Geostatistical Estimation of Besshi-Type Mineral. . . 173
3D Semivariograms
3.00
2.40
1.80
Gamma (H)
1.20
Strike direction
0.60 Dip direction
Vertical direction
0
0 30 60 90 120 150 180 210 240 270 300
Range [m]
2
Estimated grade values
-1
-4
-4 -4 -3 -3 -2 -2 -1 -1 0 1 1 2 2 3 3 4 4
Sample grade values
Geostatistics simulation was used to determine the rational SMU size. A realization
of the regionalized variable can be obtained by geostatistics simulations. The
sequential Gaussian simulation is a powerful technique (Gómez-Hernández and
Cassiraga 1994) and the most used method to simulate values. In this sense, as the
Cu values show extremely asymmetric distributions, the logarithmic transformed
values, LnCu, were simulated in a dense network in order to reproduce possible real
values. These values were computed by the average of simulated data inside of
blocks with different sizes of SMU.
The procedure consists in comparing values of the possible real values with the
ordinary kriging estimations for each SMU proposed. Next, the least mean squared
error (MSE) was used to obtain the rational size of the blocks for estimation. This
procedure guarantees a high precision for resources estimations in a local or global
scale. The rational SMU size obtained was 32 40 m.
The procedure proposed was repeated with other realizations from sequential
Gaussian simulations, and the results were similar; the 32 40 m SMU size was
obtained as the rational SMU size. Gómez-González and Cuador-Gil (2011) made a
similar research in nonmetallic deposits, based on the opinion that mining equip-
ment must not be the factor determining the SMU size but the results of applying
the right mathematics estimation methods, which take into account the natural
behavior of the phenomenon.
Finally, lognormal kriging is used for resources estimation with the rational
network obtained before. In future studies, multi-Gaussian kriging will be applied
for more accuracy of the estimations in this ore deposit and other mineral deposits
of the regions, when the logarithmic transformation does not guarantee normal
distribution and consequently stationary data. The simulation was performed using
“sgsim” program from GsLib (Deutsch and Journel 1998).
The sequential Gaussian simulation performed with transformed data, the LnCu,
is presented in Fig. 5, using the variability model represented by Eq. 1 and a
simulation network of (1 1 1 m). The simulation was verified by the basic
statistics between real and simulated values that were similar, and the histograms
and semivariograms between them show similar behavior (Fig. 6).
The simulation network was very dense in order to obtain possible real values
inside each SMU proposed (Table 4). Estimation values were obtained by ordinary
kriging for each SMU proposed. Then, errors were computed comparing the
possible real values with the estimated values for each SMU, which are shown in
Table 4 and Fig. 7. Mean squared error and relative errors were used for this
comparison. Figure 7 shows the nomograms of the mean squared errors and relative
errors vs the size of the SMU. The SMU was proposed in two forms: squared and
rectangular, to take into account the anisotropy.
The influence of the SMU form was verified (Fig. 7). The least SME is present
for 32 40 m SMU size, which is the most precise for the resources estimation.
Using the rational SMU obtained before, the estimation was performed with
Complexities in the Geostatistical Estimation of Besshi-Type Mineral. . . 175
250 1600000
Number of samples
1400000
200
1200000
150 1000000
800000
100 600000
400000
50
200000
0 0
–5.3 –4.7 –4.1 –3.5 –2.9 –2.4 –1.8 –1.2 –0.6 0.0 0.6 1.2 1.8 2.4 3.0 –5.3 –4.7 –4.1 –3.5 –2.9 –2.4 –1.8 –1.2 –0.6 0.0 0.6 1.2 1.8 2.4 3.0
lognormal kriging over the mineralized body (Fig. 5). The back transformation was
carried out by the following expression (Rivoirard 1990):
Z*ðvÞ ¼ Exp LðvÞKO þ σ 2 KO þ γ ðv; vÞ =2 μ ð2Þ
where
L(v)KO ¼ estimation of the transformed values by ordinary kriging
σ2KO ¼ kriging variance
γ(v,v) ¼ mean semivariogram block to block
μ ¼ Lagrange multiplier
Table 5 shows the basic statistics between original and estimated grade values
after back transformation. The mean values are close, and the variance has adequate
values, in correspondence with the smoothen characteristics of the interpolation. No
extreme values have been obtained in the estimations.
176 A. Dı́az-Carmona et al.
170% 180%
160% 170%
Square mean error
Relative error
150% 160%
140% 150%
130%
140%
120%
130%
110%
120%
100%
4*5
8*10
12*15
16*20
20*25
24*30
28*35
32*40
36*45
40*50
44*55
4*5
8*10
12*15
16*20
20*25
24*30
28*35
32*40
36*45
40*50
44*55
SMU SMU
Square network rectangular network Square network rectangular network
5 Conclusion
References
Clark I (1998) Geostatistical estimation and the lognormal distribution. Geocongress, Pretoria
RSA
Cressie N (1993) Statistics for spatial data Rev edn. Wiley, New York, 900 p.
Cressie N (2006) Block kriging for lognormal spatial processes. Math Geol 38(4):413–443
De Oliveira V, Kedem B, Short DA (1997) Bayesian prediction of transformed Gaussian random
fields. J Am Stat Assoc 92(440):1422–1433
Deutsch CV, Journel AG (1998) GSLIB: geostatistical software library and user’s guide, 2nd edn.
Oxford University Press, New York
Dowd P (1982) Lognormal kriging: the general case. Math Geol 14(5):474–500
Gómez-González O, Cuador-Gil JQ (2011) Determinación del tama~ no racional del bloque para la
estimación de recursos minerales en el yacimiento Mariel. Revista Minerı́a y Geologı́a 27
(4):20–39
Gómez-Hernández JJ, Cassiraga EF (1994) Theory and practice if sequential simulation. In:
Armstrong M, Dowd PA (eds) Geostatistical simulation. Kluwer, Dordrecht, pp 111–124
Isaaks EH, Srivastava RM (1989) Applied geostatistics. Oxford University Press, Oxford, p 561
Journel AG (1980) The lognormal approach to predicting local distributions of selective mining
unit grades. Math Geol 12(4):285–303
Journel AG, Huijbregts CJ (1978) Mining geostatistics. San Francisco Academic Press, New York,
p 500
Krige DG (1981) Lognormal-De Wijsian geostatistics for ore evaluation: S. African Inst. Min.
Metall, Monograph series, p. 40
Krige DG, Magri EJ (1982) Geostatistical case study of the advantage of lognormal De Wijsian
Kriging with mean for a base metal mine and a gold mine. Math Geol 14:547–555
Lee YM, Ellis JH (1997) Estimation and simulation of lognormal random fields. Comput Geosci
23(I):19–31
Marcotte D, Groleau P (1997) A simple and robust lognormal estimator. Math Geol 29
(8):993–1008
Marechal A (1974) Krigeage normal et lognormal: Ecole des Mines de Paris, Centre de
Morphologie Matematique, unpublished note, N376. p. 10
Matheron G (1974) Effet proportionnel et lognormalite ou: Le retour du serpent de mer. Centre de
Morphologie Mathematique, Ecole des Mines de Paris, Publication N-374, p. 43
Paul R, Cressie N (2011) Lognormal block kriging for contaminated soil. Eur J Soil Sci
62:337–345
Rendu JM (1979) Normal and lognormal estimation. J Math Geol 11:407–422
Rivoirard J (1990) A review of lognormal estimators for in situ reserves. Math Geol 22
(2):213–221
Roth C (1988) Is lognormal kriging suitable for local estimation? Math Geol 30(8):999–1009
Szilágyi-kishné A, Bringmark E, Bringmark L, Alriksson A (2003) Comparison of ordinary and
lognormal kriging on skewed data of total cadmium in forest soils of Sweden. Environ Monit
Assess 84:243–263
Thurston M, Armstrong M (1987) A simplification of lognormal kriging to suit moderately skew
data: in proceedings of APCOM; Johannesburg, South Africa, vol. 3, pp. 53–64
Yamamoto JK (2007) On unbiased backtransform of lognormal kriging estimates. Comput Geosci
1:219–234
Yamamoto JK, Furuie R (2010) A survey into estimation of lognormal data, S~ao Paulo, UNESP,
Geociências, v. 29, n. 1, p. 5–19
Definition of Operational Mining Unit
(OMU) Size
C. Diedrich (*)
Vale Base Metals, 337 Power Street, Door #105, Copper Cliff, Sudbury, ON P0M 1N0, Canada
e-mail: [email protected]
J.D. Reuwsaat • R. Menin • W.F. De Paula
Vale Base Metals, 63 Grajau Street, 1st floor, Carajás, PA 68516-000, Brazil
e-mail: [email protected]; [email protected]; [email protected]
1 Introduction
The ore control model estimates should be classified by ore type and sent to the
correct destination (mill, stockpile, or dump). The problem of how to make the
mineral reserve grade-tonnage curves accurately and predict the tonnes and grades
above the cutoff that will be available at the time of mining still exists. Conven-
tional SMU selection processes try to account for the impact of support and
information effect on a short-term scale; however, they are insufficient to handle
conditionally unbiased estimations and disregard potential external factors in the
mining process. As a result, operations struggle to achieve the forecasted mining
production, or, if production is achieved, the mining plan and stockpile adherence is
compromised by changes in the mining configuration in order to achieve targeted
values. Block SMU size should not be selected to only attain the mill targets; they
also need to align with the operational process that will be implemented and to
facilitate stockpiling and waste dump targets.
The methodology proposed is to incorporate the estimated ore control model,
defined by mining operating practices, in the long-term mineral reserve estimations.
Assuming that a verified regular unbiased global grade and tonnage estimation is
used, the global proportions of planned dilution incorporated in the mining operat-
ing process can be assumed as representative in terms of block estimation. The
results of applying this methodology show that commonly defined SMU size does
not correspond directly with the smallest mineable volume chosen by the mining
professionals. As mining selectivity is understood as the process of separating ore
from waste, its global and local concept is strongly related to three functions that
affect operational results. These are:
1. Geology function (Fig. 1)
2. Production rates function (Fig. 2)
3. Mining configuration function (Fig. 3)
Each of these functions can result in different local mining recovery curves. The
geology function (1), at operating scale, is presented in Fig. 1a, b. For the planned
grade control polygons in Fig. 1a1, a2 that have the same SMU block size and
operating practices, the resulting operational ore and waste materials (b1 and b2)
are different. This clearly demonstrates the impact of different local geology under
real operating circumstances. The question that arises is, for identical mining
configuration and production rates, should the long-term SMU size be the same
where there is different local geology. The answer should be no. This scenario has
been verified many times in a real mining operation where different local mining
recoveries and dilution of the defined SMU have resulted in deviating from
forecasted local and global mining recovery curves.
Changes in the local mining recovery curves were verified for different produc-
tion rates considering the same geology and mining configuration in the operational
process. The production rate function (2) at operating scale is presented in Fig. 2a,
b. The planned grade control polygons in Fig. 2a1, a2, a3, a4, a5 have the same
182 C. Diedrich et al.
Fig. 1 Geology function operational aspect: (a) planned open pit grade control polygons (a1 and
a2) and its respective (b) executed open pit grade control (b1 and b2)
Fig. 2 Production rate function operational aspect: (a) planned open pit grade control
polygons (a1 and a2, a3, a4, a5) and its respective (b) executed open pit grade control (b1 and
b2, b3, b4, b5)
Definition of Operational Mining Unit (OMU) Size 183
Fig. 3 Mining configuration function operational aspect: (a) planned open pit grade control
polygons (a1 and a2) and its respective (b) executed open pit grade control (b1 and b2)
SMU block size, geology, and mining configuration, but different production rates
and different resulting operational ore and waste proportions (b1 and b2, b3, b4,
b5). For different production rates and related grade control processes, different
mining recovery curves above cutoffs resulted. This demonstrates that SMUs used
to predict ore and waste proportions should not be the same.
Different mining configurations, applied to the same local geology and using the
production rates in the operational process, may change mining recovery curves as
well. The mining configuration function (3), at operating scale, is presented in
Fig. 3a, b. For planned grade control polygons in Fig. 3a1, a2 that have the same
SMU block size, geology, and production rates, but different mining configuration,
the resulting proportions of ore and waste (b1 and b2) are different. For different
mining configurations and related grade control processes, different resulting min-
ing recovery curves above cutoffs are obtained, and SMUs used to predict local
values should not be the same.
Long-term mine planning takes into account that each SMU will be selectively
mined as ore or waste. However, local operating particularities can result in poor
reconciliation of long-term block models, and the resulting mining recovery curves
will impact milling, stockpiling, and waste dump processes.
184 C. Diedrich et al.
The proposed approach for OMU size(s) definition uses information from the
anticipated grade control procedure based on operational mining sequencing and
the use of the long-term resource model defined by a common SMU process. The
idea is to determine the OMU(s) by local and/or global ore and waste proportions
(tonnage and grades) calculations obtained from the resulting operational mining
sequencing and proposed grade control processes. The first step is to choose a
regular SMU size based on robust grade estimation practices and to define reason-
ably representative operating grade control process according to local geology,
production rates, and mining configuration.
There are two ways of determining OMU(s) block size(s):
• Global: where all individual grade control polygons are used to determine global
ore/waste proportions in the mining sequencing and the respective global OMU
(regular/constant) size (see Fig. 9[2]) and Fig. 9[3] of case study)
• Local: where each individual grade control polygon is used to determine local
ore/waste proportions in the mining sequencing and the respective local OMUs
(regular/constant or irregular) size(s) (see Fig. 9[4] of case study)
The procedure is as follows:
1. Compute reasonable mining sequencing A (e.g., blasting polygons) based on the
expected mining operating practices using a common regular SMU size.
2. Simulate the grade control practice to arrive at ore/waste dig limits. The idea is
to mimic the real grade control process (e.g., Figs. 1, 2, and 3) that will be
implemented in the mine for each operational mining sequencing (A).
3. Then, the ore/waste dig lines are used to calculate the expected local (A) and/or
global (sum of local A) tonnes of waste (Wt) and ore (Ot), grades of waste
(Wg) and ore (Og), and metal of waste (Wm) and ore (Om). The concept is to
determine local and/or global OMU size(s) that matches these resulting values.
4. Choose the possible OMU size (s) by simply calculating the proportions of ore
and waste and its respective targeted grade control tonnes (Wt and Ot), grades
(Wg and Og), and metal (Wm and Om) results.
5. Plot the results of both grade control and OMU approach in a series of graphs
containing (1) planned dilution (PD), (2) tonnes of waste and ore versus OMU
size, (3) average grades of waste and ore versus OMU size, and (4) metal tonnes
of waste and ore versus OMU size. In each graph, the OMU results are plotted to
Definition of Operational Mining Unit (OMU) Size 185
Fig. 4 (a) Grade control polygon with 5 5 5 m SMU block size and its respective grades (b)
resulting operational grade control polygons based on (a) and (c) OMU block size that represents
the grade control execution (target)
yield a functional relationship, while the grade control values provide a single
true value that plots as a horizontal line. The OMU size is the size of which these
grade control polygon (Wt, Ot, Wg, Og, Wm, and Om) lines intersect with its
respective class of material (i.e., waste and ore). Figures 4, 5, 6, and 7 present a
schematic illustration of the OMU process definition and resulting graphs.
6. When the resulting OMU size(s) is selected, which is (are) suitable to represent
the local and/or global ore/waste proportions, the mining sequencing should be
revisited through an iterative process. This is required to check the overall ore
recovery and planned dilution compared to the initial mining sequencing and
grade control based on a preliminary SMU size (step 1). Depending on the waste/
ore proportions and the deposit characteristics, the recovery of mineralization
can be significantly impacted. An adjustment of the mining sequence selectivity
(see case study) may be required to maximize ore recovery or to change mining
sequencing to better fit OMU block configuration. A detailed critical analysis
should be made by experienced professionals.
The concept of the OMU method is to add operational factors for determining
block size(s) through the calculation of planned dilution from mining and grade
control process. Global and local approaches can be applied in different ways or
purposes in the mining sequencing. The global approach provides ore/waste
186 C. Diedrich et al.
Fig. 5 Graphs of (1) tonnage and planned dilution of grade control polygons (target), SMUs, and
OMU (top left); (2) grades and planned dilution of grade control polygons (target), SMUs, and
OMU (top right); (3) ore metal and planned dilution of grade control polygons (target), SMUs, and
OMU (center left); (4) waste metal and planned dilution of grade control polygons (target), SMUs,
and OMU (center right); (5) and (6) SMU (5 5 m) and OMU (23 23 m) tonnages and grades
above the cutoff grade (bottom left and right)
Fig. 6 (a) Grade control polygon with 5 5 5 m SMU block size and its respective grades (b)
resulting operational grade control polygons based on (a) and (c) OMU block size that represents
the grade control execution (target)
block support curves and how the ore and waste proportions with its respective
grades are achieved. In the mining operating process, resulting external waste
material incorporated as ore and ore defined as waste strongly affects the recover-
able proportions of material above and below cutoffs and may change its final
destination (mill, stockpile, waste dump, etc.). For these examples, an SMU size of
5 5 5 m, with a generic density of 2.8 t/m3, was chosen. Only tonnes and
contained metal within the polygon boundaries were considered. Tables 1 and 2
summarize the grade control polygons (target), SMUs, and OMU resulting values.
The results demonstrate how operating processes change the ore/waste proportions,
how the resulting OMU block size differs from the previously defined SMU size,
and how it is possible to achieve global and/or local operational ore and waste
targets.
In both examples presented in Figs. 4 and 6, the operational process and local
geological aspects are significant components of the ore/waste proportions in the
grade control polygons. This clearly demonstrates that operating aspects are fun-
damental in the block size definition for achieving reasonable global and local
expected mining recovery curves. Although OMU block size definition is a time-
consuming process, mining professionals should think about different ways of
better defining block volumes since there is no current definitive solution.
188 C. Diedrich et al.
Fig. 7 Graphs of (1) tonnage and planned dilution of grade control polygons (target), SMUs, and
OMU (top left); (2) grades and planned dilution of grade control polygons (target), SMUs, and
OMU (top right); (3) ore metal and planned dilution of grade control polygons (target), SMUs, and
OMU (center left); (4) waste metal and planned dilution of grade control polygons (target), SMUs,
and OMU (center right); (5) and (6) SMU (5 5 m) and OMU (45 23 m) tonnages and grades
above the cutoff grade (bottom left and right)
Table 1 Executed (target) grade control (Fig. 4 [b2]), diversified SMUs, and resulting OMU size
selection (Fig. 4 [c2])
Item Ot (t) Og (%) Om (t) Wt (t) Wg (%) Wm (t) PD (%)
Grade control polygon 75,600 1.11 837 0 – 0 36 %
SMU 5 5 m 55,650 1.45 805 19,950 0.16 32 0
SMU 10 10 m 62,300 1.31 817 13,300 0.15 20 12 %
SMU 15 15 m 66,150 1.23 814 9,450 0.25 23 19 %
SMU 20 20 m 70,000 1.17 821 5,600 0.28 16 26 %
OMU 23 23 m 75,600 1.11 837 0 – 0 36 %
Definition of Operational Mining Unit (OMU) Size 189
Table 2 Executed (target) grade control (Fig. 6 [b1]), diversified SMUs, and resulting OMU size
selection (Fig. 6 [c1])
Item Ot (t) Og (%) Om (t) Wt (t) Wg (%) Wm (t) PD (%)
Grade control polygon 59,850 1.68 1,004 15,750 0.29 46 23 %
SMU 5 5 m 48,650 2.07 1,008 26,950 0.15 42 0%
SMU 10 10 m 53,900 1.89 1,021 21,700 0.13 29 11 %
SMU 15 15 m 53,550 1.89 1,014 22,050 0.16 36 10 %
SMU 23 23 m 63,000 1.63 1,030 12,600 0.16 20 29 %
OMU 45 23 m 59,850 1.68 1,004 15,750 0.29 46 23 %
A series of mining plans used for operational grade control, based on a block model
with an SMU size of 5 5 1 m and 1.0 t/m3 density, were compiled from the
Walker Lake dataset. A series of grade-tonnage curves were analyzed at the ore
definition cutoff (5 %) to determine the ore/waste proportion changes due to regular
block support size and mining sequencing. Different production polygon configu-
rations based on a block model with an SMU size of 5 5 1 m were compiled as
follows:
• Production polygons ranging from 6 to 9 SMU blocks in cases 1, 4, and 7
• Production polygons ranging from 9 to 15 SMU blocks in cases 2, 5, and 8
• Production polygons ranging from 15 to 25 SMU blocks in cases 3, 6, and 9
These production polygons were then overlain (with minor adjustments) over
block models with SMU sizes of 15 15 1 and 25 25 1 to develop three
mining sequences and nine different mining scenarios. Figure 8 and Table 3 present
the grade-tonnage curves for the nine different mining scenarios. Cases 1–3 are
related to a 5 5 m block size, cases 4–6 are related to a 15 15 m block size, and
cases 7–9 are related to a 25 25 m block size.
The mining recovery curves (Fig. 8 and Table 3) demonstrate how the ore/waste
(cutoff grade of 5 %) proportions of the grade control polygons compare to the
relative block sizes of the mining sequences. In mining scenarios with high pro-
duction volumes, the 5 5 m block mining recovery curves tend to deteriorate and
do not represent the waste that is incorporated to the system. When block support is
increased (reblocked to 15 15 m or 25 25 m), the resulting grade control mining
recovery curves tend to approximate the blocks and polygons (tonnes, grades, and
metal), and ore recovery is strongly impacted by external material that is incorpo-
rated within the reblocked volumes. This may lead to significant changes to mining
sequencing and proportions of ore and waste that were originally based on a 5 5 m
block size plan. The original plan may not be realistic since the ore material could
be recovered even considering the waste blocks in the grade control polygons (cases
1–3).
190 C. Diedrich et al.
Table 3 Summary of mining sequences for different production rates (6–25 SMUs of 5 5 m
block size) and block sizes (5 5 m; 15 15 m; 25 25 m) as described in the text above and its
respective block and polygon values at 5 % cutoff grade
Block Block Block Poly. Poly. Poly. Planned %
tonnage grade metal tonnage grade metal dilution Diff.
Item (t) (%) (t) (t) (%) (t) (%) metal
Case1 12,184 6.8 827 13,389 6.4 852 9.9 3.1
Case2 12,410 6.8 841 13,991 6.3 878 12.7 4.4
Case3 10,604 7.0 740 14,003 6.0 837 32.1 13.1
Case4 10,646 6.5 692 10,682 6.4 680 0.3 1.8
Case5 10,646 6.5 693 11,437 6.2 705 7.4 1.9
Case6 10,681 6.5 694 11,823 6.0 708 10.7 1.9
Case7 8,709 6.4 560 8,015 6.4 509 8.0 9.2
Case8 8,686 6.4 559 8,011 6.1 491 7.8 12.1
Case9 8,687 6.4 559 8,254 6.0 493 5.0 11.8
Fig. 9 (1) Case 1 production polygons (6–9 SMU blocks, bolded lines in all figures) and
5 5 1 m (top left); (2) OMU REG, regular 15 15 1 m block size and regular block
(centroid) position (top right); (3) OMU RSIP, regular 15 15 1 m block size and irregular
block (centroid) position (bottom left); (4) OMU ISIP, irregular block size and irregular block
(centroid) position (bottom right)
The OMU REG (regular) approach (global) shows that a 15 15 m block size
fits well to the grade control curves (blue line) based on a 15 15 m regular block
size. The mining recovery curves are strongly impacted by external dilution mate-
rial if compared to the SMU grade control polygons, which are considered reason-
able for operations to mine according the Fig. 9(1). In the RSIP approach (global),
with a regular 15 15 m block size and irregular centroid position, the recovery
issue is greatly improved, and its resulting block size curves provide a near-perfect
match with the initial SMU grade control polygon curves at 5 % cutoff and above.
192
Table 4 Recoverable values at given cutoffs for grade control polygons (poly.) on case 1 (grade control target), case 1 SMU (5 5 1 m), OMU REG, OMU
RSIP, and OMU ISIP
Poly. Poly. Poly. Case1 Case1 Case1 OMU REG OMU REG OMU REG
Cutoff (%) tonnage (t) grade (%) metal (t) tonnage (t) grade (%) metal (t) tonnage (t) grade (%) metal (t)
3 16,375 6.1 999 16,050 6.2 998 11,732 6.3 739
5 13,850 6.4 883 12,500 6.8 847 10,646 6.5 692
7 3,675 8.1 298 4,375 8.4 368 2,865 8.3 238
9 675 10.0 67 1,150 10.3 119 898 9.5 85
Poly. Poly. Poly. OMU RSIP OMU RSIP OMU RSIP OMU ISIP OMU ISIP OMU ISIP
Cutoff (%) tonnage (t) grade (%) metal (t) tonnage (t) grade (%) metal (t) tonnage (t) grade (%) metal (t)
3 16,375 6.1 999 19,125 5.7 1,087 16,100 6.1 980
5 13,850 6.4 883 12,375 6.4 793 13,650 6.3 857
7 3,675 8.1 298 3,375 8.2 275 3,400 8.3 282
9 675 10.0 67 675 10.0 67 675 10.3 69
C. Diedrich et al.
Definition of Operational Mining Unit (OMU) Size 193
Fig. 10 Mining recovery curves of SMU, OMU REG, OMU RSIP, and OMU ISIP, respectively,
from top to bottom
194 C. Diedrich et al.
Fig. 11 SMU block size and sequencing (left), actual grade control polygons (right)
In the ISIP approach (local), considering irregular block sizes and centroid,
arranged through each local sequencing polygon, a near-perfect match is achieved
for the entire initial SMU grade control distribution.
A case study for a Brazilian sulfide Cu-Au mine is used to demonstrate applicability
of the method by comparing to actual operational results. The grade of waste is
assigned as “zero.” The cutoff grade of the deposit is 0.25 % Cu. Figure 11 illus-
trates the 15 15 15 m SMU block estimates (left) and the actual grade control
polygons (right). Table 5 lists resulting mining recovery values above cutoffs
(Fig. 12).
The results demonstrate how mining recovery curves are improved using oper-
ational aspects for defining the mining unit size. For the real case study, the
operational OMUs near matched the operational grade control practice. The actual
values below and above cutoffs were considerably different than those from the
SMU. Also, it demonstrated that a 30 30 m and 45 45 m block sizes should be
applied to the material with grades below and above 1.3 % Cu, respectively.
Operational aspects should be included in the mining unit block size definition
by mining professionals. Selection of the appropriate mining unit is a compromise
between getting the right mining recovery estimates for tonnages of ore and waste
and getting the right grade of these materials. This compromise depends on the
global and/or local ore and waste proportions for a predefined operational produc-
tion rates and mining configurations. The greatest impact of the support effect is
where the operational aspects are not included or checked against regular SMU
sizes defined by common geostatistical estimation practices. Significant differences
Table 5 Mining recovery values at given cutoffs for SMU 15 15 15 m, grade control polygons (target), OMU 30 30 15 m, and OMU 45 45 15 m,
respectively
Cutoff (%) SMU 15 15 Mt SMU 15 15 %Cu SMU 15 15,000 t Cu Polygon Control Mt Polygon Control %Cu Polygon Control 000 t Cu
0 104 0.68 702 139 0.51 702
0.25 89 0.77 682 93 0.70 653
0.4 67 0.91 607 66 0.86 562
0.8 34 1.25 429 32 1.16 368
1.2 18 1.50 263 14 1.37 195
Definition of Operational Mining Unit (OMU) Size
Fig. 12 Mining recovery curves of SMU 15 15 15 m, grade control polygons (target), OMU
30 30 15 m, and OMU 45 45 15 m
between grade control and operating practices to define SMU may result in poor
mining, waste dumps, and stockpile plans and strongly impact financials.
Bibliography
Deraisme J, Roth C (2000). The information effect and estimating recoverable reserves.
Geovariances
Isaaks E (2004, 2005) The kriging oxymoron: a conditionally unbiased and accurate predictor.
Proceedings of geostatistics banff, 2nd ed. Springer, Banff, pp 363–374
Jara RM, Couble A, Emery X, Varela EM, Ortiz JM (2006) Block size selection and its impact on
open-pit design and mine planning. J South African Inst Min Metall 106:205–211
Leuangthong O, Neufeld C, Deutsch CV (2003) Optimal selection of selective mining unit (SMU)
size. University of Alberta, Alberta
Optimizing Infill Drilling Decisions Using
Multi-armed Bandits: Application in a Long-
Term, Multi-element Stockpile
1 Introduction
2 Method
The proposed method selects the best infill drilling pattern from a predefined set of
patterns. The representation of each pattern for infill drilling will be called an arm in
the MAB framework. These arms will be played. When an arm is played, the
algorithm to assess the value for the pattern associated to that arm is set in motion.
The steps in this value assessment are the following (Fig. 1 demonstrates the
general flow of the algorithm):
• The grades in the drill holes of the selected pattern are drawn from an initial
simulation, which is set to be the possibly ‘true’ representation of the deposit.
• The drill holes of the pattern linked to the pulled arm are used as additional data
to the initial exploration data for the re-simulation of the deposit.
• After the re-simulation with additional data, the blocks of the re-simulation are
classified according to a material classification guide.
• The reward or contribution from pulling that arm is defined as the percentage of
blocks that have changed material classification in the re-simulation compared to
the possibly ‘true’ deposit. Optimizing for this reward is in line with the
definition of the best pattern given above.
• The Thompson sampling solution algorithm updates the current arm, and the
next arm is selected.
After convergence of the algorithm, the whole procedure is repeated for other
simulations of the deposit as a possibly ‘true’ deposit in order to test the sensitivity
of the method to this choice. The best pattern is then selected over the results of all
of these possibly ‘true’ deposits. By considering multiple simulations as possibly
200 R. Dirkx and R. Dimitrakopoulos
‘true’ deposits, the method is able to give an assessment of the performance of all
patterns under geological uncertainty, which leads to the quantification of the
upside potential and downside risk for each pattern.
All of the patterns in one MAB set-up are comparable in the sense that they
belong to what is called the same budget class, which means that they all have the
same amount of drill holes and thus represent a similar cost to be drilled. The
procedure is also repeated for different budget classes of patterns where each
budget class represents patterns with a different number of holes than in the other
classes. A separate MAB is required to optimize within each budget class, as the
rewards of patterns with more holes cannot be directly compared to the rewards of
patterns with fewer holes.
The classic and simplest MAB problem follows the four basic rules formulated
below, as in Mahajan and Teneketzis (2008). The state of an arm in a simple MAB
is the internal time, which is equal to the number of times the arm has been played.
1. Only one arm is played at each time step. The reward from this play is
uncontrolled.
2. Not played arms remain frozen, meaning they do not change state.
Optimizing Infill Drilling Decisions Using Multi-armed Bandits: Application. . . 201
with:
• Jγ , the objective function value under policy γ, defined by Ui(t).
• βt, the discount factor with time t, reflecting that earlier rewards are higher
valued than later rewards, like in an NPV calculation. The time t is a measure for
the number of iterations/plays.
• Ri(Xi(Ni(t)), Ui(t)), the reward of arm i depending on:
– Xi(Ni(t)), the state of arm i at time t, which depends on:
Ni(t), the local time of arm i at time t; the number of times arm i has been
played before time t.
– Ui(t), the policy decision, which determines whether or not arm i is played at
time t.
• Z(0), the initial state of all arms i.
The total objective function value is then found by taking the expected value
over the summations over time and the number of arms conditional to the initial
states of all arms. The goal is to find the optimal scheduling policy γ that determines
the values of Ui(t) for every arm i and time t in such a way that the total expected
reward is maximized. The MAB problem can also be used to find the best
performing arm, instead of a schedule of when each arm should be played. The
best arm is just the one that is scheduled for continuously.
The algorithm chosen to solve the MAB is Thompson sampling (Thompson
1933), which has demonstrated excellent performance for MAB optimization
(Agrawal and Goyal 2012; Scott 2010). The general idea of Thompson sampling
is always play the arm with the highest likelihood of being the best arm. This gives a
natural approach to the exploration versus exploitation trade-off in MAB problems.
In Thompson sampling each arm is represented by a distribution linked to its
reward. Usually, the prior distributions are uniform beta distributions, B
(α¼1,β¼1), which are updated each time an arm is played, and the reward is
observed. The reward percentage is used as the success probability of a Bernoulli
202 R. Dirkx and R. Dimitrakopoulos
trial. When the outcome of the Bernoulli trial is a success, the β parameter is
incremented with one; otherwise, the α parameter is incremented with one.
Thompson sampling is proven to converge for the MAB problem by Agrawal
and Goyal (2012) and May et al. (2012), but it has no predefined termination
moment. Therefore, it is required to define a convergence criterion that works for
the application in consideration. In this case a simple criterion is defined, which has
shown satisfactory results. That is, if in the last 20 % of iterations one arm is
selected to be played more than 90 % of the iterations, then this is the best arm.
The case study presented below illustrates the optimization of the infill drilling
decision for a long-term, multi-element stockpile of a gold mining complex in
Nevada, USA. This complex consists of two open-pit mines, an underground mine
and several stockpiles. The downstream processing includes an autoclave, an oxide
mill and multiple heap leaches. Deleterious compounds are co-simulated with the
gold grade because of strict grade requirements on them to guarantee efficient
processing recoveries. Knowledge of the grade of these compounds can assure
blending of ore from different sources to meet the constraints at the processing
facilities. The compounds being considered are sulphide sulphur, organic carbon
and carbonate, next to gold as the paying metal.
For this case study, three classes of patterns are considered, each corresponding
to a different budget. The first budget class has patterns with 5 holes, the next class
has patterns with 10 holes and the last class has patterns with 15 holes. Eight
patterns, based on the ideas provided to us by the mine site, are considered for the
optimization in each class. For every class the procedure is repeated for 20 simula-
tions as possibly ‘true’ stockpiles to guarantee that the results are independent from
the selected possibly ‘true’ stockpile. This also assesses the performance of the
patterns under geological uncertainty. A validation step tests all patterns with
simulations not used in the optimization procedure and shows that the results are
independent of the simulations employed.
The exploration drilling is done on a regular grid with a 40 ft spacing. In total there
are 104 drill holes with one sample each. The height of the stockpile is 20 ft. The
size is 200,000 m3, which corresponds to 400 kton. Twenty simulations are
generated, each containing 204 blocks within the shape of the stockpile. The block
size is 30 30 20 ft. This is the same as in the mines feeding the stockpile
because the same mining selectivity is used for both the stockpiles and the mines.
Optimizing Infill Drilling Decisions Using Multi-armed Bandits: Application. . . 203
Fig. 2 Spatial material variability of the stockpile simulations at the block level
All simulations below are generated with the direct block MAF simulation method
by Boucher and Dimitrakopoulos (2009).
Figure 2 shows the material variability of three stockpile simulations at the block
level. The material is classified based on the material classification guide provided
by the mining complex. The mining complex flags the stockpile under consider-
ation as medium sulphide material, but Fig. 2 shows that high- and low-sulphide
material usually dominates the material distribution in the stockpile. This trend is
carried on in the remaining simulations.
Figure 2 also shows that there is a large amount of spatial material variability in
the stockpiles, within one simulation but also from one to another simulation. This
observation is the main motivation for the application of the proposed method to
stockpiles. It shows that more information is required to assess the local-scale
variability in stockpiles. This is especially true for the stockpile in this mining
complex because, in some periods, it is the main contributor to the processed blend
at the autoclave. This stream has tight constraints on the deleterious compounds like
a maximum organic carbon content and an upper and lower bound on the sulphide
sulphur–carbonate ratio. Figures 3, 4, 5 and 6 show the spatial material variability in
two simulations for the four variables of interest. All simulations show that the
same main features are respected between two simulations but that there is also
local-scale variability within each simulation.
Figures 7, 8, 9 and 10 below show the eight patterns that are considered in the
optimization. The figures always show all 15 additional drill holes for each pattern
on top of the initial exploration drill holes which are represented as black crosses.
However, if only the red squares (‘first five holes’) are considered, the patterns
represent the budget class with five holes per pattern. If the red squares and the
green triangles (‘second five holes’) are considered, the patterns represent the
budget class with ten holes per pattern. The patterns have the same numbering
204 R. Dirkx and R. Dimitrakopoulos
Fig. 4 Sulphide sulphur content variability in two simulations at the block level
Fig. 5 Organic carbon content variability in two simulations at the block level
over the budget classes because they are always based on the same idea for the
location of the holes. For example, pattern 1 is designed as being equally spread and
pattern 2 to target the high-grade gold zone in the top of the stockpile.
Optimizing Infill Drilling Decisions Using Multi-armed Bandits: Application. . . 205
Fig. 7 Pattern 1 (left) and pattern 2 (right) with 15 additional drill holes
Fig. 8 Pattern 3 (left) and pattern 4 (right) with 15 additional drill holes. The black line indicates
that the green triangle below it belongs to pattern 3
206 R. Dirkx and R. Dimitrakopoulos
Fig. 9 Pattern 5 (left) and pattern 6 (right) with 15 additional drill holes. The black line indicates
that the green triangles below it belong to pattern 5
Fig. 10 Pattern 7 (left) and pattern 8 (right) with 15 additional drill holes
Figure 11 shows the convergence results for the algorithm over all simulations for
every pattern in each budget class. Convergence results show how often the
algorithm has converged on that specific pattern in that budget class for the
20 simulations as a possibly ‘true’ deposit. The pattern that has been converged
upon the most is called the ‘winner’ of that budget class. For the 10- and 15-hole
budget classes, there is always a clear winner in pattern 1. However, for both budget
classes, it has to be noted that pattern 7 is also a strong performing pattern,
especially for the 15-hole budget class. The observation on a winner for the five-
hole budget class is much less clear. Pattern 2 has 30 % convergence, but pattern
1, 4 and 5 each have 20 % too. This is only a small difference, and no strong, clear
decision on the winner can be made from this.
Figure 12 shows the average reward over all 20 simulations for every pattern in
each budget class. For the 10- and 15-hole budget classes, it is pattern 1 that has the
Optimizing Infill Drilling Decisions Using Multi-armed Bandits: Application. . . 207
Fig. 11 Convergence over all 20 simulations for all patterns and every budget class
highest average reward, as is expected from the convergence results in Fig. 11.
However, the results for the five-hole budget class are not as expected in the sense
that pattern 2 does not have the highest average reward. Pattern 1 has the highest
average reward for the five-hole budget class. Therefore, it is concluded that pattern
1 is more robust against geological uncertainty than pattern 2. Figure 12 strengthens
the observation of no clear winner for the five-hole budget class by demonstrating
that the average rewards for all patterns are very close to each other. It is hard for
the MAB to distinguish between patterns with such similar rewards.
The differences between the running times observed in Table 1 can also be
explained by the observations in Figs. 11 and 12 mentioned above. Because the
five-hole patterns yield similar average rewards, the algorithm has a hard time
finding the winner, and therefore, it requires more iterations, and thus time, to
evaluate the patterns over all simulations. The number of iterations to evaluate all
patterns within one budget class for all simulations is very high. In fact, it is higher
than the brute-force approach, which tests every pattern 20 times for each simula-
tion and then takes the average. This brute-force tactic would result in
(8 20 20¼) 3200 iterations to evaluate each budget class. The numbers higher
than 3200 in Table 1 come from just a few simulations with hard convergence
problems that add a lot of iterations to the total count.
Table 1 also shows the number of times the algorithm did not converge for every
budget class. The number of non-converging simulations is low for every budget
class. To get it to zero, it requires unnecessary large amounts of iterations because it
208 R. Dirkx and R. Dimitrakopoulos
Fig. 12 Average reward over all simulations for all patterns and every budget class
Table 1 Overview of the statistics of the algorithm for each budget class
Budget class Running time Number of iterations Number of times not converged
5 holesa 24 h 24 min 46 s 25,465 3
10 holes 13 h 9 min 31 s 13,085 4
15 holes 8 h 41 min 31 s 9378 2
a
A different set-up is used for the five-hole pattern optimization. This set-up allows for 5000
instead of 2000 iterations as the maximum iterations without convergence. Another addition to this
set-up is a forced exploration phase at the start of the algorithm forcing each pattern to be played
20 times before the algorithm starts doing its normal steps. Both measures are taken to overcome
convergence issues observed in initial tests
is always possible that two patterns perform very similarly for a certain simulation.
Therefore, it is not really seen as a problem because the algorithm still identifies the
two patterns that perform equally well in these cases. As a tiebreaker, the pattern
with the highest average reward is attributed the convergence. In most cases one of
the tied patterns is pattern 1. If ties are broken in the opposite way, the results
remain unchanged.
Figure 13 compares the performance of the patterns over the budget classes. This
graph demonstrates the profile (P10-P50-P90) on the best pattern’s reward over all
simulations compared to the average of the rewards of all patterns over all simu-
lations. The tenth percentile (P10) represents the downside risk, and the ninetieth
percentile (P90) represents the upside potential of the best pattern under geological
Optimizing Infill Drilling Decisions Using Multi-armed Bandits: Application. . . 209
Fig. 13 Profile of the best pattern for every budget class versus the all-pattern average for every
budget class
uncertainty. Figure 13 shows that the five-hole budget class performs worse than the
other two budget classes. The five-hole best pattern average is lower than the
all-pattern average for both other budget classes. Also, the downside risk is much
higher, and the upside potential is much lower. Therefore, in combination with the
reasons mentioned above, it is not recommended by the author to drill a five-hole
pattern in this stockpile. Purely based on performance, pattern 1 with 15 holes is
better than pattern 1 with ten holes. However, the extra five holes also cost more to
drill. Another observation is that the downside risk for the 10- and 15-hole pattern
1 is almost equal. This means that they both guarantee similar results in the worst-
case scenario. This is an extra argument in favour of pattern 1 with ten holes,
especially if there are constraints on the budget that are more important than having
the best possible knowledge.
Fig. 14 Profile of the best pattern for every budget class versus the all-pattern average for every
budget class based on the reward from the alternate set of simulations
The case study presented above shows that the proposed MAB algorithm works as
intended. It is able to select the best pattern within a set of patterns of the same
budget and provides a decision tool to select the best pattern between sets of a
different budget. Also, the case study demonstrates how the method successfully
quantifies the influence of geological uncertainty on the performance of an infill
Optimizing Infill Drilling Decisions Using Multi-armed Bandits: Application. . . 211
Acknowledgements We thank Newmont Mining Corporation for providing us with the data
necessary to conduct this research and the organizations that funded this research: the Natural
Sciences and Engineering Research Council of Canada (NSERC) Discovery Grant 239019 and the
COSMO Mining Industry Consortium (AngloGold Ashanti, Barrick Gold, BHP Billiton, De Beers
Canada, Kinross Gold, Newmont Mining and Vale) supporting the COSMO laboratory.
References
Agrawal S, Goyal N (2012) Analysis of Thompson sampling for the multi-armed bandit problem.
Proceedings of the 25th Annual Conference on Learning Theory (COLT)
Barnes RJ (1989) Sample design for geological site characterization. In: Armstrong M
(ed) Geostatistics 1988, vol 4. Springer Netherlands, Avignon, pp 809–822
Benndorf J (2015) Making use of online production data: sequential updating of mineral resource
models. Math Geosci 47(5):547–563
Boucher A, Dimitrakopoulos R (2009) Block simulation of multiple correlated variables. Math
Geosci 41(2):215–237
Boucher A, Dimitrakopoulos R, Vargas-Guzman JA (2005) Joint simulations, optimal drillhole
spacing and the role of the stockpile. In: Leuangthong O, Deutsch C (eds) Geostatistics 2004,
vol 14. Springer Netherlands, Banff, pp 35–44
Chorn LG, Carr PP (1997) The value of purchasing information to reduce risk in capital invest-
ments. SPE hydrocarbon economics and evaluation symposium. Society of Petroleum
Engineers
Desbarats AJ, Dimitrakopoulos R (2000) Geostatistical simulation of regionalized pore-size
distributions using min/max autocorrelation factors. Math Geol 32(8):919–942
Diehl P, David M (1982) Classification of ore reserves/resources based on geostatistical methods.
CIM Bull 75(838):127–136
Gershon M, Allen LE, Manley G (1988) Application of a new approach for drillholes location
optimization. Int J Surf Min Reclam Environ 2:27–31
Goovaerts P (1997) Geostatistics for natural resources evaluation. Oxford University Press,
New York
212 R. Dirkx and R. Dimitrakopoulos
Goria S, Armstrong M, Galli A (2001) Quantifying the impact of additional drilling on an openpit
gold project. 2001 annual conference of the IAMG. Cancun
Jewbali A, Dimitrakopoulos R (2011) Implementation of conditional simulation by successive
residuals. Comput Geosci 37(2):129–142
Mahajan A, Teneketzis D (2008) Multi-armed bandit problems. In: Hero A III, Castanon D,
Cochran D, Kastella K (eds) Foundations and applications of sensor management. Springer,
New York, pp 121–151
May BC, Korda N, Lee A, Leslie DS (2012) Optimistic Bayesian sampling in contextual-bandit
problems. J Mach Learn Res 13(1):2069–2106
Menabde M, Froyland G, Stone P, Yeates G (2007) Mining schedule optimisation for condition-
ally simulated orebodies. In Orebody modelling and strategic mine planning. p 379–384
Prange M, Bailey WJ, Couet B, Djikpesse H, Armstrong M, Galli A, Wilkinson D (2008) Valuing
future information under uncertainty using polynomial chaos. Decis Anal 5(3):140–156
Ravenscroft PJ (1992) Risk analysis of mine scheduling by conditional simulation. Trans Inst Min
Metall Sect A 101:A104–A108
Schlee E (1991) The value of perfect information in nonlinear utility theory. Theory Decis 30
(2):127–131
Scott SL (2010) A modern Bayesian look at the multi-armed bandit. Appl Stoch Model Bus Ind 26
(6):639–658
Switzer P, Green AA (1984) Min/max autocorrelation factors for multivariate spatial imagery.
Stanford University, Department of Statistics, Standford
Thompson WR (1933) On the likelihood that one unknown probability exceeds another in view of
the evidence of two samples. Biometrika 25(3/4):285–294
Fixing Panel Artifacts in Localized Indicator
Kriging (LIK) Block Models
Abstract Three types of panel artifacts are created by the localized indicator
kriging (LIK) methodology. Two of these artifact types can be fixed (eliminated)
by a two-step process that consists of averaging models with different panel origins
and then performing a global re-localization. The first type of panel artifact, and
probably the most important, is caused by change-of-support transformations,
which are independent and applied on a panel-by-panel basis. The second type of
artifact is the most common but is usually undetectable. This artifact is caused by
the fact that there are several possible LIK models where the only difference is the
panel origin, and although each of the resulting models has the same global
distribution, the high-grade blocks change location. The third type of artifact is
caused by using a search ellipse that is too large (nonstationary), which is more
typical in areas where drill data is sparse. Unlike the first two artifact types, this type
needs to be fixed by adjusting search parameters. Artifacts should be eliminated
from LIK models whether they are obvious as in the case of those caused by large
change-of-support transformations or are simply the added spatial noise caused by
having more than one origin.
1 Introduction
The LIK methodology uses the histograms (probability distributions) from a mul-
tiple indicator kriging (MIK) estimation, which are created using a large block size
(panels), and localizes those distributions into smaller SMU-sized blocks using an
ordinary kriging (OK) model as a localizer (Hardtke et al. 2009). The concept of
using one model to localize the distribution of a second estimation that was done at
W. Hardtke (*)
WCH Consulting LLC, 9808 S. Rosewalk Dr., Highlands Ranch, CO 80129, USA
e-mail: [email protected]
C. Wilson
Newmont Mining Corp, Elko, NV, USA
e-mail: [email protected]
a different scale was first presented by Marat Abzalov for use with uniform
conditioning (Abzalov 2006). Although LIK typically yields resource models that
have both the desired global target distribution and excellent location for the block
values within the panels, there is always at least some degree of global imprecision
in the location of the values that is caused by the panels. Any change in a block
value that is due solely to the panel location can be referred to as an artifact, but
there is no reason why they should be accepted as an innate part of a LIK model.
There are three types of artifacts and the most common type is undetectable and
in most cases it does not have to be eliminated or even addressed. This type of
artifact, which can be referred to as a change of origin artifact, is found in every LIK
model and is simply the fact that a change of panel origin changes the value of every
block in the model. Unless the exact value of an individual block is important as
perhaps may be the case in a high-grade underground mine, this type of artifact can
be ignored.
The second type of artifact is caused by change-of-support transformations and
usually needs to be addressed because this type of artifact can be incredibly obvious
to the point that they make the model unacceptable. This type of artifact arises
because the transformations are done independently on a panel-by-panel basis.
Because the transformations cause all the block values within the panel to become
closer to the mean value of a panel, adjacent panels with dissimilar mean values can
suddenly become quite visible. In general, the greater the transformation (the
amount of variance reduction), the more pronounced the panel boundaries become.
The third type of artifact that is caused by having too large a search for the size of
the panel is referred to as a non-stationarity artifact. This type of artifact is much
more obvious in areas with sparse data and usually results in a pattern where each
panel has its highest-grade block in the same relative position. In most cases this
type of artifact is only cosmetic and unimportant because it generally occurs in
areas that will be classified as inferred.
In any LIK model, there is a defined number of possible panel origins that is
equal to the number of SMU-sized blocks per panel. Each of these possible panel
origins will result in a unique LIK model where the block values vary locally but
have a nearly identical global distribution. It is analogous to a simulation where all
of realizations are equally probable and none of them is better than the others.
However, because there are a finite number of panel origins for any LIK model,
they can easily be combined into a single “best” answer.
Two approaches for combining LIK models to eliminate artifacts were tested
during this study. In the first approach, the value for each block was randomly
selected from all the possible LIK models. This resulted in a final model that
eliminated all the artifacts while retaining the original global distribution, but it
added a degree of spatial variability that probably is not real. This approach was
abandoned early on and is not presented here.
The second approach was to average all the blocks from each of the LIK models
into a single block model. This resulted in a model that had a very good spatial
distribution for the gold, but the averaging of values lowered the variance of the
global distribution. Then in order to correct the smoothing, the original distribution
Fixing Panel Artifacts in Localized Indicator Kriging (LIK) Block Models 215
was re-localized using the averaged block model as the localizer. The final
re-localized model had the best possible local distribution of gold combined with
the desired global distribution. This approach gave very good results and the
remainder of the paper assumes this methodology.
change in the panel distribution results in a very obvious panel boundary where it
comes into contact with the low-grade panel directly below it. In a case as obvious
as this, the contact between ore and waste actually changes to the panel boundary,
which would be a real problem for an engineer trying to design a mine.
Fixing Panel Artifacts in Localized Indicator Kriging (LIK) Block Models 217
Fig. 4 LIK model (no. 10) with an extreme amount of variance reduction
LIK models are very sensitive to data density, and when the panel size is too small
or the search distance is too large, strange-looking typically repetitive artifacts are
usually created. This type of artifact is most common in low-grade areas where
panels should be totally low grade, but the search ellipse is large enough to include
high-grade composites that are well outside the panel being estimated.
In Fig. 5, large search distances result in high-grade blocks being estimated in an
area that is obviously low grade (solid blue squares are low-grade drill intercepts).
This is clearly a situation where the local search ellipse is nonstationary in nature
218 W. Hardtke and C. Wilson
and can easily be fixed by changing the search parameters. The same sort of thing
happens all the time in OK models but is not easy to see because the effects of the
high-grade composites in the search are spread more evenly over all the blocks.
However, LIK partitions the misplaced high-grade values into one or two blocks per
Fixing Panel Artifacts in Localized Indicator Kriging (LIK) Block Models 219
panel making them very easy to see. In this example of an area that should be all
low grade (blue <0.2 g/t), the LIK model has created a few high-grade blocks (red
>1.0 g/t).
This type of artifact can usually be fixed by changing the kriging parameters in the
MIK model (panel size). In some cases, where the search cannot be shortened, the
panel size probably needs to be increased. The methodology of averaging and making
a global re-localization does not eliminate this type of artifact; it merely changes its
location. Figure 6 shows the same area with a more appropriate search ellipse.
block in the averaged model. The target distribution can be any 1 of the 16 LIK
models because they have the almost exactly the same distribution. Then it is a
straightforward step of sorting the blocks in both models (common blocks only) and
replacing the averaged values with the corresponding target value.
Figure 10 contains an example of the final model. It has no panel artifacts; it has
the desired distribution, and the location of the blocks is probably as good as it can
be. It looks very much like the averaged model, but after close examination it
definitely has less smoothing like the target distribution.
Figure 11 shows the distribution of blocks before and after the artifact removal,
and there is very little difference between the original distribution with artifacts
(blue) and the final distribution after artifact removal (red).
6 Summary
All LIK models have panel artifacts but they are not always important enough to
need to be removed. In most cases artifacts are just cosmetic and do not affect the
quality of the model or the subsequent mine design, but when they need to be
eliminated, it can be done with a two-step process. The first step is to make a new
Fixing Panel Artifacts in Localized Indicator Kriging (LIK) Block Models 223
localizer by averaging all of possible LIK models into a single model. The second
step is to adjust this new averaged model so it has the same global distribution as
any of the LIK models.
Bibliography
Abzalov M (2006) Localised Uniform Conditioning (LUC): a new approach for direct modeling of
small blocks. Math Geol 38:393–411
Hardtke W, Allen L, Douglas I (2009) Localised indicator kriging. 35th Apcom Symposium, Perth
WA, pp 1–4
Hoerger S (1999) Implementation of indicator kriging at Newmont gold company.. 23rd APCOM
Proceedings, pp 202–213
Implications of Algorithm and Parameter
Choice: Impacts of Geological Uncertainty
Simulation Methods on Project Decision
Making
Abstract Uncertainty in geological models and grade uncertainty are two major
contributors to the total resource uncertainty of a mining project. Previous attempts
at determining uncertainty in geological models using methods such as MPS
(multiple-point statistics), SIS (sequential indicator simulation), and multiple appli-
cations of RBF (radial basis functions) with different parameters have shown that it
is nontrivial; the uncertainty profiles are dependent on the method and the param-
eters selected. Most of the methods tested require additional information in the form
of either local probabilities or proportions derived from the existing geological
interpretation or a conceptual geological model in the form of a training image. This
makes some methods amenable to use in the early stages of a project because the
method allows for a more complete testing of different geological concepts. In later
stage projects where there is an increased level of confidence (due to the amount of
data collected) in the geologic interpretation, methods that achieve ranges of
uncertainty around the interpretation likely provide a more realistic assessment of
uncertainty. This paper details the continuation of research into geostatistical tools
suitable for the evaluation of geological uncertainty in order to further understand
the intricacies of the methods and the impact of the technique on the resulting
uncertainty profile. Suggestions of which methods to use based on the amount of
geological information available are provided.
1 Introduction
This section gives a brief description of the categorical methods applied. The
techniques were used to generate simulated models of the main mineralized domain
for two deposits. The geostatistical tools trialed include:
• Indicator-based methods (SIS-LVM) (Deutsch 2006).
• Methods based on distance functions (Munroe and Deutch 2008a, b; Wilde and
Deutsch 2012).
• Methods based on multiple-point statistics (snesim) (Strebelle 2002; Remy et al.
2009).
• Methods based on radial basis functions (Leapfrog®) (Stewart et al. 2014) using
different sets of parameters. While not a simulation method, Leapfrog® is
increasingly used to quickly build alternate models of geology and to determine
the impact of the geological interpretation on project financials.
Xn
I *LVM ðu; kÞ pk ðuÞ ¼ /¼1
μsk
α ðuα ; kÞ½iðuα ; k Þ pk ðuα Þ ð1Þ
where
ILVM (u; k) is the simple kriged estimate at location u for category k.
pk(u) is the probability of category k at location u.
μsk
α (uα; k) are the simple kriging weights for data at location uα for category k.
i(uα; k) are the indicators for category k at data location uα.
pk(uα) are the local probabilities for category k at location (uα)
The estimates for each category are performed independently which can lead to
order relations deviations. Due to the noisy nature of the simulations, a post-
processing step is usually applied to clean the simulations up (Deutsch 2006).
This method is based on the interpolation of a signed distance function built using
conditioning data. The distance function can be seen as the distance between a data
point itself and the nearest sample belonging to another domain. Positive and
negative distances are used to distinguish between inside and outside a domain.
Distances depend on the orientation, geologic shape, and extension of the rock
types. The methodology is as follows (Munroe and Deutch 2008a, b; Wilde and
Deutsch 2012):
• Code all data points as either inside or outside the domain of interest.
1 if inside domain of interest at u/
i ð u/ Þ ¼ /¼ 1, , n ð2Þ
0 otherwise
where u is the location vector, / is the sample index, and n is the number of
samples.
• Next calculate the distance for each sample to the nearest sample located in
another domain. If u/ is inside the domain, the distance is set to negative or else
the distance is positive. A factor C is added (or subtracted) to the distance
function value which increases the difference between positive and negative
values. The C factor is calibrated using a jackknifing approach.
þu/ uβ þ C if iðu/ Þ ¼ 0
df ðu/ Þ ¼ /6¼ β ð3Þ
u/ uβ C if iðu/ Þ ¼ 1
Simulate the boundary only for locations where the distance function estimates
lie between C and C.
df l ðuÞ ¼ 2CG1 yl ðuÞ C ð4Þ
This methodology is geared toward binary systems (in and out a particular
domain) and does not handle multiple domains. The calculated distance function
is nonstationary which can make variogram inference challenging. Another
approach using distance functions can be found in Cáceres et al. 2011.
structural trend which is similar to locally varying anisotropy. The approach allows
for locally varying directions of continuity. Volumes were created at desired values
to represent probability shells. Leapfrog® software was used for this approach. The
RBF interpolant is similar to the general expression of dual kriging (Stewart et al.
2014).
X XK
sðx Þ ¼ i
ωi φðjx xi jÞ þ k
ck qk ðxÞ ð6Þ
where
xi are the data locations over which the interpolation is to be constrained.
ωi are RBF coefficients (weights).
φk(x) is a spatial distance function (the RBF – from which the method takes
its name).
The term on the right refers to the set of K drift functions (qk(x)), each having a
coefficient (ck) applied globally across all data.
3 Case Studies
This section details the use of the geostatistical tools described above to build
simulated models of the main mineralized domains for the Merian deposit located
in East Suriname and for the Subika deposit located in the Ahafo Region in Ghana.
It starts with a description of the geologic settings of the two deposits. While both
deposits are orogenic in genesis, each has different types of controls on minerali-
zation from which to assess uncertainty.
The Merian deposit lies within lower Proterozoic-aged rocks of the Guiana Shield
in northeast Suriname, South America, approximately 100 km east of the capital
Paramaribo (Fig. 1). In Suriname the Guiana Shield is composed of distinct, east-
west trending belts of low-grade metamorphic rocks which are separated by large
areas of granitic rocks and gneisses. Gold mineralization within the Merian deposit
occurs as a vein-type Proterozoic lode-gold deposit; gold is found within and
immediately adjacent to quartz veins, quartz stockworks, and irregular quartz
breccia bodies. Host rocks are composed of highly folded sandstones and siltstones.
Gold mineralization at Merian occurs over a strike length of approximately 3.5 km,
elongate in a northwest-southeast direction, and over a width of 200–600 m.
The Merian geologic model was constructed using the following attributes
collected from drill core logging:
Implications of Algorithm and Parameter Choice: Impacts of Geological. . . 231
Percent quartz vein content is recognized as the primary control on the geometry
and grade of gold mineralization at Merian. Figure 2 displays three sections of the
Merian mineralized envelope as coded in the Merian resource model. In tightly
drilled areas, the drillhole spacing is approximately 25 m across and 25 m along
sections.
In order to build simulated models of the quartz vein density shape and to assess
the uncertainty in its contained volume, categorical simulations were generated
using various techniques:
• Fifty simulations using the SIS-LVM approach with local varying probabilities
were calculated using the geological interpretation and moving window sizes of
12 12 m (0.5 times the tightest drillhole spacing), 25 25 m (the tightest
drillhole spacing), 50 50 m (twice the tightest drillhole spacing), and
232 A. Jewbali et al.
Fig. 2 Main mineralized domain for the Merian deposit with drillhole data. The area within the
white boundary (right) is used for the study
Fig. 3 Local varying probabilities to be inside the mineralized domain derived from the resource
model (plan view) – high probabilities in warmer colors
75 75 m (three times the tightest drillhole spacing) (Fig. 3). It is expected that
larger filter distances will yield wider bands of uncertainty. Besides the condition-
ing data, this method uses the geological interpretation (as per the resource model).
• Fifty simulations were generated using the distance function approach. Three
C factors were used to determine the sensitivity of the uncertainty profile to the
C factor (C ¼ 20, approximately equal to the tightest drillhole spacing; C ¼ 50,
twice the tightest drillhole spacing; and C ¼ 100, four times the tightest drillhole
spacing) (Fig. 4). It is expected that larger C factors will result in wider
Implications of Algorithm and Parameter Choice: Impacts of Geological. . . 233
Fig. 4 Interpolated distance function for different C factors (red represents distances between –C
and C (plan view)
uncertainty intervals. The geological interpretation was initially used to code the
data as being inside or outside the mineralized domain. The logged information
from the drillholes could have been used instead; however, there was some
grouping of highly variable intervals into the geologic interpretation, which
would not have been accounted for.
• Fifty multiple-point simulations using snesim, where the resource model in
Fig. 2 was used as the training image (servo system factor of 0.5).
• Alternate models using Leapfrog®: in addition to the simulations, three volumes
were produced using RBF indicator probability shells derived from conditioning
data. Volumes were analyzed based on selecting volumes contoured from a range
of interpolated values (P30 to P50). Indicator statistics were analyzed to determine
the balanced shell (P38) which is the volume that includes as many indicator data
misclassified inside the shell as it excludes indicator data misclassified outside the
shell and to determine the Russian doll shell (P30) where the next larger volume
incrementally includes data where the indicator mean is less than the probability
of the shell. In order to produce a range of results, an arbitrary smaller (conser-
vative) volume was selected at P46 (Inglis 2013).
Plan views of the generated simulations for the various methods are shown in
Figs. 5 and 6. The figures show that the simulations generated by SIS-LVM are
234 A. Jewbali et al.
Fig. 6 Simulation results for snesim and the distance function approach (plan view)
Implications of Algorithm and Parameter Choice: Impacts of Geological. . . 235
noisier compared to the model. This is especially so, when a large window is used to
generate the local probabilities. As shown in Fig. 3, this creates lower probabilities
in areas where the probability to be inside the mineralized shape was previously
high (with a smaller window size). This causes the simulations to have a dispersed
appearance. After cleanup, the simulations for the 25 25, 50 50, and 75 75
still do not display the same level of continuity in the north-south direction seen in
the resource model. The snesim simulations also appear noisier compared to the
resource model. The distance function approach produces simulations that look
similar to the resource model. Some artifacts are visible especially with large
C factors which tend to create “donut holes” inside a mineralized volume.
Figure 7 displays the uncertainty profile for the mineralized volume. It shows the
smallest, largest, and average (over all 50 simulations) volume. It also shows the
width of the uncertainty interval (difference between the smallest and largest
volume). In general from these figures, the following can be derived:
• For most of the methods, the volumes fluctuate above or below the volume
defined by the resource model. Only three methods (SIS-LVM 75 75 m with
post-processing clean, Leapfrog®, and distance function C50 and C100) contain
the volume defined by the resource model in their uncertainty interval. Whether
simulation volumes are above or below, the mineralized volume defined by the
resource model depends on the chosen parameters.
• As expected, for the SIS-LVM, the wider the uncertainty window used to
generate the local probabilities, the wider the uncertainty interval. The width
of the uncertainty interval for SIS-LVM (75 75 m) is wider compared to that of
SI-LVM (50 50 m). For this approach there appears to be a bias related to the
size of the window used to create the local probabilities. For larger windows the
simulations tend to generate more mineralized volume (i.e., also more
dispersed).
• For the distance function approach, the models derived from larger C factors also
tend to have wider uncertainty intervals. For smaller C factors, the simulated
models tend to filter out small pods of mineralization (zones that one can argue are
more uncertain), which is the main reason the mineralized volume of the simulated
Fig. 7 Uncertainty profile for mineralized volume (left) and width of the uncertainty interval for
mineralized volume (right)
236 A. Jewbali et al.
models is lower compared to the mineralized volume in the resource model. For
larger C factors, this effect is negated due to the large uncertainty window.
• For snesim the width of the uncertainty interval is approximately similar to
SIS-LVM 25 25 m.
• The Leapfrog® approach delivers the widest uncertainty interval which none of
the other methods are able to match.
In general after simulation of the mineralized volume, simulations of grade are
generated in a hierarchical approach to derive the uncertainty interval for metal,
tonnes of ore, and average grade for annual/quarterly production volumes. If, for
example, the SIS-LVM 12 12 m was chosen for the geological simulation, one
would tend to think that the resource model was too optimistic. The opposite can be
said for the use of SIS-LVM 50 50 m, i.e., the model is too pessimistic. If a
distance function approach with C ¼ 50 was chosen for the geology simulations,
one would possibly conclude that the resource model is reasonable in terms of
contained volume.
The generation of stochastic simulations of geology that quantify uncertainty
can be difficult, and the end result can be heavily influenced by the parameters and
method chosen. It is therefore imperative that one understands the impact the choice
of method and parameters will have on the uncertainty profile. Most of these
methods (except for Leapfrog®) are based on information derived from the existing
resource model whether through rock-type proportions or resource model as train-
ing image. By doing this there is an implicit assumption that there is some level of
confidence in the resource model. For pre-feasibility or feasibility stage projects
where there is a lot more data available, this assumption might be justified;
however, for early stage projects with far less data, these methods might not be
applicable.
The Subika deposit is the southernmost of the known Ahafo deposits (Fig. 8) and is
hosted entirely within the granitoid package in the hanging wall of the Kenyase
Thrust. High-grade gold mineralization is focused in a dilatant fracture zone,
locally referred to as the magic fracture zone (MFZ). This zone ranges from 1 to
60 m wide with a halo of lower-grade mineralization extending out to 30 m. A
number of higher-grade ore shoots, which appear to be controlled by dilatant left-
lateral jogs in the MFZ, are recognized and plunge steeply to the southeast.
Quartz-sericite-pyrite and iron-carbonate (QSP-Fe) alteration is the dominant
alteration associated with high-grade mineralization. Alteration fluids appear to
have accessed the MFZ via a network of shallow angle, brittle fractures within an
overall steeply dipping shear zone. QSP alteration intensities are logged as 1, 2, or
3. The combined QSP 2/3 alteration forms the basis of the higher-grade population,
while QSP 1 alteration correlates well with the lower-grade population. These two
Implications of Algorithm and Parameter Choice: Impacts of Geological. . . 237
Fig. 9 Subika geology model (left) and block probabilities for alteration type (unaltered, QSP1
and QSP23) (right)
alteration categories have been used to define the geologic framework used in the
resource estimates and are the main focus of this study.
attempted to reproduce the target input proportions. For these three methods, the
simulated volumes fluctuated slightly above the volume of the resource model, with
the exception of one MPS case. The following methods/parameters were employed:
• Fifty simulated models using SIS-LVM were constructed. The node spacing
was 2 2 2 m, and the window of influence applied to the LVM was based
on the average drill spacing (approximately 35 m) within the simulated area
(Fig. 9). This parameter is based on drill spacing which allows block proba-
bilities to be calculated within ranges equivalent to the spacing of hard
information, resulting in bands of uncertainty around geologic features at a
scale similar to the drill spacing. Large filter distances associated with wide-
spaced drilling will yield large bands of uncertainty around interpreted geo-
logic features, while small filter distances associated with close-spaced drilling
will yield tight bands of uncertainty around the same features. This coincides
with the idea that uncertainty should decrease as drilling density increases. In
addition, anisotropy derived from the variogram model was applied to the
calculation of the LVM in order to preserve the preferred orientation of the
structure and mineralization. The resulting SIS with the LVM used as control
is shown in Fig. 10.
• Fifty simulated models were generated using the distance function approach
(only applied to the QSP23 volume). The method being a binary approach
cannot facilitate more than one domain. Simulated models were developed
on 6 12 6 m blocks with C factors of 10, 20, and 50 (Fig. 11). Figure 11
also displays the simulated models in 3D, which clearly indicate areas where
the mineralization is quite thin (square box). Areas with limited data also
tend to show edge effects due to extrapolation of the distance function.
• Three alternate models were generated using Leapfrog®. Three volumes were
produced using RBF indicator probability shells derived from conditioning
data. Volumes were analyzed based on selecting volumes contoured from a
range of interpolated values (P30 to P50). Indicator statistics were analyzed to
determine the balanced shell which is the volume that includes as many
indicator data misclassified inside the shell as it excludes indicator data
misclassified outside the shell and to determine the Russian doll shell where
the next larger volume incrementally includes data where the indicator mean is
less than the probability of the shell. In order to produce a range of results, an
arbitrary smaller (conservative) volume was also selected (Inglis 2013).
• Fifty simulated models were generated using snesim. This case was developed
on 6 12 6 m blocks, using the interpreted geologic model as the training
image (Fig. 12). Three different scenarios were developed, where the servo
system factor was modified in each run (0.1, 0.5 and 0.9). This parameter
controls how snesim reproduces the target input proportions from the TI. The
higher the factor, the better the reproduction of the input target proportions. The
selection of this value is somewhat subjective and should be chosen with the
quality of the geologic model in mind. There are other parameters that have an
impact on the result, but they were not tested during this exercise.
Implications of Algorithm and Parameter Choice: Impacts of Geological. . . 239
Fig. 11 Interpolated distance function for different C factors (red) represents distances between –
C and C (plan view) and 3D view of simulated models (for QSP23) generated with the distance
function approach (different C factors). Notice edge effects (circle) where there is limited data
A summary of the uncertainty profiles for QSP1 and QSP23 is shown in Fig. 13.
Note that the SIS-LVM and MPS 0.5 show similar results, while MPS 0.1 results in
a wider range of uncertainty and more volume than any of the other techniques.
MPS 0.9 on the other hand has a much smaller range of uncertainty and appears to
be biased low on volume. It is understandable that the range of uncertainty would
diminish with an increasing servo system factor because reproduction of the input
target proportions is more strictly enforced; however, the reduction in volume is not
so easily explained.
Implications of Algorithm and Parameter Choice: Impacts of Geological. . . 241
Fig. 13 Uncertainty profile for QSP1 and QSP23 (top) with the width of the uncertainty interval
(bottom)
The results from the distance function approach show larger levels of uncertainty
with larger C factors. For smaller C factors, the simulated models tend to filter out
sections where the mineralization is thinner (zones that one can argue are more
uncertain), which is the main reason the mineralized volume of the simulated
models is lower compared to the mineralized volume in the resource model.
Figure 13 also illustrates the differences between the minimum and maximum
volumes of each technique. This shows the variability of ranges of uncertainty
depending on technique and parameter selection.
All the techniques yield promising results, and given that it is reasonable to put a
high level of confidence on the geologic interpretation, it seems acceptable to try to
achieve ranges of uncertainty that fluctuate around the interpretation. In earlier
stages of the Subika project, this would not have been the case. Over time tonnages
at Subika have shown large fluctuations caused by the wide-spaced drilling and
overall complexity of the geologic framework. With drill spacing at ~35 m and a
significant amount of thought and effort placed on the geologic interpretation,
Newmont is to the point of applying these results to the risk associated with the
mine plan.
The two case studies have shown that quantification of geological uncertainty for
purposes of defining project risk is not trivial. Uncertainty profiles are highly
dependent on method and parameter choice:
242 A. Jewbali et al.
• For the SIS-LVM approach, the size of the uncertainty window has an impact on
the width of the uncertainty profile, with wider profiles resulting in wider
uncertainty windows (not necessarily centered on the input resource model)
and more dispersed simulations.
• For MPS the servo system factor also appears to control the width of the
uncertainty interval with servo system factors closer to one delivering narrower
uncertainty intervals.
• For the distance function approach, the width of the uncertainty interval is a
function of the C factors; larger C factors result in wider uncertainty intervals.
For both deposits tested, lower C factors tended to filter out the thinner more
discontinuous mineralized zones generally resulting in tighter uncertainty inter-
vals and slightly less volume compared to the resource model.
• Models built using Leapfrog® showed wider bands of uncertainty.
Most of the methods (except Leapfrog®) require additional information (i.e.,
proportions of rock types, training image, etc.) derived from the resource model
which requires some confidence in the quality of the resource model. At early stages
of a project, when there is limited information, this might not be appropriate;
instead one might also try to incorporate the uncertainty in the rock-type pro-
portions/training image in the simulated models.
Leapfrog® appears to be a better tool at early stages of a project because it allows
for a more complete testing of different geological concepts. In later stage projects
where there is an increased level of confidence (due to the amount of data collected)
in the geologic interpretation, other methods that achieve ranges of uncertainty
around the interpretation likely provide a more realistic assessment of uncertainty.
The case studies have also shown that in order to get ranges of uncertainty that
fluctuate around the interpretation, the parameters for the various methods need to
be selected carefully.
Bibliography
Cáceres A, Emery X, Aedo L, Gálvez O (2011) Stochastic geologic modelling using implicit
boundary simulation. In Beniscelli J, Kuyvenhoven R, Hoal K (eds) Proceedings of the 2nd
international seminar on geology for the mining industry
Deutsch C (2006) A sequential indicator simulation program for categorical variables with point
and block data: BlockSIS. Comput Geosci 32:1669–1681
Guardiano F, Srivastava M (1993) Multivariate geostatistics: beyond bivariate moments. In:
Soares A (ed) Geostatistics Troia 1992, vol 1. Kluwer Academic, Dordrecht, pp 133–144
Inglis R (2013) How to select a grade domain-A gold mine case study in exploration, resource and
mining geology conference, Cardiff, UK,.21–22 October
Jewbali A, Perry R, Allen L, Inglis R (2014) Applicability of categorical simulation methods for
assessment of mine plan risk. In: Proceedings orebody modelling and strategic mine planning
symposium 2014, pp 85–98 (The Australasian Institute of Mining and Metallurgy: Melbourne).
Reprinted with the permission of The Australasian Institute of Mining and Metallurgy
Lui Y (2006) Using the Snesim program for multiple point statistical simulation. Comput Geosci
32:1544–1563
Implications of Algorithm and Parameter Choice: Impacts of Geological. . . 243
Munroe MJ, Deutsch C (2008a) Full calibration of C and beta in the framework of vein type
deposit tonnage uncertainty, Center for Computational Geostatistics Annual Report 10. Uni-
versity of Alberta, Edmonton
Munroe MJ, Deutsch C (2008b) A methodology for modeling vein-type deposit tonnage uncer-
tainty, Centre for Computational Geostatistics Annual report 10. University of Alberta,
Edmonton
Remy N, Boucher A, Wu J (2009) Applied geostatistics with SGeMS: a user’s guide. Cambridge
University Press, Cambridge, UK, 263p
Stewart M, de Lacey J, Hodkiewicz PF, Lane R (2014) Grade estimation from radial basis
functions – how does it compare with conventional geostatistical estimation. In ninth interna-
tional mining geology conference 2014, pp 129–139 (The Australian Institute of Mining and
Metallurgy)
Strebelle S (2002) Conditional simulation of complex geological structures using multiple-point
statistics. Math Geol 34:1–21
Wilde JB, Deutsch C (2012) Kriging and simulation in presence of stationary domains: develop-
ments in boundary modelling. In: Abrahamsen P, Hauge R, Kolbjørnsen O (eds) Geostatistics
oslo 2012. Springer, Dordrech, pp 289–300
Approaching Simultaneous Local and Global
Accuracy
Abstract When estimating block grades for mining purposes, the currently available
methods allow us to maximize the accuracy of either global grade and tonnage curve
prediction or local block selection but not both at once. Locally accurate block
estimates provide the best result during actual selection and mining but can give
highly distorted global grades and tonnages at cutoffs above zero. Globally accurate
block estimates provide good prediction of grade and tonnage curves but perform
badly during actual selection giving much higher misclassification rates leading to
serious degradation of value of the material selected for processing. These statements
hold true in varying degrees for all scales and combinations of sample spacing and
block size. This paper puts forward a method that retains the properties of accurate
global estimation while simultaneously approaching maximum local accuracy.
The process is a simple application of rank and replace combining two estimates,
one that targets local block accuracy and one that targets actual block variability.
The method is empirically demonstrated using a case study using real data. The
conclusion, for this data set, is that local selection accuracy can be greatly improved
(but not maximized), in comparison to existing methods, while maintaining grade
and tonnage curve accuracy that results from true block variability.
Comparisons with ordinary kriging, sequential Gaussian simulation, turning
bands, local uniform conditioning, and ordinary kriging with reduced sample
numbers are presented.
1 Introduction
The method described in this paper is a remarkably simple one and it must surely have
been used or published previously; however the author cannot find any references to
anything similar. The underlying estimate is conventional ordinary kriging (OK) with
an optimized search neighborhood for the specified block size designed to maximize
local block accuracy. The OK block grades are then ranked in grade order. A second
estimate, using any method that targets true block variability (global change of
support, simulation, degraded neighborhood kriging, etc.), is made and the block
grades also ranked. The grades from the second estimate are then assigned to the
locations defined by the OK estimate by rank order. This happens to be a direct
generalization of the LUC method proposed by (Abzalov 2006). For the purposes of
this paper, we will call the general method rank and replace (RR).
Some other methods of smoothing reduction and or correction that have been
proposed are (Journel et al. 2000) using spectral methods and (Richmond et al. 2009)
using an affine correction. Topical papers around the balance between the reproduc-
tion of global variability and the reproduction of local block accuracy required for
different stages of a project are Krige (1951, 1994, 1996) and Isaaks (2005).
It will be shown in the case study that the resulting block RR estimates now have
(our best estimate from available data of) the true block variability globally as well
as local selectivity very close to that obtainable by OK and are superior to that
obtainable by any simulation methods.
Hence, in the mining context, we now have a block model that can be used for
mine planning purposes that has (our best estimate from available data of) the actual
global block grade, tonnage, and conventional profit curves as well as (our best
estimate from available data of) local block accuracy.
2 Case Study
The data set is a real gold data set from a mined-out Australian opencut mine. It
consists of a widely spaced exploration drilling data set at approximately 12 m by
25 m by 1 m spacing and a corresponding close-spaced grade control data set at
approximately 2.5 m by 4.0 m by 2.5 m spacing. The results presented are block
models calculated from the wide-spaced exploration data with block models from
the grade control data as reference.
The typical mining rule of thumb is that, for good local estimates, block sizes should
approximate or be no less than half of the data spacing. This often results in blocks
that are much larger than the anticipated selective mining unit (SMU) block size.
Kentwell (2014) describes and quantifies some of the issues around ordinary kriging,
smoothing and block size related to this and how estimates of tonnage and grade at
higher cutoffs can significantly depart from the true tonnages and grades.
In order to evaluate the potential benefits of the RR method, block models at
both SMU scale, 2.5 m by 4 m by 2.5 m, and data scale, 10 m by 24 m by 10 m, were
Approaching Simultaneous Local and Global Accuracy 247
compared during the study. In the interest of brevity, only the results for the SMU
scale models are presented in detail here.
2.3 Process
For the SMU scale block model, the estimates as in Table 1 were produced. For the
sake of direct comparability of methods, the variogram used for all estimates was
that derived from the grade control data set with ranges and relatives nuggets/sills
held constant. Both the exploration and grade control data were composited to 3 m
intervals. De-clustering weights, top cutting, and restricted searches for high grades
were used as appropriate to ensure all estimates validated and that average grades
were within an acceptable tolerance at zero cutoff. Two RR estimates are calculated
and compared, one derived from a Gaussian anamorphosis with global change of
support and one derived from a single simulation.
For the RR estimate (Ana RR) derived from the global change of support, the
individual ranked block grades were calculated by the following procedure:
1. Calculate the block histogram via Gaussian anamorphosis and global change of
support using the exploration samples and the variogram.
2. Report the tonnage curve at 200 evenly spaced cutoffs across the full range of
grades.
3. Export the cutoffs and corresponding tonnages to a curve fitting software and fit
the data. In this case a Savitzky–Golay smoothing (Savitzky and Golay 1964)
(a form of moving window polynomial) fit was used.
4. Split the fitted curve into as many tonnage increments as there are blocks in the
model. The resulting ascending cutoffs become the new ranked block grades.
5. Import these ranked anamorphosis grades back to the underlying OK block
model using the OK rank identifier to locate each new anamorphosis grade of
the corresponding rank.
This process can also be completed using the local uniform conditioning (LUC)
tools in the Isatis software package but with the panel model being a single block
covering the entire domain. While not detailed in this paper, the author has
replicated the Ana RR results using single panel LUC in Isatis.
For the RR estimate, SGS RR derived from the sequential Gaussian simulation
(SGS) the SGS grades and underlying OK grades that are both ranked and then the
SGS grades that are assigned to the OK locations based in the rank identifier. The
problem with deriving the grades from a set of simulations is choosing which
individual simulation as they will all be a little different.
With the exception of the splitting and interpolation of the tonnage curve from
the change of support anamorphosis into block size increments, all calculations for
both the Ana RR and the SGS RR were performed in the Geovariances Isatis
software package.
Even though we have a close-spaced reference data set, in this case with approx-
imately one grade control sample within one SMU sized block and the blocks sizes
and data spacing are well inside the range of the variogram, we still need to estimate
the SMU model block grades. Even with this well-informed data and block con-
figuration, significant smoothing still occurs during OK. The question is then which
estimation parameters and/or method creates the closest approximation to reality to
use as our reference model? We have the same problem as the one we set out to
solve only at a smaller scale. At the risk of getting swamped with data, we will
examine two versions of “reality.” The first is the block model that results from
grade control data estimated by OK using a degraded estimation neighborhood
selecting only 4 (GC 4) surrounding samples and intended to reproduce global
block variability. The second is the block model that results from grade control data
estimated by OK with an optimum neighborhood selecting 32 (GC 32) surrounding
samples and intended to insure local block accuracy.
2.4 Results
In comparing the results of the different estimation methods, we will talk about both
prediction and performance. These terms are specifically defined here.
Approaching Simultaneous Local and Global Accuracy 249
Prediction is what the models say will happen, for example, tonnages and grades
at different cutoffs. This can be compared to the prediction of our reference model.
Prediction is inherently global.
Performance is the interaction with our reference “reality.” In other words if a
particular method’s block model was used to actually select/reject material at
specific cutoffs, then what grades and tonnages actually result by applying that
selection to the reference model? Performance is local in that it is a direct block-by-
block comparison albeit that the net result is expressed as a single correlation
coefficient or a cutoff curve.
In the normal case of events during planning and prior to mining, we only ever
have predictions. Only after the fact and after mining and processing do we get any
real information on performance.
2.4.2 Statistics
The block model statistics of the different estimation methods and correlations with
the two grade control reference models are shown in Table 2. The results are listed
in order of their correlation with the reference models. Note that the two RR
methods produce correlations with both reference models that are very close but
just less than the correlation resulting from optimal OK. As expected degraded OK,
Table 2 Statistics for SMU model estimates
Correlation Correlation
Method Mean Variance with GC 32 with GC 4 Min Max Comment
GC 4 1.72 3.75 0.88 1.00 0.00 38.56 Reference model 1
GC 32 1.70 2.17 1.00 0.88 0.00 18.25 Reference model 2
EXP 80 1.73 0.51 0.57 0.45 0.40 10.67 Best correlation
with reference
models 1 and 2
SGS 1.73 1.66 0.56 0.45 0.01 19.12 Close to best
RR correlation with
reference
models 1 and 2
Ana 1.76 2.50 0.56 0.44 0.15 27.96 Close to best
RR correlation with
reference
models 1 and 2
EXP 6 1.67 0.89 0.51 0.40 0.07 14.06
LUC 1.74 2.13 0.41 0.33 0.07 12.61
SGS 1.70 1.57 0.27 0.22 0.01 19.12 n 15 (of 100)
Sim selected for best
n15 match to sample
mean
TB Sim 1.70 2.69 0.26 0.21 0.00 23.92
n82
250 D.J. Kentwell
AU GC 4
5 5
0 0
0 5 10 0 5 10
AU EXP 80 AU EXP 80 Ana RR
AU GC 4
5 5
0 0
0 5 10 0 5 10
AU EXP SGS [00015] AU EXP TB [00082]
AU GC 4
5 5
0 0
0 5 10 0 5 10
AU EXP 6 AU EXP 80 LUC
Total Tonnage
Ana RR
SGS
TB
1000000
LUC
500000
0
0 1 2 3
Cutoff
4
Mean Grade
3
OK EXP 80
OK GC 4
2 SGS RR
OK GC 32
OK EXP 6
1 Ana RR
SGS
TB
LUC
0
0 1 2 3
Cutoff
SGS, and TB produce much lower correlations. Interestingly the LUC correlation is
in between optimal OK and the simulations but not as good as the RR methods.
Table 3 shows the kriging regression slope averages to put the relative qualities of
the kriging estimates in perspective.
Figure 1 shows the scatterplots of the correlations relative to GC 4 together with
the first bisector, conditional expectation, and one standard deviation from the
conditional expectation (dotted lines). Note that although the correlations for
EXP 80 and Ana RR are similar, Ana RR shows a much greater spread.
2.4.3 Prediction
Although difficult to see at the scale in this document, Figs. 2, 3, 4, and 5 show the
range of grade, tonnage, metal, and profit curves bounded by ordinary OK at one
252 D.J. Kentwell
Metal Tonnage
200
OK EXP 80
OK GC 4
SGS RR
OK GC 32
100 OK EXP 6
Ana RR
SGS
TB
LUC
0
0 1 2 3
Cutoff
OK GC 32
OK EXP 6
Ana RR
200 SGS
TB
LUC
100
0
0 1 2 3
Cutoff
extreme and the GC 4 reference model at the other. For the grade and tonnage
curves, despite some crossovers, the order, getting closer and closer to GC 4, is as
follows: EXP 80, EXP 6, SGS, SGS RR, LUC, Ana RR, GC 32, and TB. The main
point to take from these results is that, as expected, every other method is a better
global predictor than optimal OK in that all other grade, tonnage, metal, and profit
curves are closer to those of GC 4 which is our proxy for reality.
The profit curve is calculated as:
Profit ¼ tonnage above cut off *ðgrade above cut off cut off gradeÞ:
Approaching Simultaneous Local and Global Accuracy 253
Fig. 6 Performance
tonnage curves OK EXP 80
OK GC 4
SGS RR
1500000 OK GC 32
Total Tonnage
OK EXP 6
Ana RR
SGS
TB
1000000
LUC
500000
0
0 1 2 3
Cutoff
Ana RR
SGS
3 TB
LUC
0
0 1 2 3
Cutoff
3 Performance
Metal Tonnage
200
OK EXP 80
OK GC 4
SGS RR
OK GC 32
100 OK EXP 6
Ana RR
SGS
TB
LUC
0
0 1 2 3
Cutoff
OK GC 32
OK EXP 6
Ana RR
200 SGS
TB
LUC
100
0
0 1 2 3
Cutoff
4 Texture
5 Conclusions
The RR method does not make small block estimation any more locally accurate
than OK estimates. Small block OK and RR estimates from widely spaced data
retain on average a similar level of confidence as measured by the kriging regres-
sion slope or kriging efficiency of the underlying OK estimate. A poor quality local
block OK estimate is still a poor quality local block RR estimate.
The advantage of the RR method is global grade, tonnage, and profit curves are
closer to reality compared to optimal OK estimates and that the estimate retains on
average a level of block selectivity performance comparable to optimal OK. Also a
256 D.J. Kentwell
single model is available for mine planners to work with that therefore has the best
available estimate both locally and globally. The RR method is better adapted to
mine planning than either optimal OK or individual simulations. The RR method is
effectively an alternative implementation of LUC using a single panel. It gives
similar global predictions to multi-panel LUC and, for this case study, local
performance which is closer to reality than multi-panel LUC.
With RR estimates the spatial “texture” of the block grades is still smooth and
does not display the true texture at block scale. Approximations of true texture are
better estimated by simulations (SGS or TB for example) with the inherent loss of
local accuracy.
Approaching Simultaneous Local and Global Accuracy 257
Acknowledgments The author would like to acknowledge Daniel Guibal’s contribution to the
ideas behind this paper. When the author originally discussed the concept for the method with
Daniel, it transpired that he had already been thinking along the same lines and suggested the much
more elegant process of taking the final grades directly from the global change of support
anamorphosis rather than from a simulation or degraded kriging estimate. This resulted in the
Ana RR methodology detailed in this paper.
Bibliography
Abzalov M (2006) Localised Uniform Conditioning (LUC): a new approach for direct modelling
of small blocks. Math Geol 38(4):93–411
Isaaks E (2005) The kriging oxymoron: a conditionally unbiased and accurate predictor (2nd edn).
In: Geostatistics Banff 2004. Springer, Dordrecht, pp 363–374
Journel AG, Kyriakidis PC, Mao S (2000) Correcting the smoothing effect of estimators. Math
Geol 32(7):787–813
Kentwell D (2014) Aligning resource estimates with mine planning. In: Orebody modelling and
strategic mine planning. The Australasian Institute of Mining and Metallurgy, Perth, 185–192
Krige DG (1951) A statistical approach to some basic mine valuation problems on the Witwa-
tersrand. J Chem Metall Min Soc S Afr 52(6):119–139
Krige DG (1994) An analysis of some essential basic tenets of geostatistics not always practiced in
ore valuations. First regional APCOM, Slovenia, pp 15–28
Krige DG (1996) A basic perspective on the roles of classical statistics, data search routines,
conditional biases and information and smoothing effects in ore block valuations. Conference
on mining geostatistics, Kruger National Park Sept. 1994. UNISA
Richmond A, Gaze A, Horton J (2009) Desmoothing block models for underground mine design.
Seventh international mining geology conference. The Australian Institute of Mining and
Metallurgy, Perth, pp 265–268
Savitzky A, Golay MJ (1964) Smoothing by differentiation of data by simplified least squares
procedures. Anal Chem 38(8):1627–1639
Geostatistics for Variable Geometry Veins
Abstract This paper presents the idea of applying a modified “moving trihedral”
borrowed from the abstract Theory of the Geometry of Riemann, allowing us to
model the random functions L2(Ω, σ, P) of geostatistics, in a special 2-D plane
within such trihedral, a feature not available in any software existing in the market.
Thus, from working in the R2 space in that trihedral plane, we obtain our results in
the R3 physical space, in which we can then derive the desired linear or nonlinear
geostatistical results. The method presented here therefore can be considered akin
to a kind of spatially adaptative geostatistics. And so, among other applications of
this method, it becomes possible to apply geostatistics simply, without the need for
a detailed, three-dimensional geological model of the mineralization, defining
instead its contours in a simple plane, a feature very useful when modelling
irregular veins.
1 Introduction
The previously presented version of the applied methodology, used in the veins of
Centro Peru 1983 and in the vein of Navalmedio de Minas Almadén-Espa~na 1984,
was programmed in Fortran and Assembler, with its respective graphic limitations.
Said version was presented at Fontainebleau’s Cycle of Formation Specialisée in
Geostatistics as a graduate course, in the presence of Dr. Georges Matheron, who
authorised its publication in 1986 (Marin 1986). The present version is based on the
use of modern graphic 3-D and 2-D software, available in the market.
This methodology was applied as a consequence of irregularities in the vein
thickness, strike, dip and local domains, resulting from existing geological controls
in use at the aforementioned deposits. This can be seen in Fig. 1.
Let us bear in mind that vein systems are very complex (Bateman 1982) and that
the construction of geological model (three-dimensional) is difficult, somewhat
inaccurate and often uncertain, due to the fact that its elaboration is subjective
and that it takes a lot of time, because of the unavailability of required drills,
necessary to identify what is happening between one section and the other, for
example, the events that take place between a geological section that corresponds to
a left chimney of a given cut and another geological section that passes through the
right chimney of the cut understudy.
Let us see Fig. 2.
Geostatistics for Variable Geometry Veins 261
Fig. 1 A geological cross-section of a vein system with its frequent complexity, (Bateman 1982)
Sample channels and/or drill samples with its grades and real vein width,
resulting from mining activity, are identified and matched with points x, y, z,
corresponding to their gravity centre.
• Capping values are treated.
• “Decluster” to solve or prevent potential.
• Grouping/clustering of samples is made.
262 A. Marı́n Suárez
P*CDEF ¼ λ1 P1 þ λ2 P2 þ λ3 P3
The average accumulation and the whole chapter (Marin 1986) (grade-thickness
product) in the CDEF block is estimated by
ðLPÞ*CDEF ¼ λ1 P1 l1 þ λ2 P2 l2 þ λ3 P3 l3
0 0 0
0 0 0
where {λ1, λ2, λ3} and {λ1 , λ2 , λ3 } are the weights calculated through the kriging
geostatistical technique.
Therefore, the estimated grade of the CDEF block (Marin 1986) is given by
ðLPÞ*ABCD
L* ¼
P*ABCD
N GALERIA SUPERIOR
BLOQUE AESTIMAR
CHIMENEA
CHIMENEA
B C
A D
GALERIA INFERIOR
T
B
B 2m C
15m Bloque a
Estimar
A D
D F
P1, l1 P2, l2 P3, l3
Sublevel
Lower Gallery
264 A. Marı́n Suárez
Fig. 5 Vein 11 estimated grade of Ag ppm, blocks 5 5, Morococha and Argentum S.A. Peru,
deposits
σ2
CM ð%qualityÞ ¼ 100 1
maxðσ 2 Þ
where:
– σ 2 is the estimated variance of a particular block.
– max(σ 2)is the maximum value of the group of variances of the estimated
blocks.
thus making it possible to have relative quality planes of the estimate, expressed
percentage-wise: see Figs. 5 and 6.
3.1 Practice
Fig. 6 Vein 11 percentage quality of the estimate of Ag ppm, blocks 5 5 Morococha and
Argentum S.A. Peru
and their workings – with even better results when these are very thin veins as,
for example, veins that average 5 cm. or so.
2. It must be pointed out that the proposed trihedral is different from Frenet and
Serret standard trihedral of manifold geometry.
3. The use of the tangent plane with its proposed mobile trihedral precludes the
building of the geological model as conceived in its classical form by three-
dimensional software available in the market. In effect, building geological
models in veins, as conceived by three-dimensional software, is difficult and
inaccurate, since appropriate information for its construction is not readily
available. In this case, geological modelling is limited to defining, in the tangent
plane, the outline of a domain, or domains corresponding to the mineralized
zone, i.e. simple lines drawn with a standard “mouse” on the tangent plane, in
other words, in a simple, immediate and interactive way.
4. It must be noticed that, here, the condition is that the tangent plane contains, in
the best way possible, the surface part of the vein, calculated by vectors, by the
minimum square method, or with the help of block models in 3-D software
available in the market.
5. This way we have the design of the methodology that will help perform
paragenesis studies and estimate or simulate short- and midterm resources so
that the mining engineer may proceed to calculate the exploitable reserves, in the
same manner as the author of this publication, acting as a consultant, applied to
mining companies. Examples of the foregoing are polymetallic veins of Cu, Ag,
Zn and Pb at Morococha of the Cia Argentum S.A. and polymetallic veins of
Raura of Cia Minsur S.A. deposit, both located in the Peruvian Andes.
266 A. Marı́n Suárez
3.2 Theory
With this work, we hope to open a field of research in theoretical and practical
applications in mining, GIS and other regionalised variable domains, reframing the
algorithms of current geostatistics with Riemannian geometry metrics and building
Riemannian geostatistics.
Differentiable Manifolds
xα : U α ℝn ! M
The set of all derivations of C1 ðMÞ at p, denoted by TpM is called a tangent space to
M at p. An element of TpM is called a tangent vector at p.
A Riemannian metric on a differentiable manifold M is a correspondence which
associates to each point p of M, an inner product h, ip on the tangent space TpM,
which varies differentiably; this metric is also called the metric tensor.
A differentiable manifold with a given Riemannian metric will be called a
Riemannian manifold (Lee JM 2013; Do Carmo 1976).
Regular Surface
Proposed Trihedron
Now let us build a trihedron, from a point on the surface and a tangent vector to the
surface at that point; this construction is a different wing of the trihedron Frenet-
Serret; in very specific cases of surface geometry, these may coincide.
Let S ℝ3 a regular surface, parametrization (x,U ) at point p 2 S and a regular
0
parametrized curve α : ε, ε ! ℝ3 content in S such that α ð0Þ ¼ v and αð0Þ ¼ p,
clearly v 2 T p S.
On the TpS tangent plane, rotate the vector v angle π/2 in a sense counterclock-
wise forming a basis for the tangent plane TpS (see figure). We now vector cross
product w ¼ v u and form the proposed trihedron (see Fig. 3).
Bibliography
Guibal D, Remacre A (1984) Local estimation of the recoverable reserves. E.N.S.M.P, France
Lee JM (2013) Introduction to smooth manifolds, 2nd edn. University of Washington, Seattle
Maréchal A, Deraisme J, Journel A, Matheron G (1978) Cours de Géostatistique non Linéaire.
C-74 E.N.S.M.P. France
Marı́n Suárez A (1978) Méthodologie de L’estimation et Simulation Multivariable des Grands
Gisements Tridimensionnels. Thèse de Docteur Ingénieur en Sciences et Techniques Minières
– Option Géostatistique, présentée a L’école Nationale Supérieure des Mines de Paris
Marı́n Suárez A (1986) Modelo Geoestadı́stico de Filones de Almadén. Minas de Almadén S.A,
Almadén
Marı́n Suárez A (2009) Dise~no de una Metodologı́a Geoestadı́stica con Geometrı́a de Variedades
en Yacimientos Filoneanos. Facultad de Geologı́a, Geofı́sica y Minas, UNSA, Arequipa
Matheron G (1962, 1963) Traité de Géostatistique Appliquée, vol 1; vol 2. Technip, Paris
Warner FW (1983) Foundations of differentiable manifolds and lie groups. Springer, New York
Drilling Grid Analysis for Defining
Open-Pit and Underground Mineral
Resource Classification through
Production Data
Abstract The varied types of mineral deposits and geological features around the
world have led to the creation of a large number of techniques, methodologies, and
definitions for mineral resource classification. The most common methods used in
the mineral industry include kriging variance, drilling spacing, neighborhood
restriction, and conditional simulations. These methods generally do not use rec-
onciled production information, only long-term borehole information based on
personal judgment for defining confident intervals/limits on the mineral resource
classification. A drilling spacing back analysis study for defining mineral resource
classification was completed considering tonnages and grades confidence intervals
related to its respective production volumes, based on short-term production rec-
onciliation of analog deposits. The definition of adequate drill holes spacing and
detailed results for classifying mineral resources are demonstrated by both an open-
pit and an underground project adjacent to an existing mining operation. This study
has considered a Brazilian sulfide deposit (Cu-Au) operating mine as analog
information.
1 Introduction
New methodologies for classifying mineral resources are proposed every year with
constant efforts for reducing ore deposits estimation uncertainties using new tech-
nologies. However, given different types of mineralization and its geological
characteristics, it becomes impractical to define a single standardized methodology
for mineral resource classification. In general the estimation classification process
is based on qualified/competent person’s (QP/CP) judgment. Whether or not the
confidence of estimates (drilling spacing, experience, back analysis, or
geostatistical methods) are based on this judgment, their validity is subject to
assumptions that should be tested and validated against actual production results
through reconciliation studies (Morley 2003).
Several authors have reported the major issues associated with mining operating
cases that have not achieved planned production targets in its first years of operation
(Burmeister 1988). The main reasons are inherent to the mineral resources and
reserve classification processes, including errors associated with an inadequate
sampling procedure and lack of local geological knowledge (Harquail 1991).
Some studies present approaches for drill hole spacing using conditional simulation
(Dimitrakopoulos et al. 2009) or estimation variance calculation (Verly et al. 2014)
from long-term (hard data) drill holes information. Also a number of authors have
published a series of recommendations for quantitative estimates of accuracy to
classify resources and reserves (Vallée 1992; Wober 1993; Stephenson 1998). A
summary used for resource classification, by compiling public disclosure of mineral
projects issued by companies listed on the Toronto Stock Exchange, presents the
most commonly applied resource classification practices (Silva 2013). In regard to
these, none of the proposed mineral resource classification methods presents a
connection to actual operating aspects or makes use of an operational reconciliation
process. Therefore, mining configuration, geological knowledge, and production
aspects can help in the decision for better defining mineral resources categories.
In this paper an operating mining case study (sulfide Cu-Au deposit) has been
developed for defining mineral resources classification by using actual production
reconciliation results. The idea is to determine drilling grid spacing that best
support mineral resource categories for both open-pit and underground operating
processes. Different production volumes related to its respective geology and
operating practices were compared to actual exhaustive values for determining
confidence levels of mineral resources categories.
Production information of a copper mine operation, based on blasting reverse
circulation drilling, was used for both an underground and an open-pit projects due
to its similarity in geological aspects and production scale.
Drilling Grid Analysis for Defining Open-Pit and Underground Mineral. . . 273
A Brazilian large copper mine was used as analog information for both an under-
ground and open-pit sulfide Cu-Au project. This analog deposit will be called as
CMC (copper mine complex) and will be divided into two data sets, CM1 and CM2.
CM1 is the one defined for the underground project (UGP). The open-pit project,
referred to as OPP in the paper, took into account the CM2 data set.
The complete database of the CMC contains over 100,000 blast holes 1214 inch
analysis ranging from approximately 5 to 6 m average drilling spacing. A series of
different drill hole spacing grids were created for each mining method using
production information. A new geological interpretation wireframe was constructed
for each new grid by vertical extrusion of interpreted horizontal sections. Ordinary
kriging routines were used to estimate block models of each drilling grid for both
case studies.
The nearest neighbor interpolation algorithm was used for the construction of
each regular drilling grid. For the selection of the drilling grid spacing (e.g.,
30 30 m), a new block model was created with same expected drilling grid
dimensions. After running the nearest neighbor interpolation, only the closest
sample to the center of each block was selected and exported to create a new
regular drilling grid with approximate 30 m spacing based on original
production data.
Figure 1 illustrates how drilling grid selection was created and a sample selec-
tion was made for a regular sampling grid. The original geographic coordinates
(X, Y, and Z) of each sample was kept.
According to the geological interpretation for long-term operational assump-
tions, drilling grid supporting each production increment (global, annual, and
quarterly) was created according to long-term geological interpretation practices
and interpreted on both vertical and plan sections. Vertical section traverses were
created for supporting geological interpretation of plan sections without consider-
ing actual exhaustive information, only the information defined on the regular grids
by the sample selection procedure.
Fifty-three vertical sections striking N-S were interpreted with distance between
sections ranging from 10 to 40 m. A total of 21 plan sections, spaced at 16 m (bench
height) and between 128 and 192 m sea level, were interpreted in CM1. In the
same way, 21 vertical sections striking NE-SW were interpreted with distance
between sections ranging from 10 to 20 m. After modeling the vertical planes and
applying traverses for plan view interpretation, nine plan sections were interpreted
for CM2 considering 16 m bench height and elevation varying between 64 m and
192 msl (sea level). Figure 2 presents interpreted plan sections (1a and 2b) and
geological wireframes (1b and 2a) created for both CM1 and CM2 data sets.
The CMC operation long-term model (10 10 16 m block size) was used in
studied cases. Figure 3 presents the 20 20 m (grid spacing) interpolation results at
CM1 taking into account the geological interpreted mineralization. When different
statistical domains or geological interpretations are considered, the final copper
274 R. Menin et al.
Fig. 2 Figures on top: (1) CM1 20 20 m interpreted plan section (left); CM1 10 10 m
geological wireframes (right); Figures on bottom: (2) CM2 30 30 m geological wireframes
(left); CM2 20 20 m interpreted plan section (right)
grades are defined by weighing the percent of each statistical domain within the
block.
A surface volumetric calculation for determining differences (tonnage, grades,
and metal) between drilling models in comparison with the production model was
Drilling Grid Analysis for Defining Open-Pit and Underground Mineral. . . 275
The level of confidence for the definition of inferred resources was applied in one
volume increment (global) for both CM2 and CM1. The tonnage produced in the
selected comparison period was obtained through volumetric calculation between
first and last year topographic surfaces. Figure 5 presents the percentage difference
Drilling Grid Analysis for Defining Open-Pit and Underground Mineral. . . 277
15.00% Tonnes
ContMetal
relative difference
0.00% Cu%
-15.00%
-30.00%
-45.00%
-60.00%
10x10 20x20 30x30 40x40 60x60 80x80
0.00%
-15.00%
-30.00%
10x10 20x20 30x30 30x40 40x40 40x80 50x50
Fig. 5 (1) Global difference in terms of metal content (light blue), grade (blue), and tonnages
(red) for CM1 (top); (2) global difference in terms of metal content (light blue), grade (blue), and
tonnages (red) for CM2 (bottom)
between metal content, tonnages, and grades for all drilling grids in comparison
with the CM1 and CM2 production actual models.
The resulting tonnages, grades, and contained metal obtained in the estimated
models, for each drilling grid, had acceptable values for inferred resource classifi-
cation in comparison with the production exhaustive models. Tables 2 and 3 present
the percentage of scenarios within acceptable stipulated limits previously men-
tioned for both CM1 and CM2:
In the CM1 case, only the 80 80 m drilling spacing is not within acceptable
limits in terms of tonnages and metal content to the established production rate. In
the CM2 case, all drilling grids are within an acceptable range (percentage differ-
ences) for defining inferred mineral resources.
278 R. Menin et al.
Table 2 Global percentage Spacing (m) Metal (%) Tonnes (%) Grade (%)
of scenarios within acceptable
10 10 100 100 100
difference for CM1
20 20 100 100 100
30 30 100 100 100
40 40 100 100 100
60 60 100 100 100
80 80 0 0 0
Table 3 Global percentage Spacing (m) Metal (%) Tonnes (%) Grade (%)
of scenarios within acceptable
10 10 100 100 100
difference for CM2
20 20 100 100 100
30 30 100 100 100
30 40 100 100 100
40 40 100 100 100
40 80 100 100 100
50 50 100 100 100
CM1 and CM2 global production were divided into five and four annual incre-
ments, respectively. This is in agreement with the annual production rates for both
open-pit (OPP) and underground (UGP) examples. The tonnage produced in the
selected comparison period was obtained through volumetric calculation between
the first and the last monthly topographic surfaces in the same year (January 1st–
December 31st).
The executed production rate on an annual basis ranged from 4.0–5.5 to
3.5–5.0 Mt for CM1 and CM2 open-pits, respectively. Figures 6, 7, and 8 present
a comparison of the percentage differences (grades, metal, and tonnages, respec-
tively) for each selected period for both CM1 (top) and CM2 (bottom), considering
the actual production information.
As commented, a level of confidence of 15 % on the annual recoverable
contained metal, tonnage, and grades is needed to support an indicated mineral
resource category. Tables 4 and 5 present the percentage of scenarios within the
acceptable stipulated limits:
The high-grade portions could not be mapped for drilling grid spacing higher
than 80 80 m in the CM1 case study. This reduced the tonnage of this material
significantly where only the grades were within acceptable threshold limits for
indicated mineral resources. Small differences in tonnages and grades were verified
in the 60 60 m drill spacing. The metal content was above the stipulated threshold
limits in two of the five annual increments. However, for these two scenarios, the
grade and tonnages were within acceptable limits.
Drilling Grid Analysis for Defining Open-Pit and Underground Mineral. . . 279
30.00%
relative difference
15.00%
0.00%
-15.00%
-30.00%
-45.00%
10x10 20x20 30x30 40x40 60x60 80x80
30.00%
Annual Difference - Grade
relative difference
15.00%
0.00%
-15.00%
-30.00%
10x10 20x20 30x30 30x40 40x40 40x80 50x50
Fig. 6 (1) Annual difference in terms of grade for CM1 (top); (2) annual difference in terms of
grade for CM2 (bottom)
45.00%
Annual Difference - Contained Metal
30.00%
relative difference
15.00%
0.00%
-15.00%
-30.00%
-45.00%
10x10 20x20 30x30 40x40 60x60 80x80
15.00%
0.00%
-15.00%
-30.00%
10x10 20x20 30x30 30x40 40x40 40x80 50x50
Fig. 7 (1) Annual difference in terms of metal content for CM1 (top); (2) annual difference in
terms of metal content for CM2 (bottom)
280 R. Menin et al.
45.00%
Annual Difference - Tonnage
30.00%
relative difference
15.00%
0.00%
-15.00%
-30.00%
15.00%
0.00%
-15.00%
-30.00%
10x10 20x20 30x30 30x40 40x40 40x80 50x50
Fig. 8 (1) Annual difference in terms of tonnages for CM1 (top); (2) annual difference in terms of
tonnages for CM2 (bottom)
Table 4 Annual percentage Spacing (m) Metal (%) Tonnes (%) Grade (%)
of scenarios within acceptable
10 10 100 100 100
difference for CM1
20 20 100 100 100
30 30 80 100 100
40 40 40 80 80
60 60 40 80 60
80 80 0 0 60
Table 5 Annual percentage Spacing (m) Metal (%) Tonnes (%) Grade (%)
of scenarios within acceptable
10 10 100 100 100
difference for CM2
20 20 100 100 100
30 30 100 100 100
30 40 100 50 100
40 40 75 75 75
40 80 25 75 50
50 50 50 100 50
In the CM2 case study, for the 40 40 m drilling spacing, slight differences in
the grade and tonnages than those stipulated were observed. However they reflected
a higher difference in the metal content compared to the values obtained by the
exhaustive production model. For two of the four annual production volumes, the
Drilling Grid Analysis for Defining Open-Pit and Underground Mineral. . . 281
30 40 m drilling grid scenario obtained tonnages which were slightly above the
stipulated limits; however, for these two scenarios, the grade and the contained
metal were within acceptable values.
CM1 and CM2 global production were divided into quarterly increments (20 and
15, respectively) which are in agreement with the quarterly production rate exe-
cuted. The tonnage produced in the selected comparison period was obtained
through volumetric calculation between the first and the last monthly topographic
surface of the quarter.
A level of confidence of 15 % on the quarterly recoverable tonnages, grades,
and metal content is required to support measured mineral resource categorization.
The quarterly production rates of CM1 case study ranged from 0.8 to 1.4 Mt. For
CM2 the ranges were from 0.75 to 1.5 Mt. The production period was divided into
actual monthly basis topography (short-term CMC) for the CM1 comparison.
Figures 9, 10, and 11 present a comparison of the percentage differences (grades,
metal, and tonnages, respectively) for each selected period for both CM1 (top) and
CM2 (bottom), considering the actual production information.
CM1 tonnages and metal content estimation results of 60 60 m and 80 80 m
drilling grid spacing are substantially different for the selected cutoff grade com-
pared to the exhaustive production model. It is more difficult to establish spatial
connectivity for the available samples when increasing drilling grid spacing, which
significantly reduces the volumes of certain portions of the deposit given the lack of
information for delineating mineralization. In the 30 30 m drilling grid spacing
scenarios, only a few quarterly production intervals lied outside acceptable metal
content limits in support of measured mineral resources. In the 40 40 m drilling
grid spacing scenario, there were several quarterly production intervals outside
acceptable tonnage and metal content threshold limits.
Minor differences in the tonnages and grades were observed in the CM2 case
study for a 30 30 m drilling grid spacing. These were verified mainly in scenarios
with lower amount of production information. The results demonstrated that dril-
ling grid spacing larger than 40 40 m does not predict within acceptable confi-
dence the production tonnages, grades, and metal content in a quarterly period.
Tables 6 and 7 summarize the percentage of scenarios within acceptable limits
defined for both cases studies.
282 R. Menin et al.
30.00%
relative difference
15.00%
0.00%
-15.00%
-30.00%
-45.00%
10x10 20x20 30x30 40x40 60x60 80x80
15.00%
0.00%
-15.00%
-30.00%
10x10 20x20 30x30 30x40 40x40 40x80 50x50
Fig. 9 (1) Quarterly difference in terms of grade for CM1 (top); (2) quarterly difference in terms
of grade for CM2 (bottom)
15.00%
0.00%
-15.00%
-30.00%
10x10 20x20 30x30 30x40 40x40 40x80 50x50
30.00%
relative difference
15.00%
0.00%
-15.00%
-30.00%
-45.00%
10x10 20x20 30x30 40x40 60x60 80x80
Fig. 10 (1) Quarterly difference in terms of metal content for CM1 (top); (2) quarterly difference
in terms of metal content for CM2 (bottom)
Drilling Grid Analysis for Defining Open-Pit and Underground Mineral. . . 283
15.00%
0.00%
-15.00%
-30.00%
10x10 20x20 30x30 30x40 40x40 40x80 50x50
45.00%
Quarterly difference on Tonnes
30.00%
relative difference
15.00%
0.00%
-15.00%
-30.00%
-45.00%
10x10 20x20 30x30 40x40 60x60 80x80
Fig. 11 (1) Quarterly difference in terms of tonnages for CM1 (top); (2) quarterly difference in
terms of tonnages for CM2 (bottom)
Table 6 Quarterly Spacing (m) Metal (%) Tonnes (%) Grade (%)
percentage of scenarios
10 10 100 100 100
within acceptable difference
for CM1 20 20 95 100 100
30 30 70 85 80
40 40 35 65 50
60 60 45 70 70
80 80 15 10 75
Table 7 Quarterly Spacing (m) Metal (%) Tonnes (%) Grade (%)
percentage of scenarios
10 10 93 100 93
within acceptable difference
for CM2 20 20 79 100 86
30 30 71 86 79
30 40 50 43 71
40 40 57 71 86
40 80 36 79 64
50 50 50 79 57
284 R. Menin et al.
6 Conclusions
CM1 (UGP) study indicates that drill holes spacing wider than 60 60 m is
insufficient to adequately represent the mineralization and to reflect production
information due to lack of information at a given cutoff grade of 0.75 % Cu. The
historic reconciliation between the executed CMC production model and the CM2
(OPP) long-term model is not within acceptable threshold limits in areas with
drilling grid spacing wider than 40 40 m (measured + indicated).
Table 8 summarizes the drill hole spacing required to support mineral resource
categories for both CM1 and CM2 case studies using CMC production information.
The random selection of information can include waste information from neigh-
borhoods that are mostly mineralized. This randomness in the choice of information
can sometimes change the proportions of economic mineralization and waste in
some local portions of the model. This is part of the “real” variability of the deposit.
The mineralization characteristics of a deposit do not define the sample spacing
required for mineral resource categorization. A realistic production volume should
be considered in the evaluation of drill spacing for supporting mineral resource
categories. Realistic volumes are determined from analog deposits where produc-
tion information is available.
As mentioned in Diedrich et al. (2016), mining selectivity is understood as the
process of separating ore from waste, and its global and local concept is strongly
related to three mining functions such as geology, production rates, and mining
configuration that affect operational results. Operational aspects should be included
in a project or operation for better defining adequate drilling spacing and for
supporting mineral resources classification.
Bibliography
Burmeister B (1988) From resource to reality: a critical review of the achievements of new
Australian gold mining projects during 1983–1987. Macquarie University, Sidney
Deraisme J, Roth C (2000) The information effect and estimating recoverable reserves.
Geovariances
Diedrich C, Reuwsaat JD, Menin R, Paula WF (2016) Selection of operational mining unit (OMU)
size, Proceedings GEOSTATS 2016. Universitat Politècnica de València, Valencia, p 16
Dimitrakopoulos R, Godoy M, Chou C (2009) Resource/reserve classification with integrated
geometric and local grade variability measures. In: Proceedings orebody modelling and
Drilling Grid Analysis for Defining Open-Pit and Underground Mineral. . . 285
strategic mine planning. The Australasian Institute of Mining and Metallurgy, Melbourne, pp
207–214
Harquail D (1991) Investing in junior mining companies. 6th Mineral Economics Symposium of
CIM. Canadian Institute of Mining, Mettalurgy and Petroleum, Montreal, p 10
Isaaks E (2004, 2005) The kriging oxymoron: a conditionally unbiased and accurate predictor. In:
Proceedings of geostatistics banff, 2nd edn. Springer, Banff, pp 363–374.
Jara RM, Couble A, Emery X, Varela EM, Ortiz JM (2006) Block size selection and its impact on
open-pit design and mine planning. J S Afr Inst Min Metall 106:205–211
Leuangthong O, Neufeld C, Deutsch CV (2003) Optimal selection of selective mining unit (SMU)
size. University of Alberta, Alberta
Morley C (2003). Beyond reconciliation: a proactive approach to using mining data. In: Large
open pit mining conference. Kalgoorlie, pp 185–192
Silva DA (2013) Mineral resource classification (NI 43–101): an overview and a new evaluation
technque. In: Centre for computational geostatistics. University of Alberta, Edmonton, p 307
Sinclair A, Blackwell G (2002) Applied mineral inventory estimation. Cambridge University
Press, Cambridge
Stephenson P (1998) The JORC code, 1987–1997, in Geology of Australian and Papua New
Guinean Mineral Deposits. In: Australasian institute of mining and metallurgy. Melbourne, pp
44–52
Vallée MA (1992) The guide to the evaluation of gold deposits: integrating deposit evaluation and
reserve inventory practices. CIM Bull 85:50–61
Verly G, Postolski T, Parker H (2014) Assessing uncertainty with drill hole spacing studies –
applications to mineral resources. In: Proceedings orebody modelling and strategic mine
planning symposium. The Australasian Institute of Mining and Metallurgy, Perth, p 10
Wober HA (1993) Classification of ore reserves based on geostatistical and economic parameters.
CIM Bull 86:73–76
A High-Order, Data-Driven Framework
for Joint Simulation of Categorical Variables
1 Introduction
Oliver 2007; Remy et al. 2009; Chiles and Delfiner 2012; Pyrcz and Deutsch 2014).
The choice of the Gaussian distribution is driven by several factors. First of all, the
Gaussian variables can be fully described by a small amount of parameters, such as
the first-order statistics (i.e., average values) and the second-order statistics (i.e.,
covariance or variogram). Secondly, the small number of parameters allows one to
run simulations on grids with several million nodes.
Natural phenomena are known to exhibit non-Gaussian distributions and have
complex nonlinear spatial patterns (Guardiano and Srivastava 1993; Tjelmeland
and Besag 1998; Dimitrakopoulos et al. 2010), which cannot be adequately
described by second-order statistics. To overcome these limitations, multiple
point spatial simulation (MPS) methods were introduced in the 1990s (Guardiano
and Srivastava 1993; Journel 1993; Strebelle 2002; Journel 2003; Zhang et al. 2006;
Chuginova and Hu 2008; Straubhaar et al. 2011; Toftaker and Tjelmeland 2013;
Strebelle and Cavelius 2014; others). The additional information is taken into
account via training images (TI), which are not conditioned on the available data
but contain additional information about complex spatial relations of the attributes
to be simulated. To retrieve this information from the training image, the similarity
between the local neighborhood of an unsampled location and the training image is
calculated in explicit or implicit form. Based on this similarity measure, the value
of a node from the training image with the most similar neighborhood is assigned to
the unsampled location being simulated. Generally, most of the multipoint simula-
tion techniques are a Monte Carlo sampling of values from the TI in some form or
another. No spatial models are used and, importantly, no spatial information from
the hard data is retrieved. As a result, simulations of attributes reflect the TI. In
cases where there are relatively large datasets, conflict between the hard data and
TI’s statistics is clearly observed and the resulting simulations do not reproduce the
spatial statistics of the hard data (Dimitrakopoulos et al. 2010; Pyrcz and Deutsch
2014).
Several attempts have been made to incorporate more information from the hard
data. Some authors suggest using replicates from the hard data in addition to TI
(Mariethoz and Renard 2010); however, in practice, it is hard to find any replicates
for three-point relations when data is sparse. Others (Mariethoz and Kelly 2011)
apply affine transformations to better condition the hard data; however, TI is still
used as the main source of information. Another approach is to construct TI based
on the hard data (Yahya 2011), but the resulting simulations may be biased from the
method chosen for the TI construction.
Mustapha and Dimitrakopoulos (2010a, b) proposed to use the high-order spatial
cumulants as the extension of variogram models to capture complex multipoint
relations during the simulation of non-Gaussian random fields. The technique
estimates the third- and the fourth-order spatial statistics from hard data and
complements them with higher-order statistics from the TI. However, this tech-
nique is based on the approximation of conditional distribution using Legendre
polynomials, which are smooth functions and limited in generating adequate
approximation of the discrete distribution of categorical variables.
A High-Order, Data-Driven Framework for Joint Simulation of Categorical. . . 289
Without loss of generality, assume that sk ¼ k. It can be shown that the probability is
equivalent to spatial indicator moment (Vargas-Guzmán 2011):
290 I. Minniakhmetov and R. Dimitrakopoulos
Consider two random variables Z i 0 and Zi1 separated by the lag h1 ¼ xi1 xi0 . Due
to the stationarity assumption, their second-order spatial indicator moment for
categories k0,k1 can be expressed as a function of the lag h1:
For the sake of demonstration, consider data from the Stanford V reservoir case
study (Mao and Journel 1999) on Fig. 1a and its categorization on Fig. 1b, in which
the size of image is NxNy pixels. Let Wi,j be a value at pixel (i,j) of the categorized
image, where i ¼ 1 Nx, j ¼ 1 Ny. If the image Wi,j describes statistical properties
of the random vector Z, then the estimation of indicator moment M ^ k0 k1 ðh1 Þ on the
lag (h1¼(h,0) can be calculated using pairs {Wi,j,Wi+h,j}. From now on, consider
that the direction of h1 is e1¼(1,0) and fixed and then Mk0 k1 ðh1 Þ is the function of
distance h.
The sections of the function Mk0 k1 ðhÞ for fixed h equal to 0, 5, and 40 pixels
which are shown on Fig. 2a–c, respectively. Figure 2d presents the sections of the
function Mk0 k1 ðhÞ for fixed values k0,k1. Each line corresponds to one of the 3 3
possible combinations of k0 and k1.
It is not hard to see that for h ¼ 0 only diagonal elements are not equal to zero:
where δk0 , k1 is Kronecker delta and Mk0 is the marginal distribution. Additionally,
for two distant locations h!1, the values Zi0 and Z i1 can be considered as
independent random variables and the indicator moment can be factorized:
A High-Order, Data-Driven Framework for Joint Simulation of Categorical. . . 291
Fig. 2 The estimation of the second-order spatial indicator moment M ^ k0 k1 ðhÞ: (a) zero-distance
h ¼ 0 indicator moments; (b) indicator moments on the lag h ¼ 5; (c) indicator moments for far
separated points h ¼ 40; (d) the sections with different combinations of k0,k1 depends on h. Each
line on (d) corresponds to value in one of 3 3 cells in (a), and its evolution across different lag
separation (b) and (c)
M k 0 k 1 ð h ! 1Þ ¼ M k 0 M k 1 : ð7Þ
That can be traced in the function behavior on Fig. 2d. These functions represent
two-point spatial cross-relations, similar to indicator covariances, and satisfy
boundary conditions (6) and (7).
292 I. Minniakhmetov and R. Dimitrakopoulos
where h\hp denotes all the lags h excluding the lag hp. Similarly for k\kp.
If the directions are quite different, then additional boundary conditions are
valid:
Mk h1 , . . . , hp ! 1, .. . , hn
ð10Þ
¼ Mk\ kp h\ hp Mkp , 8p 2 1 . . . n:
Thus, the high-order spatial indicator moments are bounded with lower-order
moments, and this information should be taken into account during simulation.
For example, in case of three-point relations, for the image Wi,j, the sampling
third-order spatial indicator moment M ^ k0 k1 k2 ðh1 ; h2 Þ of random variables separated
by the vectors h1, h2 with directions e1¼(1,0) and e1¼(1,0) can be calculated using
triplets W i, j ; W iþh1 , j ; W i, jþh2 . The indicator moment M ^ 111 ðh1 ; h2 Þ is shown on
^
Fig. 3. The values of the function M 111 ðh1 ; h2 Þ on boundaries (h1,0), (h1,50), (0,h2),
and (50, h2) correspond to two-point statistics shown on Fig. 2d.
3 Mathematical Model
In this paper, multidimensional B-spline approximation under constrains (8) and (9)
is used to model the high-order spatial indicator moments Mk(h). Consider the fixed
categories k0, kn; then, the function Mk(h) is a multidimensional function which
values are known at the limited number of points hd ¼ h1d ; . . . ; hnd , d ¼ 1 . . . m
estimated from the hard data. The calculation of sampling indicator moments M ^k
d
h is presented in the subsequent section. Then, the function Mk(h) can be
approximated using the following recursive model:
A High-Order, Data-Driven Framework for Joint Simulation of Categorical. . . 293
X
ω X
ω
δMk ðhÞ ¼ αi1 , ..., in Bi1 , r ðh1 Þ . . . Bin , r ðhn Þ; ð13Þ
i1 ¼1 in ¼1
The octant model is used (Fig. 4) to estimate the sampling indicator moments M ^k
d
h from the hard data. The neighborhood area of each hard data sample is divided
into Nϕ ¼ 8 sectors representing Nϕ directions. Then, each sector is divided into Nr
lags and forms an NrNϕ bin template. Only one point within each bin is randomly
chosen to construct a replicate. Finally, the values M^ k hd are estimated from
replicates using law of large numbers:
X
N hd
^ k hd ¼ 1
M I k0 zij0 . . . I kn zijn , d ¼ 1 . . . m; ð16Þ
N hd j¼1
where the sum is taken over all N hd replicates with the spatial configuration hd, data
samples zij0 . . . zijn in the replicate j are separated by lags hd, and d is the index of
different spatial configurations hd. It should be noted that replicates separated by at
least half the variogram range should be used for the law of the large numbers to be
applicable.
The amount of information about high-order statistics that can be retrieved from
data crucially depends on the number of categories K, the total number of data
samples N, and the level of correlation between values. It is not hard to see that the
higher order of statistics considered are the larger the number of samples available
should be. In order to have an adequate number of replicates for a particular order
m of statistics, the minimum number of replicates N repl ðmÞ N hd , 8hd is set up by
the user. However, more advanced techniques based on an entropy or information
theory should be considered (Arndt 2004). Having a minimum number of replicates
Nrepl(m) for the given order of statistics m, the optimal number of lags Nopt r is
calculated.
A High-Order, Data-Driven Framework for Joint Simulation of Categorical. . . 295
4 Simulation Algorithm
Combining all of the above, the new data-driven algorithm is formulated as follows:
Algorithm A.1
1. Starting from two-point statistics m ¼ 2 until stopping criteria is reached.
(a) Define the minimum number of replicates Nrepl(m) for given order m.
(b) Scan the hard data using the octant model (Fig. 4) with different Nr starting
with the number of the radial divisions Nr ¼ 2 and find the higher Nr for
which the average number of replicates is bigger than Nrepl(m).
(c) If Nr ¼ 2 and the average number of replicates is less than Nrepl, then exit
the loop.
(d) Save all the replicates for obtained Nr and the order of statistics m.
(e) Increase the order of statistics m¼m + 1.
2. Define a random path visiting all the unsampled nodes.
3. For each node xi0 in the path:
(a) Find the closest data samples xi1 , xi2 , . . . xin . The categories at these nodes
are denoted by k1, kn.
(b) For all k0 ¼ 1 K, calculate the high-order spatial indicator moments M ^k
d
h using formula (15) from the replicates found in step 1 and recursive
model (11–13). Note that k1, kn are fixed. For the orders higher than
maximum order m, consider δMk ðhÞ0.
(c) Calculate the conditional distribution from joint distribution:
P Z i0 ¼ k0 Zi1 ¼ k1 . . . , Z in ¼ kn ¼ AMk ðhÞ; ð17Þ
296 I. Minniakhmetov and R. Dimitrakopoulos
X
K
A ¼ 1= Mk ðhÞ ð18Þ
k0 ¼1
(d) Draw a random value zi0 from this conditional distribution (17) and assign it
to the unsampled location xi0 .
(e) Add zi0 to the set of sample hard data and the previously simulated values.
4. Repeat Steps 3a–e for all the points along the random path defined in Step 2.
5 Simulation Results
The proposed approach is tested on the data set from the Stanford V reservoir case
study (Fig. 1a). This image is discretized on categories 0, 1, and 2 and is used as a
reference image (Fig. 5a). Hard data is randomly selected from the image and
shown in Fig. 5b. This represents 520 points (5 % of the image points). The results
are compared with sequential indicator simulation algorithm (sisim; Journel and
Alabert 1990; Deutsch and Journel 1998).
The simulation results for the case with 520 data samples are shown in Fig. 6.
Neither the training image nor the reference image (Fig. 5a) is used during the
simulation and is presented herein only for the sake of comparison. Simulations are
done in two modes: using only boundary conditions (Fig. 6b) and using both
conditions (14) and (15) (Fig. 6c). In the case of using just boundary conditions,
the result is smooth and the width of channels is overestimated because all high-
order statistics are derived from the second-order statistics. However, sisim simu-
lation results (Fig. 6a) are less connected and the channels can be hardly detected.
The result obtained with the account of higher-order statistics from data (Fig. 6c)
reproduces the channels quite well with adequate dimensions of geometrical
bodies.
On Fig. 6c, the histograms for data samples, the reference image, sisim simula-
tion, and the simulation using the proposed technique are shown by blue, light blue,
yellow, and red bars, respectively. The deviations from the distribution in hard data
are small for both sisim and proposed algorithm simulations.
The second-order statistics are compared in Fig. 7a. The direction e1¼(1,0) is
used. The indicator moments M01(h1) of simulations using sisim and the proposed
algorithm are reproduced well. Nevertheless, the third-order statistics M012(h1,h2)
of the reference image (Fig. 7b), simulation using sisim algorithm (Fig. 7c), and the
result of the proposed technique (Fig. 7d) are quite different. The third-order spatial
indicator moments are calculated using directions e1¼(1,0) and e2¼(1,0). Some
similarities of patterns can be traced in the bottom part of Fig. 7b, d, which
correspond to statistics of the reference image and the simulation using the
A High-Order, Data-Driven Framework for Joint Simulation of Categorical. . . 297
Fig. 5 Case study with 520 data samples: (a) the reference image, (b) data samples
Fig. 6 Case study with 520 data samples: (a) sisim simulation result, (b) the simulation using
proposed algorithm with only boundary conditions, (c) the simulation result using both boundary
conditions and high-order statistics from data. Subfigure (d) shows histograms for data samples,
the reference image, sisim simulation, and the simulation using the proposed technique presented
by blue, light blue, yellow, and red bars, respectively
proposed algorithm. However, the point of interest is the reproduction of the high-
order spatial statistics of the hard data.
The surface on Fig. 8 is a 3D view of Fig. 7d. Dots represent the statistics
calculated from the hard data. Colors show the number of triplets used for the
calculation of the third-order spatial indicator moment M012(h1,h2). The higher the
number of triplets, the more reliable the value of the point is. The spatial indicator
298 I. Minniakhmetov and R. Dimitrakopoulos
Fig. 7 Case study with 520 data samples. (a) Second-order spatial indicator moment M01(h1)
using directions e1¼(1,0) for data samples, the reference image, sisim simulation, and the
simulation using proposed technique are presented by black dots and green, blue, red lines,
respectively. The third-order spatial indicator moment M012(h1,h2) using directions e1¼(1,0) and
e1¼(1,0) for (b) the reference image, (c) sisim simulation result, (d) the simulation using both
boundary conditions and high-order statistics from data
moment M012(h1,h2) of the simulation using the proposed algorithm tends to fit
more reliable points and is consistent with the boundary conditions (Fig. 8).
6 Conclusions
This paper presented a new data-driven, high-order sequential method for the
simulation of categorical random fields. The sequential algorithm is based on the
B-spline approximation of high-order spatial indicator moments that are consistent
with each other. The main distinction from commonly used MPS methods is that, in
the proposed technique, conditional distributions are constructed using high-order
spatial indicator moments as the functions of distances based on hard data. The
simulations herein are generated without a TI. Note that in applications with
A High-Order, Data-Driven Framework for Joint Simulation of Categorical. . . 299
Acknowledgments Funding was provided by the Natural Sciences and Engineering Research
Council of Canada (NSERC) Discovery Grant 239019 and mining industry partners of the
COSMO Lab (AngloGold Ashanti, Barrick Gold, BHP Billiton, De Beers Canada, Kinross
Gold, Newmont Mining, and Vale).
300 I. Minniakhmetov and R. Dimitrakopoulos
Bibliography
Arndt C (2004) Information measures: information and its description in science and engineering.
Springer, Amsterdam
Babenko K (1986) Fundamentals of numerical analysis. Nauka, Moscow
Caers J (2005) Petroleum geostatistics. SPE–Pennwell Books, Houston
Chilès JP, Delfiner P (2012) Geostatistics: modeling spatial uncertainty, 2nd edn. Wiley,
New York
Chugunova TL, Hu LY (2008) Multiple-point simulations constrained by continuous auxiliary
data. Math Geosci 40:133–146
Cressie NA (1993) Statistics for spatial data. Wiley, New York
David M (1977) Geostatistical ore reserve estimation. Elsevier, Amsterdam
David M (1988) Handbook of applied advanced geostatistical ore reserve estimation. Elsevier,
Amsterdam
Deutsch C, Journel A (1998) GSLIB: geostatistical software library and user’s guide, 2nd edn.
Oxford University Press, New York
Dimitrakopoulos R, Mustapha H, Gloaguen E (2010) High-order statistics of spatial random fields:
exploring spatial cumulants for modeling complex non-Gaussian and non-linear phenomena.
Math Geosci 42:65–99
Evans J, Bazilevs Y, Babuška I, Hughes T (2009) N-widths, sup–infs, and optimality ratios for the
k-version of the isogeometric finite element method. Comput Methods Appl Mech Engrg
198:1726–1741
Goovaerts P (1998) Geostatistics for natural resources evaluation. Cambridge University Press,
Cambridge
Guardiano J, Srivastava RM (1993) Multivariate geostatistics: beyond bivariate moments.
Geosatistics Tróia’92:133–144
Journel AG (1993) Geostatistics: roadblocks and challenges. Stanford Center for Reservoir
Forecasting
Journel AG (2003) Multiple-point Geostatistics: a State of the Art. Stanford Center for Reservoir
Forecasting
Journel AG, Alabert F (1990) New method for reservoir mapping. Pet Technol 42(2):212–218
Journel AG, Huijbregts CJ (1978) Mining geostatistics. Academic, San Diego
Kitanidis PK (1997) Introduction to geostatistics—applications in hydrogeology. Cambridge Univ
Press, Cambridge
Mao S, Journel A (1999) Generation of a reference petrophysical/seismic data set: the Stanford V
reservoir. 12th Annual Report. Stanford Center for Reservoir Forecasting, Stanford
Mariethoz G, Kelly B (2011) Modeling complex geological structures withelementary training
images and transform-invariant distances. Water Resour Res 47:1–2
Mariethoz G, Renard P (2010) Reconstruction of incomplete data sets or images using direct
sampling. Math Geosci 42:245–268
Matheron G (1971) The theory of regionalized variables and its applications. Cahier du Centre de
Morphologie Mathematique, No 5
Mustapha H, Dimitrakopoulos R (2010a) A new approach for geological pattern recognition using
high-order spatial cumulants. Comput Geosci 36(3):313–334
Mustapha H, Dimitrakopoulos R (2010b) High-order stochastic simulations for complex
non-Gaussian and non-linear geological patterns. Math Geosci 42:455–473
Pyrcz M, Deutsch C (2014) Geostatistical reservoir modeling, 2nd edn. Oxford University Press,
New York
Remy N, Boucher A, Wu J (2009) Applied geostatistics with SGeMS: a user’s guide. Cambridge
University Press, Cambridge
Straubhaar JRP, Renard P, Mariethoz G, Froidevaux R, Besson O (2011) An improved parallel
multiple-point algorithm using a list approach. Math Geosci 43:305–328
A High-Order, Data-Driven Framework for Joint Simulation of Categorical. . . 301
Strebelle S (2002) Conditional simulation of complex geological structures using multiple point
stastics. Math Geosci 34:1–22
Strebelle S, Cavelius C (2014) Solving speed and memory issues in multiple-point statistics
simulation program SNESIM. Math Geosci 46:171–186
Tikhonov AN, Arsenin VY (1977) Solution of Ill-posed problems. Winston & Sons, Washington,
DC
Tjelmeland H, Besag J (1998) Markov random fields with higherorder interactions. Scand J Stat
25(3):415–433
Toftaker H, Tjelmeland H (2013) Construction of binary multi-grid Markov random field prior
models from training images. Math Geosci 45:383–409
Vargas-Guzmán J (2011) The Kappa model of probability and higher-order rock sequences.
Comput Geosci 15:661–671
Vargas-Guzmán J, Qassab H (2006) Spatial conditional simulation of facies objects for modeling
complex clastic reservoirs. J Petrol Sci Eng 54:1–9
Wang W, Pottmann H, Liu Y (2006) Fitting B-spline curves to point clouds by curvature-based
squared distance minimization. ACM Transactions on Graphics 25(2):214–238
Webster R, Oliver MA (2007) Geostatistics for environmental scientists. Wiley, New York
Yahya WJ (2011) Image reconstruction from a limited number of samples: a matrix-completion-
based approach. Mater Thesis, McGill University, Montreal
Zhang T, Switzer P, Journel A (2006) Filter-based classification of training image patterns for
spatial simulation. Math Geosci 38(1):63–80
Conditional Bias in Kriging: Let’s Keep It
Abstract Mineral resource estimation has long been plagued with the inherent
challenge of conditional bias. Estimation requires the specification of a number of
parameters such as block model block size, minimum and maximum number of
data used to estimate a block, and search ellipsoid radii. The choice of estimation
parameters is not an objective procedure that can be followed from one deposit to
the next. Several measures have been proposed to assist in the choice of kriging
estimation parameters to lower the conditional bias. These include the slope of
regression and kriging efficiency.
The objective of this paper is to demonstrate that both slope of regression and
kriging efficiency should be viewed with caution. Lowering conditional bias may
be an improper approach to estimating metal grades, especially in deposits for
which high cutoff grades are required for mining. A review of slope of regression
and kriging efficiency as tools for optimization of estimation parameters is
presented and followed by a case study of these metrics applied to an epithermal
gold deposit. The case study compares block estimated grades with uncertainty
distributions of global tonnes and grade at specified cutoffs. The estimated grades
are designed for different block sizes, different data sets, and different estimation
parameters, i.e., those geared toward lowering the conditional bias and those
designed for higher block grade variability with high conditional biases.
M. Nowak (*)
SRK Consulting (Canada) Inc., Oceanic Plaza, 22nd Floor, 1066 West Hastings Street,
Vancouver, BC V6E 3X2, Canada
e-mail: [email protected]
O. Leuangthong
SRK Consulting (Canada) Inc., Suite 1300, 151 Yonge Street, Toronto, ON M5C 2W7, Canada
e-mail: [email protected]
1 Introduction
The discussion among both theoreticians and practitioners revolves around how
conditional bias affects the quality of the block grade estimates (McLennan and
Deutsch 2002). The first school of thought insists that the conditional bias should be
as small as possible and must be dealt with, and the second school of thought
believes that the conditional bias should be ignored and the variability of block
estimates should be as high as the variability of underlying true block grades.
Rivoirard (1987) suggested that the size of the kriging neighborhood should
consider the weight given to the mean. If the mean is given a large weight, then the
neighborhood should be expanded so as to increase the slope of regression and
thereby reduce conditional bias. Conversely, if the weight of the mean is low, then a
localized neighborhood is adequate. Krige introduced a metric called kriging
efficiency (Krige 1997) that correlates to the slope of regression. He contends
that one should never accept conditional bias in an effort to reduce the smoothing
effect of kriging (Krige 1997; Krige et al. 2005). Deutsch et al. (2014) further
expanded on potential sub-optimality of the estimates due to large conditional
biases reflected in slope of regression and in kriging efficiency measures. Deutsch
proposed a new expression of kriging efficiency to aid in the assessment of quality
of estimated block grades.
From a procedural perspective, Vann et al. (2003) introduced quantitative
kriging neighborhood analysis (QKNA) to optimize the estimation parameters for
selection of the minimum/maximum of number of samples, quadrant search, search
neighborhood, and block size. The proposed criteria for evaluating quality of block
grade estimates include slope of regression of true block grades on estimated block
grades, weight attached to the mean in simple kriging, distribution of kriging
weights, and kriging variance.
Conditional Bias in Kriging: Let’s Keep It 305
All of the above contributors have focused on metrics and efforts to minimize
conditional bias. On the other end of the spectrum of this discussion, Isaaks (2005)
argued that estimates cannot be both conditionally unbiased and globally accurate
at the same time. The estimates may be close to conditionally unbiased but the
histogram of block estimates is smoothed, which results in inaccurate predictions of
the recoverable tonnes and grade above cutoff grades. He advocates that a condi-
tionally biased estimate is necessary to obtain a globally unbiased recoverable
resource above cutoff grade. There is support that during the early stages of project
feasibility assessments, it is more important to accurately predict the global recov-
erable reserves than to produce locally accurate estimates (Journel and Kyriakidis
2004).
Despite valid points on both ends of the spectrum, it appears that the first school
of thought has been winning the discussion in recent years. The authors noticed a
substantial increase in application of those measures for optimization of estimation
parameters. Some of the proposed measures for optimization of kriging estimates,
such as slope of regression and kriging efficiency, are currently readily available in
most commercial software. In some organizations, this quantitative approach has
become standard in the resource estimation process irrespective of the stage of
exploration and/or development of the mineral deposit.
It appears that Isaaks’ sound argument for recoverable resources above an
economic cutoff grade, particularly in early stage projects, appears to have been
forgotten in the popularization of a quantitative approach because of software
accessibility. In the wake of convenience, we seem to have lost the idea of a fit-
for-purpose model, including consideration for the stage of the project.
The objective of this paper is to demonstrate that both slope of regression and
kriging efficiency should be viewed with caution for optimizing estimation param-
eters. Lowering conditional bias may be an improper approach to estimating metal
grades. In fact, it might be outright wrong especially in deposits for which high
cutoff grades are required for mining.
This paper presents a summary of the two typical tools, suggested for optimiza-
tion of estimation parameters, slope of regression and kriging efficiency, followed
by application of these metrics applied to an epithermal gold deposit. The case
study compares block estimated grades with uncertainty distributions of global
tonnes and grade at specified cutoffs. The estimated grades are designed for
different block sizes, different data sets, and different estimation parameters, i.e.,
those geared toward lowering the conditional bias and those designed for higher
block grade variability with high conditional biases.
306 M. Nowak and O. Leuangthong
Block ordinary kriging is one of the most common estimation methods used for
resource modeling in the mining industry (Journel and Huijbregts 1978; Sinclair
and Blackwell 2002). Each block estimate can be written as:
X
n
Z * ð uÞ ¼ λ i Z ð ui Þ
i¼1
where Z*(u) represents the estimated block grade at location vector u, and λi is the
kriging weight assigned to sample Z(ui). The resource model, comprised of esti-
mated block grade at all relevant locations, forms the distribution of estimated
grades that is the basis for a mineral resource statement.
Among other considerations, such as geologic confidence, grade continuity, and
database quality, a cutoff grade is used to differentiate between those blocks that are
reported as a mineral resource (Sinclair and Blackwell 2002). This cutoff grade is
applied throughout a project, depending on the method of mineral extraction. The
smoothness of this estimated grade distribution, relative to the cutoff grade, is then
paramount to this discussion of accurately predicting the global mineral resource
for a project.
The smoothness of the estimated grade distribution depends not only on the
quantity and location of conditioning data and the modeled variogram but also on
the estimation parameters such as minimum/maximum number of samples, size of
search neighborhood, and type of search. While there are a number of suggested
measures for assistance in the choice of optimal kriging estimation parameters, such
as slope of regression, kriging efficiency, or weight of the mean from simple
kriging, this paper will only focus on the first two due to their prominent use in
the mineral resources sector.
When kriged estimated Z* block grades are plotted on X axis and unknown true
Z block grade are plotted on Y axis, then the regression of true values given the
estimates is an indication of the conditional bias in the estimate (Journel and
Huijbregts 1978) (Fig. 1). Conditional bias takes place when the expected value
of true block grade Z conditional to estimated block grade Z* is not equal to the
estimated grade. The slope of the regression b is often used to summarize the
conditional bias of the kriging estimate:
Conditional Bias in Kriging: Let’s Keep It 307
Fig. 1 Schematic
illustration of conditional
bias (McLennan and
Deutsch 2002). The
estimates Z* are on the X
axis, and the true block
grades Z are on the Y axis
E Z Z* ¼ z a þ bz ð1Þ
Naturally, the block true values are unknown, but the slope b can be calculated
once a variogram model is known by the following formula:
cov Z; Z*
b¼ ð2Þ
σ 2Z*
To calculate the slope from formula (2), it is enough to know kriging weights
attached to samples used to estimate a block and to know covariances between
samples and samples and the block. Note that actual sample grades are not taken
into account in the calculation. The slope will be identical in both lower- and
higher-grade areas, although potentially higher conditional bias, and by extension
lower slope of regression, could be expected in the high-grade areas.
σ 2Bl σ 2kr
KE ¼ ð3Þ
σ 2Bl
Kriging efficiency values can range from negative (poor estimates) to a maxi-
mum value of 1 (very good estimates). As with the slope of regression, actual assay
data do not get used in the calculation. The results are purely dependent on a
variogram model and on data locations used to estimate the blocks. Krige mentions
that based on a number of case studies he conducted there is a correlation between
the efficiency and the slope of regression (Krige 1997). For the increased slope
value, there is also an increase in kriging efficiency.
3 Case Study
0.90
0.60
0.30
0.00
20 40 60 80 100 120 140 160 180 200
Lag Distance (h)
0.90
0.60
0.30
0.00
20 40 60 80 100 120 140 160 180 200
Lag Distance (h)
Fig. 2 Modeled correlograms from simulated grades (true data) and from exploration drilling
(Fig. 3). Drill hole spacing in the second data set is similar to what would be
typically expected in a gold deposit at the early stages of exploration.
The smaller data set was used for variogram modeling. As is common in
practice, the best continuity was modeled for directions slightly different than
those known from simulated grades. Moreover, the modeled variogram parameters
were different from those applied for the simulated reference data. Each data subset
was used to estimate gold grades by ordinary kriging for different block sizes and
different number of data allowed to estimate the blocks. Vertical block size was
kept constant at 8 m with increased sizes along easting and northing directions.
310 M. Nowak and O. Leuangthong
Fig. 3 Plan view of true block grades and data locations: (a) large data set, (b) small, more
typical, data set
For strongly conditionally biased estimates, a small number of data was used
with a minimum of 4 and a maximum of 12 composite assay grades (referred to as
the “sub-optimal” design). In an attempt to lower the conditional bias, blocks were
estimated with a large search ellipsoid with a minimum of 24 and maximum of
64 composite assays; this is referred to as the “optimal” design. For each estimated
block grade, a slope of regression and kriging efficiency was calculated. In addition,
the estimated block grades were compared to true block grades.
4 Results
Figure 4 shows how slope of regression and kriging efficiency change for different
block sizes, different data sets, and estimation parameters used. As discussed,
modification to the estimation parameters was limited to a number of data used
starting from “poorly designed” not optimal estimation procedure with small
number of data used for the estimation and ending with “well designed” optimal
process with large number of data used for the estimation. As expected, the optimal
Conditional Bias in Kriging: Let’s Keep It 311
Fig. 4 Slope of regression (a) and kriging efficiency (b) for blocks estimated from two data sets
for different block sizes and different estimation parameters
efficiency with the increased block size. Note that kriging efficiency is quite low
regardless of the estimation type. As expected, using the large number of data
(optimal case) to estimate block grades resulted in a substantial increase in the slope
of regression, i.e., it resulted in substantial decrease of conditional bias. Not
surprisingly, the slope of regression is quite high for the large data set and large
number of data used for the estimation. These graphs clearly indicate that using a lot
of data during the estimation process lowers conditional bias and increases kriging
efficiency.
Now that it has been established that applying more data to the block estimates
increases slope of regression and by extension decreases conditional bias, the next
step involved comparisons of actual true block grades with the estimated grades for
different cutoffs. Figures 5 and 6 show relative tonnage and grade differences
between true and estimated block grades at the 1.0 g/t cutoff for optimal and not
optimal estimates from the small data and the large data sets, respectively. Both
figures show that despite high conditional bias in the sub-optimal design, the
estimated tonnes and grades are closer to reference tonnage and grade in the
deposit. This is also true for the estimates from the large data set, although here
the differences between the not optimal and optimal models are smaller. Note that,
as presented in Fig. 6, at 20 m block size estimated tonnes and grade are very
similar to reference tonnes and grade. At the same time, it would be misleading to
conclude that this block size produced superior estimates.
In fact, the reported tonnage and grade from different block sizes is quite similar
(Fig. 7). It just happens that the estimated tonnes and grade in this specific deposit
are comparable to recoverable tonnes and grade at the selective mining unit (SMU)
size higher than 20 m, roughly the size of half of drill hole spacing. This is an
important observation that suggests it does not matter what block size is used for
estimating resources. Reported resource at 8 m or 20 m block size will be similar,
but the 20 m block appears to approach the size that, if successfully applied during
mining operation, would result in actually recovered tonnes and grade very similar
to those estimated. As long as there is no connection made between a block size
used and actual SMU considered for mining, there is nothing particularly wrong
with estimates based on a small block size.
An important result of the optimization of the estimation process is high
smoothing of the estimated block grades. The smoothing effect may result in
large differences between estimated and actual metal content at higher cutoff
values. For block estimates from the typical (smaller) data set, the estimated
metal content at higher cutoffs may be as much as 70–80 % lower than the actual
metal content (Fig. 8) for the “optimized” parameters, while the “sub-optimal”
model yields 50–60 % less metal content relative to the reference. Similar conclu-
sions can be made for the larger data set, with percent differences ranging from
20 to 30 % for the sub-optimal model and 30–50 % for the optimal model. There-
fore, it is obvious that the lower the conditional bias the higher the smoothing effect
that can be expected when estimating from sparsely spaced data. The interesting
trend, however, is that for both the sparsely and densely sampled data sets, the
Conditional Bias in Kriging: Let’s Keep It 313
Fig. 5 Relative tonnage (a) and grade (b) differences between estimated and true block grades at
1.0 g/t cutoff. The blocks were estimated from typical, smaller data set
sub-optimal set of parameters yield the closest estimate of contained metal for
cutoff grades above the 1.25 g/t mean grade.
The final task considered comparison of the estimation models with conditional
simulation of the small data set. The purpose of this exercise was to compare the
estimation model to a model constructed using a method that is considered to avoid
conditional bias altogether (McLennan and Deutsch 2002; Journel and Kyriakidis
2004). The reference distribution was obtained via p-field simulation wherein the
local distributions of uncertainty considered a local trend model. For this task,
sequential Gaussian simulation was performed with no consideration for any trends,
and variograms were calculated based on the small data set. As with the estimation
models, the continuity directions vary slightly from those of the reference model.
314 M. Nowak and O. Leuangthong
Fig. 6 Relative tonnage (a) and grade (b) differences between estimated and true block grades at
1.0 g/t cutoff. The blocks were estimated from large data set
Fig. 7 Estimated from small data set grade and proportion of tonnage above 1.0 g/t cutoff for
different block sizes
Fig. 8 Relative metal losses in estimated block grade models for different cutoffs in 16 m block
models
reference model at higher cutoff grades. Thirdly, the simulation approach, which is
considered to be a non-biased method, yields the distributions of uncertainty in
contained metal that encompasses the reference data. This latter observation is
important, particularly as a reference model is never available in practice for
benchmarking purposes. This indicates that a conditional simulation approach can
316 M. Nowak and O. Leuangthong
Fig. 9 Contained metal curves for estimated block models for different cutoff grades at 8 m, 12 m,
16 m, and 20 m block sizes
be used to validate the estimated tonnage, grades, and ultimately metal when
determining an appropriate set of estimation parameters.
5 Discussion
Both slope of regression and kriging efficiency assume global stationarity within a
specifically modeled domain. Mineral deposits are not stationary, even within a
specific estimation domain. In an estimation domain, there are always small regions
of high and small regions of very low grades. Slope of regression and kriging
efficiency formulas do not take into account the fact that true variance of estimation
errors depends on data values. In regions with higher grade or in regions with local
Conditional Bias in Kriging: Let’s Keep It 317
data having more variance than in the whole domain, fluctuation of errors is larger.
Disregarding these local changes in variability may lead to estimated resources that
steer away from what would be expected of a typical resource estimate.
A resource block model should not be designed to produce an inventory of
recoverable resource that is based on SMU size much larger than would generally
be considered for mining, only because for this SMU size, when mined, the
resource model will turn out to be correct. A deposit is never mined according to
a resource block model. A decision on what will be mined will be based on grade
control drilling, necessary both in open-pit and underground mining. A resource
geologist should strive to produce a block model that predicts, reasonably well, the
tonnage and grade that a mine can expect to achieve over the life of the mine, once
it has sorted out its grade control procedures. In addition, the resource block models
are often used for dilution calculations or blending issues. A block model designed
from optimizing slope of regression, or block size, will not serve this purpose.
In short, optimizing kriged block estimates with slope of regression or kriging
efficiency measures may lead to block models that do not adequately reflect true
block grades. It is tempting and easy to use slope of regression and kriging
efficiency for validation of block estimated grades. Those measures are commonly
available in commercial software packages. Although theoretically high slope of
regression, i.e., low conditional bias, is considered necessary for good quality
estimates, in practice this approach may be outright harmful if the objective of
the study is to predict global resource quantities above an economic cutoff grade.
Both measures are a reflection of a modeled variogram and data locations and do
not take into account actual assay values or their variability in the vicinity of an
estimated block. Moreover, it is often quite difficult to construct a reliable
variogram model, particularly in early exploration stages, and relying on its metrics
to design “best” resource estimates cannot be considered best practice.
Based on the presented case study, there is strong indication that it is better to
have conditionally biased block estimates for global resource quantities required for
a potential investment decision, life-of-mine planning, and/or development deci-
sions. Once a cutoff is applied to block estimated grades for reporting or further
mining studies, it is better to have unsmoothed conditionally biased block esti-
mates. In this context of achieving globally accurate predictions, it looks like the
onus is back on a resource geologist to design estimation parameters that produce a
realistic block model that reflects the underlying true block grades the best way
possible.
Bibliography
Arik A (1990) Effects of search parameters on kriged reserve estimates. Int J Min Geol Eng 8
(4):305–318
Deutsch J, Szymanski J, Deutsch C (2014) Checks and measures of performance for kriging
estimates. J South Afr Inst Min Metall 114:223–230
318 M. Nowak and O. Leuangthong
Isaaks E (2005) The kriging oxymoron: a conditionally unbiased and accurate predictor, 2nd edn,
Geostatistics Banff 2004. Springer, Dordrecht, pp 363–374
Journel A, Huijbregts C (1978) Mining geostatistics. Academic, London
Journel A, Kyriakidis P (2004) Evaluation of mineral reserves: a simulation approach. Oxford
University Press, New York
Krige D (1997) A practical analysis of the effects of spatial structure and of data available and
accessed, on conditional biases in ordinary kriging, Geostatistics Wollongong ’96, Fifth
International Geostatistics Congress. Kluwer, Dordrecht, pp 799–810
Krige D, Assibey-Bonsu W, Tolmay L (2005) Post processing of SK estimators and simulations
for assessment of recoverable resources and reserves for South African gold mines,
Geostatistics Banff 2004. Springer, Dordrecht, pp 375–386
McLennan J, Deutsch C (2002) Conditional bias of geostatistical simulation for estimation of
recoverable reserves, CIM Proceedings Vancouver 2002. CIM Proceedings Vancouver 2002,
Vancouver
Olea R (1991) Geostatistical glossary and multilingual dictionary. Oxford University Press,
New York
Rivoirard J (1987) Teacher’s aide: two key parameters when choosing the kriging neighborhood.
Math Geol 19:851–856
Sinclair A, Blackwell G (2002) Applied mineral inventory estimation. Cambridge University
Press, Cambridge
Vann J, Jackson S, Bertoli O (2003) Quantitative kriging neighbourhood analysis for the mining
geologist – a description of the method with worked case examples. 5th International Mining
Geology Conference, Bendigo, pp 1–9
Operational SMU Definition of a Brazilian
Copper Operation
1 Introduction
Estimated mineral resources models are made to predict the deposit tonnage and
grade distribution curves that will be extracted and fed into the plant (Rossi and
Deutsch 2014).
Geostatistical estimates of mining recovery curves may be somewhat optimistic,
as they do not take into account factors other than block grades, commonly
disregarding mining configuration (method and geometry), local geology in the
production environment, and production rates. Mining dilution (planned and/or
2 Dilution
The case study was based on a copper/gold operation mine in a massive deposit
with large production rate. Three block models are defined in the production
process:
1. Long-term model: used in the life of mine, 5Y, and annual mining plans
2. Short-term model: used in the quarterly and monthly plans
3. Grade control model: used in the daily mining operation
The long-term model uses only diamond drill hole (DDH) information. The
short-term and the grade control models use only blast holes data (approx. 6 6 m).
The long-term and short-term models use a 15 15 15 m SMU size. The grade
control model corresponds to the estimated blasting polygons of the mining pro-
cess. Figure 1 shows long-term block model along with the mining polygons on a
specific level.
The tonnages and grades on this study are compared between models using all
mining blasting polygons and production data existing in the mine. Therefore, all
the tonnage and grade reports correspond only to the mined portion of the deposit.
The long-term block model was updated considering production reconciliation
information (geological contacts and short-term grade variations) for this study;
thus both long- and short-term models are practically the same in this operational
example. Hence, all comparisons in the study used the long-term model.
Support dilution is taken into account in the long-term model through ordinary
kriging estimation plans, ensuring that block distribution is coherent with the
change-of-support data analysis. Contact dilution is taken into account by tonnage
weighting of the materials inside each block.
In the mining scheduling process, long-term engineers compute tonnages and
grades based on the long-term model. Mining blasting polygons, however, are the
main drivers of the operational process. This is ultimately considered as the grade
control model and has “irregular” SMU size as a result of the local mining process
and its resulting quality values. As the models (long-term compared to the grade
control model) have distinct support, they have different tonnage and grade curves.
Comparing these curves, the differences due to operational configuration (mining in
polygons) is clear. Figure 2 shows these curves, where it shows how much is not
recovered for each cutoff. This difference is a measure of the planned dilution.
Long-term mine planning take into account that each SMU will be selectively
mined as ore or waste. However, mining operational process is based on blasting
polygons and local operating particularities, which can result in poor reconciliation
of long-term block models and the mining recovery curves, indicating lesser levels
of selectivity. Hence, the importance of acknowledging planned dilution is verified.
Planning engineers usually consider planned/unplanned dilution as a constant
factor. However, dilution is not constant and depends on the spatial distribution of
the deposit. Homogeneous regions should have less dilution than regions with more
contact variability. Therefore, if the production rate remains constant, without
322 J.D. Reuwsaat et al.
0.00 - 0.20
0.20 - 0.60
0.60 - 1.00
1.00 - 2.00
2.00 - 100.00
Scale 1:3000
Fig. 1 Mining polygons with the long-term block model for a specific level
changing significantly the sizes of the ore polygons, it is logical to increase the
SMU size to account for the operational planned dilution instead of using factors.
4 Results
The SMU size definition that will better represent actual mining recovery curves
was based on an iterative long-term change-of-support process, keeping 15 m as
bench height. The models were reblocked based on the official 15 15 15 m
long-term block model. Figure 3 presents the comparison between 15 15 15
and 30 30 15 m models. Density was considered as a weighting factor in the
reblocking process. Figure 4 shows the tonnage and grade curves from all the
models. It can be observed that the grade control model (production polygons)
corresponds to a larger support size than the long-term model.
The impact of change of support is clear when computing the tonnage, grade,
and metal differences between long-term and the reblocked and grade control
models. Figure 5 presents the differences in tonnage, grade, and metal, respectively.
The grade control curve is presented by a thick black line.
The selection of the SMU size is made by a simple visual interpretation of the
metal recovery and difference curves, as it accounts for both tonnage and grade
Operational SMU Definition of a Brazilian Copper Operation 323
Fig. 2 Grade control and long-term model tonnage and grade curves
results. Figure 5 shows that the grade control model near matches the
30 30 15 m long-term model for most of the distribution curve (between
0.2 % Cu and 1.2 % Cu). Figure 6 shows 15 15 15 m and 30 30 15 m
metal and grade mining recovery curves compared to the grade control model.
The results indicated that a near-perfect match to the actual distribution curve
(grade control) is unlikely, since no irregular SMU sizes are recognized and applied
in the mining industry for different mining configurations, production rates, and
local geology. As a consequence, impacts in mining companies’ downstream
processes such as milling, stockpiling, and financials are commonly verified.
Mineral resource evaluators and mining planners should start to apply different
mining concepts for “achieving” the expected operational mining recoverable
curves at given cutoffs. Irregular local SMU sizes could be an option.
Nevertheless, in this case study, the 30 30 15 m model would better fit
operational distribution curves than the 15 15 15 m as a result of the accounted
actual planned dilution, avoiding the use of a gross dilution factor.
324 J.D. Reuwsaat et al.
0.00 - 0.20
0.20 - 0.60
0.60 - 1.00
1.00 - 2.00
2.00 -100.00
5 Discussion
A real operational database was used for the construction of the mining recoverable
curves and method definition. The method is simple and helpful in the selection of
the SMU size, accounting the planned operational dilution. Unplanned dilution was
not considered in this study; however it should be accounted as a full evaluation of
the mining process.
In cases where no actual production information is available, planning engineers
can simulate the grade control practice to arrive at ore/waste dig limits that will be
implemented in the mine for operational mining sequencing. Then engineers and
geostatisticians can verify the impact of the production scale on the dilution and
consequently on the recoverable model. Figure 7 shows the impact of the dilution
on metal distribution for several real production polygons. The difference of metal
in Fig. 7 is changed to positive for better graphic visualization. The metal loss is
higher for larger polygon sizes (larger production rate and mining configuration).
If the operational mining recovery curve does not match to any regular SMU
support, it could be the case of using irregular operational mining unit (OMU) sizes
(Diedrich et al. 2016) for defining local mining recoverable curves based on an
initial mining schedule. The idea is not to forecast future short-term SMU grade
distributions at the time of mining and not questioning estimation methods for
defining grades at SMU scales. However it is to determine a reasonable regular
and/or irregular block support for a chosen conventional SMU block model size that
represents ore/waste proportions derived from a real expected mining sequencing
and grade control process. In the OMU approach, there is no commitment that an
Operational SMU Definition of a Brazilian Copper Operation 325
6 Conclusion
Recognizing the impact of operational dilution (planned and unplanned) and the
fact that it is not constant over the deposit, operational aspects should be included in
the mining unit block size definition and, consequently, on the resulting mining
326 J.D. Reuwsaat et al.
Fig. 5 Tonnage, grade, and metal differences between the reblocked and the long-term models
compromise depends on the global and/or local ore and waste proportions for
predefined operational production rates and mining configurations. Significant
differences between operational and long-term recovery curves may result in poor
mining, waste dumps, and stockpile plans and strongly impact financials.
Bibliography
1 Introduction
2 Tools
The tools which will be used are derived from the geostatistical concept of
estimation variance and have been developed to measure the sampling density,
either with a regular sampling design or an irregular one. A reminder of these tools
is presented here. More details are available in the above references.
Let Z(x) be the regionalized variable under study, assumed to be additive.
Consider a domain V, an estimation Z(V )* of Z(V ) by inner samples of V, and its
estimation variance σ2E (V ). By definition the “spatial sampling density variance”
(ssdv) associated to this estimation Z(V )* is (Rivoirard 2013):
χ ðV Þ ¼ σ E 2 ðV ÞjV j:
X jvi j
Z ðV Þ* ¼ i jV j
Z ðvi Þ* :
If the estimation errors are uncorrelated, then the estimation variances can be
combined as:
X jvi j2
σ 2E ðV Þ ¼ σ 2E ðvi Þ
i
jV j2
so that χ is additive:
X jvi j
χ ðV Þ ¼ i jV j
χ ðvi Þ:
This is exactly the case of random stratified sampling, where each stratum vi is
sampled by samples with random locations that are uniform and independent (note
that this is true for model-based geostatistics, but the same property exists in
random sampling theory, Cochran (1977)). In practice the correlation between
errors, if not zero, can often be neglected. Matheron (1971) has developed appro-
priate principles of approximation in order to compute estimation variances by
combination of variances in different situations. In the particular case of a
two-dimensional square regular grid and usual isotropic variogram models,
V being divided into N cells with same support v, we have:
σ 2E ðV ÞjV j ¼ σ 2E ðvÞjvj;
that is, χ ðV Þ ¼ χ ðvÞ. We can deduce that the ssdv of any union V of vi is the same:
the spatial sampling defined by the regular grid pattern is characterized by this ssdv.
As a consequence the estimation variance of any set V (even not connected) made
of such vi can be derived:
jvj 2 χ
σ 2E ðV Þ ¼ σ E ðvÞ ¼ :
jV j jV j
Remark 1: The non-correlation between errors is not general and can be tested by
simulations. Having disjoint volumes estimated by inner samples does not imply
that errors are uncorrelated, even approximately (example of elongated blocks
sampled by centered samples which are close to each other). On the contrary the
correlation between errors often tends to vanish when considering bigger blocks.
Remark 2: Because the estimates such as Z(V )* make use of inner samples only,
they are less precise than kriging and so σ 2E (V ) that can be derived is pessimistic
compared to kriging variance σ 2K (V ). This is particularly the case when V is
composed of a small number of v.
332 J. Rivoirard et al.
Such specific volumes are similar to the inverse of the classical density of points
in space (number of points per unit volume). In irregular sampling pattern, when
there are variations of sampling in space (not due to known and distinctly sampled
areas), the ssdv or specific volume can be estimated and mapped just like the density
of points, that is, using a moving kernel, for instance, a parallelepiped (but similarly
the result will depend on this). In mining, this can be done by “superkriging,” that
is, kriging superblocks from inner samples only (Rivoirard and Renard 2016).
These superblocks are centered at every block of the block model (usually these
model blocks are small and most are not sampled).
In the following, the methodology is applied to different cases. The first case is
two-dimensional with a regular sampling. The deposit is divided into big blocks
which have the size of the cell and negligible errors correlation. The ssdv or the
specific volume, which measures the density of the sampling pattern, is derived
from the estimation variance of such a big block from its sample. Then they are
used to compute the coefficient of variation of the resources contained in a
production area of given size, which will fix the category of resources.
The two other cases are three dimensional with irregular sampling. A map of
ssdv or specific volume, which measures the density of the sampling pattern, is
derived from the kriging variance of a moving superblock from its samples. Then
this map can be used to make the distinction between differently sampled areas. It
also allows computing the coefficient of variation of the resources contained in a
production volume of given size, which will fix the category of resources in
each area.
From the Spatial Sampling of a Deposit to Mineral Resources Classification 333
The case of a large deposit studied in two dimensions and sampled with regular
grids is frequent and is presented here as a generic example. The two-dimensional
regionalized variable of interest can be the thickness of the deposit (typically when
the grade in metal presents small variations) or the metal accumulation.
Corresponding resources are either volumes or abundances. In the present case,
the resources will be volumes, and the variable of interest will be the thickness. It
has a mean of 10 m, a sample point variance of 10, and a variogram equal to
1 nugðhÞ þ 9 sphðh=200 mÞ.
The initial sampling grid is 100 100 m. This has been progressively centered,
giving a regular grid of 71 71 m, and finally tightened down to a pre-exploitation
grid of 50 50 m. The specific volumes will be computed to measure the spatial
pattern corresponding to each grid. We first compute the extension variance of the
grid cell centered by a sample using the variogram and divide it by the squared
mean of the variable. Equivalently we can compute directly the estimation variance
using the relative variogram (i.e., divided by the squared mean), which is
0:01 nugðhÞ þ 0:09 sphðh=200 mÞ. The specific volume (actually a
two-dimensional area) is obtained by multiplying the result by the cell area. Each
grid is characterized by its specific area, which decreases when the grid is refined
(Table 1). This provides an objective way to quantify the sampling efficiency
corresponding to a given grid for a variable having given statistics. It enables
comparison between different grids for the same deposit, as well as comparison
between the sampling of different deposits of the same type.
Let us for instance consider a second deposit of the same type as the first one, but
with a thickness having a mean of 8 m and a variogram equal to
6 nugðhÞ þ 6 sphðh=200 mÞ. Note that it is less thick and more variable than the
first deposit. Its initial sampling grid is 70 70 m, practically equal to the second
grid of the first deposit. However, because of its variability, this grid corresponds
here to a specific area of 502 m2, that is, a spatial sampling twice less good than the
100 100 m grid of the first deposit (269 m2).
To convert specific volumes or areas V0 in resource category, we must have an
idea of the exploitation rate to be considered. Let us go back to the first deposit and
consider an annual exploited area of V ¼ 100,000 m2 (i.e., 40 blocks with size
50 50 m). This corresponds to a volume of ore around 1,000,000 m3. The CoV on
the average thickness, that is, on this ore volume in such an area V, depends on the
grid through its specific volume V0 and is given by Eq. 1 (Table 1). The classifica-
tion itself finally depends on the thresholds based on the CoV that the company has
chosen for this type of resources. Suppose that the CoV on annual resources must be
less than 2.5 % for measured resources and less than 5 % for indicated resources.
Then, the resources would be inferred where the grid is 100 100 m, indicated
where it is 71 71 m, and measured where it is 50 50 m. But with a higher
production rate, resources where the grid is 100 100 m could be upgraded from
inferred to indicated, while with a smaller production rate, resources where the grid
334 J. Rivoirard et al.
Table 1 Specific area V0 for different grids, corresponding to the relative variogram equal to
0:01 nugðhÞ þ 0:09 sphðh=200 mÞ. CoV represents the coefficient of variation that can be deduced
for an estimation of the target variable within an area of 100,000 m2
Grid Vo CoV
100 100 269 m2 5.2 %
71 71 111 m2 3.3 %
50 50 46 m2 2.1 %
where the very large vertical range is purely conventional. Because of the irregular
sampling, the specific volume corresponding to the spatial sampling has been
mapped using a superkriging at each block of the model. This superkriging consists
in kriging with its inner samples within a moving superblock centered at each block.
The superblock must be large enough to avoid irregularities, but small enough to
limit the smoothing. A superblock size of 150 150 50 m is chosen. At each
From the Spatial Sampling of a Deposit to Mineral Resources Classification 335
2000 N
1000 N
0N
Unit 1 Unit 4
Unit 2 Unit 5
-1000 N
Unit 3 Unit 6
2000 E
1000 E
0E
Fig. 1 Horizontal projection of drill holes, with the pipe in the center
Fig. 2 E-W vertical section with the units. Same unit colors as in Fig. 1. Units 1 (blue) and
2 (brown) are poor
Fig. 3 E-W vertical section. Copper grade estimated by kriging. Black lines represent drillholes
close to the section
336 J. Rivoirard et al.
0.20 0.20
hor
0.15 0.15
0.10 0.10
ver
0.05 0.05
0.00 0.00
0 100 200 300
Fig. 5 E-W vertical section. Specific volume. Green lower than 15,000 m3; yellow lower than
50,000 m3
0 1000 2000
2300 2300 Measured
2200 2200
2100 2100
Indicated
2000 2000
1900 1900
0 1000 2000 Inferred
block the specific volume is obtained as the superkriging variance computed from
the relative variogram, multiplied by the size of the superblock (Fig. 5).
The map of specific volume provides a delineation of the deposit in terms of
sampling density. This can be compared to the current classification made on a
geometric criteria based on the presence of data close to each block (Fig. 6). Based
on the specific volume, measured resources would correspond to blocks in green:
From the Spatial Sampling of a Deposit to Mineral Resources Classification 337
specific volume less than 15,000 m3 with an average specific volume of 5500 m3.
Indicated resources would correspond to specific volume between 15,000 and
50,000 m3, with mean specific volume 28,000 m3 (in yellow). Other resources
(red and gray) would be inferred.
The annual production is about 50 Mt, that is, V~20 Mm3with rock density
2.6, and the metal quantity should not deviate by more than r ¼ 3 %. The
coefficient of variation of this quantity depends on the specific volume (Eq. 1).
In a Gaussian framework, the probability that a random variable with given CoV
deviates by more than r from its expectation is given by 2Gðr=CoVÞ where G is
the c.d.f. of a standard Gaussian. The CoV threshold for measured resources,
corresponding to V0 ¼ 15,000 m3, is 2.7 % (probability of 1/4, i.e., 1 year out of
4, to deviate by more than 3 %), while the CoV corresponding to the average
specific volume of measured resources (5500 m3) is 1.7 % (probability of 1/10 to
deviate by more than 3 %). Concerning indicated resources, the threshold CoV,
corresponding to V0 ¼ 50,000 m3, would be 5 % (probability of 1/1.8 to deviate
by more than 3 %), and the CoV corresponding to the average specific volume of
indicated resources (28,000 m3) is 3.7 % (probability of 1/2.4 to deviate by more
than 3 %).
Compared to other porphyry copper deposits, the specific volumes in the best
sampled areas of El Teniente are low, which means that the spatial sampling is quite
good, probably because of a higher spatial continuity of grades.
This case study is a big iron ore deposit located in Brazil. The banded iron
formation from the Quadrilatero Ferrı́fero is called itabirite, a metasedimentary
rock. The iron enrichment is given mainly by weathering which dissolves carbonate
and siliceous bands, increasing the residual concentration of iron oxides, and
decreasing the cohesion between grains, bands becoming friable in most part of
economic ore body (Ribeiro and Carvalho 2000). A number of 15 lithotypes with
distinct lithochemical characteristics, either mineralized or waste, have been dis-
tinguished. Of particular interest here will be the group of uncontaminated facies
(denoted HC + HF + IF + IC), and the group of ore facies (HC + HF + IF + IC +
HAR + IAR + IFR + IMN). The deposit is sampled irregularly with vertical or
subvertical holes. Locally a base grid of about 70 70 m exists, but the sampling
design is very often less dense and varies horizontally (Fig. 7) as well as vertically
(Fig. 8). The block sizes used in this deposit are 50 50 13 m and 10 10 13 m.
Sample data are 13 m composites, regularized by lithotype domain. Facies is
available for all samples, but this is not the case for chemical variables, in particular
for the Fe grade. This has a consequence on the estimation and classification. For
some blocks, it is possible to estimate the facies, but not the Fe grade. Classification
338 J. Rivoirard et al.
1000
-1000
-2000 -1000 0 1000 2000
Fig. 7 Holes within a horizontal section in the middle of the deposit. In yellow, the area where Fe
grade is estimated
1300
1200
1100
of blocks is made on facies, but only blocks where Fe grade can be estimated are
considered.
The regionalized variables used to measure the spatial sampling of the deposit
are the facies indicators (at samples), regularized into facies proportions
(on blocks). In such deposits, the facies are numerous and many variograms are
not well known, so that some simplifications are commonly made. Facies indicators
are supposed to be in intrinsic correlation, that is, all facies indicators (whether
individual or grouped) obey to the same variogram model (up to its vertical scale).
This variogram is determined from the variogram of the indicator of
uncontaminated facies, which is well known and is further used for all ore facies.
The reference plane is dipping 25 to the east. The variogram is isotropic within the
reference plane. The variogram model, using the reference plane, is composed of a
nugget, a short isotropic 40 m component, and a large component with anisotropy
coefficient of 4 (Fig. 9):
0.25
0.20
vert
0.15
0.10 ~hor
0.05
0.00
0 100 200 300 400 500 600 700 800
125
100
75
50
25
0
0 100 200 300
The mean indicator is close to 0.5 and the sill or variance is 0.25. A normalized
variogram γ 0(h) is obtained by dividing the variogram by 0.25:
1.0
1000 1.0
0.9
0.8
0.7
0.6
0 0.5
0.4
0.3
0.2
0.1
-1000 0.0
-2000 -1000 0 1000 2000
Fig. 11 Horizontal section in the middle of the deposit. Kriged iron ore proportion for
50 50 13 m blocks
5000000
1000 3000000
2000000
1000000
500000
300000
0 250000
200000
150000
100000
50000
-1000 0
-2000 -1000 0 1000 2000
Fig. 12 Horizontal section in the middle of the deposit. Specific volume computed from
350 350 39 m superblocks at 50 50 13 m blocks
where Ind* is the ore proportion of the block, obtained by kriging the iron ore
indicator, and σ 2IK the normalized kriging variance (i.e., using the variogram
normalized by its sill). Typically, blocks are classified as measured when RI
<0.6, indicated when 0:6 < RI < 0:8, and inferred otherwise. This RI extends
the notion of confidence to risk, as a block with a proportion of ore less than 20 %
corresponds to inferred resources, even if this proportion is perfectly known.
Another version of Risk Index can be derived using the specific volume V0 or the
normalized specific volume v0:
V0 1 v0
New RI ¼ 1 Ind *
þ ¼ 1 Ind þ
*
1
W p W
where p is the global ore proportion (63 %), and W is a constant. The classical RI is
sensitive to the size of blocks (e.g., 10 10 13 m or 50 50 13 m) through the
kriging variance. On the contrary both terms Ind* and V0 are additive, so that the
new RI does not depend on the size of blocks. It can provide the same service as
the classical one, with this additional advantage. By taking W ¼ 10,000,000 m3
(practically the annual volume of reference V ), the new RI (Fig. 14) was observed
342 J. Rivoirard et al.
1.0
1000 1.0
0.9
0.8
0.7
0.6
0 0.5
0.4
0.3
0.2
0.1
-1000 0.0
-2000 -1000 0 1000 2000
1.0
1000 1.0
0.9
0.8
0.7
0.6
0 0.5
0.4
0.3
0.2
0.1
-1000 0.0
-2000 -1000 0 1000 2000
Fig. 14 Horizontal section in the middle of the deposit. New RI for 50 50 13 m blocks
6 Conclusion
To apply the proposed classification method, one must choose the regionalized
variable corresponding to the resources. In our two-dimensional application, the
thickness was used, but in other cases the metal accumulation would be more
appropriate. In cases where the target variable is not additive, but is the ratio of
additive variables, combining variances and specific volumes could be done using
reduced differences (Journel and Huijbregts 1978). More generally, the method
addresses in situ resources rather than recoverable ones. Geological cutoff grades
on sample values can be applied, but mining cutoffs on selection blocks would
require more sophisticated tools such as simulations.
From the Spatial Sampling of a Deposit to Mineral Resources Classification 343
The originality of the method is in its two steps. The first step consists in
measuring (or mapping when it varies) the spatial sampling density in the deposit.
The tool which is used, the specific volume, can be seen as a generalization of the
density of sample points (Deustch et al. 2007), but it takes into account the
variographic structure. In the second step, the specific volume is converted into
coefficients of variation corresponding to large target volumes, typically the
expected annual volume to be mined.
The use of coefficients of variation on production units, rather than confidence
on individual blocks, is not original in resource classification, and it allows a
validation by actual production data when available. It is advocated in particular
by Dohm (2005). In her “logical approach,” she makes the link between the
coefficient of variation of “units representing likely production periods” and a
“typical coefficient of variation for blocks” through a “factor production period,”
which is determined by conditional simulations. Our approach via a specific volume
is different.
The method is simple and straightforward but it aims at giving orders of
magnitude rather than precise numbers. Coefficients of variation obtained from
the specific volume assume that the target volumes are estimated from their inner
samples only. This may lead to pessimistic values of coefficients of variation,
particularly for target volumes that are too small. The specific volume depends on
the variogram and so on its knowledge. At the early stage of systematic exploration,
when the hole spacing exceeds the variogram range and resources are expected to
be classified as inferred, a pure nugget effect may be considered as an approxima-
tion. Moreover the target volumes need not be precisely delineated, which makes
the method flexible with respect to expected production volumes. Of course the
method is not meant to replace the direct estimation of the resources that are
contained in delineated domains – such as the area that is to be mined, say, next
year – nor their confidence. Kriging and conditional simulations are the appropriate
tools for this, but are more demanding.
Acknowledgments The authors are grateful to Codelco and Vale for their support in this
research. They wish to thank the participants of the workshops on mineral resource classification
that were held on 26–28 November 2014 at Codelco and 1–2 December 2014 at Vale, where the
present methodology was discussed and put into practice. Mapping the specific volume has been
made possible, thanks to a Superkriging plug-in developed for Geovariances Isatis software.
Bibliography
Cochran WG (1977) Sampling techniques, 3rd edn. Wiley, New York, 428p
Deutsch CV, Leuangthong O, Ortiz JM (2007) Case for geometric criteria in resources and
reserves classification. Trans-Soc Min Metall Explor Inc 322:1
Dohm C (2005) Quantifiable mineral resource classification – a logical approach. In:
Leuangthong O, Deutsch CV (eds) Geostatistics banff 2004, vol 1. Kluwer Academic, Dor-
drecht, pp 333–342
344 J. Rivoirard et al.
Oscar Rondon
Abstract The estimation of ore dilution and ore loss factors is critical for the
evaluation of mining projects, with the proper estimation of these effects essential
for Ore Reserve Estimation. Unplanned ore dilution occurs when excessive
amounts of waste are mined with ore, and such dilution may result in the processing
of lower than economic cut-off grade material with the ore. Unplanned ore loss
occurs when material that is above the economical cut-off grade is hauled to waste
stockpiles, due to poor mining practices and/or poor information regarding the local
grade of the critical components.
The sources of dilution and ore loss in mining operations are many. This study
focuses on the assessment of the ore dilution and loss in a Mineral Resource model,
specifically, the proportion and average grade of blocks that are misclassified as ore
and waste, which can be referred to as Resource Model Ore Dilution and Model Ore
Loss, respectively. Unlike the well-established geostatistical conditional simulation
approach for assessing model dilution and loss, in this study analytical expressions
are derived that are defined under the theoretical framework of the discrete Gauss-
ian method for change of support to quantify the expected model dilution and ore
loss. Practical application of this method is demonstrated through a case study from
an Iron Ore deposit in Australia.
1 Introduction
Mineral Resource estimates are reported above an economic cut-off grade on the
basis of model estimated grades and not according to the real grade values. The
effect of this is that all Mineral Resource models will certainly incur misclassi-
fication of mining blocks as ore or waste.
O. Rondon (*)
CSIRO Mineral Resources, 26 Dick Perry Avenue, Kensington, WA 6152, Australia
Centre For Exploration and Targeting, University of Western Australia, Crawley, Australia
e-mail: [email protected]
Let Z(x) and Z(v) be the grade at point x and block support v, respectively. The
discrete Gaussian method (Matheron 1976) expresses Z(x) and Z(v) as function of
two standard Gaussian variables Y and Yv as
X
Z ð x Þ ¼ ϕð Y ð x Þ Þ ¼ ϕ H ðY ðxÞÞ
n n n
ð1Þ
where ϕ is the point support anamorphosis function derived from the point support
data, H n n 0 are the normalised Hermite polynomials, ϕn n 0 are the coeffi-
cients of the expansion of ϕ in terms of Hermite polynomials and
Resource Model Dilution and Ore Loss: A Change of Support Approach 347
X
Z ð v Þ ¼ ϕv ð Y v Þ ¼ ϕ r
n n
n
H n ðY v Þ ð2Þ
where ϕv is the block anamorphosis function which is derived via Cartier’s relation
by assuming that the two Gaussian variables Y(x) and Yv have joint Gaussian
distribution with correlation r > 0 (Rivoirard 1994).
The coefficient r corresponds to the variance or support correction factor from
point to block support and is chosen so as to respect the variance of Z(v) by
inverting:
X
VarðZðvÞÞ ¼ ϕ2 r 2n
n n
ð3Þ
where Yv stands for a standard Gaussian variable and ϕv is the corresponding
anamorphosis function but with correction factor s chosen so as to respect the
variance of Z*(v) by inverting:
X
VarðZ * ðvÞÞ ¼ ϕ2 s2n
n n
ð5Þ
The relation between Z(v) and Z*(v) is obtained using (2) and (4) and is given by
X
CovðZðvÞ, Z * ðvÞÞ ¼ ϕ ϕ r
n, m1 n m
s E H n ðY v ÞH m ðY *v Þ
n m
ð6Þ
A further assumption is that (Yv, Yv ) has bivariate Gaussian distribution with
correlation ρ, then E H n ðY v ÞH m Y *v ¼ 0 for n 6¼ m and E H n ðY v ÞH n Y *v ¼ ρn
(Rivoirard 1994) which implies that
X
CovðZðvÞ, Z* ðvÞÞ ¼ ϕ2 r n s n ρn
n1 n
ð7Þ
After computing the correlation ρ by inverting (7), the bivariate Gaussian distri-
bution of (Yv, Yv ) is completely specified, and the joint distribution (Z(v), Z*(v))
348 O. Rondon
can be modelled. This is the key property for modelling the information effect
(Roth and Deraisme 2000) and for the development of the analytical expressions for
the assessment of model dilution and ore loss as discussed in this study.
For the modelling of dilution and ore loss, the focus is on the true block grade Z(v)
conditioned by whether the estimated block grade Z*(v) is above or below an
economical cut-off grade z.
The tonnage associated with model dilution is
P ZðvÞ < z=Z * ðvÞ z ¼ P Y v < y, Y *v y* =P Y *v y* ð8Þ
with y ¼ ϕ1
v ðzÞ and y ¼ ϕv ðzÞ. Since (Yv, Yv ) has Gaussian distribution with
* *1
and gρ is the bivariate Gaussian density with correlation ρ. Therefore, the tonnage
associated with model dilution can be explicitly calculated as
P ZðvÞ < z=Z* ðvÞ z ¼ 1 F y* Hρ y; y* = 1 F y* ð11Þ
1
E ZðvÞ1ZðvÞ<z =Z * ðvÞ z ¼ E Z ðv Þ1 1
Z ðvÞ<z Z ðvÞz
*
1 Fð y * Þ
1
¼ E ϕv ðY v Þ1Y v <y 1Y *v y* ð13Þ
1 Fðy Þ
*
X
Eðϕv ðY v Þ1Y v <y 1Y *v y* Þ ¼ ϕr
n n
EðH n ðY v Þ1Y <y 1Y *v y* Þ
n
X Z y Z þ1 v
ð14Þ
¼ ϕ rn
n n
Hn ðuÞgρ ðu, tÞdtdu
1 y*
where g corresponds to the standard Gaussian density (Chiles and Delfiner 2008).
Substituting (15) in (14) and defining
Z þ1
Ui, j ðxÞ ¼ H i ðuÞH j ðuÞgðuÞdu i, j 0 ð16Þ
x
with δnk the Kronecker delta function representing the orthogonality property of the
Hermite polynomials
Z þ1
δnk ¼ H n ðuÞH k ðuÞgðuÞdu ð18Þ
1
Therefore, the analytical expression for computing the expected metal associ-
ated with model dilution is
1 X X
EðZðvÞ1ZðvÞ<z =Z* ðvÞ zÞ ¼ ϕ r n
ρk ½δnk U n, k ðyÞU 0, k ðy* Þ ð19Þ
1 Fðy* Þ n n k
The values U n, k ðÞ can be computed recursively (Rivoirard 1994), and therefore
once the anamorphosis modelling is completed, the metal amount can be computed.
The expected average metal associated with model ore loss can be obtained
similarly and is given by
1 X X
EðZðvÞ1ZðvÞz =Z* ðvÞ < zÞ ¼ ϕn r n k ρk ½δnk U 0, k ðy* ÞU n, k ðyÞ ð20Þ
Fðy Þ
* n
Using the derived tonnage and metal, the average grade E½ZðvÞ=ZðvÞ < z, Z* ðvÞ
z and E½ZðvÞ=ZðvÞ > z, Z * ðvÞ z associated with model dilution and ore loss,
respectively, can be computed.
The tonnage and metal associated with blocks correctly classified as waste or ore
can also be obtained. These equations are provided in the Appendix.
350 O. Rondon
The derived analytical expressions can be used in two ways: for forecasting,
from sparse drilling data, the expected misclassification of blocks at the time of
having close-spaced drilling data and smaller block sizes, and for assessing the
expected model dilution and ore loss from a given resource model. Both applica-
tions are shown below.
Resource Model Dilution and Ore Loss will inevitably occur when selecting blocks
on the basis of whether their estimated grades are above an economical cut-off z or
not. Therefore, it is instructive to study the behaviour of both for a range of cut-off
values in order to decide, for instance, whether it is preferable to accept dilution or
ore loss. As a demonstration, consider the iron grades (Fe%) in an Iron Ore deposit
with a nominal drill hole sample spacing of 100 m by 100 m that have been used to
compute the tonnage associated with model dilution and ore loss with a block
support of 50 m by 50 m by 5 m (Fig. 1). The tonnage curves reveal that, with
increasing cut-off grades, the tonnage associated with model dilution increases
(Jara et al. 2006) and that by accepting more dilution ore loss is minimised
(Bertinshaw and Lipton 2007). The reverse effect occurs if the cut-off grade is
reduced (lower dilution but higher ore loss).
Analysis of the tonnage associated with model dilution for varying block sizes
and a fixed 54 Fe% cut-off (Fig. 2) shows that dilution increases with decreasing
block sizes and that small blocks are subject to more model ore losses (Jara et al.
2006). Furthermore, the results shown provide an alternative way to reemphasise
the dangers of estimating into small blocks from sparse drilling data (Ravenscroft
and Armstrong 1985).
Iron grades Fe(%) from an Iron Ore deposit in the Pilbara region in Australia are
used to demonstrate the assessment of the model dilution and ore loss using the
method presented above.
The drill hole data are samples from Mineral Resource definition RC drill holes
within a mineralised domain with a nominal drill spacing of 25 m by 25 m which
are used to carry out a kriging estimation of Fe into a grade control block model
with blocks of dimension 12 m by 12 m by 6 m. Further to this, available Fe grades
from grade control blastholes at a nominal drill spacing of 6 m by 6 m are kriged
into the same grade control block model. These kriged grades are considered as the
true block Fe grades to be the benchmark in this practical application. Since the
blasthole data partially covers the domain under study, the comparison is limited to
blocks that are informed by both the RC and blastholes kriged estimates. In
addition, it is assumed to be no bias in either the drill hole or blasthole. The scatter
Resource Model Dilution and Ore Loss: A Change of Support Approach 351
45
40 Model dilution
30
Tonnage (%)
25
20
15
10
0
50 51 52 53 54 55 56 57 58
Fe (%)
Fig. 1 Tonnage associated with model dilution (red) and ore loss (black) as function of Fe%
cut-off grades
plot between both estimates (Fig. 3) indicates that the model will incur in misclassi-
fication of blocks when reporting at the cut-off of 60 Fe%, which corresponds to
approximately the average Fe grade.
Using the RC drilling data, the discrete Gaussian method is used for modelling
the distribution of Z(v) and Z*(v). Some close-space RC data available was used to
fine-tune the values of Var(Z*(v)) and Cov(Z(v), Z*(v)) required to derive the
correction factors s and ρ for modelling the metal Q*(z), tonnage T*(z) and average
grade above cut-off M*(z) of Z*(v).
The amount of metal Q*(z) above cut-off is given by
X
Q* ðzÞ ¼ E Z* ðvÞ1Z* ðvÞz ¼ ϕ sn U 0, n y*
n n
ð21Þ
and T * ðzÞ ¼ P Z* ðvÞ z ¼ 1 F y* which allows to compute the average
grade above cut-off M* ðzÞ ¼ Q* ðzÞ=T * ðzÞ.
The modelled tonnage T*(z) and average grade M*(z) are compared to the
corresponding estimates reported from the grade control block model to assess
the adequateness of the distribution of Z*(v) to represent the distribution of the
kriging estimates obtained from the RC drill hole data (Fig. 4). The average grade
above cut-off is closely reproduced for all cut-offs with a maximum absolute
relative error of 0.75 % at 60 Fe% cut-off. Tonnage is also reasonably reproduced
for all cut-offs with the exception being the tonnage at 60 Fe% where the absolute
relative error is approximately 20 %. Nonetheless, the global average relative error
352 O. Rondon
27
26
25
Tonnage (%)
24
23
Model dilution
22
Model ore loss
21
20
5
5
0x
0x
0x
0x
0x
0x
x8
x5
x4
x3
x2
x1
80
50
40
30
20
10
Block size
Fig. 2 Tonnage associated with model dilution (red) and ore loss (black) for different block sizes
using drill hole data with a nominal drill spacing of 100 m by 100 m and a fixed 54 Fe% grade
cut-off
5 Conclusion
This study developed analytical equations to assess the proportion and average
grade of blocks misclassified as ore and waste and presented a case study with their
application to a grade control model from an Iron Ore deposit in the Pilbara region
of Australia. The method proposed provides only global estimates of the expected
tonnage and average grade associated with these types of misclassification of
Resource Model Dilution and Ore Loss: A Change of Support Approach 353
68
66
Fe (%) estimates from blast hole data
64
62
60
58
56
54
52
50
50 52 54 56 58 60 62 64 66 68
Fe (%) estimates from RC data
Fig. 3 Comparison of Fe kriged grades obtained from RC and blasthole data at a nominal drill
spacing of 25 m by 25 m and 6 m by 6 m, respectively. Red lines indicate the 60 Fe% cut-off. Note
that the top left sector of model ore loss has many more blocks than the lower right sector of model
dilution
100 66.00
90
80 65.00
50
40 63.00
30
20 62.00
10
0 61.00
50 52 54 56 58 60 62 64
Fe (%)
Fig. 4 Comparison of modelled tonnage and average grades above cut-off (red) to corresponding
estimates from the grade control model (GCM) kriged grades using the RC drill hole data (black)
354 O. Rondon
Table 1 Comparison of true and estimated model dilution and ore loss at a 60 Fe% cut-off
Misclassification type Tonnage (%) Grade Fe%
Model dilution True 6.16 58.40
Estimated 7.38 58.60
Model ore loss True 29.85 61.46
Estimated 28.36 61.18
Table 2 Comparison of true and estimated model dilution and ore loss at a 62 Fe% cut-off
Misclassification type Tonnage (%) Grade Fe%
Model dilution True 11.15 60.77
Estimated 12.6 61.04
Model ore loss True 20.96 63.85
Estimated 21.89 62.82
blocks. Therefore, it is not possible to know the spatial location of the blocks that
are potentially misclassified when selecting on the basis of an estimated grade.
Nonetheless, these global estimates could be used to anticipate the outcome of such
selection. More advanced approaches based on conditional simulations are required
to localise the blocks and classify them as ore or waste (Verly 2005) (Deutsch et al.
2000), but these approaches involve more time and effort.
The equations proposed in this paper are derived under the framework of the
discrete Gaussian method for change of support and modelling of the information
effect. Therefore, their practical application is largely limited by the assumptions
made in both of these techniques as well as the ability of the theoretical distribution
of the estimates to resemble the distribution of the Resource Model estimates.
Acknowledgements The author wishes to thank Mark Murphy for his help during the preparation
of the manuscript.
Appendix
1 X X
EðZðvÞ1ZðvÞ<z =Z* ðvÞ < zÞ ¼ ϕ r n
ρk ½δnk U n, k ðyÞ½δ0k U0, k ðy* Þ
Fðy* Þ n n k
P Z ðvÞ z=Z * ðvÞ z ¼ H ρ y; y* = 1 F y*
1 X X
E ZðvÞ1ZðvÞz = Z* ðvÞ z ¼ ϕ r n
ρk U n, k ðyÞU 0, k y*
1 Fð y * Þ n n k
Bibliography
Bertinshaw R, Lipton I (2007) Estimating mining factors (dilution and ore loss) in open pit mines,
vol 6, Sixth Large Open Pit Mining Conference. AusIMM, Perth, pp 13–17
Chiles J-P (2014) Validity range of the discrete Gaussian change of support and its variants. J
South Afr Inst Min Metall 114(3):231–235
Chiles J-P, Delfiner P (2008) Geostatistics. Modeling spatial uncertainty. Wiley, New York
Deutsch C, Magri E, Norrena K (2000) Optimal grade control using geostatistics and economics:
methodology and examples. Soc Min Metall Explor 308:43–52
Emery X (2007) On some consistency conditions for geostatistical change of support models.
Math Geol 39(2):205–223
Huijbregts C (1976) Selection and grade-tonnage relationships. Advanced geostatistics in the
mining industry, 24. D. Reidel Publishing Company, pp 113–135
Jara R, Couble A, Emery X, Magri E, Ortiz J (2006) Block size selection and its impact on open-pit
design and mine planning. S Afr Inst Min Metall 106:205–211
Lantuejoul C (1988) On the importance of choosing a change of support model for global reserves
estimation. Math Geol 20(8):1001–1019
Matheron G (1976) Forecasting block grade distributions: the transfer functions. In: David M,
Guarasico M (eds) Advanced geostatistics in the mining industry, vol 24. D. Reidel Publishing
Company, Dordrecht, pp 239–251
Ravenscroft P, Armstrong M (1985) Kriging of block models, the dangers reemphasised, vol
2. APCOM, Berlin, pp 577–587
Rivoirard J (1994) Introduction to disjuntive kriging and non-linear geostatistics. Oxford Press,
New York
Roth C, Deraisme J (2000) The information effect and estimating recoverable reserves. In:
Kleingeld W, Krige D (eds) Geostatistics 2000 proceedings of the sixth international
geostatistics congress. Geostatistical Association of Southern Africa, Cape Town, pp 776–787
Verly G (2005) Grade control classification of ore and waste: a critical review of estimation and
simulation based procedures. Math Geol 37(5):451–475
Drill Holes and Blastholes
1 Introduction
2 Formal Link
The initial data (Fig. 1) are from an open-pit copper mine in Northern Chile where a
subdomain was chosen for analysis because it is almost homogeneously covered by
around 3000 drill hole samples (3 m long) and 13,000 blasthole samples (15 m
long). In this case study, the diameters of the drill holes and blastholes are
considered to be the same.
In a previous paper (Séguret 2015), the author showed that if we omit the
problem of the nugget effect, both blast- and drill holes can be considered a
regularization of the same phenomenon in accordance with their respective sup-
ports. But the drill holes have their own errors, independent of the blast ones, so that
they do not share the nugget effect and finally, we have
with
Drill Holes and Blastholes 359
Y(x,y,z), the point grade assumed to be isotropic and without any measurement
error;
“*” denotes a convolution product;
Zþ1
Y ðx; y; zÞ*p15m ðzÞ ¼ Y ðx; y; uÞp15m ðz uÞdu;
1
p15m ðzÞ ¼ 15
1
1½0;15 ðjzjÞ, the convolution function;
2
1½0;15 ðjzjÞ the indicator function equal to 0 outside the interval [7.5 m, 7.5 m] and
2
1 inside it;
R(x,y,z), a “white noise” residual, statistically and spatially independent of Y(x,y,z)
and representing the blast error
The variogram of Yblast(x,y,z) becomes
with
γ R(h), the nugget effect due to the blast error, with the variance σ 2R ;
γ 15m ðhÞ ¼ ðγ ∗ P15m ÞðhÞ ðγ ∗ P15m Þð0Þ, regularized variogram expressed as a con-
volution product;
γ, the point variogram, assumed to be isotropic;
P15m ðhÞ ¼ ðp15m *p_ 15m ÞðhÞ ¼ 1512 ðjhj þ 15Þ1½0;15 ðjhjÞ, function with “P” (upper
case) that regularizes the variogram, expressed as an auto-convolution product
of p15m (lower case) as previously defined by itself.
Similar equations can be established for the drill holes and the 3 m support. The
model assumes that the blast and the drill samples have the same average because
the independent residuals are of zero mean.
360 S.A. Séguret and S. De La Fuente
From Eqs. (1) and (2), some linear systems can be deduced; in the following we
propose three of them that are tested on a conditional simulation.
One can remove the blast error by “factorial kriging” estimation (Matheron 1982),
using a linear system applicable to each blast measurement and a local neighbor-
hood of surrounding blast samples. The system is presented symbolically by matrix
formalism:
γ 15m þ γ R 1 λ γ 15m þ σ 2R
¼ ð3Þ
1 0 μ 1
In this system, γR disappears from the second member of the linear system and is
replaced by σ 2R , the value of the nugget effect. Thus, we remove, from the estima-
tion, the part associated with the measurement error. This does not mean that there
is no nugget effect in the remaining part γ15m; it means that only the “natural” part
remains. In our case, the complete nugget effect has to be removed because blasts
and drills do not share any microstructure.
The result of the estimation is the average value of the grade over the blast
support at blast sample locations with no measurement error.
The difference with the previous system is that in the second member, γ 15m ¼ γ*
P15m ðγ*P15m Þð0Þ (upper case P) is replaced by ðγ*p15m Þ ðγ*p15m Þð0Þ (lower
case p). Initially developed to improve microscopy images of thin plates in the
petroleum industry (Séguret1988; Le Loch 1990).
Drill Holes and Blastholes 361
Finally, one can imagine renewing the mine planning block model locally by using
blast and drill samples together in a cokriging system with a linked mean (same
average for both measurements, Chilès and Delfiner 2012):
0 10 1 0 1
γ 3m γ 3m, 15m 1 λ γ*p3m *pV ðγ*p3m *pV Þð0Þ
@ γ 3m, 15m γ 15m þ γ R 1 A@ λ0 A ¼ @ γ*p15m *pV ðγ*p3m *pV Þð0Þ þ σ 2R A
1 1 0 μ 1
ð5Þ
As usual, the matrix on the left in the linear system concerns only the data; here, a
set of drill and blast measurements. As a consequence, it is composed on the
diagonal of submatrices where the variogram of the drill holes (regularization
over 3 m) and the variogram of the blastholes (regularization over 15 m) appear.
The cross submatrix, which concerns the link between blast- and drill holes, is
based on a regularization of the point-support variogram over both supports, which
is the reason why γ 3m, 15m ¼ γ*p3m *p15m ðγ*p3m *p15m Þð0Þ intervenes. Compared
to a usual cokriging system that one can find in the literature (Wackernagel 2003),
the linked mean constraints reduce to one line and one column, the submatrix
associated with the non-bias constraints. Without this simplification, there would
be two lines and two columns. In practice, the consequence is important because
here the drill and blast measurements play the same role, while in a normal
cokriging system, one of the measurement types would be considered auxiliary,
thus reducing its relative influence on the final result. In the right-hand vector, the
term pV appears because the objective is to estimate the average grade over a
v-sized block. This convolution is combined with previous regularizations.
4 Simulation
Fig. 2 Blast- and drill holes grid nodes involved in the simulation
For simplification, the drill samples have no errors. We add to the blast samples a
random noise with a nugget effect of 0.02, representing the blast sampling error.
The grades are realistic with a 0 minimum, 3.5 % maximum; an average of 0.63 %
as in the real deposit and the distribution is correctly skewed to the right. We verify
that the sills of the drill, blast, and block variograms obey the laws of regularization
by following the procedure presented in Séguret (2015) which is based on the charts
by Journel and Huijbregts (2003) pages 125–147. Figure 3 shows the result.
In the first test, we propose to remove the blast error by using the linear system (3).
This filter can be applied to every blast measurement, using a local neighborhood of
surrounding blast samples. The neighborhood must contain the sample from which
the noise is to be removed; otherwise, the filtering is not efficient.
For comparison, the estimation is made by kriging with no filtering, using the
same neighborhood (but without the target sample, otherwise the kriging will
obviously give back the value of the data point).
We select, among all the simulated blasts, a subset of around 1000 samples on
which estimations will be conducted, using the additional samples in case of
ordinary kriging and all the samples in case of nugget filtering.
Drill Holes and Blastholes 363
Fig. 3 From left to right, properties of point-support simulated grades, 3 m support, 15 m and
blocks. Upper figures are experimental histograms, bottom variograms
The reference is the “truth,” i.e., the blast without errors which we know because
we work on a simulation where everything is known. Figure 4 shows the results.
On these scatter diagrams, the horizontal axis represents the true blast value
without any sampling error. On the left-hand scatter diagram, the vertical axis is a
usual kriging. The correlation with the truth is 0.65. On the right-hand one, when
the filtering is activated, the correlation increases to 0.896. Why?
With filtering, the kriging neighborhood can incorporate the target point where
the filter is applied. This point takes a high kriging weight (more than 65 %).
Although noisy, this point is closer to the truth than any average based on sur-
rounding points which explains why the filter estimate is closer to the truth.
Finally, the advantage of this linear system (3) is to enable the kriging neigh-
borhood to incorporate the target point information.
6 Deconvolution by Kriging
Now we propose a second test: removing the effect of regularization on the blast
with a kriging system that estimates a “point” value for each blast measurement,
while simultaneously removing the part of the nugget effect associated with blast
errors. By this procedure, we expect to restore the initial variability of the point-
support value judged to be too strongly smoothed by the regularization. The linear
system used is (4).
364 S.A. Séguret and S. De La Fuente
Fig. 4 Scatter diagrams between the true values (horizontal axis) and estimations (vertical axis).
Left figure, ordinary kriging estimation; right figure, estimation where the blast error is removed by
kriging
A comparison is made with the true point value and with previous estimates
(estimating a blast with or without nugget effect). Figure 5 shows the results.
In the three scatter diagrams, the horizontal axis represents the true point-support
value. The upper left-hand scatter figure presents the result when a deconvolution is
made jointly with error filtering. The correlation with the true value is good at 0.85.
The upper right-hand figure presents the result of the error removal with no
deconvolution. It corresponds to the previously presented system but this time as
compared with the point-support value; this is the reason why the correlation is 0.78
and not 0.89 when compared to the blast values. In comparison with the left-hand
figure, the deconvolution increases significantly the accuracy of the estimation.
The bottom figure shows the results when neither filtering nor deconvolution is
done. The correlation is very low; it is 0.55.
As for nugget filtering, the deconvolution is efficient because the linear system
authorizes the use of the target points where the filter is applied.
The third test is made to renew the mine planning block model locally by using blast
and drill samples together. The system used is (5), a cokriging system with linked
mean because drill and blast samples have the same average, which is mandatory
for carrying out all these calculations.
Our objective is to estimate the average grade at the block scale, and we compare
it with two other systems: block grade estimate by kriging using only drill holes and
block grade estimate by kriging using only blastholes. Figure 6 shows the results.
Drill Holes and Blastholes 365
Fig. 5 Scatter diagrams between the true values (horizontal axis) and estimations (vertical axis).
Upper left figure, deconvolution together with error filtering; upper right figure, error removal
without deconvolution; bottom figure, no deconvolution, no error filtering
For the three previous scatter diagrams, the horizontal axis is the true block
grade. The upper left-hand diagram is the result obtained by ordinary kriging with
drill samples; the upper right-hand diagram is the result obtained with blast
samples. The jump by the correlation coefficient from 0.38 (OK using drills) to
0.88 (OK using blasts) is impressive. Even if the blast samples are regularized over
15 m, the fact that they are more numerous and respect the variogram (up to a
nugget effect) justifies their use when possible, in selections for mining operations,
instead of the drill samples.
The bottom diagram concerns cokriging using blast and drill samples together.
The performance is similar to ordinary kriging with blast samples only. In our case
cokriging is not useful because the blasts are so numerous that adding a drill
contribution does not improve the results. This does not mean that such a system
is not helpful, for example, in short-term planning to evaluate a domain to be
blasted where there are only drill holes.
366 S.A. Séguret and S. De La Fuente
Fig. 6 Scatter diagrams between the true values (horizontal axis) and estimations (vertical axis).
Upper left figure, estimation is kriging using drill holes; upper right figure, kriging using
blastholes; bottom figure, kriging using both blast and drill samples together
In our case, short-term planning is based on averages using the blast samples
included in the block, so the question is whether kriging can produce an improve-
ment. We are comparing three experiments:
• Ordinary kriging with 24 surrounding blast measurements (previous work)
• Moving average with the same 24 surrounding blast measurements
• Moving average with four blast measurements at the same elevation
Figure 7 shows the results.
Replacing ordinary kriging by an average reduces the correlation with the truth
from 0.88 to 0.74. This is a very large reduction which should encourage the
practitioners to use kriging instead of present practices in the company.
If practitioners do not want to change their habits, one can see that with only
4 points, the result is better than when 24 points are used because the smoothing is
weaker: the correlation with the truth increases from 0.74 to 0.84, a result still
Drill Holes and Blastholes 367
Fig. 7 Scatter diagrams between the true values (horizontal axis) and estimations (vertical axis).
Upper left, ordinary kriging with 24 surrounding blast measurements; upper right, moving average
with the same 24 surrounding blast measurements; bottom, moving average with 4 blast measure-
ments of the same elevation
below the one obtained with kriging but very close to it. Does this mean that we
recommend a moving average with only 4 points? NO! To use so few points is risky
for reasons of conditional bias. To illustrate this concept, consider again the
previous scatter diagram but this time, with the true block grade for the vertical
axis and the estimate on the horizontal axis (Fig. 8).
In the previous figures, red continuous curves represent the mathematical expec-
tations of the true values conditioned by different estimates. We focus on the most
representative [0.3 %, 1 %] range of grades.
When kriging is done with 24 points, the conditional expectation curve is close
to the first diagonal. Thus, when we select the block according to its estimation, we
obtain, on average, what we expect, with perhaps a slight tendency to underestimate
the high grades.
When we replace kriging by a moving average using 24 points, the red curve is
still close to the diagonal, with a slight tendency to overestimate the low grades and
underestimate the high ones.
368 S.A. Séguret and S. De La Fuente
Fig. 8 Scatter diagrams between the true values (vertical axis) and estimations (horizontal axis).
Upper left: the horizontal axis represents ordinary kriging with 24 surrounding blast measurements
(previous work); upper right: the horizontal axis represents a moving average with the same
24 surrounding blast measurements; bottom: the horizontal axis represents a moving average with
4 blast measurements at the same elevation
When the moving average is done with only four points, the conditional bias
appears clearly: in the range of the low grades, we systematically underestimate the
average grade of the blocks and may decide to classify as “waste” blocks that are in
reality richer than expected. Conversely, in the range of the high grades, this
moving average with only four points systematically overestimates the average
grade of the block so that we classify as “rich” blocks which must be considered
“waste.”
It is for this reason that one must use enough points in the kriging neighborhood
for grade control, and reflect on the reason why kriging and geostatistics were
created (Matheron 1971).
Drill Holes and Blastholes 369
9 Conclusion
The study of a porphyry copper deposit showed a formal link between blast- and
drill holes, leading to numerous linear systems able at least to filter blast errors,
make blast deconvolution, or build a block model using blasts and drills together,
techniques that could be used at different stages of the mining process.
Tested on a realistic simulation, these systems have proved their worth, as well
as the danger of replacing kriging by a moving average, especially with few points,
producing a strong conditional bias, and this is a useful reminder of the reason why
kriging was created.
Overall, a formal comparison between blast- and drill holes shows that in this
mine – and more generally, in this company – the quality of the blast values is as
good as the quality of the drill values, contrary to conventional wisdom.
Bibliography
Abstract Classical metal tonnage cut-off curves are an important output of min-
eral deposit evaluation. However, such curves omit local morphological variations
such as those found in sulphide stockwork-type deposits, in which sulphides and
host rock exhibit variable proportions. This paper presents a methodology for
building a stochastic model of the massive and stockwork Zambujal ore deposit
with respect to both morphology and metal content. The model allows a metal
tonnage cut-off surface to be constructed conditional both to metal content and to
local proportions of sulphides relative to host rock. Two random variables are
modelled using stochastic simulation: (i) a variable (P(x)) that represents the
local proportion of sulphides within each mining block and (ii) an auxiliary variable
(Y(x)) representing the relative contents of metal, that is, the metal contents
recalculated for the matrix sulphides. The results are validated by comparing the
global quantities obtained using ordinary kriging of the effective grades with those
obtained by modelling the variables P(x) and Y(x).
D. Silva (*)
Earth Sciences Department, Faculty of Sciences and Technology, Universidade NOVA de
Lisboa, Campus da Caparica, 2829-516 Caparica, Portugal
e-mail: [email protected]
J.A. Almeida
Earth Sciences Department and GeoBioTec, Faculty of Sciences and Technology,
Universidade NOVA de Lisboa, Campus da Caparica, 2829-516 Caparica, Portugal
e-mail: [email protected]
1 Introduction
stockwork ores than in the massive ores, but this issue increases the relevance of the
stockwork ores.
Therefore, a high-resolution morphological model was built to quantify the local
proportion of sulphides in each block of the block model (the variable P(x)),
followed by a model of relative grades (the variable Y(x)). By doing so, the metal
contents are recalculated for the sulphide fraction rather than modelling the total
metal content in the total volume. For instance, for the sample location xi with a
proportion of sulphide minerals P(xi) and effective grade of copper ZCu(xi), the
relative grade of copper YCu(xi) is expressed by:
Both P(x) and YCu(x) are regionalised variables and can be modelled by using
geostatistical tools such as variograms, kriging, and stochastic simulation. The total
amounts of metals and ores can be calculated directly from the estimated or
simulated values of these variables. Also, with models of both these variables,
parametric surfaces relating the amount of copper to both the proportion of sul-
phides (morphological constraint) and the grade of metal (grade constraint) are
readily constructed. The ability to construct such surfaces in a straightforward way
is an important objective and outcome of this paper.
The proposed methodology contains the following steps:
1. Build a two-region 3D solid model of the Zambujal ore deposit that encompasses
the boundaries of the deposit and the transition between the massive ores in the
top part (RM) and the stockwork ores in the bottom part (RS). Convert the solid
model into a voxel model (mining blocks).
2. Using the densities and metal grades measured at the boreholes and knowledge
about the paragenesis of the Zambujal deposit, evaluate optimal solutions for the
sulphide proportion random variable P(x). In the present case, ten solutions were
selected for each sample.
3. Estimate local cumulative distribution functions (cdfs) of the random variable P
(x) for both region RM and region RS by using indicator kriging (IK). Build a
high-resolution morphological model of P(x) using Direct Sequential Simulation
(DSS) (Soares 2001) conditional to the computed local cdfs and thus conditional
to the low-resolution model.
4. Compute YCu(x) at the sample locations. Estimate local cdfs of YCu for both
region RM and region RS using IK. Simulate the relative grades using DSS for
regions RM and RS simultaneously.
5. Analyse the results involving an evaluation of ore and metal tonnages, and a
comparison of the tonnages with those obtained using ordinary kriging (OK).
Validate the variograms of the simulated images and histograms. Map the
uncertainty.
6. Build a global surface function of copper quantity conditional to the copper
grades and the local proportions of sulphides relative to host rock.
374 D. Silva and J.A. Almeida
It is important to note that variables P(x) and copper grades (both the relative and
effective grades) are not stationary within the entire deposit; rather, they are
constrained to the massive and stockwork ores. For this reason, it is essential to
delimit the two regions (massive and stockwork ores) and to constrain the simula-
tions with the cdfs of the studied variables. A modified version of DSS in which the
simulated values are resampled from the local distributions with local cdfs is used to
simulate the entire deposit (Charifo et al. 2013; Matias et al. 2015; Roxo et al.
2016).
2 Case Study
The upper and lower limits of the massive and stockwork regions were digitised
according to the borehole data and expert information in several parallel cross
sections. Then, three surfaces were constructed using linear interpolation: the top of
the massive ore, the bottom of the massive ore/top of the stockwork ore, and the
bottom of the stockwork ore. The final step for building the low-resolution mor-
phological model was the conversion of the surfaces into a block model (block size
2 2 2 m) of the two major ore types (Fig. 1).
Before simulations were performed, the variable P(x) was evaluated for all samples
by combining the measured values of density D(x) and metal grade Z(x).
The evaluation of P(x) for core samples was made using a classification proce-
dure because this variable is not usually measured in the laboratory. To make this
estimate at the sample locations, the measured levels of the most abundant chemical
elements (Fe, Cu, Zn, Sn, Pb, and S) were used, as well as the densities of the
sampled cores and the paragenesis of the deposit (pyrite, chalcopyrite, galena,
blend, or host rock). It is important to emphasise the difference between the density
of the host rock (about 2.88 t/m3) and the densities of the minerals listed (all >4 t/m3).
A lookup table listing all admissible combinations of these minerals and the host
rock was made (with a resolution of 1 % for proportions), and, for each combina-
tion, the theoretical values of the element grades (based on the chemical formulae)
and the composite density were computed. For example, considering a sample
comprising 30 % host rock, 30 % pyrite, and 40 % chalcopyrite (P(x) ¼ 0.7), the
theoretical key element grades and density should be Fe ¼ 26.14 %, S ¼ 30.01 %,
Cu ¼ 13.85 %, and D ¼ 4.10 t/m3. Then, each laboratory-measured value of grade
Building a Tonnage-Surface Function for Metal Grades and Geological Dilution. . . 375
900m
700m
0m
500m
0m
300m
Massive ores 350m
Stockwork ores
Fig. 1 Binary morphological model of the two regions RM (massive ores) and RS (stockwork ores)
in a voxel structure
and density was compared with each item of the lookup table, and a set of the closest
combination of minerals plus host rock was extracted. Ten theoretical mixtures of
minerals plus host rock were considered sufficient to cover the uncertainty, and the
selected measurements were in the main very close to each other. Each selected set
of ten values constitutes the local cdf of P(x) at the locations of the boreholes. These
cdfs were then extended to the entire deposit using DSS (Soares 2001). The model of
P(x) is a detailed morphological model and must be used, in a second step, to
condition the construction of the model of metal grades.
In summary, at this point, a cdf of P(x) is constructed for each sample. Figure 2
shows a 3D view of a single realisation of P(x) at the locations of the boreholes
converted to the grid of blocks. In the next section, P(x) is simulated for the entire
deposit.
The proportion of sulphide ore minerals P(x) was simulated using DSS for the two
regions (RM and RS) simultaneously (Ruben and Almeida 2010). Instead of using a
global cdf, the simulation of P(x) was performed using a modified version of DSS
that uses local cdfs for each region, P(x)|RM and P(x)|RS (Charifo et al. 2013). For
declustering purposes, P(x)|RM and P(x)|RS were estimated using IK of 20 P(x)
376 D. Silva and J.A. Almeida
Fig. 2 3D view of a realisation of P(x) at the locations of the boreholes converted to the grid of
blocks. Red colours represent higher values of P(x) (massive ores), and blue-based colours refer to
lower values of P(x) (stockwork ores)
classes. Thirty realisations of P(x) were generated. Figure 3 shows the average
image of P(x) (top diagram) and the 10 % of blocks with lowest uncertainty (bottom
diagram).
Higher values of P(x) are located at the top of the structure, and lower values at
the bottom (see Figs. 1 and 3), and the transition between the two regions is sharp.
Histograms are well reproduced by DSS, as are the variograms of the data (Silva
2015).
For simulating metal grades, the proportion of ore minerals (P(x)) was combined
with borehole grades to evaluate the relative metal grades (Y(x)). For instance, if a
sample had a grade of 2 % copper and a value of P(x) ¼ 0.7, then the relative grade
of copper would be equal to 2 %/0.7 ¼ 2.86 %. The simulation of Y(x) follows the
same approach as taken for P(x), namely, the use of IK to estimate conditional cdf
histograms of Y(x)|RM and Y(x)|RS and DSS with local histograms. Three
realisations of grades were generated conditional to each of the 30 realisations of
P(x), giving a total of 90 realisations for YCu(x). Again, histograms are well
reproduced by DSS, as are the variograms of the data (Silva 2015).
Figure 4 shows the average relative copper grades YCu(x) (top diagram) and the
actual copper grades (bottom diagram).
It is important to note that the values obtained for the relative grades of copper
and for the actual grades obtained through the quotient of the relative grades and
P(x) are theoretically valid. This theoretical validity is a good indicator that the
proposed methodology can be applied; for example, a maximum of 32 % for
relative copper grades is plausible. Figure 4 also shows that the relative grades in
the stockwork ores are higher than those in the massive ores. The transition of
Building a Tonnage-Surface Function for Metal Grades and Geological Dilution. . . 377
Fig. 3 3D representation of (top) the average values of the sulphide proportion variable P(x);
(bottom) the 10 % of mining blocks with the lowest uncertainty (in blue), also showing boreholes
(lines) and the bottom surface of the stockwork ores (in grey)
grades between regions is smooth, which is in accordance with the variability of the
grades along the boreholes.
Finally, Table 1 compares the average of grades obtained using OK and DSS for
the massive ore region, the stockwork ore region, and the regions combined.
Deviations are minimal, confirming that DSS with local histograms generates
unbiased realisations.
378 D. Silva and J.A. Almeida
Fig. 4 (top) 3D representation of average relative copper grades (%). Higher relative grades are in
red, lower values in blue; (bottom) 3D representation of average copper grades. Higher grades are
in red, lower values in blue
Using the simulated values of the variables P(x) and YCu(x), average estimates for
[P(xi)]*, [YCu(xi)]* and ½Z Cu ðxi Þ* ¼ ½Pðxi Þ* :½Y Cu ðxi Þ* were computed for each
mine block xi. The tonnage of copper (TCu) within each mine block xi with volume
V and ore density ρ is given by
Building a Tonnage-Surface Function for Metal Grades and Geological Dilution. . . 379
Table 1 The proportion of sulphide minerals and the relative grades of Cu for massive (RM) and
stockwork (RS) ores
RM RS RM + RS
Morphological model [P(x)] OK 0.755 0.184 0.350
DSS 0.768 0.176 0.348
% deviation 1.6 4.7 0.7
Copper [YCu(x)]OK 1.465 3.289 2.763
[YCu(x)]DSS 1.482 3.339 2.795
% deviation 1.1 1.5 1.1
3 Final Remarks
The proposed methodology highlights the importance of the stockwork ores in the
Zambujal deposit, and as mineral concentrations are more efficient in stockwork
ores, metal cut-offs should take into account the proportion of sulphides relative to
host rock. Modelling the proportion variable and the relative metal grades adds
detailed information to each mining block and enables metal tonnage cut-off
surfaces combining metal grades and the proportion of sulphides relative to host
rock (geological dilution) to be generated in a straightforward way.
Acknowledgements The authors are grateful to Lundin Mining for providing the data, informa-
tion, and assistance required for the present work and to Midland Valley for providing an academic
licence for Move® software. This work is a contribution to Project UID/GEO/04035/2013 funded
by FCT-Fundaç~ao para a Ciência e a Tecnologia, in Portugal.
380 D. Silva and J.A. Almeida
P(x)
1.0
0.5
0.0
0% 1.6% 5% 10%
Cut-off grade of Cu (%)
Fig. 5 Copper resources as a function of copper grade and the proportion of sulphides P(x). The
amount of copper increases from blue to red
Bibliography
Almeida JA (2010) Modelling of cement raw material compositional indices with direct sequential
cosimulation. Eng Geol 114(1):26–33
Charifo G, Almeida JA, Ferreira A (2013) Managing borehole samples of unequal lengths to
construct a high-resolution mining model of mineral grades zoned by geological units. J
Geochem Explor 132:209–223
Gerst MD (2008) Revisiting the cumulative grade–tonnage relationship for major copper ore
types. Econ Geol 103:615–628
Goovaerts P (1997) Geostatistics for natural resources evaluation. Oxford University Press,
New York
Matias FV, Almeida JA, Chichorro M (2015) A multistep methodology for building a stochastic
model of gold grades in the disseminated and complex deposit of Casas Novas in Alentejo,
Southern Portugal. Resour Geol 65(4):361–374
Nunes R, Almeida JA (2010) Parallelization of sequential Gaussian, indicator and direct simula-
tion algorithms. Comput Geosci 36(8):1042–1052
Relvas J, Jorge R, Pinto A, Matos J, Rosa C, Barriga F (2002) The neves–corvo deposit, Iberian
Pyrite Belt, Portugal. In: Impacts and future, 25 years after the discovery, vol. Special
Publication 9. Society of Economic Geologists, pp 155–176
Rossi ME, Deutsch CV (2014) Mineral resource estimation. Springer, Dordrecht (Vol. Science +
Business Media)
Roxo S, Almeida JA, Matias FV, Mata-Lima H, Barbosa S (2016) The use of sensory perception
indicators for improving the characterization and modelling of total petroleum hydrocarbon
(TPH) grade in soils. Environ Monit Assess 188(3):1–19
Building a Tonnage-Surface Function for Metal Grades and Geological Dilution. . . 381
Abstract The main objective of this work is to develop a model of the morphology
and coal-quality variables of a sub-region of the Ncondezi coal deposit, Mozam-
bique, with data supplied by the Ncondezi Coal Company. This coal deposit is
characterized by a complex stratigraphy, with fine coal layers interleaved with
layers of shales and sandstones. The morphological model consists of the spatial
characterization of the random variable P(x) (proportion of coal). To model the
coal-quality variables and because of the large number of variables and given that
some pairs of variables exhibit high correlations, an approach involving simulation
of principal components as secondary variables is used. A set of selected principal
components are simulated using direct sequential simulation, and the coal-quality
variables are then co-simulated conditional to the previously simulated principal
component values. This approach works as a global co-regionalization model. The
results of the morphological model and of the model of the coal-quality variables
allow global coal reserves to be calculated as well as the quantities of coal to be
parameterized as a function of the coal-quality variables.
1 Introduction
Modelling coal deposits usually involves a two-step approach: first, the creation of
a model of the morphology and second, the construction of a model of the
properties or coal-quality variables (Cornah et al. 2013; Srivastava 2013). Coal
resources are commonly estimated and classified based on drill-hole data, which are
often multivariate (Tercane et al. 2013). Kriging is the preferred geostatistical
method in the coal industry for estimating and assessing coal resources.
Regarding the construction of the morphological model, if coal seams have
considerable thicknesses, an indicator variable I(x) can be adopted and modelled
using geostatistical indicator kriging or indicator simulation (Almeida 2010a).
Where the coal seams are considerably thinner (only a few or tens of centimetres
thick), as in the present case study, it is not possible to use an indicator variable, but
it is possible to use a variable that quantifies the volumetric proportion of coal in
each block, P(x). This variable proportion of coal can be modelled by estimation
and/or simulation. However, for complex and heterogeneous deposits, kriging is
insufficient and should be complemented with stochastic simulation to provide
uncertainty assessment (Charifo et al. 2013; Cornah et al. 2013; Matias et al. 2015).
Of the various papers that have addressed the modelling of coal variables,
Tercan and Sohrabian (2013) used direct sequential simulation (DSS) (Soares
2001) to characterize the quality of a coal deposit in Turkey in a two-step approach.
In that work, principal component analysis (PCA) (Davis 1986) is applied first to
the three coal-quality variables (ash, calorific value, and inherent moisture). DSS is
then used to generate simulated images of the principal components, after which the
simulated values of the major principal components are transformed into the values
of the initial variables. Although this approach enables a multivariate set of linearly
correlated variables to be simulated and the multivariate matrix of correlations to be
generated, it does not impose both the histograms and the variograms of each
variable. Also, anomalous combinations of results (not observed in the borehole
data) can be obtained.
This paper presents a methodology to build a model of the morphology and coal-
quality variables of a sub-region of the Ncondezi coal deposit, based on data from
surveys conducted by the Ncondezi Coal Company. The deposit is located in the
sub-basin of Moatize-Minjova, Tete Province, Mozambique (Fig. 1) (Johnson et al.
1996).
The Ncondezi deposit is a stratified coal deposit that was developed under the
influence of tectonic process that played a major role in determining the strati-
graphic sequence of the basin and the quality of the coal. The extensional regime
experienced by the basin during its formation led to a fragmentation of blocks in a
horst-graben system, which greatly influenced the lateral extent of the coal seams.
This fragmentation and the resultant low levels of lateral continuity of strata would
appear to be the primary reason for the difficulty of correlating coal seams between
boreholes. Even between relatively close boreholes separated by distances of less
than 300 m, coal-seam correlation is weak. Moreover, the tectonic instability of the
basin caused variations in the depositional environment, particularly fluctuations in
the groundwater level, which explain, for example, the high levels of ash in
Ncondezi coal.
This coal deposit comprises a heterogeneous package of thin coal lenses, each
measuring a few tens of centimetres or less in thickness, interbedded with non-coal
lithologies. The total thickness of the deposit, where not eroded, is several tens of
An Application of Direct Sequential Simulation and Co-simulation for. . . 385
Fig. 1 Location of the three Zambezi River Valley Basins (Adapted from Fernandes et al. 2015)
386 S.F. Sokhin et al.
metres. Lakshminarayana (2015) described this type of coal deposit and the prac-
tical implications for their extraction and referred to such deposits as “barcode”
deposits.
2 Methods
Geostatistical co-simulation
Geostatistical simulation of
Geostatistical simulation of of quality variables using
independent coal-quality
P(x)using DSS Co-DSS conditioned to PC
variables using DSS
uncertainty and the probability of occurrence of extreme values are tools of prime
importance for the planning and timing of resource exploitation with respect to
optimizing the proportion (mixture) of raw materials coming from different mining
stopes.
388 S.F. Sokhin et al.
3 Case Study
The results of the univariate and bivariate statistical analysis of the studied vari-
ables are presented in Tables 1 and 2, respectively. In Table 2, the highest
correlations (greater than 0.6 or less than 0.6) are depicted in bold. This matrix
shows that the Pearson and Spearman correlations between the variables are of
similar magnitude and that the variables IM and TS are independent of the
remaining variables.
The loadings of PCA and the projection of observed individual values onto the
three main principal components PC-1/PC-2 and PC-1/PC-3 are presented in Fig. 3
and show the following:
Table 2 Correlation coefficient matrix for Pearson (bottom) and Spearman (top) of the coal-
quality variables
IM Ash VM FC CV TS ARD
IM 0.186 0.275 0.116 0.217 0.033 0.197
Ash 0.100 0.717 0.878 0.976 0.194 0.898
VM 0.230 0.677 0.353 0.693 0.074 0.628
FC 0.054 0.873 0.238 0.877 0.261 0.818
CV 0.153 0.977 0.647 0.868 0.233 0.909
TS 0.029 0.173 0.067 0.187 0.201 0.240
ARD 0.109 0.883 0.574 0.789 0.881 0.196
Fig. 3 (Top) Loadings between the PCA components PC-1/PC-2 and PC-1/PC-3 and the initial
coal-quality variables. (Bottom) Projection of observed individual values onto the three main
principal components PC-1/PC-2 and PC-1/PC-3
390 S.F. Sokhin et al.
1. The initial seven variables can be summarized in three PCs that account for
87.5 % of the variation.
2. The variables Ash, FC, CV, and ARD are explained by PC-1, as is the variable
VM but to a lesser extent. The variables Ash and ARD lie on the positive axis of
PC-1 and the remaining variables on the negative axis. These results corroborate
the bivariate analysis, in which these five variables are correlated with each
other, some negatively.
3. The variables IM and TS are explained by PC-2 and PC-3, respectively, which
means that these variables have a very high degree of independence, both with
each other and with other variables. This result also confirms the results of the
bivariate analysis and also reflects that these two variables have a more asym-
metric distribution and are closer to lognormal type compared with the other five
variables.
4. The projection of observed individual values does not differentiate populations.
Some individual values are represented in dispersed patterns, which mean
abnormal records.
These results suggest that the set of four variables (Ash, CV, FC, and VM)
should be co-simulated with Co-DSS via the PC-1 axis and that the remaining
variables IM and TS should be simulated independently. ARD is redundant and
does not need to be simulated.
3.2 Variography
Fig. 4 Horizontal (left) and vertical (right) experimental variograms of (top) P(x), (middle) PC-1,
and (bottom) calorific values (CV), together with the respective models (blue lines) and sills (red
lines)
The uncertainty is zero at the borehole locations and increases away from
boreholes depending on the distance to the nearest borehole and the heterogeneity
of the closest values. It should be noted that the variograms show a relatively low
range (spatial continuity) in the study area, and the spacing between boreholes
should be tightened to improve the quality of the estimation for the blocks with the
highest uncertainties.
Figure 6 shows the results for the CV variable for section OX 18/63, and Fig. 7
shows in 3D the locations of the best coal (CV >12.5 MJ/kg) for the whole deposit.
To validate the results of the multivariate co-simulation, including the repro-
duction of the correlations between variables as observed in the initial data,
calculations of the average Pearson correlation coefficient between homologous
392 S.F. Sokhin et al.
Fig. 5 Representation of the results for P(x) in one section (OX 18/63): (a) one realization; (b)
another realization; (c) the average image of the simulated set of images of P(x); and (d) the image
of the local variance constructed with the 500 simulated images of P(x)
images (# 1/# 1, . . . # 105/# 105, and so on) were achieved by sampling. In all,
2,500 pairs of simulated images were selected from the set of all variables. The
results are shown in Table 3. Overall, the values of the correlation model are similar
to the observed data.
The results of the simulations enable the resource to be parameterized by coal
quality. For illustrative purposes, Fig. 8 presents the tonnages of coal conditional to
the variable CV. A density of 1.8 t/m3 for coal was used for the calculations.
An Application of Direct Sequential Simulation and Co-simulation for. . . 393
Fig. 6 Results for the coal-quality variable CV in one section (OX 18/63): (a) one realization of
PC-1; (b) one realization of CV, co-simulated to the realization of PC-1; (c) the average image of
the simulated set of images of CV; and (d) the image of the local variance constructed with the
500 simulated images of CV
4 Conclusions
Fig. 7 3D view of the locations of coal resource with CV values higher than 12.5 MJ/kg for the
whole deposit
Table 3 Correlation matrix of observed data (bottom) and simulated images (top) of coal-quality
variables
IM Ash VM FC CV TS
IM 0.005 0.005 0.005 0.005 0.001
Ash 0.100 0.598 0.865 0.972 0.010
VM 0.230 0.677 0.317 0.602 0.015
FC 0.054 0.873 0.238 0.864 0.013
CV 0.153 0.977 0.647 0.868 0.010
TS 0.029 0.173 0.067 0.187 0.201
Fig. 8 (Left) Curves of the tonnage of coal with respect to the coal-quality variable CV. (Right)
Moving average of CV calculated for a 1-m depth interval
An Application of Direct Sequential Simulation and Co-simulation for. . . 395
Acknowledgements The authors are grateful to Ncondezi Coal Company for providing the data
and information and to Midland Valley for providing an academic licence for Move® software.
This work is a contribution to Project UID/GEO/04035/2013 funded by FCT-Fundaç~ao para a
Ciência e a Tecnologia, in Portugal.
References
Johnson MR, Vuuren CJV, Hegenberger WF, Key R, Shoko U (1996) Stratig-raphy of the Karoo
Supergroup in Southern Africa: an overview. J Afr Earth Sci 1:3–15
Lakshminarayana G (2015) Geology of barcode type coking coal seams, Mecondezi sub-basin,
Moatize Coalfield, Mozambique. Int J Coal Geol 146:1–13
Matias FV, Almeida JA, Chichorro M (2015) A multistep methodology for building a stochastic
model of gold grades in the disseminated and complex deposit of Casas Novas in Alentejo,
Southern Portugal. Resour Geol 65(4):361–374
Nunes R, Almeida JA (2010) Parallelization of sequential Gaussian, indicator and direct simula-
tion algorithms. Comput Geosci 36(8):1042–1052
Rossi ME, Deutsch CV (2014) Mineral resource estimation. Springer, Berlin, 332 pp
Soares A (2001) Direct sequential simulation and cosimulation. Math Geol 33(8):911–926
Srivastava RM (2013) Geostatistics: a toolkit for data analysis, spatial prediction and risk
management in the coal industry. Int J Coal Geol 112:2–13
Tercan AE, Sohrabian B (2013) Multivariate geostatistical simulation of coal quality by indepen-
dent components. Int J Coal Geol 112:53–66
Tercan AE, Ünver B, Hindistan MA, Ertunç G, Atalay F, Ünal S, Kıllıo glu SY (2013) Seam
modelling and resource estimation in the coalfields of Western Anatolia. Int J Coal Geol
112:94–106
Castelo de Sonhos: Geostatistical
Quantification of the Potential Size of a
Paleoproterozoic Conglomerate-Hosted Gold
Deposit
Abstract Castelo de Sonhos, a gold deposit in Pará State, Brazil, has seen several
phases of exploration since the mid-1990s. These programs have provided drill hole
data, surface mapping of outcrops, geophysical surveys, geochemical surveys of
soil samples, and preliminary metallurgical test work. All available data from these
exploration programs have been integrated with recent advances in paleo-plate
reconstructions, in modeling sedimentary depositional systems, in geostatistical
simulation, and in data mining. This integration of ideas and methods from petro-
leum geostatistics, from classical statistics, and from plate tectonics makes it
possible to predict the range of the project’s potential tonnage and grade and to
assess the project’s upside and downside risk. This leads to an exploration target
range that is probabilistically quantified, that is well grounded in data, in field
observations and science, and that is testable through drilling. Not only does this
quantitative risk assessment improve analysis of the project’s technical and eco-
nomic viability but also, importantly, it builds confidence among investors whose
support is critical for advancing the project.
1 Introduction
Rising several hundred meters above the cattle-grazing lands of southern Pará State
in Brazil (Fig. 1) is a plateau cut by a ravine (Fig. 2). In the 1970s and 1980s,
garimpeiros (artisanal miners) worked the alluvial deposits on the flanks and the toe
of the plateau, recovering hundreds of thousands of ounces of gold using manual
methods to excavate and concentrate ore.
In 1995, Barrick Gold, reasoned that the likely source of the alluvial gold was a
hard-rock deposit on top of the plateau. They launched an exploration program,
initially searching for a granite-hosted gold deposit, but quickly recognizing that
the host was a band of conglomerates that rims the plateau. After a small drill
program, an extensive soil geochemistry survey, and a trenching program, Barrick
closed the exploration program in 1996 and relinquished the mineral concessions.
Garimpeiros then moved onto the plateau, digging trenches and pits by hand
near Barrick’s trenches, and near the up-dip projection of the best intervals in
Barrick’s drill holes. Over the next decade, garimpeiros excavated near-surface
mineralization along strike for several kilometers, to a depth of 12–15 m. In places,
tunnels were dug into the face near the base of trenches, following well-mineralized
bands 50–70 m into the wall (Fig. 3). The garimpeiro workings (garimpos) were not
limited by lack of gold but by the difficulty of extending hand-dug slots and tunnels
below the water table. The back-breaking and life-shortening work became phys-
ically impossible without access to explosives and mechanized equipment. By the
late 2000s, garimpeiro activity had dwindled to a few aging “hobby farmers” who
continued to be able to recover a few grams a day from the more prolific garimpos.
The past several years have been difficult for the mining industry. The price of
gold, which flirted with $1,800/oz in 2012, had dropped to barely $1,000/oz by late
2015. The collapse in commodity prices made many once-promising projects
unprofitable and caused funding for exploration to dry up. Mineral exploration is
Fig. 2 Perspective view of the plateau, facing northeast, with a 10:1 vertical exaggeration
a risky and costly business; with most mineral prospects never going into produc-
tion, exploration can be justified only when commodity prices are so strong that the
cost of the many misses is covered by the few hits.
By late 2015, the Castelo de Sonhos project was in jeopardy of being shut down a
second time. More than $1,000,000 was needed just to make the final land payments
that would secure TriStar 100 % ownership of the project, and much more was
needed to continue exploration. Even though the drilling continued to show prom-
ising potential, lack of funding limited the areal extent of TriStar’s drilling; by
2014, the last year of drilling, only 25 % of the 16 km of conglomerate outcrop and
soil anomaly had been drilled (Fig. 4).
Although there were good showings of gold throughout the 250–300 m thickness
of the conglomerate band, the drill holes targeted sweet spots, hunting for good
news that was necessary to feed a cash-starved project. Very few holes penetrated
400 R.M. Srivastava et al.
Fig. 4 Map showing the conglomerate outcrop (hatched), the gold anomaly in soil samples (red-
orange-yellow), the location of drill hole collars (white dots), and the footprint of the 2014
resource estimate (blue)
the entire thickness of the conglomerate, leaving untested resource potential even in
the limited areas that had been drilled.
Attracting more funding was made more difficult by a 2014 technical report that
established a resource estimate of 180,000 oz of gold in indicated resources and
100,000 oz in inferred resources (Mello 2014). At this scale, the project became
regarded as small, likely too small to ever become a modern industrial mine at
current gold prices. It was difficult to communicate to weary and skeptical investors
that this first resource estimate was “just a start” and that the project had great
promise and potential.
The mining industry is littered with promises of pots of gold at the end of
rainbows, with tales of the discovery of El Dorado, the fabled city of gold that
the Spaniards never found. Mining promoters traffic in hyperbolic claims and
wishful thinking. This is, in fact, a necessary part of sustaining investment through
the high-risk phase of exploration. Many projects move forward only through sheer
optimism. Exploration geologists have a tendency to fall in love with their discov-
eries, always maintaining faith that their project has much more potential than
anyone has been able to document. Next year’s drilling will prove up that potential
and vindicate their faith . . . if only someone would fund the project.
Against this backdrop of boundless optimism, discussions of a project’s potential
are understandably met with a degree of cynicism and doubt. Arm-waving pro-
jections of blue-sky possibilities are rarely convincing to investors who have
Castelo de Sonhos: Geostatistical Quantification of the Potential Size of a. . . 401
jurisdictions where the partners are headquartered, and where they operate, the
rules that govern reporting of mineral resources have gravitated toward a global
norm. Although there are differences between the details of the securities regula-
tions in each country, these differences are becoming smaller as many countries
adopt rules that are similar to the JORC Code developed by the Australian Institute
of Mining and Metallurgy and to Canada’s National Instrument 43-101.
At the same time that securities regulators require adherence to specific rules and
definitions, they also encourage the use of best practice guidelines developed by
professional organizations. In Canada, for example, National Instrument 43-101
recommends that the best practice guidelines of the Canadian Institute of Mining
and Metallurgy (CIM) be followed.
Although the geostatistics community has developed many tools for building
stochastic models that span a space of uncertainty, these have not found broad use
in the mining industry because securities regulations for mineral resource reporting
do not embrace the possibility of different but equally likely versions of a deposit’s
grade and tonnage. National Instrument 43-101, for example, is unambiguous in its
requirement that when multiple versions of a deposit’s grade and tonnage are
presented, the report must make clear which of the alternatives is the single one
being presented as the mineral resource estimate.
The single-estimate tradition is reinforced by best practice guidelines that
present advice that focuses on the use of single-estimate interpolation procedures
like kriging and inverse distance weighting. With decades of effort having gone into
harmonizing regulations and guidelines in different countries, it is very unlikely
that the single-estimate tradition of the mining industry can be undone.
2.1 Classification
Classification is the one purpose for which conditional simulation has gained some
traction in mineral resource estimation. Regulations require that mineral resource
estimates be classified into three categories that reflect different levels of confi-
dence: “measured,” “indicated,” and “inferred,” from most confident to least.
Several mining companies and mining engineering consulting firms have devel-
oped a standard practice of classifying resources according to fluctuations in grade,
tonnage, and metal content observed in multiple realizations from a conditional
simulation study. But even when used for this purpose, conditional simulation is
very rarely the direct basis for the reported estimate of the mineral resource; it
informs only the choice of appropriate confidence categories for a resource estimate
developed using a single-estimate procedure.
Castelo de Sonhos: Geostatistical Quantification of the Potential Size of a. . . 403
Canadian and Australian regulations for the reporting of mineral resources explic-
itly permit the reporting of an “exploration target range.” Canada’s National
Instrument 43-101 does not define what this is, a puzzling omission in a document
that insists on adherence to standardized definitions and terminology. Fortunately,
Australia’s JORC Code does give a definition, one that captures the sense of the
term as it is generally understood throughout the mining world:
An Exploration Target is a statement or estimation of the exploration potential of a mineral
deposit in a defined geological setting where the statement or estimate, quoted as a range of
tonnes and a range of grade, relates to mineralization for which there has been insufficient
exploration to estimate a Mineral Resource.
In the syngenetic view, free gold grains would have been moved by water and
settled to the bottom when the velocity of the current was not sufficient to keep
them in suspension, or to keep them as part of the bed load that rolls along the
bottom. The primary direction of continuity would be stratigraphically horizontal,
subparallel to bedding, and aligned with the current direction. Large-scale changes
in gold grade would follow lithology, tending to be higher where large pebbles are
more frequent and tending to be lower where the sediments are finer. These
lithologic trends are due to the fact that the hydrodynamic conditions that cause
gold grains to settle will also cause large pebbles to settle. The large-scale trends in
gold grade would go from high in the proximal direction (nearer the source of the
eroded gold grains) to low in the distal direction (further away from the source).
In the epigenetic view, gold would follow cracks and fissure, the high perme-
ability pathways through which mineralized fluids could travel. Directions of
maximum continuity would be aligned with these fractures; for near-surface
deposits, these would be subvertical because the lack of a confining load on the
top makes horizontal stress greater than vertical stress, and fractures tend to
propagate perpendicular to stress. Large-scale trends in gold grade would reflect
proximity to faults and fractures, with gold grades tending to be higher in brecciated
rocks in fracture zones and lower as one moves into less fractured rock away from
those zones.
There is geological evidence that supports both styles of mineralization. The
coincidence of the conglomerate outcrop with the soil anomaly (Fig. 4) supports the
syngenetic view because it suggests that the conglomerate was always the host rock
and not merely a convenient stopping point where gold later came out of solution
from mineralized fluids. Direct observations of thin gold films on fracture surfaces
are consistent with the epigenetic view, as is anecdotal commentary on the presence
of hematite alteration with high gold grades. Statistical analysis of the available
data, however, confirms that the primary controls on mineralization are those of
placer environment and that although gold was subsequently remobilized, it has
moved only a short distance from the location where it was originally deposited
among pebbles and gravels, likely in flowing water.
Figure 5 shows a schematic of the main lithologies in the Castelo de Sonhos
Formation, which is composed of a conglomeratic band, 250–300 m thick, that sits
between thicker units of arenite above and below. Within the main conglomerate
band, there is a mixture of pebble-supported conglomerates (mC1), matrix-
supported conglomerates (mC2), a micro-conglomerate (mC3), and conglomeratic
arenites (mAC). The conglomeratic arenites are sandstones in which the occasional
pebbles are so widely spaced that they don’t look like true conglomerates. In places
where the pebbles are very widely spaced, the conglomeratic arenite looks exactly
like the upper or lower arenite (mA).
Figure 6 shows side-by-side box plots of the gold grade distributions in the five
lithologies, with the progression from left to right being proximal to distal: pebble
conglomerates will occur closer to the source and the finer-grained arenites will
occur further from the source. The gradual progression from high grade in the
pebble-rich lithologies to low grade in the pebble-poor lithologies is consistent with
Castelo de Sonhos: Geostatistical Quantification of the Potential Size of a. . . 405
the view that the hydrodynamic processes that caused pebbles to settle to the bottom
also caused gold grains to settle at the same time.
Recursive partitioning (Breiman et al. 1984) provides additional statistical
support for the view that the deposit is primarily an ancient placer. Figure 7
shows the recursive partitioning tree that separates high grade from low grade in
the drill hole intervals coded as mC3, which is referred to as a “micro” conglom-
erate because it is composed entirely of small particles of gravel. The single most
important factor in separating high-grade mC3 from low-grade mC3 is the thick-
ness of the mC3 layer. Layers of micro-conglomerate less than 2 m thick have much
higher gold grades (nearly 10) than the thicker mC3 layers. This speaks to the
likelihood that the micro-conglomerate acted like a natural sluice box, trapping
gold that settled from the bed load above. When the same mass of gold grains is
distributed in a thick mC3 layer, the gold grade ends up being lower than it does in a
thin mC3 layer. For the thinner mC3 units, the next most important factor that
separates higher grade from lower grade is the lithology of the overlying layer. If it
is a matrix-supported conglomerate (mC2), the average grade is 3 that of the
micro-conglomerates that are not overlain by a matrix-supported conglomerate. In
order for bottom gravels to trap gold grains in a placer environment, the water
velocity has to be sufficient to carry gold grains in the bed load, but not so high that
the gold grains will be in suspension in the water column. Although the pebble-
supported conglomerates likely reflect a higher water velocity, it is the matrix-
supported conglomerates that would have been deposited in water that was flowing
quickly, but not so quickly that there was a minimal bed load rolling at the bottom
of the water column.
Other factors that support the syngenetic view include high recoveries of free
gold using gravity methods in the preliminary metallurgical test, no gold in the
lower arenite and very little gold in the upper arenite and only immediately above
the conglomeratic band, and low variance in field measurements of the azimuths of
paleo-current directions, typically from cross-bedding.
406 R.M. Srivastava et al.
Fig. 6 Side-by-side box plots of gold grade distributions in the five lithologies
Although there are some factors that support the epigenetic view (visible gold on
fracture surfaces and association of gold with hematite alteration), there are several
factors that are inconsistent with an epigenetic explanation. There is no correlation
between gold grades and proximity to faults or fractures, and the deposit does not
contain any of the sulfide minerals, like pyrite and chalcopyrite that commonly
occur with epigenetic deposits.
Placer deposits can form in many different depositional environments, including
alluvial fans, braided streams, meandering rivers, fluvio-deltaic systems, and in
near-shore marine environments. There is not yet enough field evidence to provide
a strong opinion on the specific details of the depositional environment. Occasional
hematite rims on quartz pebbles are consistent with subaerial deposition commonly
seen in alluvial fans; but the low variance of paleo-current directions and the
statistical evidence of subaqueous deposition (Fig. 7) point to something other
than an alluvial fan. The lack of very fine-grained sediments is consistent with
the separation of silts that occurs in the surface layer in a near-shore marine
environment where fresh water from land meets salt water from the ocean.
Despite the lack of specific details on the depositional environment, several
broad characteristics remain clear. The direction of maximum continuity will be
Castelo de Sonhos: Geostatistical Quantification of the Potential Size of a. . . 407
Fig. 7 Recursive partitioning tree for gold grades in drill hole intervals coded as mC3
Figure 9 shows a reconstruction of the continental plates 2.05 billion years ago
(Eglington 2015), when a large supercontinent known as Nuna formed near the
South Pole. In the 2 billion years since, the continental crust of Nuna has separated
into four continental plates, two of which are now part of South America (the green
plates on Fig. 9) and two of which are now part of Africa (the blue plates). There are
408 R.M. Srivastava et al.
Fig. 8 Photographs showing sinuous channel geometries at several scales. From left to right: at
the 10 km scale in the desert near Hamra Al Drooa in Oman (Kjell Lindgren, International Space
Station, NASA); at the 100 m scale in the Thjorsa River near the Hofsj€ okull glacier, Iceland
(Olivier Grunewald); and at the meter scale in heavy minerals in sand at Ventry Beach, Ireland
(Jessica Winder)
Fig. 9 Paleo-reconstruction of Nuna, approximately 2.05 billion years ago, from data provided by
Bruce Eglinton (University of Saskatchewan)
several known lode gold deposits that date to 2.05 Ga or older and that can be traced
back through time to their locations on Nuna (the orange triangles). These formed
an arc that stretched across the continent, likely a ridge of volcanic mountains much
like the modern Andes.
Zircon dates from Castelo de Sonhos place the formation of the conglomerates at
2.0–2.1 billion years ago, approximately the same age as two other paleo-placer
gold deposits, Tarkwa in Ghana (2.1 Ga) and Jacobina in Brazil (2.0 Ga), both of
which are now large operating mines. At the time that all three of these quartz-
pebble conglomerates were forming, they sat near the shore, below lode gold
deposits at higher elevations. Streams and rivers carrying eroded gold grains
would have deposited their gold where the water velocity dropped, either in the
alluvial plains or in the fluvio-deltaic region where rivers meet the sea.
5 Structural Geology
From its original flat-lying configuration, the band of conglomerates has been
folded and faulted. The axial plane of the fold has been tilted to the northwest
(in current UTM coordinates), and the hinge line has been tilted so that it dips to the
Castelo de Sonhos: Geostatistical Quantification of the Potential Size of a. . . 409
Fig. 10 Vertical and horizontal proportion curves for the five lithologies
outcrop, the proportions of the high-energy lithologies (mC1 and mC2) will tend to
decrease, and the gold grades will also tend to decrease.
The SIS procedure is able to use locally varying directions of maximum conti-
nuity that are subparallel to simulated channel centerlines. These centerlines were
created by borrowing a tool developed for oil and gas applications. Srivastava et al.
(2013) presents a grid-less method for simulating the geometry of the centerlines of
sinuous fluvio-deltaic channels, conditioned by lithology observations in well
bores, with the conditioning data providing information on the proximity to sand
channels. For Castelo de Sonhos, lithologies in drill holes provide information on
the proximity of channels. The five lithologies were coded as integers from 1 to 5 in
proximal-to-distal order: mC1 ¼ 1, mC2 ¼ 2, mC3 ¼ 3, mAC ¼ 4, and mA ¼ 5.
Within each drill hole, within each sequence, the average of these integer codes
is a numerical summary of the energy of the environment. High-energy environ-
ments will have a low average, and low-energy environments will have a high
average. The spatial trends in these average lithologies provide clues to the prox-
imity to channels within each sequence.
Figure 11 shows simulated channel centerlines for the six sequences. The
colored dots on this figure show the average lithology indicator for each drill
hole, within the sequence boundaries. Red-to-yellow colors denote high-energy
environments (low values of the average lithology integer), while green-to-blue
colors denote low-energy environments (high values of the average lithology
integer).
Figure 12 shows an example of the SIS lithology simulation on a stratigra-
phically horizontal slice at the middle of the conglomerate band, using the propor-
tion curves (Fig. 10) for SK means, and using simulated channel centerlines
(Fig. 11) to define local directions of maximum continuity.
Once lithology simulations have been created, grade simulations are done with
SGS, using only the drill hole data that fall within the same lithology. Following the
grade simulation, SIS is used to simulate broad zones of weak, moderate, and
intense hematite alteration that slightly modify the gold grade. Figure 13 shows
Castelo de Sonhos: Geostatistical Quantification of the Potential Size of a. . . 411
Fig. 11 Simulated centerlines of channel, conditioned by a proxy for the distance to channels: the
average lithology indicator calculated from drill hole samples within each sequence
Fig. 12 An example of lithology simulation using SIS with proportion curves and locally varying
directions of maximum continuity
an example of the grade simulation and the alteration simulation. The left frame in
Fig. 13 shows that the gold grades decrease in the down-dip (southwest) direction.
This is due to the linkage between the samples and the lithologies; when the grade
within the mAC region is estimated using only the mAC samples from nearby drill
holes, the fact that the gold grades are generally low in mAC (Fig. 6) entails that the
simulated gold grade will also be low.
412 R.M. Srivastava et al.
Fig. 13 Simulated gold grades (left) and simulated intensity of hematite alteration (right)
7 Results
The procedure described in the previous section was used to create 100 realizations
of lithology, of gold grade and of hematite alteration. For each of these, the gold
that could be extracted profitably using open-pit mining methods was calculated.
Figure 14 shows the histogram of the metal content of the 100 realizations. Table 1
shows the exploration target range based on the realizations corresponding to the
10th percentile (P10) and 90th percentile (P90) of the distribution of metal content.
This conceptual exploration target range is not a mineral resource estimate.
Further drilling is required to calculate a mineral resource estimate that complies
with the requirements of National Instrument 43-101. If this additional drilling is
done, the eventual future resource may not fall within the range expressed in
Table 1.
Despite being conceptual in nature, the quantitative risk assessment (QRA) that
leads to the exploration target range was an important step in advancing the project.
In the months since the results of the QRA have been available, the project has been
able to inspire confidence in a large and growing group of people who have invested
several million dollars in the project. In the current market, when there is little
funding available for mineral exploration, a project like Castelo de Sonhos could
not have been advanced without the sound technical analysis of future potential that
geostatistical simulation provides.
Castelo de Sonhos: Geostatistical Quantification of the Potential Size of a. . . 413
Fig. 14 Histogram of
100 outcomes of gold
contained in an
economically viable
open-pit
Bibliography
Breiman L, Friedman J, Stone CJ, Olshen RA (1984) Classification and regression trees. Chapman
and Hall, New York
Eglington B (2015) Supercontinent cyclicity: relevant data, constraints, limitations and aspects
requiring particular attention, AGU Fall Meeting. American Geophysical Union, San Francisco
Mello R (2014). Mineral resources estimation for the Castelo de Sonhos project. TriStar Gold Inc:
filed on www.sedar.com
Srivastava RM, Maucec M, Yarus J (2013) Grid-less simulation of a fluvio-deltaic environment,
CSPG GeoConvention. Canadian Society of Petroleum Geologists, Calgary
A Hybrid Model for Joint Simulation
of High-Dimensional Continuous
and Categorical Variables
1 Introduction
2 Methodology
Compositional data are multivariate data where the components represent some
part of a whole. They are measured on the same scale and are constrained by a
constant sum property. A compositional data set fZðxα Þ ¼ ½z1 ðxα Þ, . . . , zD ðxα Þ
jzi ðxα Þ 0, i¼1, . . . , D;α¼1, . . . , k;g with D components and k observations can
be represented by the equation:
XD
z ðx Þ
i¼1 i α
¼ 100%, α ¼ 1, . . . , k: ð1Þ
Compositional data raise some challenges for statistical and geostatistical ana-
lyses. Firstly, they are relative values and not free to vary in ð1, þ 1Þ. Secondly,
the constant sum constraint forces at least one covariance to be negative, causing
spurious correlations (Aitchison 1986). Therefore they are often transformed via a
log-ratio transform (Aitchison 1986; Pawlowsky-Glahn and Olea 2004). Several
418 H. Talebi et al.
transforms are available, including additive log-ratio (alr), centred log-ratio (clr),
and isometric log-ratio (ilr) (Egozcue et al. 2003). In this study the alr will be used.
The alr transformation is defined as
z1 ðxα Þ z2 ðxα Þ zD1 ðxα Þ
ζ ðxα Þ ¼ alrðZðxα ÞÞ ¼ ln , ln , . . . , ln ð2Þ
zD ðx α Þ zD ðx α Þ zD ðxα Þ
Its inverse,
known as agl transformation, is able to recover Z(xα) from
ζ ¼ ζ 1 ; ζ 2 ; . . . ; ζ D1 and is defined as
The Murrin Murrin nickel (Ni) laterite deposit is located in Western Australia.
Laterite deposits are formed during chemical weathering of ultramafic rocks near
the surface of the earth. At Murrin Murrin, the nickel laterite deposits occur as
laterally extensive, undulating blankets of mineralisation with strong vertical
anisotropy covering basement ultramafic rocks (Murphy 2003). From the bottom
of the deposit through to the top, the following layers can normally be recognised:
A Hybrid Model for Joint Simulation of High-Dimensional Continuous and. . . 419
ultramafic, saprolite, smectite, iron oxide, and/or clay layers with a hard ferruginous
cap at the top. In this study area, there are four main geological units that control the
spatial distribution of the Ni and Co. The following are short geological descrip-
tions for each unit (Markwell 2001):
• Ferruginous zone (FZ): This zone is mainly composed of goethite and clay
(kaolin). Although FZ tends to be less enriched in Ni, it can host significant
amounts of Co.
• Smectite zone (SM): Consists mainly of smectite and is confined to the shoul-
ders saprolite domes. SM is enriched in Ni and Co and depleted in Mg.
420 H. Talebi et al.
• Saprolite zone (SA): Consists mainly of lizardite and smectite, and high amount
of Mg is present in this zone.
• Ultramafic zone (UM): At the base of the deposit, there is a thin layer of
unweathered ultramafic rock.
In total, 5145 samples, located in a volume of 1300 500 40 m (Fig. 2), make up
the database for this study with information on the rock types (FZ, SA, SM, and
UM), five major elements (Ni, Co, Mg, Al, and Fe), and five secondary elements
(Cr, Mn, Zn, Cu, and As). These ten elements plus the filler variable (i.e. continuous
variables) form the compositional data, with rock type the only categorical variable.
All samples were regularised to a bench height of 1 m.
Figure 3 shows a cross section of the rock types with Ni and Co distributions
near boreholes for northing 180 m. It can be observed from this figure that the
highly mineralised zones occur at the transitions between FZ, SA, and SM domains.
This illustrates the importance of joint modelling of grade and geological units, as
prior domaining may result in misclassification in the transition zones. Figure 4 and
Table 1 depict the histograms and descriptive statistics of the continuous variables
globally. Bimodal distributions can be recognised in Al, Fe, Filler, and Mg, while
the other variables are strongly positively skewed.
According to the transition probability matrices and the geology of the deposit, UM
domain can only be in contact with SA, while SA, SM, and FZ are mutually in
contact. To define the truncation rule, two independent GRFs, {Y1(xα), Y2(xα)}, and
three thresholds, {y1, y2, y3}, are required (Fig. 5). Accordingly, the rock type
prevailing at a given spatial location xα is defined in the following fashion:
• Location xα belongs to UM domain , Y 1 ðxα Þ < y1
• Location xα belongs to SA domain , y1 Y 1 ðxα Þ < y2
• Location xα belongs to SM domain , Y 1 ðxα Þ y2 and Y 2 ðxα Þ y3
• Location xα belongs to FZ domain , Y 1 ðxα Þ y2 and Y 2 ðxα Þ < y3
The threshold values are determined in agreement with the domain proportions
calculated from the drill hole data. For variogram analysis, the two GRFs are
assumed to be independent and their variograms are determined through their
impact on the variograms of the domain indicators. Table 2 provides the parameters
A Hybrid Model for Joint Simulation of High-Dimensional Continuous and. . . 421
Fig. 2 Perspective view of samples showing a different rock types and b nickel grade
Fig. 3 Rock types (coloured data), nickel (left), and cobalt (right) distributions for the cross
section with north coordinate 180 m
for the variogram models obtained for the two GRFs. In order to reproduce a regular
boundary between different domains, cubic variogram models which are isotropic
horizontally have been used for the GRFs, since they are smooth at the origin and
associated with regular boundaries. Gaussian data are generated at sample locations
via the Gibbs sampler algorithm as described in Armstrong et al. (2011) and
plurigaussian model parameters conditional to categorical information at sample
locations.
422 H. Talebi et al.
Table 2 Parameters of variogram models of GRFs for the plurigaussian model (the anisotropy
ranges are long, middle, and short range, respectively)
GRFs Nugget Direction Type Sill Range (m)
Y1 0 N0 Cubic 1 250,250,20
Y2 0 N0 Cubic 1 200,200,20
transformation to normal scores is required prior to the simulation. The alr data are
transformed to Gaussian space via Gaussian anamorphosis.
The final phase of this study is to achieve realisations of the whole composition of
11 components and rock types on selective mining units (SMU) of size
10 10 2 m. Based on the separation between drill holes, these SMUs are treated
as point data. The Gaussian data from all previous preprocessing steps (two
Gaussian variables from the Gibbs sampler algorithm and ten Gaussian variables
from the alr-transformed data) are transformed to orthogonal factors via the MAF
algorithm. Through testing several lag separations, a lag of 30 m has proven to be a
practical choice. Experimental variograms of the 12 factors were calculated using
10 lag distance classes at a spacing of 30 m and towards several directions in
horizontal plane. In the vertical direction, 25 lags at a spacing of 1 m and with an
angular tolerance of 10 were used. For the first factor, which has the greatest
contribution on variation, there is a presence of minor anisotropy in the horizontal
plane with N-115 as the direction with greatest continuity. Table 3 shows the
variogram model parameters for the MAF factors derived from the Gaussian data,
while Fig. 6 shows the experimental variograms and fitted model for the first factor.
Conditional turning bands simulation was applied on the punctual grid and
100 simulations were generated for each factor.
424
Table 3 Variogram model parameters for the MAF factors derived from conditional Gaussian data (the anisotropy ranges are long, middle, and short range,
respectively)
Factors Nugget Direction Type Sill Range (m) Sill Range (m) Sill Range (m)
Factor 1 0.08 N115 Spherical 0.37 50,50,12 0.44 185,150,35 0.28 600,350,350
Factor 2 0.10 Exponential 0.55 60,90,15 0.40 650,300,1 0.15 2,500,350,1
Factor 3 0.10 N115 Exponential 0.79 90,90,20 0.19 750,350,30
Factor 4 0.10 N120 Exponential 0.45 50,50,15 0.45 175,150,20 0.10 2,500,750,100
Factor 5 0.10 N0 Exponential 0.60 35,40,7 0.20 150,185,25 0.15 600,270,25
Factor 6 0.10 Exponential 0.72 35,35,10 0.21 500,250,13
Factor 7 0.10 N115 Exponential 0.70 25,10,8 0.12 200,55,45 0.13 350,350,1
Factor 8 0.08 N0 Exponential 0.64 30,30,10 0.30 85,85,17
Factor 9 0.05 N0 Exponential 0.60 15,15,10 0.35 25,25,20
Factor 10 0.05 Exponential 0.96 20,20,16 0.20 1,1,25
Factor 11 0.05 N0 Exponential 1.00 27,16,14 0.15 1,1,14
Factor 12 0.08 N0 Spherical 0.90 35,35,27 0.90 1,1,27
H. Talebi et al.
A Hybrid Model for Joint Simulation of High-Dimensional Continuous and. . . 425
Fig. 6 Experimental
variograms and fitted model
for the first factor
1.00
Variogram : MAF1
0.75
0.50
0.25
0.00
0 50 100 150 200 250
Distance (m)
4 Discussion
Figure 7a, b depicts perspective views of one realisation of Ni grade and mean of
the simulated Ni grade, respectively. A gradual transition from low-grade zones to
high-grade zones can be seen in these figures (soft boundaries). Figure 7c, d shows
one realisation of the rock types and the most probable simulated rock type,
respectively. A comparison of Figs. 2 and 7 indicates a good fit with the condition-
ing data and geology of the deposit. Quantile-quantile plots of realisations of the
major elements versus the sample data in Fig. 8 show that the global experimental
distributions and simulations are reasonably well reproduced especially for nickel
which is the main target economical element. However overestimation can be
recognised for high Ni grades. Figure 9 shows box plots of realisation proportions
for the four rock types. Compared to the proportions of the rock types in the sample
data, the method has overestimated the proportions of SA and UM units and
underestimated those of FZ and SM. Simulated proportions for FZ unit are close
426 H. Talebi et al.
Fig. 7 Perspective view of (a) one realisation of Ni grade, (b) mean of the simulated Ni grade, (c)
one realisation of rock types and (d) most probable simulated rock type
to the proportion in the exhaustive data. Figure 10 shows the experimental cross
variograms between rock-type indicators and Ni grade, calculated from the sample
data and the simulated model. The relationships between rock type and grade are
reproduced qualitatively, with high spatial cross-correlation between SA and Ni as
well as SM and Ni. For FZ and UM, there are negative correlations to the Ni grade.
The underestimation of SM proportions can also be seen in the lower sill of the
experimental cross variograms of the realisations in Fig. 10. Sensitivity analysis on
PGS parameters might improve the performance of the joint simulation.
FZ and SA domains occur with the greatest proportions in sample data and
simulated models. From Fig. 7c, d it is observed that most of the transitions between
rock types in the study area occur between these two rock types. From Fig. 3 it can
be seen that although SA domain is highly mineralised, high Ni grade
A Hybrid Model for Joint Simulation of High-Dimensional Continuous and. . . 427
Fig. 8 Q-Q plots of realisations of the major elements against sample data
mineralisation occurs at the transition between SA and FZ. Figure 7a, b shows that
the proposed joint simulation algorithm has reproduced both the highly mineralised
zone at the transition area and the soft transition in realisations of Ni across
geological boundaries. The ability of the method to reproduce soft transition of
continuous data across geological domains is emphasised in the contact analysis
428 H. Talebi et al.
Fig. 10 Experimental cross variograms between rock-type indicators and Ni grade, for sample
data (black line) and simulated realisations (dashed line)
Fig. 11 (a) Contact analysis between FZ and SA domains for sample Ni grade (black graph),
mean of simulated Ni grade (continuous red graph), and a realisation of Ni grade (dashed red
graph). (b) Average prediction error for Ni grade compared with exhaustive data set
diagram of Fig. 11a. The Ni grade increases gradually from the FZ domain towards
the SA domain, and the highly mineralised zone occurs at the transition zone
between these two geological domains. Figure 11b depicts the average of prediction
error for Ni grade compared with the exhaustive data set. Even though there is
uncertainty in the exact locations of geological boundaries, the prediction errors are
close to zero at the transition zone. The prediction errors are high at the boundaries
of the deposit which shows that the method is sensitive to extrapolation.
variables was presented based on the plurigaussian model and min/max autocorre-
lation factors. The proposed algorithm is able to simulate several constrained
continuous (compositional data) and categorical variables simultaneously. Uncer-
tainty measurement of the exact location of geological boundaries, incorporation of
the spatial dependence between continuous and categorical variables, reproduction
of spatial correlation of continuous data across geological boundaries, and the
ability to simulate high-dimensional and even constrained data are some of the
advantages of the proposed model. In addition it is faster and easier to apply than
co-simulation algorithms based on an LMC. The performance of the method was
tested on a real large mining data set producing results that are satisfactory from a
practical point of view. Considering the effect of the compositional data on Gibbs
sampler algorithm and sensitivity analysis on PGS parameters might further
enhance accuracy of the simulated model. These issues will be addressed in
future work.
Bibliography
Aitchison J (1986) The statistical analysis of compositional data. Chapman and Hall Ltd, London
Armstrong M, Galli A, Beucher H, Loc’h G, Renard D, Doligez B, . . . Geffroy F (2011)
Plurigaussian simulations in geosciences. Springer, Berlin
Bandarian E, Mueller U (2008) Reformulation of MAF as a generalised eigenvalue problem. In:
Ortiz J, Emery X (eds) Proceedings Eighth International Geostatistics Congress. Santiago, pp
1173–1178
Chilès JP, Delfiner P (2012) Geostatistics: modeling spatial uncertainty, 2nd edn. Wiley,
New York
Desbarats AJ, Dimitrakopoulos R (2000) Geostatistical simulation of regionalized pore-size
distributions using min/max autocorrelation factors. Math Geol 32(8):919–942. doi:10.1023/
A:1007570402430
Egozcue JJ, Pawlowsky-Glahn V, Mateu-Figueras G, Barceló-Vidal C (2003) Isometric logratio
transformations for compositional data analysis. Math Geol 35(3):279–300
Emery X (2007) Simulation of geological domains using the plurigaussian model: new develop-
ments and computer programs. Comput Geosci 33(9):1189–1201
Emery X, Silva DA (2009) Conditional co-simulation of continuous and categorical variables for
geostatistical applications. Comput Geosci 35(6):1234–1246
Goulard M, Voltz M (1992) Linear coregionalization model: tools for estimation and choice of
cross-variogram matrix. Math Geol 24(3):269–286. doi:10.1007/BF00893750
Jones P, Douglas I, Jewbali A (2013) Modeling combined geological and grade uncertainty:
application of multiple-point simulation at the apensu gold deposit, Ghana. Math Geosci 45
(8):949–965. doi:10.1007/s11004-013-9500-3
Maleki M, Emery X (2015) Joint simulation of grade and rock type in a stratabound copper
deposit. Math Geosci 47(4):471–495. doi:10.1007/s11004-014-9556-8
Markwell T (2001) Murrin Murrin Ni/Co resource estimation: MME resource modelling report.
Anaconda, p 18
430 H. Talebi et al.
Cansın Y€
uksel and J€org Benndorf
1 Introduction
One of the main challenges in lignite mining, similar to other branches of mining, is
the waste intrusions in lignite seams. These marine and fluvial sand impurities can
lead to high ash values (e.g., more than 15 % ash) and cannot be localized
completely by exploration data and captured in the predicted deposit models.
Utilizing online sensor techniques for coal quality characterization in combina-
tion with rapid resource model updating, a faster reaction to the unexpected
C. Yüksel (*)
Resource Engineering Section, Department of Geoscience & Engineering, Delft University of
Technology, Stevinweg 1, 2628 CN Delft, Netherlands
e-mail: [email protected]
J. Benndorf
Institute for Mine Surveying and Geodesy, Faculty of Geosciences, Geoengineering and
Mining, University of Technology Bergakademie Freiberg, Reiche Zeche, Fuchsmühlenweg 9,
09599 Freiberg, Germany
e-mail: [email protected]
For rapid updating of the resource model, sequentially observed data have to be
integrated with prediction models in an efficient way. This is done by using
sequential data assimilation methods, namely, the EnKF-based methods.
With the goal of a continuously updatable coal quality attributes in a resource
model, a framework based on the normal-score ensemble Kalman filter (NS-EnKF)
(Zhou et al. 2011) approach was tailored for large-scaled mining applications. The
Performance Analysis of Continuous Resource Model Updating in Lignite Production 433
Fig. 1 Configuration of the real-time resource model updating concept (Modified from Wambeke
and Benndorf 2015)
The aim of the case study presented here is to analyze the performance of the
resource model updating framework method by performing sensitivity analyses on
main parameters, including the ensemble size, the localization and neighborhood
strategies, and the sensor precision.
The first sensitivity analyses focus on investigating the optimal realization number
(subsequently used as ensemble size) by performing resource model updating
experiments with different-sized ensembles. Defining the ensemble size that will
fully represent the ore body is a very delicate problem. A lot of research in literature
(Houtekamer and Mitchell 1998; Mitchell et al. 2002) focuses on the optimum
ensemble size investigation and usually concludes that the analysis error decreases
as the number of ensembles (realizations) increases. Contrary, the computational
costs increase with the ensemble size. Therefore a sensible size of the ensemble is
required.
3.1.2 Localization
The final sensitivity analyses focus on testing the effect of the sensor precision. In
most cases errors are involved when taking measurements, due to calibration issues
of sensor technologies. For each experiment, different amounts of standard error are
added to the actual measurement values. The standard error can be calculated as
σ
SEx ¼ pffiffiffi ð1Þ
n
where σ is the standard deviation of the actual measurements and n is the size
(number of observations) of the actual measurements. For this study, the size of the
actual measurement data set contains 700 observations, which values correspond to
coal extracted from 28 mining blocks. This leads to approximately 25 actual
measurement data per block. Therefore, where the added standard error is 0.1 %
ash, the absolute standard deviation will be 0.5 % ash, and the variance will be
0.25 %2 ash.
Similarly, when the added standard error is 0.2 % ash, the standard deviation will
be 1 % ash, and the variance will be 1 %2 ash. The variance of the actual measure-
ments will be 6.25 %2 ash, and the standard deviation will be 2.5 % ash when the
added standard error is 0.5 % ash. The variance will be 25 %2 ash when the added
standard error is 1 % ash. The variance of the averaged prior model for 48 ensembles
is calculated as 0.99.
To give a clear view, mentioned standard deviations are converted as the relative
error of the measurements. The average measurement value is calculated as 12 %
ash. This leads around 4 % ash relative error in measurement values when the added
standard deviation is 0.5 % ash. Similarly, when the added standard deviation is 1 %
ash, this indicates around 8 % ash relative error in measurement values. In the same
way, when the added standard deviation is 2.5 % ash, this indicates around 20 % ash
relative error in measurement values. Finally, when the added standard deviation is
5 % ash, this indicates around 40 % ash relative error in measurement values.
values are then merged with the previously defined coal seam. The block model
realizations are now ready to be imported into the algorithm as the first input.
Figure 2 illustrates the prior model of 48 simulations, the averaged ash values of
those simulations and related sensor measurement values, per block. A significant
underestimation of the actual measurement data is observed in the prior model. This
is because the prior model is created based on the drill hole data, where the local
sand intrusions are not fully captured. True variability of the coal seam is captured
by the online sensor measurements.
Predicted measurements are obtained by averaging the simulated ash values
from each simulation set, which falls into the defined production block boundaries.
The online sensor measurement data, namely, the Kohle OnLine Analytics (KOLA)
data, are provided for the defined time period. KOLA system applies X-ray dif-
fraction in order to accurately assess the components of the produced lignite. In
order to determine the location of the received KOLA data, in other words to track
back where the measured material comes from, the GPS data is matched with the
measurement data based on the given timecodes. The located measurements in coal
seam are then imported into the previously defined block model.
The second input file for the algorithm is written to a file containing the
following information: the block ID, the central block location (X, Y, Z coordi-
nates), and a series of real and predicted measurements.
A study bench is produced for a defined time period by considering all the
available data (topography, RGI, GPS, and production data). Later, the study bench
is divided into so-called production blocks. This was necessary to reproduce the
excavated production blocks. The horizontal divisions (or production slices) are
applied based on the movements of the excavator during production, provided by
GPS data. The vertical divisions are based on the changes in the Z coordinates in the
GPS data and capture a typical extraction sequence of bucket-wheel excavator
operations. In total, the defined production bench is divided into 28 blocks and
5 slices, which gives 140 production blocks. Once the study bench is divided both in
vertical and horizontal, the production blocks are now ready to be updated.
Performance Analysis of Continuous Resource Model Updating in Lignite Production 437
First the first block of the second slice will be updated, based on the KOLA
measurements taken from that block. The series of updating experiments included
seven updating experiments and continued until the ninth block (since there are no
KOLA data obtained on sixth and seventh block, seven experiments are performed
to update until the ninth block). In each updating experiment, only one block is
updated based on the related measured KOLA value.
An empirical error measure so-called mean square difference or mean square
error (MSE) is used in order to present results of the performed experiments. MSE
compares the difference between estimated block value Z∗ ðxÞ and actual KOLA
measurement v values per block, and it can be calculated as
1 XN * 2
MSE ¼ z ðxi Þ νi ð2Þ
N i¼1
where i ¼ 1, . . . , N is the number of blocks. The mean square error graphs are
calculated relative to the averaged prior model of 384 ensembles, in order to make a
good comparison.
With a view toward the real-time application of the updating resource model, the
industrial case presented by Yüksel et al. (2016) focused on small- and moderate-
sized ensembles (24). For the investigation of the optimum ensemble size, updating
experiment series are performed with 24, 48, 96, 192, and 384 ensembles. All of the
simulations are created by using SGS with same seed number and same variogram
parameters.
3.3.2 Localization
localization, with varying localization and neighborhood sizes. For the experiments
where the localization option was used, the localization neighborhood was assumed
as half of the defined neighborhood size, except the tenth experiment. In the tenth
experiment, in the X and Y direction, localization sizes were assumed as half of the
defined neighborhood size. In the Z direction, the localization size remained the
same. Reasons of this preference will be explained in the discussion chapter.
For each experiment, different amounts of standard error are added to the actual
KOLA measurement values. In total, five experiments are performed, where the
relative measurement error varied between 4 %, 8 %, 20 %, and 40 %.
4 Results
Figures 3 and 4 present results of the updating process from the first block until the
ninth block, for some of the representative ensemble sizes. For these experiment
series, the localization strategies were applied; the neighborhood size was 225, 225,
6 m for X, Y, Z directions, respectively, and no sensor error is assumed.
Performance Analysis of Continuous Resource Model Updating in Lignite Production 439
It can be seen that the average of the prior simulations substantially underesti-
mates the actual KOLA measurements. This is caused by the data effect. The prior
simulations are created based on the coal samples from drill holes spaced multiple
hundred meters apart, while the KOLA measurements measure more higher ash
values due to the sand intrusions in the coal seam. Integrating the KOLA measure-
ment to the first nine blocks updates the neighborhood blocks to some relatively
higher values. As expected, the update effect decreases while moving away from
the last updated block, block 9.
For all different ensemble sizes, a clear improvement is observed toward the
KOLA data when considering the average of the initial simulations, so-called prior
model.
Figure 5 presents the relative MSE values to the prior model for each experiment
performed with different ensemble sizes. The biggest reduction of the error occurs
in the update of the first block. While the skewness behavior of each MSE graphs is
similar, the biggest error behavior to the smallest is as follows: 48 ensembles,
96 ensembles, 192 ensembles, 384 ensembles, and 24 ensembles. Except for the
results from 24 ensembles, the rest of the listing supports the literature. It is
expected to observe a decrease in the MSE values while the ensemble size gets
440 C. Yüksel and J. Benndorf
larger since the representativeness gets higher. However, to increase the computa-
tional efficiency and to apply the updating framework in real time during produc-
tion, an economic ensemble size is required.
At first glance, higher initial variance of the 24 ensembles explains the very low
MSE values. Nevertheless, a further investigation is performed in order to under-
stand the phenomenon better. Five different sets of newly derived sets of 24 ensem-
bles are generated with SGS, by using different random seeds for each set. New
series of updating experiments are performed with the new series of 24 ensembles,
and the results are compared. The comparison shows a high variety among results.
MSE values obtained from the ninth block’s update varied between 0.52 and 0.69.
In addition, the new sets of MSE values were equal to, lower or higher than the
48 ensembles, 96 ensembles, 192 ensembles, and 384 ensembles. This big variety,
which is caused by different seed numbers, shows that 24 ensembles were not
sufficient to represent a statistical stable estimate of the mentioned lignite seam.
When considering the 48 ensembles, even though the 48 ensembles have the
highest MSE values by comparing to the 96 ensembles, 192 ensembles, and
384 ensembles, the MSE dropped from 1.0 to 0.64. In his research, Yin et al.
(2015) found that improvements while using larger ensemble sizes (after the
optimum ensemble size) are relatively insignificant. Likewise, the improvements
between 48, 96, 192, and 384 ensembles are obvious, yet not very significant. For
this reason, this study concludes that the optimal ensemble size for this specific
study is 48 ensembles.
Figures 6, 7, 8, and 9 present results of the updating process from the first block
until the ninth block, for different localization strategies and neighborhood sizes.
Experiments 2 (Fig. 3) and 6 (Fig. 6), Experiments 7 (Fig. 7) and 8 (Fig. 8), and
Performance Analysis of Continuous Resource Model Updating in Lignite Production 441
Fig. 7 Experiment 7 – localization option on (225,225,3 m), neighborhood size: 450, 450, 6 m
Fig. 9 Experiment 11 – localization option on (450,450,6 m), neighborhood size: 900, 900, 6 m
Fig. 10 Comparison graph for different localization and neighborhood strategies experiments
Experiments 9 (Fig. 9) and 10 (Fig. 10) are comparable to each other when
investigating the localization option. Experiments 6 (Fig. 6), 8 (Fig. 8), and 9 are
comparable to each other when investigating the neighborhood size.
Figure 10 compares all of the experiments performed in this section by plotting
MSE values of each. Higher MSE values are observed when localization strategies
are applied, and the neighborhood size is defined as 225, 225, 3 m. The MSE values
become lower when the neighborhood size is increased and localization option is
not used. This is expected because the neighborhood size was initially defined as
450, 450, 6 m based on the variogram, so performing the experiments with
225, 225, 3 m-sized neighborhood was not enough to cover the seam continuity.
Minor changes are observed between the MSE values of 450, 450, 6 m
neighborhood-sized experiment and 900, 900, 6 m-sized experiment due to no
spatial correlation between the attributes.
The reason that applying the localization strategies did not provide any improve-
ment in our case is due to the definition of the localization function.
Figure 14 illustrates the currently used function. Since the production block size
is varying for each block, sometimes the plateau phase of the used function cannot
cover a full block which is in the neighborhood. This creates un-updated values in a
block and consequently the updating process of the entire block fails. For this
Performance Analysis of Continuous Resource Model Updating in Lignite Production 443
reason, better results are obtained while the localization strategies were not in use.
The future study will improve this drawback by developing the localization func-
tion in a way that it can define the block boundaries and act according to those
distances.
Experiment 10 uses the localization option with the following dimensions:
450, 450, 6 m in X, Y, Z directions. The used neighborhood size was 900, 900,
6 m. As mentioned before, the initial intention was to use a localization size half the
size of the neighborhood size. Yet, since the depth of a production block is 6 m,
limiting the localization by 3 m decreased the expected improvements. By running
the same experiment, only changing the Z localization size parameter from 6 to 3 m,
the same results as found in Experiment 6 (Fig. 6) are obtained. This can be
observed in Fig. 10, by comparing the related MSE values.
Figures 11, 12, and 13 present the final results of the updating process from the first
block until the ninth block, for different relative sensor errors. For all the experi-
ments performed in this section, the average prediction quality gets better in the
sense that they become closer to the KOLA measurement values.
When the relative sensor error gets higher, the posterior variance appears to
increase significantly. This is mainly because the KOLA measurement values are
almost out of the range of the prior model (Fig. 2), and the variance of the prior
model significantly underestimates the KOLA measurement values. By integrating
the KOLA measurements which have lower precision (applied relative error
varies between 4 and 40 % ash), the algorithm opens up the option whether the
KOLA data can be right or the prior model. Subsequently, this inflates the posterior
uncertainty.
This study analyzes the performance of the resource model updating method by
performing sensitivity analyses on main parameters, including the ensemble size,
the localization and neighborhood strategies, and the sensor precision in lignite
Performance Analysis of Continuous Resource Model Updating in Lignite Production 445
Bibliography
Benndorf J (2015) Making use of online production data: sequential updating of mineral resource
models. Math Geosci 47(5):547–563
Benndorf J, Yueksel C, Shishvan MS, Rosenberg H, Thielemann T, Mittmann R, Lohsträter O,
Lindig M, Minnecker C, Donner R (2015) RTRO–coal: real-time resource-reconciliation and
optimization for exploitation of coal deposits. Minerals 5(3):546–569
Gaspari G, Cohn SE (1999) Construction of correlation functions in two and three dimensions. Q J
Roy Meteorol Soc 125(554):723–757
Horn RA, Johnson CR (1985) Matrix analysis. Cambridge University Press, Cambridge
Houtekamer PL, Mitchell HL (1998) Data assimilation using an ensemble Kalman filter technique.
Mon Weather Rev 126(3):796–811
Mitchell HL, Houtekamer PL, Pellerin G (2002) Ensemble size, balance, and model-error repre-
sentation in an ensemble Kalman filter*. Mon Weather Rev 130(11):2791–2808. doi:10.1175/
1520-0493(2002)130<2791:ESBAME>2.0.CO;2
Sakov P, Bertino L (2011) Relation between two common localisation methods for the EnKF.
Comput Geosci 15(2):225–237
Wambeke T, Benndorf J. (2015) Data assimilation of sensor measurements to improve production
forecasts in resource extraction. Paper presented at the IAMG, Freiberg (Saxony) Germany
446 C. Yüksel and J. Benndorf
Yin J, Zhan X, Zheng Y, Hain CR, Liu J, Fang L (2015) Optimal ensemble size of ensemble
Kalman filter in sequential soil moisture data assimilation. Geophys Res Lett 42
(16):6710–6715
Yüksel C, Thielemann T, Wambeke T, Benndorf J (2016) Real-time resource model updating for
improved coal quality control using online data. Int J Coal Geol 162:61–73, http://dx.doi.org/
10.1016/j.coal.2016.05.014
Zhou H, Gómez-Hernández JJ, Hendricks Franssen H-J, Li L (2011) An approach to handling
non-Gaussianity of parameters and state variables in ensemble Kalman filtering. Adv Water
Resour 34(7):844–864
Part IV
Petroleum Geoscience and Engineering
Geostatistics on Unstructured Grids,
Theoretical Background, and Applications
P. Biver (*)
Geostatistics and Uncertainties Team, Total SA, CSTJF avenue Larribau, 64000 Pau, France
e-mail: [email protected]
V. Zaytsev
Geostatistical Team, Geosciences Centre, Mines Paris Tech. Sponsored by Total SA,
CSTJF avenue Larribau, 64000 Pau, France
e-mail: [email protected]
D. Allard
Biostatistics & Spatial Processes, INRA, 84914 Avignon, France
e-mail: [email protected]
H. Wackernagel
Geostatistical Team, Geosciences Centre, Mines Paris Tech, 35 rue Saint Honoré,
77305 Fontainebleau, France
e-mail: [email protected]
1 Introduction
For more than 40 years, geostatistical estimations and simulations have been
performed on regular so-called “sugar-box” grids. This is mainly due to historical
reasons. The technology was emerging first from the mining industry, to estimate
grades for open pit blocks, and it was a reasonable choice to use such support of
information.
These regular grids have been kept for a long time as they allow following
stratigraphy in corner point grids geometry commonly used in petroleum industry;
they were also a convenient format to optimize algorithms of various kinds
(sequential simulations, simulation with fast Fourier transform, multiple-point
statistics simulations, etc.).
However, several new grid geometries have emerged in the last decades: tetra-
hedral meshes in hydrogeology and Voronoi grids with local grid refinements for
petroleum industry. These grids are more convenient to solve the physical equations
of flow and transport in porous media. Moreover, these grids are emerging in the
geo-modeling processes with a relevant formulation of the depositional UVT
coordinates system (see Mallet 2004); a dual grid approach was used to address
both the dynamic simulation and the geostatistical characterization (flow simulation
grid or FSG to solve the physical problem and the geological grid or GG to perform
geostatistics).
The dual grid approach has several drawbacks: the geological grid resolution is
driven by the smallest cells of the flow simulation grid, and, moreover, an upscaling
technique is necessary to transfer information from GG to FSG; this upscaling
technique needs to be general enough for the considered topologies.
As a consequence, it was necessary to adapt the geostatistical processes to use
directly the flow simulation grids; but, due to the various size and geometry of
elementary grid cells, it is mandatory to take into account the support size effect.
We present a technique based on the formalism of the discrete Gaussian model (see
Emery 2009 and Chilès and Delfiner 2012). A review of other solutions for
geostatistical simulations on unstructured grids can be found in Zaytsev et al.
2016. To simulate directly on unstructured reservoir grids, an alternative method
proposed by Boucher A. and Grosse H. (2015) will be also commented regarding
implementation aspects.
The presentation of the discrete Gaussian model (DGM) can be found in Chilès and
Delfiner (2012). It can be described as follows:
– Each block of the grid vp is attributed a parameter rp E (0,1) which is called the
change of support coefficient for this block.
– We work on the Gaussian transform Y of the variable of interest Z.
Geostatistics on Unstructured Grids, Theoretical Background, and Applications 451
X
1
Z ðxÞ ¼ φðY ðxÞÞ ¼ φ i χ i ðY ðx ÞÞ ð1Þ
i¼0
– Using the Cartier’s relation (Chiles and Delfiner 2012 p. 441), we can derive the
block support distribution Z(v) since it can also be represented in the same
polynomial basis using the decomposition (1) and the block change of support
coefficient r:
X
1
Z ð v Þ ¼ φv ð Y v Þ ¼ φi: r i χ i ðY v Þ ð2Þ
i¼0
– By double volumetric integration of the point support covariance C(x,x0 ) over the
volume of interest ν, we can derive the change of support coefficients rp:
Z Z X
1
1
Var Z vp ¼ 2 Cðx, x0 Þdx dx0 ¼ φ2i r 2i ð3Þ
vp i¼1
p
v v
ð4Þ
– Once the change of support is known for each grid cell and covariance is known
between each pair of Gaussian random variables characterizing the volumes of
each grid cell, we are back to a classical problem of generating a multivariate
Gaussian random function with a given covariance matrix which can be solved
by classical methods such as SGS.
Different formulations are available for the discrete Gaussian model; in this
paper, these aspects will not be addressed; for such discussions, we refer to Zaytsev
et al. (2016).
452 P. Biver et al.
To apply the theory of DGM to unstructured grids, key additional issues need to be
addressed.
It is mandatory to integrate efficiently point support covariance (Eqs. 3 and 4)
over cells that are not usual octahedrons. This issue is solved with an efficient
lumping of each grid cell (see Korenblit and Shmerling 2006) followed by pseudo
Monte Carlo integration using Sobol sequence of quasi-random points in the
six-dimensional space of integration. This methodology is recommended for high
dimension space because of the convergence speed.
Figure 1 illustrates the advantage of Monte Carlo methods for the problem of
computing the variance of a block average value. In this figure, several integration
methods are compared (subsequent Gauss quadrature integration, approximating
the block with regularly spaced points, Monte Carlo integration). Clearly, Monte
Carlo and related techniques are much more efficient.
It is also important to have an efficient procedure to navigate in the topology of
the grid. The definition of searching neighborhood needs to be addressed in a
general and efficient way. We propose to address this issue with a k-d tree efficient
search (Bentley 1975). Expressed in the asymptotic notations for comparison of the
algorithm performance (Cormen 2009), for a grid of Nb blocks, the k-d tree gives
performance O(log Nb) for the neighborhood search operation, which is much faster
than the naı̈ve approach of looking through all the blocks of the grid which is at
least Ω(Nb).
Fig. 1 Comparison of different integration methods to estimate variance of a typical grid cell
(Gauss integration, regular spacing, Monte Carlo with Sobol quasi-random sequence)
Geostatistics on Unstructured Grids, Theoretical Background, and Applications 453
Our tests indicate that when the Sobol quasi-random sequence of discretizing points
is used, a relatively small number of points (between 50 and 100; see Fig. 2) are
sufficient to satisfy (5).
The area covered by the grid is 70 by 90 km2. The previous model was built with
constant values over very large domain; it was not representative of the variability
of facies and porosity that can occur in this domain.
A new facies model (using the method in Gross and Boucher (2015)) was built.
Three facies are modeled (shale, shaly sand, and massive sand); the target pro-
portions over the entire grid are 20 % for shales, 45 % for shaly sands, and 35 % for
massive sands; the covariance function used for the underlying Gaussian field in the
truncated Gaussian simulation is a spherical model with areal ranges of 800 m by
250 m (azimuth of maximum range is 55 ); the vertical range is 100 m. The
coordinate system is a UVT coordinate system, built from the relevant horizons.
The facies simulation results are illustrated in Fig. 3. The simulated proportions
over the grid are clearly illustrative of the support size effect (less variations in large
cells, larger variations in small cells).
The important issue of modeling facies and related proportions is illustrated in
Fig. 4. The dominant facies (most likely facies regarding proportions) is
represented. In the area modeled with large cells, the shale facies is never dominant,
and proportional modeling procedure is the only way to keep that facies into
account; in the area of local grid refinement, this aspect is less important, and the
traditional truncated Gaussian simulation picture is observed.
The porosity model has been built using DGM assumptions. As each facies can
occur in each cell, it is therefore important to perform full field porosity model for
each facies. Point-scale distributions for each facies are provided in Table 1
(we used beta distributions with p and q referring to the shape parameters); these
porosity distributions are clearly different according to facies classification. Point-
scale normal score variogram for porosity is a spherical model with areal ranges of
600 m by 200 m (azimuth of maximum range axis is 55 ); the vertical range is 80 m.
The porosity model inside each facies is represented in Fig. 5. The support size
effect is clearly visible with small variations in large cells and larger variations in
smaller cells.
Geostatistics on Unstructured Grids, Theoretical Background, and Applications 455
F0
F1 point proportions
F2
regional
model
1 1 1
local refined model
pF0
pF2
pF1
0 0 0
Fig. 3 Facies modeling on field X, proportions maps for silt (blue), silty sand (pink), and massive
sands (yellow). The simulated facies proportion map is sharp in the region of the local grid
refinement and becomes smoother in the regions of coarse blocks
ΦF1
ΦF2
0 0 0
From these facies porosities, it is possible to derive equivalent porosity for each
cell simply by weighting them according to simulated proportions. This result is
represented in Fig. 6; the support size effect is still visible.
Geostatistics on Unstructured Grids, Theoretical Background, and Applications 457
Acknowledgments The authors would like to thank Total SA for sponsoring this research and
authorizing publication.
Bibliography
Bentley JL (1975) Multidimensional binary search trees used for associative searching. Commun
ACM 18:509–517
Chiles J-P, Delfiner P (2012) Geostatistics: modeling spatial uncertainty, 2nd edn. Wiley,
New York
Cormen TH (2009) Introduction to algorithms. MIT Press
Deutsch CV, Tran TT, Pyrcz MJ (2002) Geostatistical assignment of reservoir properties on
unstructured grids, SPE annual technical conference, SPE-77427-MS
Emery X (2009) Change-of-support models and computer programs for direct block-support
simulation. Comput Geosci 35:2047–2056
Emery X, Ortiz JM (2011) Two approaches to direct block-support conditional co-simulation. C R
Geosci 37:1015–1025
Gross H, Boucher AF (2015) Geostatistics on unstructured grid: coordinate system, connections
and volumes, petroleum geostatistics 2015, EAGE
Korenblit M, Shmerling E (2006) Algorithm and software for integration over a convex polyhe-
dron. In: Mathematical Software-ICMS 2006. Springer, pp 273–283
Mallet J-L (2004) Space time mathematical framework for sedimentary geology. Math Geol
36(1):1–32
458 P. Biver et al.
Noetinger B, Hass A (1996) Permeability averaging for well tests in 3D stochastic reservoir
models, SPE annual technical conference, SPE-366-53-MS
Sobol IM (1967) On the distribution of points in a cube and the approximate evaluation of
integrals. Zhurnal Vychislitel’noi Matematiki i Matematicheskoi Fiziki 7:784–802
Zaytsev V, Biver P, Wackernagel H, Allard D (2016) Change-of-support models on irregular grids
for geostatistical simulation. Math Geol l48(4):353–369
Using Spatial Constraints in Clustering
for Electrofacies Calculation
Abstract Petroleum reservoir geological models are usually built in two steps.
First, a 3-D model of geological bodies is computed, within which rock properties
are expected to be stationary and to have low variability. Such geological domains
are referred to as “facies” and are often “electrofacies” obtained by clustering
petrophysical log curves and calibrating the results with core data. It can happen
that log responses of different types of rock are too similar to enable satisfactory
estimation of the facies. In such situations, taking into account the spatial aspect of
the data might help the discriminative process. Since the clustering algorithms that
are used in this context usually fail to do so, we propose a method to overcome such
limitations. It consists in post-calibrating the estimated probabilities of the presence
of each facies in the samples, using geological trends determined by experts. The
final facies probability is estimated by a simple kriging of the initial ones. Mea-
surement errors reflecting the confidence in the clustering algorithms are added to
the model, and the target mean is taken as the aforementioned geological trend.
Assets and liabilities of this approach are reviewed; in particular, theoretical and
practical issues about stationarity, neighborhood choice, and possible generaliza-
tions are discussed. The estimation of the variance to be assigned to each data point
is also analyzed. As the class probabilities sum up to one, the classes are not
independent; solutions are proposed in each context. This approach can be applied
for extending class probabilities in 3-D.
1 Introduction
Petroleum reservoir geological models are usually built in two steps. First, a 3D
model of geological bodies is computed, within which rock properties are expected
to be stationary and have low variability. Such geological domains are referred to as
“facies.” In a second step, petrophysical properties are distributed within each
facies.
A common issue in geological modeling is to determine such facies along each
well. It is quite easy in cored wells, where it is possible to describe the rocks and
make precise analyses and petrophysical measurements. Note that issues related to
measurement errors, interpretation mistakes, or poorly representative sampling that
will lead to an uncertainty on facies determination go beyond the scope of the
present paper and are not discussed here. Defining facies is more difficult in
uncored wells where only electric logs are available. The issue is usually solved
by defining “electrofacies,” which are classes created from clustering petrophysical
log curves. Such electrofacies are then calibrated with core data to ensure their
geological consistency. Electrofacies calculation results are not always of satisfac-
tory and constant quality, which may lead to inconsistencies.
Facies are never randomly distributed in a reservoir, and accounting for the
sedimentological conceptual model at the electrofacies calculation phase may
improve classification performance, leading to more realistic results. Such a sedi-
mentological model is commonly used in geological modeling for guiding facies
spatial distribution simulations, through its numerical representation which is the
facies proportions model calculated from local vertical proportion curves (VPC),
the facies trends.
The goal of this paper is to detail several methods that take advantage of the
facies proportions model at the early stage of electrofacies calculation. It is assumed
here that classification is made in a supervised mode, to account for visual facies
defined by the geologist; classes are defined by more than mathematical criteria. In
these methods, facies proportions coming from geological analyses will be consid-
ered as the expected average probability to belong to a given facies in the neigh-
borhood of each sample. Conditions of use of such methods are discussed, and some
illustration examples are proposed.
Using Spatial Constraints in Clustering for Electrofacies Calculation 461
Spatial information can be taken into account at each target sample without
considering spatial correlation with neighboring data points, using different
approaches. This particular case is relevant for improving classification results
when some classes cannot be easily discriminated.
A first method consists in adding all the a priori probabilities to belong to a class as
additional variables for a standard classification, using usual classification tech-
niques. Calculations are then made with several logs plus as many a priori average
probabilities as facies. As it is common to work with three to five logs and four to
six facies, classifications with about ten variables can be expected.
It is a very simple approach, but it has some practical limitations. In particular,
classifying with about ten variables will require a lot of data to get robust results. In
linear discriminant analysis, for example, a multivariate density function has to be
estimated from available samples. With ten variables, it requires a lot of samples to
Using Spatial Constraints in Clustering for Electrofacies Calculation 463
get a significant and robust estimate of the density function. In the oil and gas
industry, the required number of samples will be rarely reached, especially in
geological modeling for rock-typing purposes.
In addition, logs and a priori probabilities are not at the same scale and giving the
same status to the two types of information in the classification process is a risky
option. It may add some noise or imprecision. This approach is certainly not the
most appropriate for rock typing.
3.2.1 Principle
Accounting for spatial information at a target sample location can be done in a very
simple way, by calculating a weighted average of local probability to belong to a
facies (from classification based on electric logs) and global probability (from a
priori model built by the geologist). Consider n samples (n > 1), K facies (K > 1)
and denote by Pi ¼ (Pi,1,. . .,Pi,K) (resp. Qi ¼ (Qi,1,. . .,Qi,K)) the set of probabilities
calculated from the electric logs (resp. from a geological model) at sample i. The
updated probability that sample i belongs to facies k is written as
i for some facies k, all Zj þ Ej in the sum would be neighboring probabilities Pj,k
and m ¼ Qi,k the local mean. In particular, if the kriging neighborhood is restricted
464 J.-M. Chautru et al.
to the target sample itself, then the simple kriging system can be developed very
easily, resulting in:
1 Vi
P∗ i, k ¼ Pi , k þ Q ; ð3Þ
1 þ Vi 1 þ V i i, k
where Vι is the ratio between σ 2 and the modeled variance of the proportion Pi,k
without measurement errors. This is equivalent to Eq. 1 with β ¼ 1þV 1
i
. Therefore, β
can be viewed as a level of error in the initial classification.
More spatial information can be added if neighbor probabilities are also included
in the kriging formula. However, the kriging system would need to be altered to
ensure that the updated probabilities P*i,1,. . .,P*i,K are in [0,1] and sum up to 1. Such
extensions are discussed in the next section “Accounting for neighbor wells.”
The following application examples are based on the dataset used in Garner
et al. (2015).
It can be observed in Fig. 1 that the final facies is intermediate between the initial
electrofacies and the facies corresponding to the highest a priori probability, as
expected. The lower the uncertainty assigned to the electrofacies, the more similar
to the electrofacies is the final result. When the variance characterizing this
uncertainty increases, the final result becomes closer to the a priori geological trend.
It can be noted in Figs. 1 and 2 that the geological trend is not identical to the
original reference facies. In particular, some shoulder effects exist at each facies
border, which are due to the fact that both data are not at the same scale. The
geological trend is less precise than the facies defined from core description. In
some cases, like at the right side of Fig. 2, some contradictions between the original
reference facies and the geological trend can be observed. Such discrepancies may
occur, due to local heterogeneities or mistakes in the estimation of the sedimento-
logical trends. In that case, combining electrofacies and geological trend cannot
restore the original facies.
Figures 3 and 4 show a typical practical issue in electrofacies determination. One
of the clusters in the multivariate distribution based on NPHI, RHOB, DT, and GR
logs is not very efficient for discriminating the existing facies in a supervised
classification, as shown in Fig. 3. The points in the cluster correspond to all of
the four facies, which may lead to classification mistakes. Such an effect is shown in
Fig. 4 in one well (W13) which is fully included in the selected cluster (see purple
points in the figure). The initial electrofacies calculation misses the facies 2, which
is actually dominant, and includes occurrences of facies 3 which is actually not
present in this well. This misclassification can be partially fixed by accounting for
spatial trend.
Using Spatial Constraints in Clustering for Electrofacies Calculation 465
Fig. 1 Sensitivity of updated facies to “measurement error” variance values. From left to right, in
both tables: the reference facies, the electrofacies, the geological trend, and the result of the
approach described above
Fig. 2 Examples of updated facies in different wells. From left to right, in both tables: the
reference facies, the electrofacies, the geological trend, and the result of the approach described
above
466 J.-M. Chautru et al.
125 100 75 50 25
2.6
RHOB
2.4
GR
2.2
2.0
DT NPHI
1 2 3 4
The selected cluster (in green)
0.4 0.3 0.2 0.1 0.0
Facies code
Fig. 3 Ambiguous cluster in logs leading to potential classification errors
With this method, each updated facies probability is between 0 and 1 and their sum
is equal to 1, the two parts of the linear combination honoring this constraint.
It is recommended to apply this method on a well-by-well basis, with specific
variance of “measurement error” for each well. The method is useful when several
facies cannot be discriminated by the classification technique but can be distin-
guished according to the location in the reservoir. The main practical issue will be
the determination of the most appropriate value for the variance representing the
relative level of confidence on the two sources of data. It is a user-defined value, but
some simple rules can be considered:
• With a null variance of “measurement error,” only the classification results are
considered.
• With a variance of “measurement error” equal to 1, classification results and a
priori geological trend have the same weight.
• It is interesting to compare the expected facies coming from the geological trend
with the initial electrofacies determined by the classification technique. In case
of strong discrepancies, it is highly recommended to revisit the geological trend,
to be sure of its validity, and/or to check in which cluster is the well under study.
If it is located in an ambiguous cluster, unable to discriminate the facies, then
electrofacies are uncertain, and it is better to put more weight on the geological
trend. By the way, it can be noted that comparing electrofacies and geological
trends is a way to QC the data.
Using Spatial Constraints in Clustering for Electrofacies Calculation 467
• In a cored well, it is easy to check the similarity of the final facies profile with the
initial description. In uncored wells, this check is impossible, but it is possible to
compute the proportions of each facies in the final result and compare it with the
average facies proportions at the well location. It is a way to check the geological
consistency of the result, assuming that the geological trend is reliable.
It is important to keep in mind that the geological trend which is used in this
method results from the quantification of large-scale trends defined by the sedi-
mentologist. Therefore, it is semi-qualitative information, which may be affected
by a significant level of uncertainty. If it is defined from many cored wells, the
accuracy of this trend is high, but if only few cored wells are available, the
corresponding facies average probability of presence have to be used with care.
Ultimately, the trend is a modeling choice, and its accuracy must be taken into
account in the definition of the linear combination weights.
A method to account for geological trend uncertainty and to quantify the
uncertainty associated with the final electrofacies consists in defining ranges of
variation for the average probability of the presence of each facies. Several
468 J.-M. Chautru et al.
4.1 Principle
In many cases, electrofacies cannot be calculated in all the wells in the same
conditions, due to data issues:
• The optimal set of logs may be available in some wells only, and the
electrofacies in the other wells have to be calculated with fewer logs and/or
with different logs;
• Due to strong differences in the date of logs acquisition, the accuracy of the data
used for classification may vary a lot. Therefore, the quality of the classification
results varies from one well to another.
The consequence of these data issues is that some classes which are properly
discriminated in some wells are difficult or even impossible to distinguish in other
wells. These issues are practical and can be aided by database preparation and
thorough log normalization procedures used by petrophysicists.
A way to enhance electrofacies determination in a well where data quality is
questionable is to account for the neighbor wells where the data are of high quality,
as illustrated in Fig. 5. It assumes that the underlying geological phenomenon is
quite continuous, at the well-spacing scale.
In this method, the different variables (probabilities to belong to class) are linked
together as they sum up to 1. As several samples are now included in the kriging
neighborhood, this constraint is not automatically honored, and normalization pro-
cedures or specific kriging approaches have to be considered.
A first simple approach consists in performing successive independent kriging
estimations for each variable, followed by a rescaling. On a numerical point of
view, the constraint on the sum of probabilities will be honored. On a geological
and mathematical point of view, considering as independent for calculations vari-
ables which are not independent is a strong approximation which may lead to
significant inconsistencies.
An alternative approach could be to use kriging or cokriging of compositional
data, where all the probabilities to belong to a facies are grouped in vectors. This
approach requires the use of preliminary transformations before applying kriging or
cokriging. Such transformations are generally based on logarithms, such as the
additive log-ratio transformation of Aitchison (1986) detailed in Pawlowsky et al.
(1994). It has been noted (Lan et al. 2006) that using log ratios in multiscale facies
modeling leads to issues, mainly related to null probabilities. In our context, such
occurrences are frequent, some facies having null proportions at different locations
in the reservoir.
Compositional kriging as introduced by Walvoort and de Guijter (2001) can be
considered instead. It extends ordinary kriging to the specific context of composi-
tional data by simply adding the non-negativity and constant sum constraints to the
minimization of the kriging mean squared error. Precisely, going back to Eq. 2, the
compositional kriging weights are determined by solving the following optimiza-
tion problem in vectorial form:
470 J.-M. Chautru et al.
min E Z* Z 22 subject to E Z * Z ¼ 0, Z* 0 and Z* 1 ¼ 1
In the present context, Z* stands for the full vector P*i at a sample of interest i,
predicted by cokriging with neighboring vectors Pj. There, Z represents the ideal
value Pi that would have been obtained if the initial classification procedure was
perfect. In contrast to the additive log-ratio transformation techniques, no restric-
tion on the values of the involved vectors is necessary; some components can be
null without impeding estimation. Moreover, all components of the target vector are
assessed simultaneously with this procedure.
Initially, the authors chose not to take cross correlations into account; in the
present context, this means that the probability of the presence of a facies at a target
point is not influenced by those of the other facies at other locations. Thus, each
type of rock is compared to all others as a group instead of individually. When such
an assumption is considered too restrictive, given the depositional patterns, it is
possible to generalize the method to the context of cokriging. This would be the
case, for instance, when two types of rock are commonly found next to each other; a
sample consisting of the first would increase the probability of finding the second in
a close neighborhood. Even if such types of dependence are often nonlinear,
including extra facies in the cokriging formula can help capture at least part of
the depositional dynamic. Like before, expert knowledge about geological trends
can be included in the model through the vector of mean class probabilities at the
target sample. Because the latter vary through space, so proceeding has the advan-
tage of relaxing the stationarity hypothesis. The covariance matrix of the data can
also be altered to reflect the level of confidence in the statistical clustering results; in
practice, this amounts to supposing that the data contain measurement errors. The
main drawback of picking cokriging over kriging is the amount of extra parameters
to estimate: estimation efficiency decreases when the latter increase. It also requires
stipulating a cross-correlation model, which can be difficult when samples are
limited.
Fig. 7 Facies estimation in W7 from electrofacies in W10, W14, and W15, accounting for
geological trend in W7. Result is the last column on the right
parameters. They have to be picked with care, maybe using core data in a super-
vised approach.
Several options for using electrofacies defined from the results of classification
based on well logs are proposed in Garner et al. (2015) to build 3-D facies models:
• Plurigaussian simulations (PGS) using electrofacies as input data
• 3-D simulation of electrofacies probabilities, known at wells from classification,
the final electrofacies at each point corresponding to the highest probability
• 3-D co-simulations of the well logs and application of the well-derived classi-
fication as a post-processing
The methods detailed in the current paper lead to an update of electrofacies
probabilities to account for geological trends. Therefore, the three facies modeling
methods will be affected, but at different levels.
For PGS, accounting for geological trends in electrofacies definition will lead to
a modification of input data (electrofacies). The lithological column may change in
some wells, which can lead to local adjustments of the vertical proportion curves
and to variogram fitting updates. It must be noted here that in PGS, the 3-D
proportion model used as soft input data is usually calculated from the electrofacies
at wells. The a priori geological trend used to modify classification results, which is
similar to the usual 3-D proportion model, is based on core data and on sedimen-
tologist interpretation. Both models are actually very close, and the one used to
update classification results could be used in PGS as an alternative choice.
The input data conditioning the 3-D simulation of electrofacies probabilities are
modified and include a contribution of the a priori geological trend. The whole
methodology shown in Fig. 8 can remain unchanged.
Details can be found in Garner et al. (2015) where this methodology is fully
described. It can be noted that the first steps of the procedure (normal score trans-
forms and MAF) can be applied to the geological trend itself. The results can be
used as external drifts for constraining the turning bands simulations. It is a way to
force this method to account for facies proportions in the whole 3-D model.
The first step of the third method, based on well logs co-simulations, will not be
affected by the classification results update. This part of the procedure remains
unchanged, but the next step, which consists in applying the classification technique
to the 3-D model of logs, will be improved. Classification results can be updated, in
the model, to account for the a priori geological trend. Here, the auxiliary spatial
information is introduced at the last moment in the calculation.
Using Spatial Constraints in Clustering for Electrofacies Calculation 473
6 Conclusion
Bibliography
Aitchison D (1986) The statistical analysis of compositional data. Chapman & Hall, London
Garner DL, Yarus J, Srivastava M (2015) Modeling three ways from electro-facies – categorical,
e-facies probabilities and petrophysics with assignment. Petroleum Geostatistics 2015,
7–11 September. Biarritz
474 J.-M. Chautru et al.
Lan Z, Leuangthong O, Deutsch CV (2006) Why Logratios are a bad idea for multiscale facies
modeling, Centre for computational geostatistics report 8. University of Alberta, Canada, p 211
Pawlowsky V, Olea RA, Davis JC (1994) Additive logratio estimation of regionalized composi-
tional data: an application to the calculation of oil reserves. In: Dimitrakopoulos R
(ed) Geostatistics for the next century. Kluwer, Dordrecht, pp 371–382
Walvoort D, de Guijter J (2001) Compositional kriging: a spatial interpolation method for
compositional data. Math Geol 33(8):951–966
Pore Network Modeling from Multi-scale
Imaging Using Multiple-Point Statistics
1 Introduction
T. Chugunova (*)
Geostatistics and Uncertainty Service, Total SA, av. Larribau, CSTJF, 64000 Pau, France
e-mail: [email protected]
H. Berthet
Digital Rock Physics Laboratory, Total SA, av. Larribau, CSTJF, 64000 Pau, France
e-mail: [email protected]
2 Case 1
A 4-mm cylindrical sample of Castle Gate rock (outcrop sandstone, Utah, USA)
was cored and placed on the sample holder of an X-ray microtomograph
(Versa520XRM, Zeiss). A first image was obtained at a voxel resolution of
1.5 μm (field of view ~33 mm3). In order to access a higher level of details, we
carried out a series of imaging at a voxel resolution of 503 nm (Fig. 1), now
reconstructing a sub-volume of the initial image (so-called interior tomography or
region of interest). Two images were acquired at two successive vertical positions
and stitched together to study a larger imaged volume (~0.13 mm3).
Fig. 1 2D images extracted from the 3D-reconstructed image at 503 nm voxel resolution. Large
grains (a) and micro-porosity (b) are present in the structure. The edges of large grains display a
phase contrast artifact due to the very high resolution of the image
478 T. Chugunova and H. Berthet
between 0 and 1. First of all we improve the contrast of the image to better
distinguish the grains and the pores. Now we clearly see that the image contains
two main spatial structures corresponding to large grains and small particles.
Second, we found that even the smallest particles are represented by scores of
pixels in the initial image. The efficiency of MPS algorithm is not optimal when that
most of the neighbors in the template are wasted to capture same values and not
spatial heterogeneity. The image is rescaled (3 3 average mean upscaling) in
order to optimize the representation of the given heterogeneities. Global statistics
analysis confirms an absence of bias neither the loss of information after upscaling.
Fig. 5 Training image of categorical variable (six classes) and three DS realizations. TI is a
second image from the left
from large or small structures. Practically, it means that the same combination of
neighbors corresponding to a current simulation step may be compatible with both
small and large structures. Consequently, the motifs started as large structures can
be continued with a pattern corresponding to the small structures and inversely,
which in its turn explains the quality of the results.
3 Case 2
A 11-mm diameter cylindrical rock sample was cored from a reservoir sandstone
petrophysics plug. The rock was chosen for its large porosity (25.4 %) and perme-
ability (300 mD) and the presence of very small pore structures (mainly clay). A
first large 3D image was obtained by successively imaging the sample at seven
different vertical positions and by stitching all the images together. The result field
of view is around 1.6 cm3 at a voxel resolution of 3.1 μm (later referred to as the low
resolution). A region of the rock containing large and small structures was located
in the reconstructed volume and subsequently imaged at a voxel resolution of
905 nm (field of view ~5.9 mm3, later referred to as the high-resolution image).
Our second case study is illustrated Fig. 6. A large low-resolution (LR) image exists
for a statistically representative sample. A high-resolution (HR) image is available
only for a small part of the low resolved one. This HR part contains several types of
heterogeneities (grains, pores, different forms of clay). Calculated only on this
small HR part, the quantity of each type of heterogeneities is not reliable, but the
shapes are correctly represented. The idea is then to extrapolate different types of
the heterogeneities seen on the HR image to the initial large sample. From the
geostatistical point of view, it remains to apply a downscaling to the LR image.
To do so, we use a DS technique constrained by an auxiliary variable. The idea
of using an auxiliary variable to characterize a TI and to constrain an MPS
simulation was first proposed by Chugunova and Hu (2008) in its collocated
co-simulation form in order to deal with nonstationary TI with a tree-based MPS
algorithm. Then, the use of auxiliary variable was generalized to a DS algorithm
with non-collocated co-simulation (Mariethoz 2009). Recently, the auxiliary vari-
able approach was used to capture multiple-point statistics of geological heteroge-
neities on several scales in the reservoir modeling context (Doliges et al. 2014). In
this work, we propose to use auxiliary variable technique in order to link statisti-
cally two different scales represented by HR and LR images.
The principle of this technique can be described as follows. Let’s take HR TI
image for a small sample and its equivalent in LR (TI and Auxiliary TI in Fig. 7).
By extracting the multiple-point statistics simultaneously from both images, we
may statistically estimate a relation between the patterns in the HR and the
corresponding values of the LR. Then, for the simulation step, we impose the
482 T. Chugunova and H. Berthet
Fig. 6 LR image capturing large sample (on the left) and HR image capturing small sample (on
the right)
In this downscaling test, we transform the HR raw image into the TI in the similar
way as in the Case 1: by improving contrasts and segmenting into several facies.
Note that for the Case 2 in HR image, the smallest heterogeneities are resolved by
only several pixels. The upscaling process will induce the loss of information and
will reduce the accuracy of the petrophysical properties calculation. Thus, the
resolution of the HR image was considered as the optimal, and no upscaling was
applied.
To illustrate the results, we show one part of the target simulation grid which has
approximately the same size as a HR TI. Intentionally, this illustrated zone contains
a part of HR TI (Fig. 7). Thereby, we could compare the DS simulation in two
zones: one covering the LR patterns with their exact replicates in the TI and another
where only similar patterns exist.
Pore Network Modeling from Multi-scale Imaging Using Multiple-Point Statistics 483
both grain and clay structures, whereas the DS algorithm has no criteria to discrim-
inate them. As statistically it is more probable to find grain pattern, that is what
happens the most during the simulation process.
A way forward can be envisioned to improve clay heterogeneities simulation.
Knowing that the regions inside grains and pores are already clearly indentified in
the LR image, we may reduce our TI image as well as the simulation zone only to
the heterogeneities of interest such as pore edges and clay regions.
4 Conclusions
In this work we propose two approaches to tackle the problem of multi-scale and
multi-support pore network reconstruction with MPS. Two case studies applica-
tions illustrate these approaches.
In Case 1 we deal with relatively simple spatial structures but containing two
scales of heterogeneities: large grains and small particles. By applying an image
treatment (contrast improving, segmentation in several classes), we transform the
raw X-ray image in the reliable TI. Applying DS technique with flexible template,
we demonstrate a good-quality internal structure reproduction. The practitioner
should nevertheless pay attention to the ergodicity aspects. Indeed, the sole TI of
case 1 is non-ergodic, and it is recommended to use multiple TIs (or 3D TI) to cover
a representative area of sample and to obtain statistically robust results.
In Case 2 we deal with two different supports (resolutions) but are also
confronted with practical difficulties. The rock sample is characterized by two
images with different resolutions and cover different volumes. The TI contains
several very different spatial structures. The structures of interest are minatory
comparing to the most represented patterns of pores and grains. To reconcile two
different resolutions, we propose to apply the DS co-simulation with auxiliary
constraint where HR image will stand for a TI and the LR image – for a constraint.
We illustrate the results of DS application using the entire TI. The unresolved
Pore Network Modeling from Multi-scale Imaging Using Multiple-Point Statistics 485
features like vague edges or filamentary clays are correctly resolved by this
approach. But some of features instead of being simulated as clay heterogeneities
are simulated as grains. This phenomenon is explained by the most frequent
patterns in the TI which are the grains. As a way forward to solve these practical
difficulties, we propose to reduce the TI scan area and the simulation area to the
uncertain zone where the downscaling is necessary and so increase the frequency of
rare clay patterns in the TI.
Having learned from these two cases, the following can be drawn. In case of
heterogeneities on several scales and a spatial organization between scales (Case 1:
small particles appear where the boards of large grains become broken), we
recommend the first approach. In case where the spatial structure dramatically
changes while changing a resolution of scan and a precision is needed for more
accurate calculation (Case 2), we recommend the second approach.
Bibliography
Blunt MJ, Bijeljic B, Dong H, Gharbi O, Iglauer S, Mostaghimi P, Paluszny A, Pentland C (2013)
Pore-scale imaging and modeling. Adv Water Res 51:197–216. doi:10.1016/j.advwatres.2012.
03.003
Chugunova T, Hu L (2008) Multiple-point simulations constrained by continuous auxiliary data.
Math Geosci 40(2):133–146
Doligez B, Ravalec M, Bouquet S, Adelinet M (2014) Multiscale reservoir geological modeling
and advanced geostatistics. In: Proc. Gussow conference “Closing the Gap II”, Sept. 22–24
Banff, Canada
Emery X, Lantuejoul C (2013) Can a training image be a substitute for a random field model. Math
Geosci SI. doi:10.1007/s11004-013-9492-z
Hu L, Chugunova T (2008) Multiple-point geostatistics for modeling subsurface heterogeneities: a
comprehensive review. Water Resour 44(11) (W11,413), doi:10.1029/2008WR006993
Leu L, Berg S, Ott H, Amstrong RT, Enzmann F, Kersten M (2014) Sensitivity study of an image
processing workflow on synchrotron μ-CT images of Berea sandstone. Geophys Res Abstr 16:
EGU2014–EGU10431
Mariethoz G, Lefebvre S (2014) Bridges between multiple-point geostatistics and texture synthe-
sis: review and guidelines for future search. Comput Geosci. doi:10.1016/j.cageo.2014.01.001
Mariethoz G, Renard P, Froidevaux R. (2009) Integrating collacated auxiliary parameters in
geostatistical simulations using joint probability distributions and probability aggregation.
Water Resour Res 45(W08421)
Mariethoz G, Renard P, Straubhaar J (2010) The direct sampling to perform multiple-point
geostatistical simulation. Water Resour Res 46(W11536), doi 10.1029/2008WR007621
Okabe H, Blunt MJ (2005) Pore space reconstruction using multiple-point statistics. J Pet Sci Eng
46-2005 p. 121–137 Elsevier
Strebelle S (2002) Conditional simulation of complex geological structures using multiple-point
statistics. Math Geol 34(1):1–21. doi:10.1023/A:1014009426274
Bernstein Copula-Based Spatial Stochastic
Simulation of Petrophysical Properties Using
Seismic Attributes as Secondary Variable
1 Introduction
2 Methodology
As stated in the introduction, the main goal of this work is to show the application of
a Bernstein copula-based spatial co-simulation method for petrophysical property
predictions using seismic attributes as secondary variables and its comparison with
the classical sequential Gaussian co-simulation method. In what follows, a brief
description of both methods and a general workflow outline are presented.
2014) and has been mainly applied in one dimension for petrophysical properties at
well-log scale.
The method basically consists of establishing a dependence model between a
primary and a secondary variable and then uses this model in conjunction with the
spatial dependence structure (variogram) of the primary variable to predict the first
one using the second one as a conditioning variable. This can be done in a global
optimization framework using simulated annealing method, but other methods,
such as genetic algorithms, could also be applied.
The modern way to analyze dependencies is by copula approach (Joe 1997).
Copula approach assumes neither a predetermined nor a priori type of dependency,
but from the data one tries to establish the best model that represents the existing
dependence on them.
In particular, here it is preferred to use a completely nonparametric approach to
modeling the dependence by using Bernstein copulas, which gives name to the
method. However, other approaches, parametric (Dı́az-Viera and Casar-González
2005) and semi-parametric (Erdely and Diaz-Viera 2010), are also possible. The
Bernstein copulas introduced by Sancetta and Satchell (2004) are nothing more
than an approximation of the sample copula by Bernstein polynomials. Its main
shortcoming is the curse of dimensionality, as it quickly becomes computationally
prohibitive for more than two dimensions. Alternatives have been proposed using
vine copulas (Erdely and Diaz-Viera 2016).
In summary, the algorithm consists of two stages:
1. A dependence model, using a Bernstein copula, is established from which a
number of sample values are generated (see Appendix A).
2. A stochastic spatial simulation is performed using a simulated annealing method
with a variogram model and a bivariate distribution function as objective
functions (Deutsch and Cockerham 1994; Deutsch and Journel 1998).
Additional details about the mathematical formulation of the method and its
computational implementation can be found on Hernández-Maldonado et al. (2012,
2014).
because it is largely dependent on the data. Here we will use the median as
upscaling procedure.
The general workflow is as follows: (1) univariate data analysis, (2) bivariate
dependence analysis, (3) variography analysis, and (4) simulations.
3 Case Study
Data used in the case study are from a marine reservoir in the Gulf of Mexico. The
reservoir is siliciclastic and it is formed mainly by alternating sequences of sands
and shales.
The data consist of a total porosity well-log from a single well and seismic attribute
(P-impedance) obtained in a vertical (inline) section. The well log has a sample
interval of 0.1 m. The section has a length of 412.5 m and covers an interval of
336.4 m in depth and was chosen so that the well was located in the middle of it (see
Fig. 1). Seismic grid is made of 33 intervals of 12.5 m in X direction and 60 intervals
of 5.5 m in depth direction.
At well-log scale, the P-wave impedance log is obtained from the product of
P-wave velocity and density logs. At the seismic scale, a seismic inversion method
was used, based on the “LP sparse spike” approach by Li (2001). Proper care was
taken in incorporating low-frequency impedance trend so that the impedance from
the log and the impedance from the seismic section are equivalent around the well.
The impedance in general depends on the type of rock and its petrophysical
properties as well as the containing fluid types and their saturations. It is very
common in reservoir geophysics to take advantage of dependency relationships
between petrophysical properties (for instance, total porosity and P-impedance) at
well-log scale to predict the former ones (total porosity) using seismic attributes
(P-impedance) at the seismic scale.
In particular, in this work the total porosity was considered as primary variable
(variable to predict) and P-wave impedance as secondary variable (conditioning
variable). As mentioned in the previous section, three scales are considered: a well-
log, a seismic, and an additional intermediate “one-meter” scale. Hereafter, the
following notation will be used:
– PhiT_T and Ip_T for total porosity and P-impedance from original well logs
(well-log scale)
– PhiT_T_1m and Ip_T_1m for total porosity and P-impedance from original well
logs subsampled every meter (1-m scale)
Bernstein Copula-Based Spatial Stochastic Simulation of Petrophysical. . . 491
Fig. 1 Vertical (inline) section with P-impedance as a result of seismic inversion. The color scale
represents impedance values. In the middle of the section, two logs are plotted along a well: in
yellow P-impedance and in green total porosity
In Fig. 2 are shown histograms and boxplots for PhiT and Ip at the three scales, and
in Tables 1 and 2 a summary of corresponding basic univariate statistics.
Note that median and mean are pretty close for 1-m scale and seismic scales,
while Ip_inline_U and Ip_inline have very consistent statistics.
In Fig. 3 are given the scatterplots with marginal histograms and boxplots for PhiT
vs. Ip (a) at well-log, (b) at 1-m, and (c) at seismic scale, respectively, while in the
Fig. 3d the corresponding scatterplot for a non-conditional Ip-PhiT simulation using
a Bernstein copula at 1-m scale is shown. In Fig. 4 are given pseudo-observations
(sample copula) scatterplots for Ip-PhiT, (a) at well-log scale, (b) at 1-m scale, (c) at
492 M.A. Dı́az-Viera et al.
Fig. 2 Histograms and boxplots for PhiT and Ip at well-log scale (a, b), at 1-m scale (c, d), and at
seismic scale (e, f), respectively
Bernstein Copula-Based Spatial Stochastic Simulation of Petrophysical. . . 493
Table 2 Statistics summary of median upscaled well logs and Ip at seismic scale
Statistics PhiT_T_med Ip_T_med Ip_inline_U Ip_inline
n 60 60 60 1980
Minimum 0.110 5730.23 5940.26 5940.16
First quartile 0.239 6074.77 6080.53 6083.72
Median 0.273 6426.24 6350.74 6340.04
Mean 0.267 6619.13 6623.50 6602.87
Third quartile 0.296 6967.11 6896.91 6893.98
Maximum 0.340 10,430.77 8726.18 8773.20
Variance 0.002 630,041 577,250 485,570
seismic scale, and (d) for a non-conditional bivariate simulation with a Bernstein
copula at 1-m scale. In Table 3 a summary of corresponding correlation (Spearman
and Pearson) coefficients is given. It can be observed that the dependence is
weakened with the increasing of the scale.
In Fig. 5 are shown estimated variograms (a, b) and best-fit variogram models (c, d)
for PhiT and Ip at seismic scale in depth direction. As is evident in Fig. 5b, the
sample variogram of Ip_inline_U shows a typical behavior related with the pres-
ence of trend, which means that at least the intrinsic hypothesis is not satisfied.
Then, the trend, which in this case was of second order, was estimated and removed,
resulting a new variable Ip_inline_U_r2 without trend. The same previous proce-
dure was applied to Ip_Inline and a resulting detrended variable was named
Ip_inline_r2. Note, in Fig. 5d the variogram was obtained after removing trend
from Ip_inline_U. While in the Fig. 6 are displayed estimated variograms and best-
fit variogram models for impedance at seismic scale in (a) X and (b) depth direc-
tions, respectively, after removing trend from Ip_inline.
494 M.A. Dı́az-Viera et al.
Fig. 3 Scatterplots with marginal histograms and boxplots for Ip vs. PhiT, (a) at well-log scale,
(b) at 1-m scale, (c) at seismic scale, and (d) a non-conditional bivariate simulation with a
Bernstein copula at 1-m scale
Because of the lack of data for total porosity in the X direction, the same
variogram structure of the impedance in this direction is adopted, considering that
they show almost the same structure in the depth direction (see Fig. 5). A variogram
model for porosity at seismic scale is proposed so that the total variance of PhiT_T
median is preserved which basically it is to consider a correlation range equal to the
impedance variogram in the X direction (see Fig. 6). For both simulation methods,
the following variogram model for porosity is used: model¼spherical,
nugget ¼ 0.0002, and structure contribution ¼ 0.0016; ranges, maximum ¼ 160,
medium ¼ 50, and minimum ¼ 1; angles, x ¼ 90, y ¼ 0, and z ¼ 0.
Bernstein Copula-Based Spatial Stochastic Simulation of Petrophysical. . . 495
Fig. 4 Pseudo-observations (sample copulas) scatterplots of Ip-PhiT data ranks, rescaled to [0,1]
(a) at well-log scale, (b) at 1-m scale, (c) at seismic scale, and (d) a non-conditional bivariate
simulation with a Bernstein copula at 1-m scale
Table 3 Summary of correlation coefficients for Ip vs. PhiT at well-log, 1-m, and seismic scales
and for a non-conditional bivariate Ip-PhiT simulation with a Bernstein copula at 1-m scale
(BCS_1m)
Correlation coefficients Well-log scale One-meter scale Seismic scale BCS_1m
Spearman 0.589 0.477 0.361 0.576
Pearson 0.711 0.657 0.529 0.703
496 M.A. Dı́az-Viera et al.
Fig. 5 PhiT and Ip estimated variograms (a, b) and best-fit variogram models (c, d) at seismic
scale in depth direction. Note, in d variogram after removing trend from Ip
Fig. 6 Estimated variograms and best-fit variogram models for Ip at seismic scale after removing
trend in (a) X and (b) depth directions, respectively
The resulting simulation is named PhiT_SGC and its map in the vertical (inline)
section is given in the Fig. 8a.
Fig. 7 (a) Scatterplot with marginal histograms and boxplots and (b) pseudo-observations
(sample copulas) scatterplot for 40,000 conditional bivariate simulations using a Bernstein copula
at 1-m scale
Fig. 8 Maps for (a) a PhiT sequential Gaussian co-simulation, and (b) a PhiT Bernstein copula-
based co-simulation in the vertical (inline) section
In comparison with the SGCS method, the BCSCS method better reproduces the
statistics in terms of variance and extreme values (see Fig. 9), and both methods
reproduce quite well the spatial structure (see Fig. 10), but the sequential Gaussian
co-simulation shows spurious correlation dependence, which does not exist in the
original data, highlighted in red color in Fig. 11. This is the main reason of the
difference between Fig. 8a and 8b.
Bernstein Copula-Based Spatial Stochastic Simulation of Petrophysical. . . 499
Fig. 9 PhiT histograms and boxplots for (a) a sequential Gaussian co-simulation and (b) a
Bernstein copula-based co-simulation, respectively
Fig. 10 Ip vs. PhiT scatterplots with marginal histograms and boxplots for (a) a sequential
Gaussian co-simulation and (b) a Bernstein copula-based co-simulation, respectively. Simulated
values with spurious dependence, which does not exist in the original data, are highlighted in red
color
Fig. 11 Estimated variograms and best-fit variogram models in X and depth directions for a
sequential Gaussian co-simulation (a, b) and a Bernstein copula-based co-simulation (c, d),
respectively
linear dependence between variables; on the contrary, it can capture and reproduce
the existing dependence between them. The method is nonparametric which means
that it does not need a specific distribution function. Since the Bernstein copula is
based on the sample distribution function, it may reproduce the variability and the
extreme values. Another advantage is that there is no need to make back trans-
formations, which are potentially biased, since copulas are invariant under strictly
increasing transformations.
Instead of using a single seismic attribute, it could be used the best combination
of them depending on the primary (explanatory) variable by applying standard
multivariate statistical procedures such as principal component and factorial
Bernstein Copula-Based Spatial Stochastic Simulation of Petrophysical. . . 501
analysis. Another option would be using a multivariate copula with three or more
variables to directly exploit their dependencies.
This work can be easily extended to 3D problems but it depends on the
computing power available. Although in this work the aim was to show the
performance of the simulation method, a simpler and efficient alternative, perhaps
more convenient for 3D large problems, could be the median regression approach
already shown in previous works (Erdely and Diaz-Viera 2010; Hernández-
Maldonado et al. 2012) in 1D.
Acknowledgments The present work was supported by the IMP project D.61037 “Interpretación
Sı́smica Cuantitativa Guiada por Litofacies para la Caracterización de Yacimientos.”
CXY is called copula function associated to (X, Y) and contains information about
the dependence relationship between X and Y, independently from their marginal
probabilistic behavior. CXY is uniquely determined on Ran FX Ran FY , and
therefore, if FX and FY are continuous, then CXY is unique on [0, 1]2. Among several
properties of copula functions, see Nelsen (2006), we have the following:
• Cðu; 0Þ ¼ 0 ¼ Cð0; vÞ
• Cðu; 1Þ ¼ u, Cð1; vÞ ¼ v
• Cðu2 ; v2 Þ Cðu2 ; v1 Þ Cðu1 ; v2 Þ þ Cðu1 ; v1 Þ 0 if u1 u2 , v1 v2
• C is uniformly continuous on its domain [0, 1]2.
• The horizontal, vertical, and diagonal sections of a copula C are all
nondecreasing and uniformly continuous on [0, 1].
502 M.A. Dı́az-Viera et al.
X
n
^ ðu; vÞ ¼ 1
C 1fuk u , vk vg
n k¼1
Xn Xn
^ ðxÞ ¼ 1
F ^ ðyÞ ¼ 1
1fxk xg F 1fy yg
X Y
n k¼1 n k¼1 k
Now the set of pairs fðuk , vk Þ ¼ F^ ðxk Þ, F
^ ðyk Þ : k ¼ 1, . . . , ng is referred to as
X Y
^ ðxk Þ ¼ 1 rankðxk Þ
copula pseudo-observations. It is straightforward to verify that F X n
^ ðyk Þ ¼ 1 rankðyk Þ. In this case the concept of empirical copula, see Nelsen
and F Y n
Bernstein Copula-Based Spatial Stochastic Simulation of Petrophysical. . . 503
Bibiography
Bardossy A, Li J (2008) Geostatistical interpolation using copulas. Water Resour Res 44:1–15
Chiles JP, Delfiner P (1999) Geostatistics: modeling spatial uncertainty. Wiley series in probabil-
ity and statistics. Applied probability and statistics section. Wiley, New York
Deutsch CV, Cockerham PW (1994) Geostatistical modeling of permeability with annealing
cosimulation (ACS). SPE 69th Annual Technical Conference and Exhibition. Society of
Petroleum Engineers, New Orleans, p 523–532
Deutsch CV, Journel AG (1998) GSLIB: geostatistical software library and user’s guide, 2nd edn.
Oxford University Press, New York
Dı́az-Viera M, Casar-González R (2005) Stochastic simulation of complex dependency patterns of
petrophysical properties using t-copulas. Proc IAMG’05: GIS Spat Anal 2:749–755
Erdely A, Diaz-Viera M (2010) Nonparametric and semiparametric bivariate modeling of
petrophysical porosity-permeability dependence from well-log data. In: Jaworski P,
Durante F, Härdle W, Rychlik T (eds) Copula theory and its applications. Lecture notes in
statistics 198. Springer, Berlin, pp 267–278
Erdely A, Diaz-Viera M (2016) A vine and gluing copula model for permeability stochastic
simulation. In: Bozeman J R, Oliveira T, Skiadas CH (eds) Stochastic and data analysis
methods and applications in statistics and demography, pp 199–207
Hernández-Maldonado V, Dı́az-Viera M, Erdely A (2012) A joint stochastic simulation method
using the Bernstein copula as a flexible tool for modeling nonlinear dependence structures
between petrophysical properties. J Petrol Sci Eng 92–93:112–123
Hernández-Maldonado V, Dı́az-Viera M, Erdely A (2014) A multivariate Bernstein copula model
for permeability stochastic simulation. Geofis Int 53(2):163–181
Joe H (1997) Multivariate models and dependence concepts. Chapman & Hall, London
Kazianka H, Pilz J (2010) Copula-based geostatistical modeling of continuous and discrete data
including covariates. Stoch Environ Res Risk Assess 24(5):661–673
Li Q (2001).LP sparse spike impedance inversion. Hampson-Russell Software Services Ltd. –
CSEG
Nelsen RB (2006) An introduction to copulas (Springer series in statistics), 2nd edn. Springer,
New York
Parra J, Emery X (2013) Geostatistics applied to cross-well reflection seismic for imaging
carbonate aquifers. J Appl Geophys 92:68–75
Remy N, Boucher A, Wu J (2009) Applied geostatistics with SGeMS: a user’s guide. Cambridge
University Press, New York
Sancetta A, Satchell S (2004) The Bernstein copula and its applications to modeling and approx-
imations of multivariate distributions. Economet Theor 20(3):535–562
Sklar A (1959) Fonctions the repartition a n dimensions et leurs marges. Publ Inst Stat 8:229–331
Robust MPS-Based Modeling via Spectral
Analysis
consistency between the training image and conditioning data in order to develop
robust models that have improved predictive ability. In case MPS-based simula-
tions are performed accounting for the uncertainty in TI, the method can be used to
rank the prior TIs so as to yield robust estimates for uncertainty.
1 Introduction
One of the most important issues in reservoir modeling is to generate maps of flow
and geologic attributes that play a great role in hydrocarbon production. Mostly, the
primary source of information is borehole measurements that are only available at
sparse locations. Integrating the available data along wells within robust models for
spatial continuity, multiple reservoir models can be generated that can be used to
quantify the uncertainty associated with reservoir predictions. Investigation of new
methods to integrate information from multiple sources that are at different reso-
lution and precision accuracy is still an active area of research.
Different algorithms have been developed to stochastically simulate reservoir
properties using sparse measured data. All these methods aim to draw realizations
of the random function Z(u) based on the joint distribution function describing the
random function. Some popular methods are sequential simulation (Journel 1983;
Isaaks and Srivastava 1989; Goovaerts 1997; Chiles and Delfiner 1999), iterative
approaches (Maksimov et al. 1993), etc. that have become the core tools in many
current geostatistical applications.
These simulation methods are based on variogram functions – a measure of
spatial variability based on a two-point spatial template. It has been pointed out that
in order to ensure legitimacy of the kriged/simulated values, the first step should be
to model the inferred variograms in such a way that the covariance matrix is
guaranteed to be positive definite. To achieve this, the variogram functions need
to be modeled using some well-known positive definite functions. Alternatively,
Yao (1998) proposed that the covariance function could be modeled
non-parametrically in Fourier domain to satisfy the positive definiteness assump-
tion. This overcomes the limitation of parametric modeling of the covariance
function, where one is limited to a few families of positive definite functions for
covariance modeling, which limits our ability to model complex heterogeneities.
The major drawback of this traditional variogram-based geostatistical modeling
is that they are not able to reproduce complex spatial patterns such as fluvial
channels. The variogram is inadequate to capture complex curvilinear features.
To reproduce the curvilinear structures and pattern continuity, the anisotropy
direction of variogram must be changed locally (Deutsch and Lewis 1992; Xu
1996; Boisvert and Deutsch 2011). One could also correct for additional connec-
tivity of the geological patterns by modifying the variogram ranges (Gringarten and
Deutsch 2001).
Robust MPS-Based Modeling via Spectral Analysis 507
All the previous algorithms are based on two-point connectivity functions only
and cannot reproduce complex curvilinear objects such as fluvial channels. Initially
Srivastava (1992) proposed to go beyond bivariate moments by introducing
multiple-point geostatistics that considers variability at more than two locations
taken jointly. Caers and Journel (1998) proposed the idea of borrowing conditional
probabilities directly from a training image (TI), allowing the use of higher-order or
multiple-point statistics to reproduce geological structures and patterns.
Strebelle (2000) developed the first structured multiple-point statistics algorithm
of “single normal (extended) simulation” (called snesim) for simulating categorical
variables. The approach of snesim is based on inferring the probability of various
outcomes at the central node of a spatial template based on the pattern of outcomes
on the remaining nodes of the template from the training image. In order to reduce
the computational complexity of computing the conditional probabilities, instead of
scanning the training image for each conditioning data template, it stores all
probabilities in a search tree by a one-time scanning of the training image. The
search tree data structure allows a fast retrieval of all required conditional proba-
bilities during the simulation. One of the shortcomings of snesim, however, is that
the training image needs to be categorical, and it does not work with continuous
variables.
Several other algorithms for mps simulation have been proposed. GrowthSim
(Eskandari and Srinivasan 2007; Huang and Srinivasan 2012) introduced the notion
of simulating multiple-point simulation event conditioned to multiple-point data
event in the vicinity of the simulation node. This method is in contrast to traditional
multiple-point statistics algorithms where the simulation progresses one node at a
time. Huang and Srinivasan (2012) demonstrated the GrowthSim algorithm for
developing the reservoir model for a deepwater turbidite system. They also showed
the capability of GrowthSim algorithm to represent nonstationary features.
Cumulants are combinations of statistical moments, such as mean or variance
that allow the characterization of non-Gaussian random variables (Rosenblatt
1985). Higher-order spatial cumulants can capture the complex geological features
and geometrical shapes of the physical phenomena. Dimitrakopoulos et al. (2010)
first used the concept of cumulants in the spatial context to characterize nonlinear
stationary and ergodic spatial random fields.
They showed that higher-order cumulants are related to the orientation of the
spatial template. Each geological process requires its own choice of cumulants for
optimal pattern analysis. They also showed that cumulants up to and including fifth
order are sufficient for efficiently characterizing complex spatial geometries
observed in training images.
Mustapha and Dimitrakopoulos (2010b) provided a computer code for calculat-
ing higher-order spatial cumulants. Furthermore, spatial cumulants were used as the
basis for the simulation of complex geological phenomena by Mustapha and
Dimitrakopoulos (2010b). The simulation takes advantage of spatial cumulants in
the high-dimensional space of Legendre polynomials in a sequential framework,
called hosim. It proceeds by randomly choosing a spatial node u, estimating the
conditional probability of the random variable given the neighboring data and
508 M. Elahi Naraghi and S. Srinivasan
previously simulated nodes and, finally, drawing a value for that node from the
distribution. The process repeats until all the nodes of the grid have been visited.
The process is similar to sequential simulation; the only difference is the method of
deriving the analytical expressions for the local probability density functions. In
hosim, Legendre coefficients are inferred using multiple-point spatial templates and
used to derive the expressions for multivariate conditional distributions. The pro-
posed cumulant-based method is assumed to be less dependent on the training
image statistics than snesim method but more data-driven; in the sense that it first
tries to infer the multiple-point statistics from the data, and only if not enough
replicates could be found, the training image will be used for inference.
Many pattern-based algorithms such as Filtersim (Zhang 2006), direct sampling
(Mariethoz and Renard 2010), Wavesim (Chatterjee et al. 2012), dispat (Honarkhah
and Caers 2010), CCSIM (Tahmasebi et al. 2012), and Bunch-DS (Rezaee et al.
2013) have been proposed to improve the computational efficiency of the simula-
tion algorithm over previously introduced ones.
However, all the proposed algorithms rely heavily on training images and the
patterns depicted in the TI. We discuss the development of an MPS algorithm that
can be applied even if an exhaustive training image is not available. We discuss the
inference of higher-order moments from available data.
This paper mainly focuses on selecting the best training image from a set of
multiple scenarios, which is the problem addressed in (Pérez et al. 2014) with
another point of view. This paper is organized as follows: we first explain what
higher-order cumulants and polyspectra are and describe their features. We present
how they can be calculated and what features can be inferred from them. We
subsequently discuss the inference of polyspectra from sparse conditioning data.
We demonstrate the application of the proposed algorithm on different reservoir
models such as channelized reservoir, fractured reservoirs, and models with elliptic
objects with different sizes and orientations. We study the sensitivity of the
algorithm to the available conditioning data. We show that the algorithm performs
well even when the conditioning data is very sparse.
2 Method
The modeling algorithm presented in this paper is based on two steps: (1) calculat-
ing the polyspectra from training images and conditioning data and (2) feature
extraction from polyspectra (size, orientation, and shape of the objects). These steps
are explained in the next sections.
Robust MPS-Based Modeling via Spectral Analysis 509
Z 1
The rth (r 0) moment of Z is Momr ½Z ¼ E½zr ¼ zr f Z ðzÞdz. Provided that
1
the moment-generating function M has a Taylor expansion about the origin:
X X
1 ωr zr 1 ω Mom½Z
r
M½ω ¼ E½eωz ¼ E ¼ ð2Þ
r¼0 r! r¼0 r!
Giannakis 1995; Giannakis and Mendel 1989). Many symmetries exist in argu-
ments of Mk, z ðh1 ; h2 ; . . . ; hk1 Þ making their calculation manageable.
By plugging Eq. 3 in Eq. 4, the power spectrum, bispectrum, and trispectrum can
be calculated by
where F (w) is the Fourier transform of Z (u) and F* is its conjugate transform. As
can be seen in Eqs. 5 and 6, the power spectrum is related to the amplitude of the
Fourier transform, but bispectrum is related to both amplitude and the phase of the
Fourier transform.
If one aims to infer the moments from a training image, the Fourier transform of
the training image can be used in Eq. 8 to calculate the polyspectra and use inverse
Fourier transform to calculate the moment as in Eq. 4. This would be computation-
ally more efficient than the case that the pairs with prescribed lag size are searched
in the training image to calculate the moments. For example, for a simple 1-D
training image with N data points, calculating covariance by calculating power
spectrum density will be done in O (N log N ) operations instead of O(N2)
(by searching the pairs in the training image and calculating covariance function
directly). In addition, this allows us to model the connectivity function without
constraining to a few parametric functions. For example, it will be shown in the
results section that using the covariance inferred from power spectrum instead of
using covariance models based on positive definite functions improves the repro-
duction of connectivity in the simulated image.
However, when the training image is not available, calculating the Fourier
transform from scatter data points (conditioning data) and inferring the
polyspectrum will be challenging. This in fact is the main reason that when the
training image is not available, one is only limited to variogram-based (two-point
statistics) simulation techniques. In order to solve this issue, we propose a compu-
tationally efficient and robust method for calculating higher-order spectra of a
random function from scattered data points using a nonuniform fast Fourier trans-
form (NUFFT) (Fessler and Sutton 2003; Keiner et al. 2009; Khalighi et al. 2015;
Drach et al. 2015).
Robust MPS-Based Modeling via Spectral Analysis 511
This problem is ill-posed and has multiple solutions. However, a unique solution
can be found by imposing a sparsity constraint on x, by solving the following l1
norm optimization problem:
This is the basis pursuit problem in the field of signal processing (Chen and
Donoho 1994). It should be noted that because of the spatial correlation of the
images, x (the Fourier transform) has a sparse representation. Another approach for
solving this problem is to change the problem to a regularization problem and solve
the following (Zhdanov 2002):
512 M. Elahi Naraghi and S. Srinivasan
Usually three alternatives for the objective function norms lq and lp in the above
equation are considered: the linear least squares (LLS) [p¼q ¼ 2], the least absolute
deviation (LAD) [p¼q ¼ 1], and the least mixed norm (LMN) [p ¼ 2, q ¼ 1].
As mentioned before, the Fourier transform coefficient of images is i sparse;
therefore, in our case, x is the Fourier transform coefficient of the geologic map, A is
the inverse Fourier transformation matrix, and y is the scattered measured geology
variables. By solving the optimization problem, Fourier transform can be calculated
and used to estimate the higher-order moments.
This idea has also been used for geostatistical simulation (Jafarpour et al. 2009)
where the Eq. 11 has been used to estimate a Fourier-based transform (DCT in that
paper) and the reservoir properties are calculated by inverse transform. However,
the proposed algorithm can only generate one realization and uncertainty cannot be
quantified.
In this section, we present how the polyspectra can be used to extract information
about size, shape, and orientation of features in images. We focus on the properties
of power spectrum and bispectrum.
One of the properties of Fourier transform is that if an image is rotated, its Fourier
transform also rotates (Fig. 1) (Reddy and Chatterji 1996). Therefore, by applying a
rotational transform in the Fourier transform space, one can account for changes in
the azimuth direction of continuity in the spatial domain. Since the amplitude of the
Fourier transform is related to power spectrum as in Eq. 4, this allows us to find the
orientation of the objects from its power spectrum. For instance, Fig. 2 shows three
different maps of channelized reservoir with different orientations, and their power
spectrum is shown in Fig. 3. As can be seen in Fig. 3, the angle subtended by the
main axis of the power spectrum represents the orientation of the objects. Figures 4
and 5 make the same point for different type of reservoirs with ellipsoid objects.
Another useful property of Fourier transform is that if the Fourier transform F{g(t})
is denoted as G(f), i.e.,
Robust MPS-Based Modeling via Spectral Analysis 513
Fig. 1 The effect of rotation in the spatial domain on the frequency characteristics: The top row
images are the original image (left) and its power spectral density, and the two bottom images are
the rotated image and its power spectral density. It can be seen that the Fourier transform has been
also rotated (http://www.cse.iitd.ac.in/~parag/projects/DIP/asign1/transrotfft.shtml)
Fig. 2 Three different maps of facies type for channelized reservoirs with different orientation of
channels
G cf
FfgðtÞg ¼ Gðf Þ, then FfgðctÞg ¼
jcj
Fig. 3 The power spectrum of the facies map shown in Fig. 2. The white dashed line shows the
main orientation of the power spectrum. It can be seen that the azimuthal orientation of continuity
shown by the spatial objects is the same as that shown by the power spectrum
Fig. 4 Three different facies models with ellipsoidal objects in different orientations
Fig. 5 The power spectrum of the facies map shown in Fig. 4. The white dashed line shows the
main orientation of the power spectrum. It can be seen that the rotation of the objects shows the
same rotation as the power spectrum
size of the objects. Figures 8 and 9 make the same point for different reservoir
models with ellipsoidal objects.
As shown previously, the size and orientation of the objects can be inferred from the
bandwidth and orientation of the power spectrum. However, the power spectrum
Robust MPS-Based Modeling via Spectral Analysis 515
Fig. 6 Two different models for channelized reservoirs with different sizes of channels
Fig. 7 The power spectrum of the facies map shown in Fig. 6. The white dashed line shows the
bandwidth of the power spectrum. It can be seen that the size of the objects corresponds to the
bandwidth of the power spectrum
Fig. 8 Two different reservoir models with ellipsoidal objects of different dimensions
516 M. Elahi Naraghi and S. Srinivasan
Fig. 9 The power spectrum of the facies models shown in Fig. 8. The white dashed line shows the
bandwidth of the power spectrum. It can be seen that the size of the objects corresponds to the
bandwidth of the power spectrum
cannot help with distinction of the object type. In this section, we present how to
distinguish between different objects using the bispectrum.
The bispectrum is a triple product of Fourier coefficients,
S3, z ðω1 ; ω2 Þ ¼ Z ðω1 ÞZ ðω2 ÞZ∗ ðω1 þ ω2 Þ, and is a complex-valued function of
two frequencies, unlike the power spectrum, which is a second-order product of
Fourier coefficients and a function of only one frequency. Unlike the power
spectrum, the bispectrum retains information about the phase of the Fourier trans-
form of the sequence. The phase of the Fourier transform is a nonlinear function of
frequency, and this nonlinearity is extracted by the biphase (the phase of the
bispectrum). For example, left and right asymmetric sequences will have opposite
signs for the biphase. These properties form a basis for the use of the bispectrum for
extracting features from patterns. Additionally, parameters can be defined from the
bispectrum that are invariant to translation, scaling, and amplification. In particular,
the phase of the integrated bispectrum along a radial line of slope (see Fig. 10)
satisfies these properties (Chandran and Elgar 1993; Chandran et al. 1997).
Parameters
I i ð aÞ
PðaÞ ¼ arctan ð13Þ
I r ð aÞ
where
Z 1=ð1þaÞ
I ð aÞ ¼ I r ð aÞ þ i I i ð aÞ ¼ Bðf 1 , af 1 Þdf 1 ð14Þ
f 1 ¼0
for 0<a <1 can serve as features for patterns. The variables Ir and Ii refer to the real
pffiffiffiffiffiffiffi
and imaginary parts of the integrated bispectrum, respectively, and i ¼ 1. It has
been shown in Chandran and Elgar (1993) that these parameters satisfy the desired
Robust MPS-Based Modeling via Spectral Analysis 517
Fig. 10 Region of
computation of the
bispectrum. Features are
obtained by integrating the
complex bispectrum along a 2 2 = 1
radial line with slope ¼ a
(dashed line). The phase of
this integral is translation
and scale invariant
1
invariance properties. It means that scaling and rotating will not change this
identifier. For example, scaling the original sequence results in an expansion or
contraction of the Fourier transform that is identical along the f1 and f2 directions.
Thus, the bispectral values along a radial line in bifrequency space map back onto
the same line upon scaling. The real and imaginary parts of the integrated
bispectrum along a radial line are multiplied by identical real-valued constants
upon scaling, and therefore, the phase of the integrated bispectrum is unchanged
(Nikias and Raghuveer 1987; Chandran and Elgar 1993; Chandran et al. 1997).
Since this identifier is not sensitive to rotation and scaling, it can be used as an
identifier to distinguish between the shapes of the objects within the image.
To demonstrate the feasibility of this application, we run an example on 4 1-D
signals. These signals are shown in Fig. 11. We also generate different signals with
different scaling factors as shown in Fig. 12. To show the sensitivity of the identifier
to noise, we also added Gaussian noise to each signal. We then calculated P(a) for
each signal as a function of a. Figure 13 shows the scatter plot of P(1) versus P
(1/16) for each signal. As shown in Fig. 13, given P(1/16) and P(1), the two similar
looking bolts clearly can be distinguished.
It should be noted that this example is just to demonstrate the feasibility of the
bispectrum identifier for shape classification. This method can be easily applied to
2-D images. The only difference is that the scalar a becomes a vector of size
2. Later in the results section, we will show the applicability of the algorithm for
2-D training images in the results section.
3 Results
In this section, we present the results of the explained methodology. First, we show
the results of calculating higher-order spectra from both TIs and conditioning data.
Then, we present how the best set of TIs in terms of shapes of the objects can be
detected using the phases of the bispectrum. Finally, we show how the size and
orientation of the objects can be corrected using the features from power spectrum.
518 M. Elahi Naraghi and S. Srinivasan
Fig. 11 Profiles of two types of bolts with left and right orientation
Fig. 12 Examples of the 1-D signal with different scaling factor (1, 0.5, 2 from left to right)
-0.1
-0.2
-1 -0.5 0 0.5 1
Feature 1 of 16 ( P(1/16) )
Robust MPS-Based Modeling via Spectral Analysis 519
In this section, we present the results for feature extraction from bispectrum for 2-D
training images. We selected three different training images for different geologic
scenarios as shown in Fig. 21. Then we generated several images with different
orientation and geobody size. Figure 22 shows four examples of these images. We
extracted four sets of scattered data from four random images of the data set and
520 M. Elahi Naraghi and S. Srinivasan
Fig. 15 Covariance map calculated by directly searching training image (left) and by using FFT
(right)
Fig. 16 Three-point connectivity inferred using different spatial templates (from left to right: L
shape, 45 degree, x-x and y-y connectivity)
calculated the biphases of the bispectrum of all the images and the scattered data
sets. The biphases computed only using the sparse conditioning data are
superimposed (diamond points) on the scatter observed on the basis of the exhaus-
tive images. As can be seen, the three different scenarios can be successfully
distinguished using these features from both scattered data sets and training images.
Therefore, we can find the correct training image (in terms of the object shape) by
selecting the image with the closest feature value to the one of the scattered data set
(Fig. 23).
In the previous section, we presented how the bispectrum can be used to distinguish
among different training images with different objects. However, the orientation
and scale of the objects cannot be detected from bispectrum. In this section, we
present how the power spectrum can be used to distinguish the orientation and the
scale of the objects. We generated the rotated realization from a TI (45 ) and used it
as the correct map of the reservoir. We then sampled some conditioning data from
the correct map as shown in Fig. 24. Then, we calculated the power spectrum of
both available TI and the scattered data set. Figure 25 shows the covariance map of
Robust MPS-Based Modeling via Spectral Analysis 521
Fig. 17 The third-order moment of the image shown in Fig. 14. The left column is calculated by
calculating bispectrum, and the right column is calculated by direct search of pairs in the image
500
0
0 500 1000 1500
X(m)
0 0.1
-200
-400 0
Fig. 20 Calculated three-point connectivity with different configurations using scattered data in
Fig. 18 (the templates are, from left to right, L shape, 45 degree, xx configuration, and yy
configuration, as shown in Fig. 16)
Robust MPS-Based Modeling via Spectral Analysis 523
Fig. 21 The training images used in this study with different objects and geologic scenarios
Fig. 22 Different rotated and scaled images from the reference training images in Fig. 21
Fig. 23 The values of a feature vector consisting of {P(1,1), P(1/16,1/16)} for all of the different
images with different scale and orientations
direct sampling (Mariethoz and Renard 2010). As can be seen in Fig. 26, the results
of the simulation using the rotated TI is much closer to the correct map indicating
the importance of the orientation of objects within TI in simulation.
524 M. Elahi Naraghi and S. Srinivasan
Fig. 24 The available TI (in the left), the corrected map with different orientation (in the middle),
and the scattered data sampled from the correct map (in the right)
Fig. 25 The covariance map of the available TI (top left), the covariance map of the scattered data
(top right), the power spectrum of the available TI (bottom left), and the power spectrum of the
scattered data (bottom right). Dashed line represents the main orientation of the power spectrum
We also generated another realization form a TI with different geobody size, and
used it as the correct map of the reservoir. We sampled some conditioning data from
the correct map as shown in Fig. 27. Then, we calculated the power spectrum of
both available TI and the scattered data set. Figure 28 shows the covariance map of
the TI and conditioning data as well as the corresponding power spectrum. As
shown in Fig. 28, the bandwidth of the two power spectrums are different. Thus, we
Robust MPS-Based Modeling via Spectral Analysis 525
Fig. 26 The available TI (top left), the corrected TI (top right), the simulation using the available
TI (bottom left), and the simulation using the corrected TI (bottom right)
generated another TI to account for the mismatch of the size of the channels. To
show the importance of the geobody size of TI in simulations, we used both original
and the corrected TIs to perform simulation using direct sampling. As can be seen in
Fig. 29, the results of the simulation using the corrected TI are much closer to the
correct map indicating the importance of the size of objects within TI in simulation.
Fig. 27 The available TI (in the left), the corrected map with different geobody size (in the
middle), and the scattered data sampled from the correct map (in the right)
Fig. 28 The covariance map of the available TI (top left), the covariance map of the scattered data
(top right), the power spectrum of the available TI (bottom left), and the power spectrum of the
scattered data (bottom right). Dashed line represents the bandwidth of the power spectrum
Robust MPS-Based Modeling via Spectral Analysis 527
Fig. 29 The available TI (top left), the corrected TI (top right), the simulation using the available
TI (bottom left), and the simulation using the corrected TI (bottom right)
Bibliography
Chiles JP, Delfiner P (1999) Geostatistics modeling spatial uncertainty. Wiley, New York
Deutsch CV, Lewis RW (1992) Advances in the practical implementation of indicator
geostatistics. 23rd international APCOM symposium
Dimitrakopoulos R, Mustapha H, Gloaguen E (2010) High-order statistics of spatial random fields:
exploring spatial cumulants for modeling complex non-Gaussian and non-linear phenomena.
Math Geosci 42:65–99
Donoho DL (2006) Compressed sensing. IEEE Trans Inf Theory 52:1289–1306
Drach A, Khalighi AH, ter Huurne FM, Lee CH, Bloodworth C, Pierce EL, Sacks MS (2015)
Population-averaged geometric model of mitral valve from patient-specific imaging data. J
Med Devices 9:030952
Eskandari K, Srinivasan S (2007) Growthsim–a multiple point framework for pattern simulation.
EAGE Conference on Petroleum Geostatistics
Fessler JA, Sutton BP (2003) Nonuniform fast fourier transforms using min-max interpolation.
IEEE Trans Sig Process 51:560–574
Gan L (2007) Block compressed sensing of natural images. 15th International Conference on
Digital Signal Processing. IEEE, pp 403–406
Giannakis GB, Mendel JM (1989) Identification of nonminimum phase systems using higher order
statistics. IEEE Trans Acoust, Speech Sig Process 37:360–377
Goovaerts P (1997) Geostatistics for natural resources evaluation. Oxford Press, New York
Gringarten E, Deutsch CV (2001) Teacher’s aide variogram interpretation and modeling. Math
Geol 33:507–534
Hall TE, Giannakis GB (1995) Bispectral analysis and model validation of texture images. IEEE
Trans Image Process 4:996–1009
Honarkhah M, Caers J (2010) Stochastic simulation of patterns using distance-based pattern
modeling. Math Geosci 42:487–517
Huang YC, Srinivasan S (2012) Efficient conditional simulation of spatial patterns using a pattern-
growth algorithm. Geostat Oslo. Springer, Dordrecht, pp 209–220
Isaaks EH, Srivastava RM (1989) An introduction to applied geostatistics. Oxford University
Press, New York, p 561
Jafarpour B, Goyal VK, McLaughlin DB, Freeman WT (2009) Transform-domain sparsity
regularization for inverse problems in geosciences. Geophysics 74:R69–R83
Journel A (1983) Nonparametric estimation of spatial distributions. Math Geol 15:445–468
Keiner J, Kunis S, Potts D (2009) Using NFFT 3–a software library for various nonequispaced fast
Fourier transforms. ACM Trans Math Softw (TOMS) 36:19
Khalighi AH, Drach A, ter Huurne FM, Lee CH, Bloodworth C, Pierce EL, . . . & Sacks MS
(2015, June) A comprehensive framework for the characterization of the complete mitral valve
geometry for the development of a population-averaged model. In International Conference
on Functional Imaging and Modeling of the Heart (pp. 164–171). Springer International
Publishing
Lustig M, Donoho DL, Santos JM, Pauly JM (2008) Compressed sensing MRI. Sig Process Mag,
IEEE 25:72–82
Maksimov MM, Galushko VV, Kutcherov AB, Surguchev LM, Bratvold RB (1993) Parallel
nested factorization algorithms. SPE Symposium on Reservoir Simulation
Mariethoz G, Renard P (2010) Reconstruction of incomplete data sets or images using direct
sampling. Math Geosci 42:245–268
Mustapha H, Dimitrakopoulos R (2010a) High-order stochastic simulation of complex spatially
distributed natural phenomena. Math Geosci 42:457–485
Mustapha H, Dimitrakopoulos R (2010b) A new approach for geological pattern recognition using
high-order spatial cumulants. Comput Geosci 36:313–334
Nikias CL, Raghuveer MR (1987) Bispectrum estimation: a digital signal processing framework.
Proc IEEE 72:869–891
Pérez C, Mariethoz G, Ortiz JM (2014) Verifying the high-order consistency of training images
with data for multiple-point geostatistics. Comput Geosci 70:190–205
Robust MPS-Based Modeling via Spectral Analysis 529
Reddy BS, Chatterji BN (1996) An FFT-based technique for translation, rotation, and scale-
invariant image registration. IEEE Trans Image Process 5:1266–1271
Rezaee H, Mariethoz G, Koneshloo M, Asghari O (2013) Multiple-point geostatistical simulation
using the bunch-pasting direct sampling method. Comput Geosci 54:293–308
Rosenblatt M (1985) Stationary sequences and random fields. Birkhuser, Boston
Srivastava RM (1992) Reservoir characterization with probability field simulation. SPE Form Eval
7:927–937
Strebelle S (2000). Sequential simulation drawing structures from training images. Ph. D. Thesis,
Stanford University
Tahmasebi P, Hezarkhani A, Sahimi M (2012) Multiple-point geostatistical modeling based on the
cross-correlation functions. Comput Geosci 16:779–797
Wang Y, Yang J, Yin W, Zhang Y (2008) A new alternating minimization algorithm for total
variation image reconstruction. SIAM J Imaging Sci 1:248–272
Xu W (1996) Conditional curvilinear stochastic simulation using pixel-based algorithms. Math
Geol 28:937–949
Yao T (1998) SPECSIM: a Fortran-77 program for conditional spectral simulation in 3D. Comput
Geosci 24:911–921
Zhang T (2006) Filter-based training pattern classification for spatial pattern simulation. Ph. D.
Thesis, Stanford University
Zhdanov MS (2002) Geophysical inverse theory and regularization problems. Elsevier,
Amsterdam
Efficient Uncertainty Quantification
and History Matching of Large-Scale Fields
Through Model Reduction
1 Introduction
Uncertainty quantification (UQ) and history matching (HM) have become a regular
routine for efficient reservoir management and decision-making in petroleum indus-
try. There are many types of uncertain parameters (e.g., permeability, porosity,
geometry, aquifer, oil-water contact, relative permeability, PVT parameters, etc.)
that have been found to have evident impacts on reservoir performance. Among
them, spatially correlated reservoir geo-models (e.g., permeability) have been a
J. Fu (*)
Institute for Water and Environmental Engineering (Currently with Chevron),
Universitat Politècnica de València, Valencia, Spain
e-mail: [email protected]
X.-H. Wen
Chevron ETC, 1500 Louisiana St., Houston, TX 77002, USA
e-mail: [email protected]
S. Du
Petroleum Engineering Department (Currently with Chevron), Texas A&M University,
College Station, TX, USA
e-mail: [email protected]
challenge for UQ and HM when combining with other parameters. These geo-models
are usually created by geostatistical tools with a high dimensionality. Traditionally,
this type of correlated parameters is reparametrized using a zonation method with one
lumped constant or multiplier specified for each zone such that the dimensionality of
problems can be reduced and the relative significance of these parameters for reservoir
performance prediction can be studied in comparison with other types of parameters.
However, this ad hoc method faces a challenge to find the optimal zones such that the
real impact can be accurately quantified. Moreover, it may fail to honor the geological
(or geostatistical) features after the lumped constants or multipliers are applied.
Several techniques are available to address this problem. For low-dimensional
problems (e.g., thousands of parameters), principal component analysis (PCA) is
used to efficiently reduce the dimensionality of geostatistical models for accurate
UQ and HM. For high-dimensional problems (e.g., large-scale correlated perme-
ability fields), the distance-based kernel PCA is used to reparametrize the field for
UQ, and geostatistical features can be preserved during HM. These approaches can
reduce the dimension of parameters yet preserve the key spatial geostatistical
features in the parameter fields. These methods have been implemented in our
in-house software platform. In this paper, we present a simple synthetic example to
demonstrate the effectiveness of this methodology.
2 Methodology
of the original large-scale covariance matrix can be recovered from these small-
scale kernel matrices and the associated snapshots. These eigenvectors and eigen-
values can be used to reconstruct the images of correlated random fields. Unfortu-
nately, the eigenvectors of the covariance matrix cannot be accurately recovered in
terms of magnitude with this method. Another challenge is how to generate random
models that are consistent with the existing information for HM.
To address these challenges, we implement a distance-based kernel PCA
(DKPCA) for large-scale problems (Scheidt and Caers 2009; Park 2011). The
re-parameterization procedure of DKPCA is the same as SPCA except the way to
reconstruct new models. The concept “distance” between random realizations is
used such that the new models are generated through an optimization process that
minimizes the HM error through the linear combination of the ensemble realizations
under distance space (Kwok and Tsang 2004; Park 2011). By doing so, the stochastic
models generated always share the same geostatistical features as the given ensem-
ble and thus are “geologically” consistent. With the DKPCA, moreover, the condi-
tioning data can be always honored as long as all of the individual realizations
respect the hard data, which is straightforward for many existing geostatistical tools.
3 An Illustrative Example
The model presents an oil-water two-phase black-oil reservoir and has 1681
(41 41 1) active cells with each one 50 50 50 [ft3]. The porosity is assumed
to be homogeneous and is equal to 0.2. The isotropic permeability (e.g., xperm) is
heterogeneous with a multi-Gaussian model specified for lnk. The reservoir is
initially produced with five wells: four producers (PROD1, PROD2, PROD3, and
PROD4) located in the four corners and one injector (INJW1) in the center. Figure 1
shows the reservoir and well configuration.
All of the four wells are put into operation at the same time. The injector is
operated by a specified bottom-hole pressure, while the producers are constrained
by a constant bottom-hole pressure subject to a maximum oil rate. The constants are
fixed for a certain period but may vary over time. The simulation time spans
2000 days. The data from the first 1000 days are used for history matching, while
results from the second 1000 days will be used for checking the quality of pre-
dictions of the HM models. The production data include both water production rate
(WPR) and oil production rate (OPR). We are going to generate stochastic realiza-
tions for xperm by matching the historical production data at four producers.
Two scenarios of synthetic experiments are performed for HM. The first one is a
so-called zonation method, i.e., to divide the field into several regions (patches or
zones) such that the field can be reduced to small number of parameters for HM. The
second one is the DKPCA-based method for preserving geological consistency in
the models. In the two scenarios, we will use the same number (6) of parameters for
534 J. Fu et al.
comparison purpose. Now that the large-scale HM problem (with 1681 parameters)
is reduced to a very small-scale one (with six parameters), a variety of optimization
methods can be used to create multiple history-matched models. The particle swarm
optimization (PSO) is employed in this work as search engine to minimize the HM
error (Isebor 2013).
The results of oil production rate (OPR) before and after HM with the first 1000-day
data are compared in this section. The predictions of the second 1000 days with the
existing wells are also plotted and compared.
Figure 2 plots the OPR profile of PROD1 by comparing the individual initial guesses
(the gray lines) generated by experimental design with the history-matched results (the
color lines) using the traditional zonation method. Obviously, the results from initial
guesses significantly deviate from the historical data, while the historical data of the first
1000 days (the magenta dots) can be reasonably reproduced by history-matched models.
Although not shown in this case, it was also observed in other cases that an improper
zonation may lead to a poor HM or even fail to do the job. It also shows that the models
could also reasonably predict OPR for the second 1000 days for well PROD1.
Figure 3 plots the OPR result for PROD1 from the models generated by
DKPCA-based method. Obviously, the final-matched models are able to match
Efficient Uncertainty Quantification and History Matching of Large-Scale. . . 535
Fig. 2 Oil production rate of PROD1 from the zonation method: color lines (the final matched),
gray lines (the initial guess), and magenta dots (historical and reference data, only the first 1000-
day data are used for history matching)
Fig. 3 Oil production rate of four producers from the DKPCA method: color lines (the final
matched), gray lines (the initial guess), and magenta dots (historical and reference data, only the
first 1000-day data are used for history matching)
536 J. Fu et al.
the data quite well. The predictions of OPR for the second 1000 days using the
DKPCA-matched models capture the reference data well, while the initial guesses
fail to do so.
Results from Scenario 1 show that both the zonation and DKPCA methods are able
to history match the first 1000-day’s data equally well and are able to provide
reasonable predictions of the second 1000-day production with the existing wells. A
more interesting test is to check the predictability of the HM models for new well
locations where no production data are used during the HM process.
The drilling of infill wells represents a very common practice for staged oil field
development. Prediction of performances for these new wells under changing
production conditions is challenging. In some cases, the historical data from
existing wells may contain useful information to identify reservoir parameters
that impact performances of these new wells, but some history-matched models
may achieve data reproduction by locally tuning their parameters near well-bore
only, leading to lack of precision in predictions. This is the underfitting issue. In
other cases, the reservoir properties near new wells remain largely unsampled, and
the historical data contain very limited information to infer them. However, some
history-matched models may bring false information from optimization, resulting
in biased forecasts for those new wells. This is the overfitting issue. Both
underfitting and overfitting models are not robust for forecast of long-term reservoir
performances. Therefore, prediction of new wells may serve as a metric to measure
the robustness of reservoir simulation models.
Specifically in this case, a campaign with two new infill production wells
(PROD12 and PROD41 in Fig. 1) is run, and the results are compared in this part
so as to further check the prediction capability of history-matched models generated
before subject to new production conditions. Note that the history-matching sce-
nario is the same as the previous part, but in the prediction scenario, the two new
wells (PROD12 and PROD41) replace PROD1, PROD2, and PROD4 to produce
the reservoir. Although the well-level performances have the similar behavior, the
metric that we are looking at in this scenario is the field’s OPR, which plays a
critical role in finance and operations for efficient reservoir management and asset
development.
Figures 4 and 5 compare the OPR of field with the new production campaign.
Note that the first 1000-day’s data were used in history matching while the second
1000-day for prediction. The magenta dots show the reference data that are created
from the “actual” model, while the gray lines plot the prediction from the history-
matched models. The zonation method (Fig. 4) displays a clear deviation from the
reference data. The DKPCA-based method (Fig. 5) obviously is able to obtain a
better result: the reference data well reside in the envelope of predictions from the
DKPCA models.
Efficient Uncertainty Quantification and History Matching of Large-Scale. . . 537
Fig. 4 OPR prediction of the field from the zonation method: The first 1000-day data are used for
history matching, while the second 1000-day for prediction. Note that the gray lines represent the
results from history-matched models and the magenta dots are historical (the first 1000-day) or
reference (the second 1000-day) data
Fig. 5 OPR prediction of the field from the DKPCA method: The first 1000-day data are used for
history matching, while the second 1000-day for prediction. Note that the gray lines represent the
results from history-matched models, and the magenta dots are historical (the first 1000-day) or
reference (the second 1000-day) data
538 J. Fu et al.
Fig. 6 Typical lnk images: (a) reference field, (b) zonation, (c) DKPCA. Note that the unit of k in
the color bar is in [md]
4 Discussion
5 Conclusion
This paper compares a PCA-based method with a traditional zonation method for
history matching. We have observed that the proposed method is efficient for
history matching large-scale heterogeneous fields. We also demonstrate the value
of the proposed method in building simulation models for improved forecasts
owing to a full consideration of the spatial heterogeneity pattern, a dominant factor
for reservoir management and asset development in reservoir engineering.
Bibliography
Cardoso MA, Durlofsky LJ, Sarma P (2009) Development and application of reduced-order model-
ing procedures for subsurface flow simulation. Int J Numer Methods Eng 77(9):1322–1350
Clark MP, Vrugt JA (2006) Unraveling uncertainties in hydrologic model calibration: assessing
the problem of compensatory parameters. Geophys Res Lett 33:L06406
Doherty JE, Christensen S (2011) Use of paired simple and complex models to reduce predictive
bias and quantify uncertainty. Water Resour Res 47:W12534
Fu J (2008) A Markov Chain Monte Carlo method for inverse stochastic modeling and uncertainty
assessment, Unpublished Ph. D. thesis, Universidad Politecnica de Valencia, Valencia, Spain,
p 140
Fu J, Gómez-Hernández JJ (2008) Preserving spatial structure for inverse stochastic simulation
using blocking Markov chain Monte Carlo method. Inverse Probl Sci Eng 16(7):865–884
Fu J, Gómez-Hernández JJ (2009a) A blocking Markov chain Monte Carlo method for inverse
stochastic hydrogeological modeling. Math Geosci 41(2):105–128
Fu J, Gómez-Hernández JJ (2009b) Uncertainty assessment and data worth in groundwater flow
and mass transport modeling using a blocking Markov chain Monte Carlo method. J Hydrol
364:328–341
540 J. Fu et al.
Isebor OJ (2013) Derivative-free optimization for generalized oil field development, Ph. D. thesis,
Stanford University, p 169
Kwok JT-Y, Tsang IW-H (2004) The pre-image problem in kernel methods. IEEE Trans Neural
Netw 15(6):1517–1525
Park K (2011) Modeling uncertainty in metric space, Ph. D. thesis, Stanford University, p 229
Refsgaard JC, Van der Sluijs JP, Brown J, Van der Keur P (2006) A framework for dealing with
uncertainty due to model structure error. Adv Water Resour 29(11):1586–1597
Scheidt C, Caers J (2009) Representing spatial uncertainty using distances and kernels. Math
Geosci 41(4):397–419
Scholkopf B, Smola A, Muller KR (1998) Nonlinear component analysis as a kernel eigenvalue
problem. Neural Comput 10:1299–1319
Scholkopf B, Mika S, Burges CJC, Knirsch P, Muller KR, Ratsch G, Smola A (1999) Input space
versus feature space in kernel-based methods. IEEE Trans Neural Netw 10(5):1000–1017
Sirovich L (1987) Turbulence and the dynamics of coherent structures, Part I: Coherent structures.
Q Appl Math XLV(3):561–571
Taguchi G, Clausing D (1990) Robust quality. Havard Bus Rev 1990:56–75
Vo HX, Durlofsky LJ (2014) A new differentiable parameterization based on principal component
analysis for the low-dimensional representation of complex geological models. Math Geosci
46(7):775–813
White JT, Doherty JE, Hughes JD (2014) Quantifying the predictive consequences of model error
with linear subspace analysis. Water Resour Res 50:1152–1173
Revealing Multiple Geological Scenarios
Through Unsupervised Clustering
of Posterior Realizations from Reflection
Seismic Inversion
1 Introduction
A widely used approach for inversion of reflection seismic data is based on a linear
forward model and Gaussian assumptions about the prior distribution of the elastic
parameters and the noise model (Buland and Omre 2003; Jullum and Kolbjørnsen
2016; Hampson et al. 2005). The solution to such inverse problems is Gaussian
distributed, which is simply a single-modal distribution.
In reality, the Gaussian prior assumption and the linear forward model may not
adequately describe the information available. In order to handle this, these prob-
lems can instead be formulated probabilistically, which leads to a solution
described by a typically unknown and non-Gaussian posterior distribution, (Larsen
et al. 2006; Ulvemoen and Omre 2010; Zunino et al. 2014; Bosch et al. 2007). In
this way, the solution to the inverse problem can be characterized by a sample from
the posterior distribution, which will represent a set of realizations (i.e., possible
solutions) that are all consistent with observed data and prior information
(Mosegaard and Tarantola 1995). The posterior distribution of such non-Gaussian
inverse problems may be highly multimodal: i.e., realizations from the posterior
probability distribution may be located at isolated clusters in the high dimensional
space spanning all possible earth models. If such clusters exist, each cluster will
correspond to a group of realizations with similar appearance, and each cluster may
represent one geological scenario. For example, one cluster may represent a
commercially viable reservoir, while another cluster may represent a
non-commercially viable reservoir. Using a probabilistic approach to inverse prob-
lem theory, the relative probability of each local scenario can easily be quantified.
This may be hugely beneficial in characterizing reservoir models of all kinds. It
should be noted that this study does not intend to analyze the different geological
scenarios but demonstrates a technique to reveal their nature, which would be
helpful in this respect.
In the following we will analyze the result of a probabilistically formulated
inversion of a normal incidence seismic data set from the southern part of Jutland,
Denmark. Initially we will demonstrate that the posterior sample contains many
quite different realizations. Then we use a clustering algorithm to locate multi-
modal clusters (i.e., similar posterior realizations). This analysis will be done both
with respect to the realizations representing lithological units and with respect to
the realization’s corresponding acoustic impedance parameterization. Finally the
analysis obtained from the two different ways of parameterizing data will be
compared.
Figure 1 shows 26,000 realizations resulting from sampling of the posterior distri-
bution for a probabilistically formulated inversion of a reflection seismic data set
Revealing Multiple Geological Scenarios Through Unsupervised Clustering of. . . 543
Fig. 1 26,000 1D lithology profiles resulting from an inversion of reflection seismic data. The
colors blue, turquoise, green, yellow, and orange represent the lithology units halite, limestone,
lamination, 50/50 anhydrite and limestone, and anhydrite, respectively
from of the Zechstein unit in the southern part of Jutland, Denmark (Cordua et al.
submitted). This distribution is based on a prior distribution that is statistically
consistent with geological scenes observed in borehole logs. Hence, the posterior
realizations are all consistent with both the observed geological scenes and the
observed seismic data (within the expected data uncertainty). A presentation of the
inversion method used to obtain the posterior realizations is outside the scope of
this paper. However, it can be assumed that the method in fact samples the true
posterior distribution. Later we actually show that the method in fact revisits the
same modes of the posterior distributions several times during the sampling period.
In Fig. 1 it can be seen that this highly underdetermined seismic inverse problem
has several possible solutions of different nature. Figure 1 further suggests some
clusters of similar posterior realizations. In this study, we try to quantify this
apparent clustering to get a better understanding of how the realizations actually
are distributed. This is done using the unsupervised clustering method called
K-means. In addition to provide a better visual representation of the results, the
clustering algorithm allows quantifying the probability of the possible geological
scenarios.
544 M.L. Gulbrandsen et al.
Figure 1 represents one way of displaying the posterior realizations. Each color
represents a lithological unit. Another way could be to parameterize these units with
their elastic property, acoustic impedance (AI). Table 1 shows the relation between
the lithological units and their corresponding AI values, as used in the inversion
algorithm (Cordua et al. submitted).
The task is now to group all the individual posterior realizations into different
clusters. A general description of a cluster is difficult, and no unique definition
exists. However, it can be valuable to think of clusters as groups of which all
members are closer together with the other members from the same group than
members from other groups, even though this not always has to be a mathematical
fact. There are several ways of defining these mutual distances and several different
clustering algorithms exist. There is no correct algorithm, since the different
methods represent different ways of analyzing the data and the different methods
should be decided based on the problem. In this study, the K-means clustering
algorithm will be used to group the posterior realizations both with respect to the
lithological units and with respect to the acoustic impedance values.
Assuming the number of clusters, K, is given, the aim is to group the D-dimensional
data x ¼ fx1 , x2 , . . . xN g into the K clusters. The clusters are represented by the D-
dimensional vectors μk, where k ¼ 1. . . K, and each vector μk is a prototype vector
for the kth cluster (Bishop 2006). More specifically, the prototype vectors represent
the centers of each cluster, and hence, the K-means clustering model is a centroid
model. The objective goal of the method is to minimize the sum of the square
distances between all points and their closest prototype vector. Mathematically this
can be described as minimizing the objective function J (Bishop 2006):
Revealing Multiple Geological Scenarios Through Unsupervised Clustering of. . . 545
XN XK
J¼ n¼1
r kxn
k¼1 nk
μk k2 ; ð1Þ
where rnk is a set of binary indicator variables, rnk 2{0, 1}, each associated with one
data point xn, and stating which cluster each point is assigned to. In order to find the
values for rnk and μk that minimizes Eq. 1, an iterative procedure is performed.
Every iteration consists of two steps. The first step optimizes on the indicator
variables and the next step on the cluster centers. Prior to the simulation,
K prototype vectors are randomly chosen among the D-dimensional points x
(note that there are several other ways to choose the starting points as well). In
the first step of the iterations, each point is assigned a cluster based on the prototype
vector, i.e., cluster center, they are closest to. This is in fact optimizing J with
respect to rnk, since each data point is independent and taking the smallest distance
between xn and μk will therefore minimize J. For the realizations of continuous
variables, i.e., acoustic impedance, the Euclidean distance is used, and for categor-
ical variables, i.e., the lithological units, the Hamming distance1 is used. The next
step is to optimize on μk, and since J is a quadratic function of μk, the optimal μk is
found by setting the derivative of the objective function with respect to μk to zero:
XN
2 r ðx
n¼1 nk n
μk Þ ¼ 0 ð2Þ
From Eq. 3 it is seen that μk represents the mean of all points within each cluster.
The second step of each iteration is hence to compute the mean of all points
assigned to each cluster and define that new mean as the optimized cluster proto-
type. This procedure is repeated until the values converge. It should be noted that
depending on the data set, two different simulations (i.e., two different starting
positions) might not converge to the same local minimum. You can never be sure if
you reach the global minimum of the objective function, so it might be a good idea
to run the simulations more than once.
1
The hamming distance is the number of position (coordinates/elements) that differs between two
vectors of the same length.
546 M.L. Gulbrandsen et al.
approaching this issue could simply be to try different values and visually examine
the results. Another way could be to use a statistical approach to get an idea of
which number of clusters that best represents the sample. In this study an analysis of
the gap statistic (Tibshirani et al. 2001) is done. The idea of the gap statistic is to
find the number of clusters which standardize the comparison of log (Wk) with a
distribution of no obvious clustering, when Wk is defined as the within-cluster
dissimilarity:
XK 1
Wk ¼ Dr ; ð4Þ
r¼1 2N r
where Nr is the number of points within clusters and Dr is the distance between all
points within cluster r. In this analysis the Euclidean distance is used for Dr when
the AI data is analyzed, and the Hamming distance is used for the lithological units.
The distribution with no obvious clustering is computed by taking the average of
20 simulations of uniformly distributed data within the range of the sample.
Computing the within-cluster dissimilarity of the uniformly distributed data, Uk,
the gap statistic is defined as
The optimal number of clusters is then the smallest number of k, which fulfill the
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1
criteria GðkÞ Gðk þ 1Þ sfkþ1g , where s{kþ1} is stdðlogðU k ÞÞ 1 þ 20 and “std”
denotes the standard deviation. Figure 2 shows the logarithm of the within-cluster
dissimilarity for both the sample (blue) and for the simulated data (red). This
simulation is done with respect to the lithological units. As expected the blue
curve decreases with an increasing number of clusters, while the red curve is
more or less flat. In Fig. 3 the gap curve is plotted. Based on this plot, it is seen
that the optimal choice of clusters for this simulation is 15 (this is the first k where G
(k þ 1) – 1std is less than G(k)).
This method is based on simulations, and the optimal number of clusters may
therefore vary a little for different simulations. It should also be mentioned that the
method is totally general and independent on any clustering method, so this analysis
can be done with different clustering techniques.
3 Results
Figures 2 and 3 show the results from an analysis of the gap statistic of the posterior
realizations in Fig. 1 with respect to the lithological units. The gap statistic
simulations are however ambiguous, and the information from one simulation is
hence not enough to base our choice of the optimal number of clusters. Figure 4
shows the histograms of ten simulations of the gap statistic using the K-means
Revealing Multiple Geological Scenarios Through Unsupervised Clustering of. . . 547
algorithm on both data sets. It is seen that the “optimal number of clusters” (Kopt)
varies between 6 and 20. It should be noted that the upper boundary of the
simulation was set to 20 clusters; hence, a result of Kopt ¼ 20 indicates that no
optimal number of cluster was found within the test range (i.e., the optimal number
of cluster may exist outside of this range). (Tibshirani et al. 2001) states however
that it can be important to examine the whole gap curve and not only the optimal
choice. Figure 5 shows an example of a gap curve representing a simulation where
no optimal cluster numbers were found within the test range. It can however be seen
that the criteria for concluding on an optimal number of clusters almost are met,
both at k ¼ 10 and k ¼ 16.
All the gap curves for the ten simulations for both data sets have been analyzed.
This analysis combined with a visual examination of the clusters has resulted in the
choice of running the K-means clustering algorithm with respect of grouping
548 M.L. Gulbrandsen et al.
Fig. 4 The histograms of optimal number of clusters from 10 runs of the gap statistic method
using the K-means clustering algorithm on the AI data (left plot) and the lithological unit data
(right plot), respectively, with respect to the optimum choice of K (x-axis)
Fig. 5 The gap curve from one simulation of the gap statistic analysis using the K-means
algorithm with respect to the AI data
Revealing Multiple Geological Scenarios Through Unsupervised Clustering of. . . 549
Fig. 6 The 15 clusters resulting from the K-means clustering of the lithology profiles in Fig. 1.
The left panel shows which models among the realizations belong to each cluster and the right
panel shows these models put together. The “probability of occurrence” represents the number of
models in each cluster relative to the whole sample. The lithology units are represented with the
same colors as in Fig. 1
550 M.L. Gulbrandsen et al.
Fig. 6 (continued)
15 clusters for both parameterizations. These results are presented in Figs. 6 and 7.
The left panels of the two figures show which realizations from Fig. 1 that is
grouped in each cluster, and the right panels show all these realizations put together
to see the overall structure of each cluster. Note that the size of each plot in the right
panel of the two figures does not represent the actual size of the clusters. The
relative sizes are printed above each plot and represent the percentage of models
belonging to those specific clusters.
Analyzing the plots in the left panels of Figs. 6 and 7, we can see that the
realizations constituting the different clusters more or less are evenly distributed
throughout the set of all realizations. Even though the inversion algorithm itself is
outside the scope of this study, it can be mentioned as a curio that the distributions
of models (seen in Figs. 6 and 7) from the same clusters indicate that the sampling
algorithm actually visits the same modes of the posterior distributions several times
during the sampling period. This is an underlying assumption of sampling the
posterior distribution using the Metropolis algorithm. However, Figs. 6 and 7
demonstrate that this actually is the case.
Fig. 7 The 15 clusters resulting from the K-means clustering of the AI realizations converted from
the realizations shown in Fig. 1, using the relation in Table 1. The left panel shows which models
among the realizations belong to each cluster and the right panel shows these models put together.
The “probability of occurrence” represents the number of models in each cluster relative to the whole
sample. The lithology units are represented with the same colors as in Fig. 1
552 M.L. Gulbrandsen et al.
Fig. 7 (continued)
4 Discussion
Fig. 8 The percentage of models overlapping between the 15 clusters of the AI realizations
(y-axis) and the 15 clusters representing the lithological units (x-axis)
will be presented here since the comparison of the different methods is outside the
scope of this paper. Distribution models such as the expectation maximization
(assuming data can be described as a Gaussian mixture model) are not considered
in this study since they are unsuited for clustering categorical variables.
The importance of being able to cluster the data in the lithological unit (LU) domain
is illustrated in Fig. 8. The figure displays the percentage overlap between realiza-
tions belonging to the different clusters arising from applying the K-means clus-
tering algorithm on the AI data (vertical axis) and the lithological units (horizontal
axis), respectively. As an example, it is seen from Fig. 8 that cluster 15 in the AI
domain is pretty well represented in the lithological unit (LU) domain. 80.9 % of all
realizations in cluster 6 from the AI domain are grouped together in one cluster,
namely, number 9, in the LU domain. The rest of the realizations are mainly
distributed between cluster 14 and 15, with 6,8 % and 9,1 %, respectively. If we,
however, look at cluster 9 in the AI domain, we can see that the models belonging to
this cluster are split between two different clusters in the LU domain, namely,
554 M.L. Gulbrandsen et al.
Fig. 9 The upper plot shows the models representing cluster 9 in the AI domain. The middle and
lower plot shows cluster number 9 and 14 from the LU domain
cluster number 9 and 14 with 48,5 % and 47,8 %, respectively. These clusters are
plotted in Fig. 9. From the deeper part of the realizations displayed in Fig. 9, it is
clear that the clustering algorithm distinguishes between the blue (halite) and the
turquoise (limestone) when applied to the LU domain but cluster realizations with
both these sections together in the AI domain. This is because these models are
much closer together in the continuous parameter space, than in the discrete
parameter space, where the distance is the same between all the categorical vari-
ables. It should be noted that the clustering simulations for the two data sets have
the same starting points, i.e., the simulations start in the same random realizations.
This is important to emphasize, since it suggests that the differences illustrated in
Fig. 8 are actual differences of the two domains and not differences due to two
different local minima of the objective function (Eq. 1).
5 Conclusion
We have shown that K-means clustering can be used to get a better understanding of
how the multimodal landscape representing the solution space of a 1D reflection
seismic inversion is distributed. By clustering the posterior realizations, we get a
Revealing Multiple Geological Scenarios Through Unsupervised Clustering of. . . 555
much clearer picture of the potential geological scenarios and the probabilistic
distribution between them. The combination of probabilistic inverse problems and
cluster analysis thus allows us to perform scenario-based inversion of reflection
seismic data.
It is however important to emphasize that the posterior probability distributions
change for the different parameterizations, which in turn can result in very different
multimodal landscapes. Clustering the same realizations with different parameter-
ization will hence result in different clusters, and it is important to know what is
analyzed. In this study we show that one cluster in the AI domain actually is
represented in two different geological scenarios, and it is hence important to
analyze the domain representing the parameterization of interest.
Bibliography
Abstract Although object models are popular with geologists due to their ability to
control the geometries that are produced, they tend to have convergence issues
when conditioning on complex well patterns. In this paper, we present a new well
conditioning algorithm that utilizes more local data when generating channels. We
show that this algorithm performs better than the currently commercially available
state-of-the-art object model and thus makes object models viable in modern
mature field well settings.
1 Introduction
Object models are one of the earliest geostatistical approaches to facies modeling,
originating with Bridge and Leeder (1979). This model lends itself nicely to both
geometrical descriptions of geology and Bayesian statistical modeling. A mathe-
matically rigorous description can be found in Holden et al. (1998). Other
approaches can be found in Deutsch and Wang (1996) and Viseur et al. (1998).
In the unconditional case, object models are simple to handle and can create very
realistic geology. They are the primary choice for generating training images for
multipoint methods such as Strebelle (2002). Low well densities are also easy to
handle both stringently and ad hoc, but as the wells get dense compared to the
object size, these models tend to have conditioning problems. This is seen either as
overshooting of target volume fractions or by lack of conditioning in some well
observations.
As the well density increases, the first noticeable problem is generally that each
object is not able to cover enough observations. Thus, more objects are needed to
to fit well observations. However, this non-iterative approach is not easy to fit into a
Bayesian framework.
A recent interesting approach, which also argues that the algorithms are the
problem, is that of Boisvert and Pyrcz (2014). Here, they look at the well condi-
tioning problem as an optimization problem and get good results. However,
establishing a sampling algorithm from an optimization is generally a difficult
problem, and we do not see how this can be generalized for any explicitly defined
model.
The next section describes the model we use, followed by a section with the
sampling algorithm for this model. The sampling algorithm is the new contribution
in this paper. Finally, we present some examples, followed by concluding remarks.
2 Object Model
We will here introduce our object model and its key parts. We also take a brief look
at some key differences between an object model and a multipoint model.
In this paper we use the same model as in Holden et al. (1998), but we will focus
only on the parts that are important for well conditioning. For a deeper understand-
ing of the full model, we refer to Holden et al. (1998). Choosing this point of view
means that we consider channel objects in a shale background; however, the
methodology is easily generalized.
The probability density for a realization r given well data d and volume fraction
target v is given by
f ðr jd, vÞ / f g ðr ÞI d ðr ÞI v ðr Þf i ðr Þ: ð1Þ
The first term here, fg, is the geological prior, controlling the shape of objects. Id
and Iv are indicators showing that all wells are conditioned and the volume fraction
is correct. Finally, fi is an interaction term, ensuring that objects do not get too close
to each other.
For a completely determined model, fg should consist of an intensity term, giving
the probability of the number of objects n in r, and the product of geometry priors
for each of the objects. We have found that in practice the distribution for n is very
troublesome; hence, we use
560 R. Hauge et al.
Y
n
f g ðr Þ ¼ f o ðoi Þ;
i¼1
where fo is the geometry distribution for a single object, i.e., the distribution for size
and shape. The reason for ignoring the intensity term is that we do not know the
distribution for the number of objects in a realization. This distribution must be
estimated from the volume fraction. However, complexities in shape, interactions,
and erosion patterns can lead to inconsistencies in this estimate, giving a model
where the volume fraction term is very difficult to satisfy given the distribution for
the number of objects. Hence, we rely only on the volume fraction to control the
number of objects. This works since the interaction keeps the objects from hiding
inside each other.
The well conditioning indicator, Id, is 1 if all well observations are satisfied and
0 otherwise. This means that only realizations which honor the well data have
positive likelihood.
The volume fraction indicator, Iv, is 1 if all volume fraction criteria are fulfilled.
In this paper, we will only consider a global volume fraction, but any volume
fraction criteria that can be given by an indicator are valid. When working with
large objects, it is important to have enough tolerance on the volume fraction, since
adding or removing one channel can change the volume significantly.
We will not discuss the interaction term, fi, here since it is included only to create
a bound on the number of objects.
Each channel is parameterized along a line, which is denoted as the channel line.
The direction of the channel line is determined by the azimuth and dip, which are
typically given a normal prior distribution. The location within the simulation area
is most often uniform.
Let us for simplicity consider the case with no dip, where the lateral extent of the
object is determined in the xy-plane. The channel line defines a local x-axis where
y ¼ 0. We let the channel line and the local coordinate system of the channel start at
the point where the normal to the line in the xy-plane is tangential to the simulation
area, setting x ¼ 0 here. Similarly, the channel line ends at the other end of the
simulation area, at x ¼ L, where L is the length of the line. See Fig. 1.
We then add a 1D Gaussian field, mh, to this line, to define the local center of the
channel laterally. Another 1D Gaussian field, w, is defined, independent of mh. This
field is the local width of the channel.
Thus, the left and right channel edges, l(x) and r(x), at local x-coordinate x are
independent and are given by
Object-Based Modeling with Dense Well Data 561
becomes a rectangle, and we can generate 2D Gaussian fields on it. The top and
base of the channel in local position (x,y2D) are given by
A notable difference between object models and grid-based models such as indi-
cator simulation or multipoint is that for some combinations of input parameters
and well data, object models explicitly fail, either to condition all well data or to
achieve the correct volume fraction. The reason is that object models are
implemented in a way that preserves a minimum of geometrical integrity. All
Object-Based Modeling with Dense Well Data 563
objects generated will have positive probability density under the model. This also
implies that some facies realizations will have zero probability density, as they
cannot be generated by the given objects, so there is a true null space.
Pixel-based models will never assign zero probability to a realization. Standard
multipoint implementation, as in Strebelle (2002), shows the pragmatic stance
taken here: Initially, probabilities are assigned to the patterns seen in the training
image, based on how often they are seen. No assumptions are made for patterns that
are not seen; however, we know that some patterns should be strictly forbidden,
whereas others should have a positive probability. The next step is to fill in facies so
that the patterns that have assigned probabilities are preserved locally. Inevitably,
this creates conflicts elsewhere, and we run into unseen patterns. This is then solved
by assigning probabilities to the patterns with similarities to the ones that got an
initial probability. This entire process is ad hoc and simulation dependent; however,
it ensures that the simulation never stops. A result is always achieved, with correct
volume fraction and well conditioning. When a multipoint model meets a compli-
cated well set, it will thus sacrifice geometry, to a degree that object models
never will.
The challenge for object models is whether the null space is too large for them to
be useful in densely drilled reservoirs. However, what we argue here is that there is
still quite a bit to gain with better algorithms. As well data gets denser and more
complicated, we must be able to explore less likely realizations, as this is where the
data will take us.
Again, we refer to Holden et al. (1998) for details of how to implement the MH
algorithm in this setting. The most important aspect of the MH algorithm is that it
generates a Markov chain that converges to the desired distribution. All we need to
provide is a Markov chain sampling algorithm that covers the state space. We must
be able to compute the sampling densities and the true density of the realization up
to a scaling constant.
This allows us to manipulate the drawing algorithm any way we want. A
drawback is inherent in the fact that we need to use this algorithm: We are not
able to draw directly from the true distribution but must rely on the iterative
564 R. Hauge et al.
convergence of the MH algorithm. This means that object models based on this
algorithm will be slow compared to single pass grid-based algorithms.
There are two extremes for the choice of the sampling algorithm. One is to go for
an algorithm that is very easy to compute so that proposals can be generated very
fast. The idea here is that a low acceptance rate does not matter as long as each
iteration is very fast. This is more or less the approach in Deutsch and Wang (1996),
where they do not look at well data at all but generate unconditional channels. The
idea is that discretized on a grid, there is a positive probability of getting the correct
facies in all wells. This has extremely slow convergence in complex well situations.
The other extreme is going for very computer-intensive proposals with high accep-
tance rates. The algorithms described in Hauge et al. (2007) and Skorstad et al.
(1999) move in that direction. In this paper we take that approach even further.
We first draw the channel line for the object and transform the relevant well data to
local coordinates. Each object is parameterized at regular intervals along the
channel line, called sections. Lateral conditioning points are channel edge points
drawn in areas where the channel is close to wells to ensure consistency. These are
put in the sections on both sides of the observation, to ensure conditioning. See
Fig. 4 for an illustration. Sections without conditioning points are later filled
through simulation and kriging. Initially, the algorithm is a 2D version of the one
in Skorstad et al. (1999).
We divide the facies observations into four categories: Negative observations
are shale observations, and we must avoid them. Positive observations are sand
observations that we will condition with this channel. We often build a channel
around an observation, and the user may also have specified other observations
belonging to the same object. Possible observations are other channel observations
that may or may not be conditioned by this object, and irrelevant observations are
channel observations conditioned by other channels or missing observations.
This is done based on the probability of having the channel edges encompass both
the possible observation and existing positives, as well as the likelihood of the top
and base fields matching the end points of the possible observation. There is also a
stochastic element in this choice, to span the realization space. So an observation
that fits nicely into the expected channel shape will have a high probability of being
Object-Based Modeling with Dense Well Data 565
Fig. 4 Three vertical wells in an xy cross section and the corresponding conditioning point for
edges. Given these, positive observations are inside the channel, and negative are outside
When drawing conditioning points in a section, we know the distribution for the
lateral edges, the vertical center displacement, and the vertical thickness. We have
chosen to condition these on the three nearest previously drawn conditioning points,
if such exists.
First, we find the expected vertical position of the object center. If we have
positive observations, we condition on these. Otherwise, this is done from the prior.
Given the vertical location, we are ready to map out the lateral possibilities.
We then reduce the well data to what we need to consider for this particular
object. Since we are now generating the lateral conditioning, we remove all well
data that is too far above or below the current vertical location. This distance
depends on the vertical standard deviation for the center but is also a tuning
parameter. So far, we have chosen to only take a narrow interval close to the center
and assume that the top and base fields can handle the rest. The wells are divided
into facies observations and projected down to the xy-plane of the object.
At this stage, all observations are already categorized as positive (object must
condition), negative (object must avoid), or irrelevant (object may pass through). In
a well, a positive observation is typically adjacent to negative observations, and
these are set to be irrelevant here, since we must partially cover them laterally.
566 R. Hauge et al.
Fig. 5 Cross section in the xy-plane showing two sections. The edges are known in the left section
and projected through the well observation points
Fig. 6 Cross section in the xy-plane showing two sections and two well observations at relevant
depth. From the projections, we see that the channel edges in section j must lie so that the L1–R3
interval is completely covered, while none of the interval R4–L4 is covered
At first, a starting point along the channel is selected, typically where we have
one or more positive observations between section j-1 and j. Assume that the
channel edges in section j-1 are known. We then create a map of legal locations
for the edges in section j by drawing straight lines from the edges in section j-1
through all the well observations and onto section j; see Fig. 5. Given the two edge
projections through a point, the interpretation is simple: If any part of the interval
these generate in section j lies between the channel edges in section j, the point is
inside the channel laterally.
This means that all edge projections together create a simple map for section j,
from where the channel edges must be drawn based on the following criteria: We
must avoid the intervals generated by negative points and cover (at least parts of)
the intervals generated by positive points. This is illustrated in Fig. 6.
So far, the algorithm is a slight simplification of the one described in Skorstad
et al. (1999). However, we now extend the algorithm by also creating another map,
looking at the transition from section j to j + 1.
Object-Based Modeling with Dense Well Data 567
Given all the conditioning points, which are symmetrically located on the right and
left edge, it is trivial to simulate and krige the channel edges.
The vertical conditioning is simpler than the lateral and is why we try to do much of
the conditioning here. First of all, we can disregard all well observations that are
outside the channel edges. We then identify the transition points at the ends of
positive observations. We also identify points where negative observations are
close to or inside the expected object and points where positive observations are
close to or outside the top or base of the expected object.
The set of transition points are possible exact observations of the top or base of
the object. We assume that a well going downward passes in through the top of an
object and out through the base and opposite for a well moving upward. This is a
slight approximation but greatly simplifies the problem. If an observation goes up
through the top into another channel, it is considered internal, and not an edge, since
the channel above has eroded into this channel.
Given the observations of top and base, we draw additional conditioning points
for top and base that ensures correct conditioning where we have high probability of
errors. At first we draw at the most crucial locations, that is, where the probability of
well conflict is largest. Multiple points close to each other are grouped to draw
conditioning points that satisfies them all. Figure 8 illustrates vertical conditioning
points. Based on the vertical conditioning points, the top and base of the object are
simulated and kriged to fit the observations.
7 Examples
We present three examples here. In two of them, we show classic failure cases for
object models and how our new algorithm handles them, whereas the third case
shows the improved trade-off between prior model and data in our new algorithm.
In this example, we have a densely drilled field. The dimensions of the field are
1,000 m by 1,000 m, and the well spacing is 50 m in both directions, as shown at the
left in Fig. 9. We have generated an unconditional realization with two channels and
sampled these into the wells, as facies observations. Trying to run a commercial
state-of-the-art object model on this case fails, meaning that we could not obtain a
single realization even after hours of run time. With our new algorithm, we are able
to generate a realization in less than a minute on an ordinary PC. Since both use the
same model, the improvement is due to the algorithm. A realization generated by
the new algorithm is shown to the right in Fig. 9.
Not only do we generate realizations, but by monitoring our proposals, we see
that we also generate almost 100 % valid proposals. That is, we almost never have
to abandon a channel proposal due to well issues not handled by our algorithm. This
is also an important speedup, since time wasted on generating invalid objects either
leads to more or slower iterations. On the downside, there seems to be some kriging
overshoot in the upper channel, with the spike coming out. This is due to densely
drawn points not following the variogram completely, here a function of us
Fig. 9 At the left is the dense vertical well pattern with the observed channels, and at the right a
simulation of two channels conditioned to these well observations
570 R. Hauge et al.
Fig. 10 At the left is the well pattern from a drilling pad, where we see one channel in black. At
the right is a channel conditioned to these observations with our new algorithm
conditioning only on the three nearest points. Kriged points may then generate
extremes as here.
Although there are fewer wells in this example, the general pattern is more
complex, as all wells are deviating from a common center. We only have a single
channel in our unconditional simulation; again, this is sampled into the wells as
facies observations. The well pattern and a realization generated from by the new
algorithm are shown in Fig. 10.
The results are the same as before. Our new algorithm is able to generate valid
channels in almost all proposals, although we note that the issue with overshooting
in the kriging of the edge fields is more severe here. We believe this is a tuning issue
that can be solved by choice of variogram and distance between conditioning
points. The commercially available algorithm fails to condition this pattern, mean-
ing that not even one valid proposal was generated when run for hours.
The two previous cases have been very clinical and only looked at the improvement
in complex cases where all parameters were set correctly. But the first indication of
failure in an object model is that it starts to sacrifice its prior distribution. We
believe this happens too fast in existing models and has created a test example to
explore this.
We have taken a set of vertical wells and sampled a realization into these, as
shown in Fig. 11. This is our data set, and so far, it is similar to the ones above.
Object-Based Modeling with Dense Well Data 571
Table 1 The number of channels originally, and the mean and standard deviation (in parentheses)
for the number of channels with 10 runs of our new and the commercial algorithm
Original New algorithm Commercial
# of channels 40 25.9 (0.53) 29.4 (0.79)
8 Concluding Remarks
Given the space available, we were not able to present the full model, which also
includes levees and crevasses, with full conditioning of the entire hierarchy. Even
with the focus on well conditioning, what is presented here is only the key idea in a
large and complex framework. As indicated by one of the examples, the tuning is
still not optimized. But already, we see significant advances in the complexity of
what can be handled. This means that object models can play an important role in
fields where the well data amounts previously was not handled by such models.
572 R. Hauge et al.
Although we only present the basic channel and shale framework here, it should
be clear how this can be generalized to any object that can be straightened out in a
local coordinate system, since this is where the well conditioning occurs. Objects
built around piecewise linear backbones would thus fit right into this conditioning,
whether they are object- or rule-based.
Bibliography
Boisvert J, Pyrcz M (2014) Conditioning 3D object based models to a large number of wells: a
channel example. Mathematics of planet earth: proceedings of the 15th annual conference of
IAMG. Springer, p 575–579
Bridge J, Leeder M (1979) A simulation model of alluvial stratigraphy. Sedimentology
26:617–644
Deutsch CV, Wang L (1996) Hierarchical object-based stochastic modeling of fluvial reservoirs.
Math Geol 28:857–880
Hastings WK (1970) Monte Carlo sampling methods using Markov chains and their applications.
Biomterika 57:97–109
Hauge R, Holden L, Syversveen A (2007) Well conditioning in object models. Math Geol
39:383–398
Henrion V, Caumon G, Cherpeau N (2010) ODSIM: an object-distance simulation method for
conditioning complex natural structures. Math Geosci 42:911–924
Holden L, Hauge R, Skare Ø, Skorstad A (1998) Modeling of fluvial reservoirs with object models.
Math Geol 30:473–496
Skorstad A, Hauge R, Holden L (1999) Well conditioning in a fluvial reservoir model. Math Geol
31:857–872
Strebelle S (2002) Conditional simulation of complex geological structures using multiple-point
statistics. Math Geol 34:1–21
Vargas-Guzman J, Al-Quassab H (2006) Spatial conditional simulation for facies objects for
modelling complex clastic reservoirs. J Petrol Sci Eng 54:1–9
Viseur S, Shtuka A, Mallet J (1998) New fast, stochastic, boolean simulation of fluvial deposits.
Proceedings, 1998 SPE ATCE, New Orleans, USA. SPE 49281
Machine Learning Methods for Sweet Spot
Detection: A Case Study
Abstract In the geosciences, sweet spots are defined as areas of a reservoir that
represent best production potential. From the outset, it is not always obvious which
reservoir characteristics that best determine the location, and influence the likeli-
hood, of a sweet spot. Here, we will view detection of sweet spots as a supervised
learning problem and use tools and methodology from machine learning to build
data-driven sweet spot classifiers. We will discuss some popular machine learning
methods for classification including logistic regression, k-nearest neighbors, sup-
port vector machine, and random forest. We will highlight strengths and shortcom-
ings of each method. In particular, we will draw attention to a complex setting and
focus on a smaller real data study with limited evidence for sweet spots, where most
of these methods struggle. We will illustrate a simple solution where we aim at
increasing the performance of these by optimizing for precision. In conclusion, we
observe that all methods considered need some sort of preprocessing or additional
tuning to attain practical utility. While the application of support vector machine
and random forest shows a fair degree of promise, we still stress the need for
caution in naive use of machine learning methodology in the geosciences.
1 Introduction
In petroleum geoscience, sweet spots are defined as areas of oil or gas reservoirs
that represent best production potential. In particular, the term has emerged in
unconventional reservoirs where the reserves are not restricted to traps or struc-
tures, but may exist across large geographical areas. In unconventional reservoirs
the sweet spots are typically combinations of certain key rock properties. Total
organic carbon (TOC), brittleness, and fractures are some of the properties influenc-
ing possible production. In identifying these sweet spots, the operators face the
challenge of working with large amounts of data from horizontal wells and model-
ing the complex relationships between reservoir properties and production.
In general, a more data-driven approach for sweet spot detection allows for a
more direct use of less costly reservoir data, such as seismic attributes. Moreover,
such an approach may potentially avoid parts of the expensive reservoir modeling.
In particular, the time-consuming computations needed to build a full reservoir
model can be avoided. Fast and reliable classification of the sweet spots is of high
significance, as it allows for focusing efforts toward the most productive areas of a
reservoir. This makes machine learning algorithms desirable, since these are typ-
ically fast to train, often easy to regularize, and have the ability to adapt and learn
complex relationships.
The use of machine learning methodology for predicting and detecting potential
areas of interest is gaining attention and is not new to the geosciences. A multidis-
ciplinary workflow in order to predict sweet spot locations is presented in Vonnet
and Hermansen (2015). An example of support vector machine application on well
data for prediction purposes is given in Li (2005). In Wohlberg et al. (2006), the
support vector machine is demonstrated as a tool for facies delineation, and in
Al-Anazi and Gates (2010), the method is applied for predicting permeability
distributions.
In this paper we continue this exploration and view sweet spot detection in a
machine learning setting, framed as a traditional supervised learning problem, i.e.,
classification. These are data-driven algorithms that aim to learn relationships
between the reservoir properties and sweet spots from labeled well-log training
data. We illustrate different popular machine learning algorithms through a case
study, considering a real and challenging data set with a weak signal for sweet
spots. The algorithms we consider and compare are logistic regression, k-nearest
neighbor (kNN), support vector machines (SVMs), and random forest.
We will emphasize a more moderate and cautious approach to uncritical use of
machine learning for classification, wherein the awareness of what we can learn is
of significance for interpreting the results. The main challenge here is related to the
low data quality and the limited evidence for sweet spots (see Sect. 2). In such
cases, the focus should be on the confidence of evidence of sweet spots, despite a
potentially low discovery rate. There is usually a high cost associated to exploration
and development of a field. It is therefore generally better to sacrifice some sweet
spots (i.e., detection rate) in order to gain accuracy and precision. This is our main
focus and we compare the ability of these machine learning algorithms to learn
from a weak signal. We show how a simple modification can be used to improve
such methods and how this improves recovering of the potential and providing
sufficiently confident evidence of sweet spots. We also discuss the inadequacy of
simple summary statistics for model validation and show that generally a more
detailed investigation is needed in order to assess the actual performance.
In Sect. 2 we describe our real data set and set the sweet spot detection in a
binary classification setting. Next, in Sect. 3, we discuss the machine learning
algorithms used in this case study. Section 4 outlines the setup for training and
Machine Learning Methods for Sweet Spot Detection: A Case Study 575
validating the machine learning methods, before the numeric results are presented
and discussed. Lastly, Sect. 5 concludes the case study.
The training and validation of machine learning methods and the predictions and
numeric comparisons are carried out in R, using the package e1071, class, and
randomForest.
The case study consists of labeled observations from four vertical blocked wells in a
reservoir, providing a total of 315 observation points. For each observation point,
there are six reservoir properties available for training, henceforth referred to as
features. These are the seismic attributes P-wave velocity (Vp), S-wave velocity
(Vs), density, acoustic impedance (AI), 4D residual of pre-stack time migration and
average magnitude of reflectivity. In addition, total organic carbon (TOC) and
gamma ray (GR) are provided in the wells, which are used to set the labels, i.e.,
sweet spots. See Fig. 1 and Table 1 for details regarding the number of observations
and fraction of sweet spots to non-sweet locations in the wells.
Note that the first four features (Vp, Vs, density, and AI) have been corrected for
a depth trend. Thus, the a priori background model for the parameters has been
removed, since this introduced a systematic bias in the predictions.
To illustrate the complexity of this data set and the weak relationship in the
underlying relationship between sweet spots and reservoir properties, we plot four
cross plots of a selection of pairwise combinations of the six features in Fig. 2.
These plots indicate a quite strong linear correlation between Vp and density and
also between density and AI. Moreover, there is no clear relationship between AI
and 4D residual and 4D residual and the average magnitude of reflectivity. This
seems to be caused by the high level of noise in the measured 4D residuals. In all
four cross plots, there is no trace of geometric delimitation of the sweet spots.
Indicating that the well data is not easily linearly separable in the feature space,
hence we expect a complex, or limited, relationships.
The reservoir model used for predicting sweet spots has dimensions 280 by
350 by 100 cells. Figure 3 shows the top layer and a vertical slice of the acoustic
impedance. Note that the upper left and lower right corners of the lateral view do
not contain defined values. Six wells are marked with a circle in the lateral plot.
Wells 1, 2, 5, and 6 have given the features defining the sweet spots. The two
additional wells in the reservoir, Wells 3 and 4, lack values for total organic carbon
and gamma ray. They cannot be used to define the sweet spots and are therefore
excluded from further analysis.
The S-wave velocity in the labeled data set appears to be artificially constructed
from P-wave velocity, as the estimated correlation between the two is above 0.99,
which is also verified by plotting. Traditionally, we would be inclined to exclude
one of these variables in the statistical analysis, e.g., to avoid collinearity. However,
576 V.L. Hauge and G.H. Hermansen
we will keep both features in our training data to test and illustrate the robustness of
the (probabilistic) model-free machine leaning methods.
The sweet spot classification is a binary classification problem, where we
identify the two classes: sweet spots and non-sweet spots. In a binary classification
there are four possible outcomes summarized below:
Fig. 2 Cross plots of a selection of pairwise combinations of the six features. All features are
plotted in the normalized domain, hence no units along the axis. Again, red color marks obser-
vations defined as sweet spots; blue color marks non-sweet spots
Fig. 3 Top layer of the acoustic impedance (left). The wells are numbered from 1 to 6, of which
Wells 3 and 4 do not have defined sweet spots. A vertical slice of the acoustic impedance along the
dashed line marked in the top layer (right). Note the values are corrected for a depth trend
578 V.L. Hauge and G.H. Hermansen
The TDR is a measure of the recall (or sensitivity) of the classification and
describes how well the classification method correctly detects the sweet spots that
actually are sweet spots. The TPR is a measure of precision and gives the proportion
of predicted sweet spots that are actual sweet spots. To combine the measure of
recall and precision, we will use the Fβ-score defined as the weighted harmonic
mean of recall and precision:
1 þ β2 TPR TDR 1 þ β2 TP
Fβscore ¼ ¼ ð3Þ
β2 TPR þ TDR 1 þ β2 TP þ β2 FP þ FN
In the following, we will use the balanced weighting with β ¼ 1, denoted F1 score.
The more general Fβ score will become of value for tuning the algorithms.
Especially for the SVMs this score will be used as a mean of favorable balancing
of TPR and TDR to avoid overfitting and collapsing the model to the uninteresting
solution of predicting all locations as either sweet or non-sweet spots.
In the sweet spot setting, we argue that TPR is of most importance, as an
assurance of correct sweet spot predictions. On the other hand, a carefully balanced
focus on the TDR will ensure that more sweet spots are found, at the cost of
including misclassified sweet spots. Again, care is needed when tuning methods
against these measures.
Moreover, we expect that there is an overrepresentation of sweet spots in the
data. This seems obvious, since the initial or any wells are not placed randomly into
the field, but they are placed exactly where the developers expect they have the
greatest potential for success, i.e., in the sweet spots. This suggests that there is
most likely a confounding, or omitted, variable not observed. The information and
process underlying the positioning of wells can be thought of as an unobserved (and
highly complex) variable influencing both the response and the explanatory vari-
ables. This may in turn result in an unbalanced data problem (too many sweet spots)
and introduce potentially complex correlations among the explanatory variables
and the response; see, among others, He and Garcia (2009) and King and Xeng
(2001) for additional discussion. It is generally hard, or even impossible, to correct
for such; see Li et al. (2011) for an attempt to correct the support vector machine.
The logistic regression model is particularly sensitive; see also Mood (2010). As a
final remark, if we consider the overall reservoir from which well logs are collected,
we will expect a minority of the sweet spots, causing an additional imbalance, this
time in the opposite direction. Proper treatment of such effects and possible
extensions are outside the scope of this paper.
Machine Learning Methods for Sweet Spot Detection: A Case Study 579
In general, machine learning refers to algorithms and statistical methods for data
analysis. Here, we will focus on machine learning methodology for prediction of
binary class labels, i.e., two class problems. It should be pointed out that all
methods discussed can easily be generalized to multiclass problems. We will
consider four common and popular supervised learning algorithms, which are the
logistic regression, random forest, k-nearest neighbor (kNN), and support vector
machine (SVM).
Random forest is the ensemble of multiple decision or classification trees; see, e.g.,
Hastie et al. (2009). A decision tree is a greedy approach that recursively partitions
the feature space. A single decision tree will easily overfit the training data to the
580 V.L. Hauge and G.H. Hermansen
test data and has potentially a large bias. In particular, with noisy data, the
generalization of a single decision tree is poor. To avoid overfitting, the ensemble
of decision trees, i.e., random forest, averages multiple decision trees based on
different resampling of training data. Each of the trees in the ensemble has
potentially a high variance, and the averaging of the ensemble reduced this vari-
ance. In general, random forest is computationally efficient and is easily
interpreted. For more details, we refer the reader to Breiman (2001).
The k-nearest neighbor (kNN) algorithm is one of the simpler and more robust
supervised learning algorithms. An introduction can be found in any introductory
textbook in machine learning. The algorithm classifies a new observation, or
location, by comparing it with the k-nearest observations in the training set and
classifies the new observation according to the dominant class. This algorithm is
completely model-free and nonparametric. However, each new prediction needs a
unique nearest neighbor search. This makes the algorithm less efficient for large
data sets and prediction grids. The best choice of the number of neighbors, k,
depends upon the data. In our case we perform a cross validation to find this
parameter. In general, small values of k may result in noisier results. Larger values
of k reduce the effect of noise, but make boundaries between classes less distinct.
This algorithm will always improve with more data, and the method is known to
work well in simpler classification problems; see also Beyer et al. (1999).
Lastly, support vector machine (SVM) classifies data by finding a hyperplane that
separates the data. In the case of linearly separable data in two dimensions, the
separating hyperplane is a separating line. Figure 4a shows an illustration of a
linearly separable case with the separating line marked as the black line and the data
points (support vectors) defining the line marked with circles. The dashed lines
mark the margins, i.e., the distance from the separating line to the nearest data
points.
For data sets that are not completely separable, the concept of soft margin is
introduced to allow some data to be within the margin. SVM now attempts to find a
hyperplane that separates the data as cleanly as possible, however, not strictly
enforcing that there are no data in the margin (hence the term soft margin). The
soft margin is controlled through a regularization parameter, often referred to as
C. A large value for this regularization parameter aims for a smaller soft margin and
fewer misclassified points. On the other hand, a small value for the regularization
parameter aims for a larger soft margin, allowing more points to be misclassified
Machine Learning Methods for Sweet Spot Detection: A Case Study 581
Fig. 4 Illustration of SVM for (a) a linearly separable data set, (b) a non-separable data set with
soft margins, and (c) a nonlinear separating hyperplane
and yielding a smoother decision boundary. Figure 4b has interchanged three points
between the blue and red classes, making the data set linearly inseparable. This
figure shows the separating plane as the black line, support vectors again marked
with circles, and we observe that some points are allowed to appear within the
margins (dashed lines).
The SVMs handle nonlinear classification by applying the so-called kernel trick,
which allows for nonlinear decision boundaries, while the algorithm for the linear
SVM still can be applied for determination of the hyperplane. The kernel trick can
be thought of as mapping the observation points into some higher-dimensional
space, in which an optimal separating hyperplane is found. Projecting the hyper-
plane back to the original space yields a nonlinear decision boundary. A typical
choice for the kernel function applied is the radial basis function; see, e.g., Hastie
et al. (2009). The radial basis function kernel is a scaled version of the Gaussian
kernel, in which the squared Euclidean distance between two features is scaled by a
free parameter. In the following, we will denote this kernel parameter γ. Adjusting
these parameters allows the decision boundary to go from finely detailed decision
boundary to a coarser distinction between the classes. Figure 4c shows a nonlinear
separating boundary.
The use of SVMs is of great interest as a sweet spot classifier, as it is known to
perform well in classification problems where the decision regions of the feature
space are of a smooth geometric nature, as we expect to be the case in several
applications in the geosciences. The SVM is often referred to be the “out-of-the-
box” classifier and is known to be of high accuracy and has the ability to deal with
high-dimensional data, i.e., usually no preselection of features is needed.
For a more extensive introduction to SVMs, the reader is referred to Bishop
(2006) and Cortes and Vapnik (1995).
582 V.L. Hauge and G.H. Hermansen
4 Numeric Comparisons
In the following, we first outline the setup for validating the various machine
learning methods. Next, we report results of several comparisons. Along with the
discussion of the results, we present additional tuning of the methods to sharpen and
balance the performances.
To evaluate the machine learning methods, we use the labeled data and carry out a
fitting (training and testing) and validation setup. In 100 rounds of validation, we
assign 30–70 % of the labeled data set (randomly) for validating. The rest is left for
fitting. In the validation, the fitted methods are applied on the validation data set,
and F1 score, True Prediction Rate (TPR), and True Detection Rate (TDR) are
recorded. For fitting of the methods (training and testing), again 30–70 % is
assigned (randomly) for testing the methods, leaving the rest of the data set for
training. In both training and testing, cross validation is used to obtain optimal
parameters for the algorithms. Here we have focused on maximizing mainly the
TPR value, but also various Fβ-scores. After the 100 rounds of training, testing, and
validating, we average the obtained performance measures.
Note that when validating the methods, we randomly choose the observations
from all of the four wells. We also consider a more real-case predicting study,
where we sequentially hold out one well, fitting the methods on the remaining three
wells, and investigate performance on the held-out well.
The optimal parameters, found by cross validation, refer to the parameters
yielding, e.g., the largest TPR score. For the kNN we find the optimal number of
nearest neighbors 0 < k <30. For the SVM we cross validate for the regularization
parameter 25 < C < 210 and kernel parameter 210 < γ < 25 . For the random
forest algorithm, ensembles of up to a couple of 1000 trees were tested. Interest-
ingly we saw no significant change in performance for ensembles of more than
100 trees.
4.2 Results
Table 2 summarizes the performances of random forest, kNN, and SVM applied as
described above. We report the obtained performance measures from the fitting, as
well as on the validation set (in bold). The last column in Table 2 reports perfor-
mance of a random classifier, which randomly (with equal probability) assign
predictions as sweet or non-sweet spots. For the logistic regression, we were not
able to obtain any results better than TPR of 0.10. The failure of the logistic
Machine Learning Methods for Sweet Spot Detection: A Case Study 583
Table 2 Summary of the performance of random forest, kNN (tuned k), and SVM (tuned C and γ)
Random forest kNN SVM Random
F1 0.24 0.23 0.51 0.50 0.50 0.49 0.40
TPR 0.33 0.32 0.35 0.34 0.36 0.34 0.33
TDR 0.21 0.19 1.00 1.00 0.88 0.89 0.50
All tuning is optimized for TPR. The first column for each method is the measures obtained on the
testing sets. The second column, marked with bold, is the measures obtained on the validation sets.
Last column is the performance of a random classifier
Table 3 Summary of the performance for random forest (excluding the features 4D residual and
average magnitude of reflectivity), kNN (also threshold tuned), and SVM (also threshold tuned)
Random forest kNN SVM
F1 0.32 0.34 0.41 0.38 0.21 0.27
TPR 0.38 0.40 0.38 0.35 0.49 0.44
TDR 0.30 0.33 0.61 0.58 0.20 0.26
All tuning is optimized for TPR. Columns are for test and validation sets as given for Table 2
Comparing reported results in Table 3 with Table 2, we see that (prior) feature
selection for random forest increases both precision and detection and seems to be
the winner among the three. Additional tuning provided no significant improve-
ments to the kNN algorithm, suggesting that sophisticated versions of kNN are
required, e.g., the popular (Friedman 1994) or the more involved (Goldberger et al.
2005). The SVM algorithm received a considerable increase in the TPR score,
indicating a good potential for additional fine-tuning of the SVM toward the most
important properties (e.g., a predefined balance between TPR and TDR).
To evaluate and to get a better understanding of how the obtained performance
measures will transfer to the real field, we now fit the models by sequentially
holding out one of the wells. Firstly, Fig. 5a shows predictions in all four wells
using random forest with four features as specified in Table 3. Here we get a visual
impression of how well sweet spots are predicted. We note several missing sweet
spots in the predictions, as well as sweet spots detected where the labeled data show
non-sweet spots. We accompany the plots of predictions with Table 4, reporting
obtained performance measures in the wells.
Next, Fig. 5b shows predictions in all four wells using kNN, with only tuning of
the number of neighbors, k (as specified for Table 2). This poor performance is
included to illustrate how “good” performance measures indeed transfer to real
field. Although we might be led to believe in the predicting power of kNN from
Table 2, here kNN is either useless (as in Wells 1 and 6) or yields quite noisy
predictions (as in Wells 2 and 5). Also, note that for Well 5 the performance
measures in Table 4 are indeed the same as one would expect from a random
classifier.
Acknowledging the need for additional balancing of TPR and TDR for a best
trade-off, we introduced for SVM additional tuning of the weighting of TPR and
TDR. This is done by optimizing the parameters, by cross validation, against the
Fβ-score, Eq. 3, for different values of β. The developments of the three perfor-
mance measures TPR, TDR, and Fβ score, as a function of the weight β, for the four
wells are shown in Fig. 6. Note that a TDR of 1.0 corresponds with predicting all
points as sweet spots, hence detecting all, at the cost of a large amount of
misclassifications.
Selecting an appropriate weight β for each well yields predictions in the wells as
displayed in Fig. 5c. By appropriate we here refer to the weights that best balance
TPR and TDR, typically at the point where TPR and TDR cross in Fig. 6. Here, the
“optimal” balance point is determined by inspecting of Fig. 6. Table 4 reports the
Machine Learning Methods for Sweet Spot Detection: A Case Study 585
Fig. 5 Prediction of sweet spots in the four wells using (a) random forest with four features and
(b) kNN, (c) fine-tuned SVM, and finally (d) random forest with four features and without
corrections of the depth trend. Leftmost well column for each well is the labeled data, while the
rightmost well column for each well is the prediction
Table 4 Obtained performance measures when sequentially holding out one well at a time
Random forest SVM kNN
TPR TDR TPR TDR β TPR TDR
Well 1 0.55 0.15 0.60 0.30 0.30 0.39 1.00
Well 2 0.25 0.33 0.00 0.00 0.40 0.19 0.78
Well 5 0.54 0.18 0.50 0.37 0.30 0.38 0.55
Well 6 0.53 0.42 0.38 0.53 0.45 0.31 1.00
For the random forest and kNN, the two columns are TPR and TDR. For the SVM, we report, in
addition to TPR and TDR, the weight β used in the optimization for parameters
586 V.L. Hauge and G.H. Hermansen
Fig. 6 Development of the three performance measures TPR (solid), TDR (dashed), and Fβ score
(gray) as a function of the weight β for the four wells. Left column shows the development on the
testing set, while the right column shows the validation set
weight β used for each of the wells. In general, we now observe in Fig. 5c that the
detection has increased compared to random forest, as well as the precision is kept
at an acceptable level, indicating good generalization potential.
Note again that predictions in Well 2 fail – more or less – for all methods.
Extracting wells as validation sets introduces a grouping of the observations. There
is reason to be skeptical regarding the results for Well 2. Figure 7 shows the
pairwise cross plots of some of the features, distinguished by color on the four
different wells. We observe that for Well 2 (red color) the features do not coincide
with the three other wells. Therefore, this well can be interpreted as significantly
noisier or to be representing something different. It is, of course, generally hard for
a predictor to predict something it has never seen before. On the other hand, a
Machine Learning Methods for Sweet Spot Detection: A Case Study 587
Fig. 7 Cross plots of three pairs of features, distinguished on color for the four wells. Red color is
for Well 2. Values are in the normalized domain, hence no units on the axis
simple linear classifier may still provide reasonable results, depending on the
structure of the underlying problem.
As pointed out earlier, four of the features in the data set have been corrected for
a depth trend. Figure 5d displays the obtained predictions applying random forest
by including depth as an independent feature. We observe a seemingly good match,
indicating possible spurious relationship. In the well data in our case study, the
majority of the defined sweet spots are indeed located toward the bottom of the
reservoir. However, none of the other methods performed acceptably with the depth
trend; results were indeed worse.
5 Conclusion
Acknowledgment We thank Arne Skorstad and Markus Lund Vevle, both at Emerson Process
Management Roxar AS, for the data set and for answering questions related to it.
588 V.L. Hauge and G.H. Hermansen
Bibliography
Al-Anazi A, Gates I (2010) A support vector machine algorithm to classify lithofacies and model
permeability in heterogeneous reservoirs. Eng Geol 114(3–4):267–277
Beyer K, Goldstein J, Ramakrishnan R, Shaft U (1999) When is “nearest neighbor” meaningful?
In: Database theory — ICDT’99, vol 1540. Springer, Berlin, pp 217–235
Bishop CM (2006) Pattern recognition and machine learning (Information science and statistics).
Springer, New York
Breiman L (2001) Random forest. Mach Learn 45(1):5–32
Cortes C, Vapnik V (1995) Support-vector networks. Mach Learn 20(3):273–297
Friedman J (1994) Flexible metric nearest neighbor classification. Stanford University
Goldberger J, Roweis S, Hinton G, Salakhutdinov R (2005) Neighborhood components analysis.
Adv Neural Inf Process Syst 17:513–520
Hastie TJ, Tibshirani R, Friedman JH (2009) The elements of statistical learning: data mining,
inference, and prediction. Springer, New York
He H, Garcia E (2009) Learning from imbalanced data. IEEE Trans Knowl Data Eng 21
(9):1263–1284
King G, Xeng L (2001) Logistic regression in rare events data. Polit Anal 2:137–163
Li J (2005) Multiattributes pattern recognition for reservoir prediction. CSEG Natl Conv
2005:205–208
Li L, Rakitsch B, Borgwardt K (2011) ccSVM: correcting support vector machines for
confounding factors in biological data classification. Bioinformatics 27(13):i342–i348
Liaw A, Wiener M (2002) Classification and regression by randomForest. R News 2(3):18–22
Menard S (2002) Applied logistic regression analysis. Sage, Thousand Oaks
Mood C (2010) Logistic regression: why we cannot do what we think we can do, and what we can
do about it. Eur Sociol Rev 26(1):67–82
Platt JC (1999) Probabilistic outputs for support vector machines and comparisons to regularized
likelihood methods. In: Advances in large margin classifiers. MIT Press, Cambridge, pp 61–74
Vonnet J, Hermansen G (2015) Using predictive analytics to unlock unconventional plays. First
Break 33(2):87–92
Wohlberg B, Tartakovsky D, Guadagnini A (2006) Subsurface characterization with support
vector machines. IEEE Trans Geosci Remote Sens 44(1):47–57
Theoretical Generalization of Markov Chain
Random Field in Reservoir Lithofacies
Stochastic Simulation
1 Introduction
metric. Transition probabilities can closely link reservoir categorical variables and
show the spatial (asymmetric) correlation relationships. As a measure for charac-
terizing spatial variability (Huang et al. 2016b), transition probabilities can be used
as fundamental tools for the spatial Markov chain theory.
Markov chain model was introduced by Markov in 1906 and applied to stratig-
raphy in the 1940s. One-dimensional (1-D) Markov chains (or transition probabil-
ities) have long been used in geosciences; see Krumbein and Dacey (1969) and
Carle and Fogg (1997) for some reviews in geology and soil science.
Multidimensional (M-D) Markov chain models can be traced to Lin and Harbaugh
(1984) in geology for modeling lithological (or sedimentological) structures. A
spatial hidden Markov chain was introduced by Huang et al. (2016a) for estimation
of petroleum reservoir categorical variables in M-D space. The M-D Markov chain
model proposed by Elfeki and Dekking (2001) and Li et al. (2012) is considered
nearest known neighbors in cardinal directions with the fully independent assump-
tion. This M-D Markov chain is composed of multiple fully independent 1-D
Markov chains, and they are forced to move to the same location with equal states.
The fully independent assumption causes the small-class underestimation problem.
Li (2007b) solved this problem based on Markov chain random field (MCRF)
theory and proposed a single spatial Markov chain (SMC) that moves in an M-D
space, with its transition probabilities at each given location entirely depending on
its nearest neighbors in different directions under the conditional independence
assumption (Li 2007a, b; Huang et al. 2016c). Although the conditional indepen-
dence assumption has been widely used to construct M-D models, the rationality to
use this assumption has not been well discussed. The MCRF theory was not
sufficiently described mathematically in previous publications (Li and Zhang
2013). The problem is discussed and solved in this paper on the basis of the
equivalence of Markovianity (Markov property) and conditional independence.
We first put forward the background information of M-D spatial Markov chain
models in Sect. 2. The fully independent assumption and conditional independence
assumption are described in Sects. 2.1 and 2.2, respectively. The Equivalence of
Markovianity and conditional independence is then derived explicitly based on the
Bayes’ theorem (Sect. 3), which gives a solid logical proof for MCRF theory. In
Sect. 4, we use MATLAB programming to give a simple reservoir lithofacies
simulation example to compare the performance of conditional independence-
based MCRF method and fully independent-based model. Finally, Sect. 5
concludes.
Oil-gas reservoir field can be divided into discrete gridding units, and the
corresponding states (lithofacies) can be regarded as a sequence of random category
variables F1, F2, . . ., Fn defined on the n ordered spatial site set S ¼ f1; 2; . . . ; ng, in
which each random variable Fs takes a state value fs in the m state set
Theoretical Generalization of Markov Chain Random Field in Reservoir. . . 591
where s1, s2, . . ., sl is its nearest known locations of current unknown site s in
different directions. In order to study the complex distribution of subsurface
reservoir categorical variables, the SMC models, based on the fully independent
and conditional independence assumptions, respectively, are proposed and system-
atically discussed in this paper.
where
" #1
X
m
C¼ Pr f s f s1 Pr f s f s2 Pr f s f sl ; ð3Þ
f s ¼1
X
m
in order to ensure that Pr f s f s1 ; f s2 ; . . . ; f sl ¼ 1.
f s ¼1
If two-point conditional probabilities are replaced by transition probabilities
pfrs f s , Eq. 2 can be expressed as
r
where pfrs f s denotes a transition probability in the rth direction from state f sr to state
r
fs.
From Eq. 4, the conditional probabilities of two- and three-dimensional Markov
chain models are derived as follows:
592 X. Huang et al.
h v
puk ptk
Pr Z iþ1, jþ1 ¼ kZi, jþ1 ¼ u, Z iþ1, j ¼ t ¼ X
m ð5Þ
h v
puk ptk
k¼1
Pr Ziþ1, jþ1, rþ1 ¼ kZi, jþ1, rþ1 ¼ u, Z iþ1, j, rþ1 ¼ t, Z iþ1, jþ1, r ¼ q
h v y
puk ptk pqk
¼X
m ð6Þ
h v y
puk ptk pqk
k¼1
where Zi,j and Zi,j,r are the coupled Markov chain (CMC) and triplex Markov chain
(TMC), respectively, all defined in the state space Ω ¼ f1; 2; . . . ; mg, including k,t,
q, and u. pvtk , phuk , and pyqk represent three transition probabilities in the vertical z,
horizontal x, and y directions. Equations 5 and 6 can be found in Elfeki and Dekking
(2001) and Li et al. (2012), which have been applied in reservoir modeling.
However, the transition in Eq. 2 requires moving to the same location with equal
states, which causes big-class-overestimation or small-class-underestimation prob-
lem (Li 2007b).
where ηs denotes the non-neighboring set of site s. The categorical variables Fs1 ,
Fs2 , . . . , Fst corresponding to the neighboring sites s1, s2, . . ., st may be condition-
ally dependent or conditionally independent. At most q sites in s1, s2, . . ., st are
non-neighbors, i.e., for 8 si , i 2 1, 2, . . . , t, the other sites sj satisfy
sj2
= ηsi , ðj 2 1, 2, . . . , tÞ \ ðj 6¼ iÞ. Assuming that the q sites are s1, s2, . . ., sq and
corresponding categorical variables are Fs1 , Fs2 , . . . , Fsq . The conditional indepen-
dence assumption is given as
Theoretical Generalization of Markov Chain Random Field in Reservoir. . . 593
Pr f s1 , f s2 , . . . , f sq jf s ¼ Pr f s1 jf s Pr f s2 jf s Pr f sq jf s : ð8Þ
Let A, B, and C be three mutually disjoint subsets of the spatial site set
S ¼ f1; 2; . . . ; ng, suppose A and B are non-neighbors, i.e., for 8s 2 A, there does
not exist i 2 B and i 2 ηs . FA, FB, and FC are three mutually disjoint subsets of
categorical random variables in F and are defined on the sets A, B, and C,
respectively. We have that Prðf A f B jf C Þ ¼ Prðf A jf C ÞPrðf B jf A f C Þ:
By Markovianity, we obtain
594 X. Huang et al.
Prðf B jf A f C Þ ¼ Prðf B jf C Þ:
Therefore,
We find that FA and FB are conditionally independent given FC. The conditional
independence in the random field F implies that if A, B, and C are any three disjoint
subsets of the spatial site set S, A and B are non-neighbors; we have that FA and FB
are conditionally independent given FC.
We have proven that if the Markovianity holds, then the conditional indepen-
dence holds. Now we assume that the conditional independence holds, it needs to be
proven that the Markovianity holds. Take an arbitrary s 2 S, and let
A ¼ fsg, C ηs , B ηs . Obviously, A, B, and C are three mutually disjoint subsets
of S; in addition, A and B are non-neighbors. Because the conditional independence
in the random field F holds, we have that FA and FB are conditionally independent
given FC, i.e.,
Prðf C ÞPrðf A f B jf C Þ
Prðf A jf B f C Þ ¼ ¼ Prðf A jf C Þ; ð11Þ
Prðf B f C Þ
hence the Markovianity holds. Then we have the equivalence between the
Markovianity and conditional independence. By using the definition of conditional
probability, the general expression of the conditional probability at any location s in
a MCRF is expressed as
Prðf s1 Þ
Pr f s f s1 ; f s1 ; f s2 ; . . . ; f sl ¼ pf s1 f s p1f f p2f f pfl f
Pr f s1 ; f s1 ; f s2 ; . . . ; f sl s s1 s s2 s sl
ð12Þ
where pfrs f s denotes a transition probability in the rth direction from state fs to f sr ,
r
and pf s1 f s denotes a transition probability along moving direction of the spatial
Markov chain from state f s1 to fs.
Li (2007b) proposed the general expression of the SMC in an MCRF theory,
which is given as follows:
Theoretical Generalization of Markov Chain Random Field in Reservoir. . . 595
pf s1 f s p1 f s f s1 p2 f s f s2 . . . pl f s f sl
Prðf s jf s1 , f s1 , f s2 . . . , f s1 Þ ¼ X
m ð13Þ
pf s1 f s p1 f s f s1 p2 f s f s2 . . . pl f s f sl
f s ¼1
Note that Eqs. 12 and 13 are essentially equivalent. We derive Eq. 12 by using the
equivalence of Markovianity and conditional independence. The general solution
given by Li (2007b) was using the Bayes’ theorem (or the definition of conditional
probability).
Some special SMC models are provided and can be found in Li (2007a). For
example, the conditional probabilities of two- and three-dimensional Markov chain
models are provided as follows:
h v
puk pkt
Pr Z iþ1, jþ1 ¼ kZi, jþ1 ¼ u, Z iþ1, j ¼ t ¼ X
m ð14Þ
h v
puk pkt
k¼1
Pr Ziþ1, jþ1, rþ1 ¼ kZi, jþ1, rþ1 ¼ u, Z iþ1, j, rþ1 ¼ t, Z iþ1, jþ1, r ¼ q
h v y
puk pkt pkq
¼X
m ð15Þ
h v y
puk pkt pkq
k¼1
These models are simple and useful in modeling categorical variables without the
overestimation or underestimation problem.
4 Case Study
The data we used for our research are gathered from Tahe area of the Tarim Basin in
Xinjiang Uyghur Autonomous Region, China. There are three major lithofacies in
this work area: mudstone, sandstone, and conglomerate. The conglomerate is
relatively low in content. We have got four wells’ lithologic data in the three-
dimensional space (Fig. 1). Three wells are located in the corners of this work area;
another well is located inside. The distance in east-west direction of the two wells is
6000 m and 8000 m in south-north direction, the simulated space is split into a
60 80 100 grid system, and each cell is a 100 100 1 m cuboid. Note that the
coordinate values represent the number of grids.
596 X. Huang et al.
Fig. 1 The 3-D work area with four wells (S66, S67, S75, S92), x axis and y axis indicate east-
west direction and south-north direction, respectively. z axis indicates vertical direction. Note that
the coordinate values represent the number of grid
According to the four wells and step size we choose, the unknown grids in this work
area are being simulated by using MCRF algorithm, which has been discussed in
Sect. 3. The four realizations of lithologic stochastic simulation are shown in Fig. 2.
These realizations generally reflect the basic rule of sand body development in this
area. The sandstone thickness is mostly in 1–5 m and presents a thin layer of output.
It is continuous in horizontal direction, which can extend to several kilometers.
Mudstone is the background lithofacies, and it is widely distributed in this area.
Conglomerate is not well developed in this work area, which is not continuous
neither in vertical direction nor horizontal direction. For one thing, we can safely
draw a conclusion from Fig. 2a–d that the conglomerate can only extend 200–300 m
in horizontal direction; the average thickness of this type of lithofacies is no more
than 3 m. For another, the conglomerate is more likely to appear in sandstone
distribution area, which means that the transition probability between conglomerate
and sandstone is higher than that of mudstone. Mudstone and sandstone are more
continuous than conglomerate in horizontal direction, and the occurrence frequency
of mudstone is significantly greater than the other lithofacies in vertical direction.
We also use the method based on fully independent assumption for comparison.
The simulation results has been shown in Fig. 3. Compared with Fig. 2, the
sandstone layer is thinner in vertical direction and less continuous in horizontal
direction. We also find that conglomerate is rarer distributed, which means that this
fully independent assumption may underestimate the small class proportion. This
conjecture is verified in Table 1. In our analysis, the well data (Table 1) are used for
validation. It is obvious that, compared with model based on fully independent
assumption, MCRF, a method based on conditional independence assumption, is
relatively better at maintaining the percentage composition of each lithofacies.
Theoretical Generalization of Markov Chain Random Field in Reservoir. . . 597
Fig. 2 (a) the first simulation result; (b) the second simulation result; (c) the third simulation
result; (d) the fourth simulation result
Fig. 3 (a) the first simulation result; (b) the second simulation result; (c) the third simulation
result; (d) the fourth simulation result
598 X. Huang et al.
Table 1 Lithofacies proportions in well data and averaged from four simulated realizations
Simulation method
Lithofacies Well data MCRF Fully independent assumption
Mudstone 78.70 % 79.81 % 83.14 %
Sandstone 15.75 % 16.02 % 14.83 %
Conglomerate 5.55 % 4.17 % 2.03 %
5 Conclusions
Acknowledgments This study is funded by the Fundamental Research Funds for the Central
Universities of Central South University (No. 2016zzts011) and the National Science and Tech-
nology Major Project of China (No. 2011ZX05002-005-006). We thank Dr. Dongdong Chen for
the helpful discussion regarding the three-dimensional stochastic simulation.
References
Carle SF, Fogg GE (1996) Transition probability-based indicator geostatistics. Math Geol 28
(4):453–476
Carle SF, Fogg GE (1997) Modeling spatial variability with one and multi-dimensional continu-
ous-lag Markov chains. Math Geol 29(7):891–917
Elfeki A, Dekking MA (2001) A Markov chain model for subsurface characterization: theory and
applications. Math Geol 33(5):568–589
Huang T, Lu D, Li X, Wang L (2013) GPU-based SNESIM implementation for multiple-point
statistical simulation. Comput Geosci 54(4):75–87
Huang X, Li J, Liang Y, Wang Z, Guo J, Jiao P (2016a) Spatial hidden Markov chain models for
estimation of petroleum reservoir categorical variables. J Petrol Explor Prod Technol. doi:10.
1007/s13202-016-0251-9
Huang X, Wang Z, Guo J (2016b) Prediction of categorical spatial data via Bayesian updating. Int
J Geogr Inf Sci 30(7):1426–1449
Huang X, Wang Z, Guo J (2016c) Theoretical generalization of Markov chain random field from
potential function perspective. J Cent South Univ 23(1):189–200
Journel AG (2002) Combining knowledge from diverse sources: an alternative to traditional data
independence hypotheses. Math Geol 34(34):573–596
Theoretical Generalization of Markov Chain Random Field in Reservoir. . . 599
Krumbein WC, Dacey MF (1969) Markov chains and embedded Markov chains in geology. Math
Geol 1(1):79–96
Li W (2007a) A fixed-path Markov chain algorithm for conditional simulation of discrete spatial
variables. Math Geol 39(2):159–176
Li W (2007b) Markov chain random fields for estimation of categorical variables. Math Geol 39
(3):321–335
Li W, Zhang C (2013) Some further clarification on Markov chain random fields and transiograms.
Int J Geogr Inf Sci 27(3):423–430
Li J, Yang X, Zhang X, Xiong L (2012) Lithologic stochastic simulation based on the three-
dimensional Markov chain model. Acta Pet Sin 33(5):846–853 (in Chinese)
Lin C, Harbaugh JW (1984) Graphic display of two and three dimensional Markov computer
models in geology. Wiley, New York
Liu W (2008) Geological modeling technique for reservoir constrained by seismic data. Acta Pet
Sin 29(1):64–68 (in Chinese)
Weissmann GS, Fogg GE (1999) Multi-scale alluvial fan heterogeneity modeled with transition
probability geostatistics in a sequence stratigraphic framework. J Hydrol 226(1):48–65
Zhang R, Zhang S, Chen Y, Chen B, Hou Y, Huang J (2010) Stochastic simulations of clastic
reservoirs in East China. Acta Pet Sin 31(5):787–790 (in Chinese)
Deepwater Reservoir Connectivity
Reproduction from MPS and Process-
Mimicking Geostatistical Methods
R. Kaplan (*)
Reservoir Management Unit, Chevron Energy Technology Company, 1500 Lousiana St,
Houston, USA
e-mail: [email protected]
M.J. Pyrcz • S. Strebelle
Strategic Research Unit, Chevron Energy Technology Company, 1500 Louisiana St, Houston,
USA
e-mail: [email protected]; [email protected]
1 Introduction
and other process-mimicking approaches, the reader can refer to various publica-
tions including Pyrcz and Deutsch (2005), Wen (2005), Miller et al. (2008),
Michael et al. (2010), Sylvester et al. (2010), and Pyrcz et al. (2015).
This technique allows for inclusion of realistic lobe stacking patterns and
internal lobe heterogeneity, which is not typically captured by conventional model-
ing approaches. Stacking is a fundamental characteristic of lobes, as well as the
internal architecture (proximal, dominated by amalgamated sands, to distal, dom-
inated by non-amalgamated sands).
This study endeavors to quantify the added value of utilizing a process-
mimicking approach compared to conventional modeling approaches. A simple
multiple-point statistics (MPS) model is utilized as a benchmark, because it repre-
sents the common modeling approach in this setting. To achieve our objective we
propose to generate typical deepwater lobe reservoir models using archetypal well
log data and perform flow diagnostics to estimate the impact recovery. Previous
studies have shown the importance of reservoir connectivity and heterogeneity on
reservoir performance (e.g., Larue and Hovadik 2006). Our results quantify the
incremental impact of connectivity (lobe stacking) and heterogeneity (within-lobe
trends) for different net-to-gross scenarios modeling with process-mimicking
models benchmarked with standard MPS workflow. This is useful to justify the
additional effort to adopt emerging process-mimicking methods rather than utiliz-
ing widely available MPS workflows.
604 R. Kaplan et al.
2 Methodology
Fig. 2 Dip section of a simple conceptual model of within-lobe heterogeneity for a simple lobe
geometry
the computed flux does not account for the full physics of the flow between two
wells as the calculation of flux within the model makes several assumptions. Firstly,
the flow is assumed to be steady state and single phase; therefore, the phase of the
flow will not change due to buoyancy, well controls, or fluid mobility changes.
Secondly, a main assumption is that the flow is incompressible and well-driven
flow; any independent reservoir compartment or fault block in the model must have
at least one producer and one injector well to initiate some flow.
This approach yields a number of outputs including dynamic Lorenz coefficient
(DLC) and time of flight information. Once the flux across the volume and time of
flight are calculated, the F-Φ relationships can be calculated. When plotted this
generates a curve representing the relationship between normalized, cumulative
flow capacity (F) against storage capacity (Φ). Specifically, the curve explains the
ratio between the volume of injected and swept reservoir pore volume. Twice the
area under this curve and above the 1:1 line is the DLC and can be used to rank the
heterogeneity of multiple models (see Fig. 7). This is a good quantification of flow
608 R. Kaplan et al.
3 Results
Fig. 8 Time-of-flight represents the time it takes (in pore volumes injected) for injected fluid to
travel from an injector to a given point in the reservoir
Deepwater Reservoir Connectivity Reproduction from MPS and Process-Mimicking. . . 609
Fig. 9 Graph showing the DLC at 40 %, 60 %, and 80 % net-to-gross. It is apparent that as we add
compensational stacking and within-lobe trends, we see an increase in the DLC
Fig. 10 Graphs showing the proportion of DLC increase due to lobe stacking pattern and within-
lobe internal architecture
that connectivity has more of an effect on the DLC; hence more emphasis should be
placed on modeling shale placement (e.g., drapes, baffles, etc.) within the reservoir.
As the net-to-gross increases, the modeling of internal trends (lobe architecture)
becomes more critical (see Fig. 10).
610 R. Kaplan et al.
4 Discussion
This work has demonstrated that the additional reservoir model complexity,
namely, lobe stacking patterns and detailed within-lobe trends, impacts reservoir
connectivity heterogeneity. This motivates adoption of new process-mimicking
technology that integrates these geologic features, depending on modeling objec-
tive. For example, this approach will be more appropriate for evaluating the
detailed flow behavior of a deepwater lobe reservoir in the presence of mud drapes
and within-lobe trends.
Conversely, this work suggests that current widely available reservoir modeling
approaches (such as MPS) may underestimate reservoir fluid flow complexity,
without the inclusion of addition constraints to directly capture these heterogene-
ities. In our experience, this may translate into overestimation of recovery factor
and time to water breakthrough. Generating more realistic models may assist in
managing and mitigating risk that may arise with conventional modeling methods.
Future work includes performing a full flow simulation which provides us with
measures of heterogeneity other than the DLC.
5 Conclusions
Deepwater lobes are important reservoir targets. New insights from geological
characterization suggest a hierarchy of heterogeneities including compensationally
stacked lobes and within-lobe reservoir property trends. The incremental impact of
these heterogeneities is determined through three workflows: (1) a traditional MPS
workflow, (2) a process-mimicking workflow that captures lobe stacking patterns,
and (3) a process-mimicking workflow that captures lobe stacking patterns and
within-lobe trends.
Each of these additional complexities has a significant impact on flow hetero-
geneity as represented by DLC.
Bibliography
Larue DK, Hovadik J (2006) Connectivity of channelized reservoirs: a modeling approach. Pet
Geosci 12(4):291–308
Michael HA, Li H, Boucher A, Sun T, Caers J, Gorelick SM (2010) Combining geologic-process
models and geostatistics for conditional simulation of 3-D subsurface heterogeneity. Water
Resour Res 46(5):1–20
Miller J, Sun T, Li H, Stewart J, Genty C, Li D, Lyttle C (2008) Direct modeling of reservoirs
through forward process-based models: can we get there. International Petroleum Technology
Conference proceedings. pp 259–270
Møyner O, Krogstad S, Lie K-A (2015) The application of flow diagnostics for reservoir man-
agement. Soc Pet Eng J 20(2):306–323
Prélat A, Hodgson DM, Flint SS (2009) Evolution, architecture and hierarchy of distributary deep-
water deposits: a high-resolution outcrop investigation from the Permian Karoo Basin,
South Africa. Sedimentology 56(7):2132–2154
Prélat A, Covault JA, Hodgson DM, Fildani A, Flint SS (2010) Intrinsic controls on the range of
volumes, morphologies, and dimensions of submarine lobes. Sediment Geol 232(1–2):66–76
Pyrcz MJ, Deutsch CV (2005) Conditional event-based simulation. Quant Geol Geostat 135–144
Pyrcz MJ, Deutsch CV (2014) Geostatistical reservoir modeling, 2nd edn. Oxford University
Press, New York, p 448
Pyrcz MJ, Leuangthong O, Deutsch CV (2005) Hierarchical trend modeling for improved reser-
voir characterization. International Association of Mathematical Geology Annual Conference,
Toronto
Pyrcz MJ, McHargue T, Clark J, Sullivan M, Strebelle S (2012) Event-based geostatistical
modeling: description and applications. In: Abrahamsen P et al (eds) Geostatistics Oslo 2012
Pyrcz MJ, Sech R, Covault JA, Willis BJ, Sylvester Z, Sun T (2015) Stratigraphic rule-based
reservoir modeling. Bull Can Pet Geol
Shahvali M, Mallison B, Wei K, Gross H (2012) An alternative to streamlines for flow diagnostics
on structured and unstructured grids. SPE J 17(3):768–778
Shook GM, Mitchell KM (2009) A robust measure of heterogeneity for ranking earth models. The
F PHI Curve and Dynamic Lorenz Coefficient: Society of Petroleum Engineers
Strebelle S (2002) Conditional simulation of complex geological structures using multiple-point
statistics. Math Geol 34(1):1–21
Sullivan MD, Foreman JL, Jennette DC, Stern D, Jensen GN, Goulding FJ (2004) An integrated
approach to characterization and modeling of deep-water reservoirs, Diana Field, Western Gulf
of Mexico. AAPG Mem 80:215–234
Sylvester Z, Pirmez C, Cantelli A (2010) A model of submarine channel-levee evolution based on
channel trajectories: implications for stratigraphic architecture. Mar Pet Geol 28(3):716–727
Wen R (2005) SBED studio: an integrated workflow solution for multi-scale geo modelling.
European Association of Geoscientists and Engineers conference proceedings
Modeling of Depositional Environments:
Shoreline Trajectory – The Link Between
Sequence Stratigraphy and Truncated
Gaussian Fields
Abstract The key to understanding the reservoir and fluid properties of any
hydrocarbon system, clastic or carbonate, is to understand the depositional envi-
ronment. The distribution of sediments, reservoir quality, as well as source and seal
is controlled by well-understood geological processes that can be interpreted and
described. By understanding these processes, it is possible to make qualified pre-
dictions of the consequential distribution of sediments by proactive geostatistical
modeling. This paper directly links the main components of the depositional
systems, in a sequence stratigraphic context, to the controlling parameters of the
Truncated Gaussian Fields algorithm and uses this to predict sediment distribution
in time and space. The approach is demonstrated at several scales, all the way from
seismic-scale basin models to reservoir models, ensuring a systematic geological
modeling approach to the entire value chain, from exploration, through appraisal,
field development, and into production.
1 Introduction
Up until the late 1970s, the understanding of sediment distribution was confined to
recognition of lithologically restricted packages of sediments (litho-stratigraphy),
where interpretation was based on correlating the same lithology between data
points. Sequence stratigraphy (Mitchum et al. 1977) developed this by stating that
logical predictions of sediment distribution could be made by correlating uncon-
formities. This fundamentally changed the understanding of reservoir architecture.
Being highly qualitative it did however not easily lend it selves to quantitative
predictions of sediment distribution.
This happened in parallel with the introduction of geostatistical methods to
populate reservoir models (Matheron et al. 1987). The main objective of
a geostatistical reservoir model was to reproduce a static image of the reservoir
architecture based on conceptual geological models and statistical measures. There
was little focus on replicating the result of the depositional processes. As a result,
the geostatistical models were able to introduce heterogeneity into reservoir
models, but not to predict the results of the sedimentary process. In other words,
if the conceptual model was wrong, the geostatistical model would not reflect this.
Furthermore, the geostatistical model could not predict changes in the stratigraphy
over larger areas.
This paper aims to demonstrate how Truncated Gaussian Fields can be used to
model the results of sedimentary processes. It will be demonstrated how this
methodology directly can be used to logically predict sediment distribution away
from points of “hard” information such as an outcrop or well at any scale (Siggerud
et al. 2015).
Fig. 1 The conceptual model elucidating the relationship between changes in shoreline trajectory
and the extension of the facies belts and subsequent reservoir architecture. Note the use of a clock
to denote the angle; a common mistake is in trying to depict the exact angle while in fact it is the
relative change (from, e.g., 10 to 12 o’clock) that is important, modified after (Helland-Hansen and
Gjelberg 1994)
Fig. 2 Eocene outcrop example in NE Spain (above top) where the dipping foresets of the
low-angle shallow shelf clinoforms can be observed as shown schematically in the figure (yellow
color); note also the stacking light-colored sands and red beds in the skyline reflecting the aggrada-
tion (high-angle trajectory) of the transgressive systems tract (shown in dark red and green)
While most common use of geostatistical geomodeling has been limited to fields
and reservoir scale (few kilometers), this paper demonstrates the extended use of
the ideas outlined above. While the ideas of shoreline trajectory were originally
related to shallow marine clinoform developments (sequence scale, which is hun-
dreds of meters; Burgess et al. 2008), it is equally applicable on megasequence or
basinal scale (tens to hundreds of kilometers: Siggerud 2012).
Drawing on extensive knowledge from several decades of fieldwork in the
Arctic islands of Svalbard and work in the Norwegian Barents Sea, a megasequence
subdivision of the post-Permian succession was established, using the facies
stacking pattern to establish an understanding of the depositional trajectory for
each megasequence. In addition, an interpretation of lithology, sedimentary facies,
and facies associations was undertaken for 255 outcrops and wells, serving as a
basis for the sedimentological sequence stratigraphic model. A megasequence is
here defined as a seismic-scale sequence, bounded by regional peak transgressions,
ranging from a few hundreds to more than 1,000 m in thickness and spanning
several million years of time (Steel 1993). In this case the near 6000 m thick
Modeling of Depositional Environments: Shoreline Trajectory – The Link. . . 617
Fig. 3 Outcrop example from Svalbard demonstrating changes in shoreline trajectory from
intermediate to very low- and high-angle trajectory (above top). The lower image illustrates the
variability around the shoreline captured by variograms
Fig. 4 The concept of shoreline trajectory as shown in Fig. 1 here visualized for a low angle (a),
high angle (b), low to high angle (c), and a “negative” angle (forced regression) following a
relative sea level fall (d). In the latter case note how the system begins to detach landward, as can
be observed for the quaternary deposits along the Norwegian coastline
618 L.E.R. Kjellesvik et al.
Fig. 5 Bird’s-eye view of the 650,000 km2 basin-wide geomodel based on the 12 megasequences.
Inserted is the sequence-scale Hoop Fault Complex model, maintaining the overall trend while
elucidating the details on semi-regional scale, while even more detailed prospect-scale model
sedimentary succession was split into 12 megasequence that was identified and
correlated across the entire Barents Sea, covering some 650,000 km2, probably
making it the largest 3D geomodel ever built (Siggerud et al. 2015) (Fig. 5).
Un-eroded thickness maps were created using the stratigraphy to establish the
sedimentary sequences at the time of deposition in order to clarify the logical
orientation of the sedimentary system for each megasequence (the trend in the
Truncated Gaussian Field) (Fig. 6). This, combined with the shoreline trajectory for
each megasequence, enabled proactive modeling of the sedimentary system
depicting extension and limitations of all lithologies across the entire Barents Sea
(including Svalbard shelf) Fig. 7. Secondly, while not only illustrating sedimentary
facies distribution and volumes of deposited sediments, it also enabled proactive
modeling of source rock presence and improved migration probabilities using the
sand probability distribution derived from the geomodel (Throndsen et al. 2016). As
such the sedimentary distribution within the Norwegian Barents Sea is no longer
random but can be logically tested regardless of sedimentological understanding
and model.
A third aspect of this methodology is in the scalability of the approach. It is
possible to zoom in on an area of particular interest within the original
megasequence model, while carrying the regional information. This is illustrated
with the prolific Hoop Fault Complex within the central Barents Sea, cf. Fig. 5. A
detailed geomodel for this area was constructed while carrying all the main trends
and architecture from the main megasequence model. In the past this has proven to
be a challenge, because of the lack of a logical regional distribution of the
sedimentary system. Most other paleoreconstructions depict a somewhat arbitrary
range of shoreline locations for the Middle Triassic within the Hoop area (Riis et al.
Modeling of Depositional Environments: Shoreline Trajectory – The Link. . . 619
Fig. 6 Example on the construction of un-eroded and current day (eroded) thickness maps for
megasequence two (Middle Triassic). The dashed green line on the left picture depicts the position
of the basin floor, while the orange dotted line elucidates the shoreline transition, thereby
illustrating the wide shallow shelf in the north as evident from seismic data
Fig. 7 Example of a vertical section from the Barents Sea Geomodel. Colors reflect the main
depositional environment as shown in the legend, and the shoreline trajectory is illustrated
schematically. Note the variability in the facies belts and that there will be higher-order variations
within the megasequences which reflect variations on sequence scale
620 L.E.R. Kjellesvik et al.
Fig. 8 Detailed time section of one layer of the of the Hoop Fault Complex geomodel depicting
the main facies associations (depositional environments) where the shallow marine (best reservoir
potential) is shown highlighted in yellow
2008; Matheron et al. 1987) (Beucher, Galli, Le Loc’h, and Ravenne). Our model,
using the understanding of the shoreline trajectory while honoring the well data and
sedimentary understanding gleaned from the seismic data, gives the logical distri-
bution of the Middle Triassic depositional system as elucidated in Fig. 8.
A fourth point is that alternative scenarios from different sedimentological
models can be tested by coherently altering the angle of the trajectory. This gives
a true handle on the uncertainty of the interpretation and provides highly valuable
feedback to the geologist as to the physical probability of the proposed depositional
model. What this means is that while in a correlation panel any “irregularity” in the
sedimentological interpretation can be “accounted” for, the proactive modeling will
provide “bulls-eyes” where the sedimentary system proposed is not physically (and
mathematically) possible.
A fifth point to be remembered is that the proactive use of truncated Gaussian
field modeling is not dependent on an enormous database of well points and/or
seismic. Remembering the relationship as outlined above between the truncated
Gaussian field and the sequence stratigraphy and shoreline in particular, any
depositional system can readily be proactively modeled, whereby (again) the
logical and quantitative consequences can be elucidated. An example is work
undertaken by the authors on basin scale across the Arctic islands in Svalbard.
Here the lack of extensive outcrop data and basic structural mapping in the past had
not yielded a satisfactory understanding of the reservoir distribution. The shoreline
trajectory enabled the construction of different detailed sequence stratigraphical
models, as shown in reduced uncertainty and enabled systematic testing of alter-
native hypothesis as seen in Fig. 9.
Modeling of Depositional Environments: Shoreline Trajectory – The Link. . . 621
Fig. 9 Two depositional models as they appear in the model (a) thickness fence diagram (b)
facies distribution: gray ¼ offshore, yellow ¼ shallow marine, and light green ¼ nonmarine (c)
show only the shallow marine deposits in the left and an alternative hypothesis to the right
As mentioned in the introduction, the aim of this paper has been to elucidate how
Truncated Gaussian Fields are suited to proactively model the consequences of our
geological understanding of any depositional environment or the sediment distri-
bution in time and space. The direct link between the parametrization of the
truncated Gaussian field and the parameters controlling and steering the sedimen-
tary succession is demonstrated with the principle of the shoreline trajectory.
Consequently, this allows an unbiased prediction of the distribution of sediments
and uncertainty, constrained by the available data.
Armed with this understanding, one can logically transfer from basin to field
scale maintaining the overall sedimentary system while capturing the details on
reservoir scale. While being systematic and highly predictive, it also is quantitative,
622 L.E.R. Kjellesvik et al.
Fig. 10 Bird’s-eye overview of two time slices depicting sediment distribution for the prospect
area within the Hoop semi-regional geomodel (see Fig. 5). The images show the occurrence of
fluvial channels correctly honoring the width, orientation, interconnectedness, and frequency of
the channels in addition to the shoreline location as evident from the regional model
Bibliography
Patruno S, Hapson GJ, Jackson CA-L (2015) Quantitative characterization of deltaic and sub-
aqueous clinoforms. Earth-Sci Rev 142:79–119
Riis F, Lundschien BA, Høy T, Mørk A, Mørk MB (2008) Evolution of the Triassic shelf in the
northern Barents Sea region. Polar Res 27:318–338
Siggerud EI (2008) Application of shoreline trajectory in geomodelling, cretaceous helvetiafjellet
formation, Arctic Norway, AAPG Extended Abstract, AAPG International Conference and
Exebition. AAPG, Cape Town
Siggerud EI (2012) In: Siggerud EIH, SGS Petroleum Geological Excursions (eds) Helvetiafjellet
formation – a discussion, vol 2. SGS Publications, Trondheim
Siggerud EI, Steel RJ (1999) Architecture and trace fossil characteristics of a 10.000–20.000 year,
fluvial-marine sequence, SE Ebro Basin, Spain. J Sediment Res 69:365–387
Siggerud EI, Ballardini L, Kjellesvik LE (2015) A quantitative tool for predicting sediment
distribution, exemplified from the post-permian succession Svalbard & Barents Sea. Recent
Advances in Exploration Technology, NCS-Conference and prospect-fair. Fornebu
Steel RJ (1993) Triassic-Jurassic megasequence stratigraphy in the Northern North Sea: rift to
post-rift evolution. In: Parker J (ed) Petroleum geology of NW Europe; Proceedings of the 4th
Conference. Geological Society, London, pp 299–335
Throndsen T, Kjellesvik LE, Siggerud EI (2016) 4D facies modelling in petroleum system analysis
of the Barents Sea TFO area. In: Carstens H (ed) Hydrocarbon habitats. GeoNova, Oslo, pp 6–7
Wenlong X, Journel AG (1993) GTSIM: Gaussian truncated simulations of reservoir units in a
West Texas carbonate field. SPE
Facies Inversion with Plurigaussian
Lithotype Rules
L. Li (*)
Department of Energy Resources Engineering, Stanford University, Stanford, CA 94305, USA
e-mail: [email protected]
S. Xu
Shell International Exploration and Production Inc, Houston, TX 77082, USA
P. Gelderblom
Shell Global Solutions International BV, Rijswijk, The Netherlands
1 Introduction
Seismic inversion is the process by which acoustic impedance and other elastic
properties are derived from the seismic traces. The results of seismic inversion
should honor geological knowledge as well as observed data, such as well logs. One
of the aims of seismic inversion is to characterize the spatial variability of the
reservoir, by extracting subsurface information from the seismic data and other data
sources. Promise (Leguijt 2001, 2009) is the Shell proprietary model-based seismic
inversion tool. It can perform trace-by-trace inversion and has been designed to
naturally honor the prior models. Moreover, geological constraints have been
deployed aiming to improve the geological realism of inversion results
(Gelderblom and Leguijt 2010). In the trace-by-trace inversion, the algorithm
generates an ensemble of local model realizations at each trace location, taking
the seismic data and the well data into account. If a geologically constrained
inversion is performed, an ensemble of reservoir-size model realizations will be
generated taking also the lateral continuity of properties into account. Applications
of the algorithm have been performed on continuous variables, such as NTG,
porosity, and layer thickness. However, a reservoir is normally characterized by
lithological facies. A lithological facies prediction algorithm based on small exten-
sions to the current algorithm is presented in this paper.
In our methodology, we invert for a pair of continuous variables and apply a
classification technique inspired by Plurigaussian simulations (Armstrong et al.
2011) to obtain the inverted facies. Seismic inversion with the application of
geostatistics has been studied and published before (Doyen 2007; Larsen et al.
2006; Ulvmoen and Omre 2010; Grana and Della Rossa 2010; Rimstad and Omre
2010; Gelderblom and Leguijt 2010; Gunning et al. 2014). Our work is different in
the sense that this method is a combination of model-based inversion in depth
domain, vertical and lateral continuity constraints, and facies modeling, which did
not appear simultaneously in any previous publication. In comparison to the
popular multistep inversion methods (Dubois et al. 2007), the facies classification
in stochastic inversion methods is performed inside the inversion process, as
opposed to as a post-processing procedure. In multistep methods, acoustic imped-
ance values are first determined before facies probabilities and properties are
known. Next, the facies are chosen on the basis of the acoustic impedance values.
Since seismic inversion is generally performed at the acoustic scale, which corre-
sponds to the vertical resolution of the seismic dataset, a downscaling process is
needed to generate a fine-scale static model for dynamic reservoir simulation
(Hesthammer et al. 2001; Doyen et al. 1997). In our algorithm, seismic inversion
can directly produce fine-scale model realizations, in which the subseismic scale
properties are constrained by geological and lithological knowledge. The specifics
of our inversion algorithm will be described in the next section, followed by a
description of a methodology to derive necessary input parameters. Finally, the
workflow will be demonstrated on two cases, a synthetic wedge model and a field
case from a deepwater offshore reservoir.
Facies Inversion with Plurigaussian Lithotype Rules 627
2 Plurigaussian Inversion
Fig. 1 Overall workflow for fine-scale facies inversion. The inversion process solves for a pair of
continuous guide variables that will be classified into facies. A prior geological and reservoir
property model is the required input for the process and is used for constructing the initial guess for
the guide variables and their respective spatial continuities
Fig. 2 Illustration of Plurigaussian simulation. Two coregionalized guide variables are jointly
thresholded to obtain a facies classification. This can be viewed as a lithofacies rule map, where
each axis represents the CDF of a guide variable. For a location x in the geological model, this
lithofacies rule map (LRM) can be used to “look up” the facies type for that location: if g1 and g2
are the realizations of the “guide” variable at x, then the color of the point (Φ(G1), Φ(G2)) in the
LRM defines the facies (Φ(.) is the standard normal CDF)
We will assume that prior facies probability maps and rock physics models are
provided, and a preliminary static geological model is available to derive other
inputs. The following subsections will discuss how we can derive the lithotype rule
and variograms.
Facies Inversion with Plurigaussian Lithotype Rules 629
Fig. 3 Illustration of how a geological model can be used to construct a lithofacies rule. The
global proportions of each facies are used to determine the area of the rectangles in the rule map,
while the frequency of contacts is used for placement of the areas
1
γ Fi ðx, x þ hÞ ¼ Var ½Fi ðxÞ Fi ðx þ hÞ
2
1n h i o ð1Þ
¼ E ðFi ðxÞ Fi ðx þ hÞÞ2 ðE½Fi ðxÞ Fi ðx þ hÞÞ2
2
630 L. Li et al.
where the location and offset vectors are x, h 2 ℝ3 and Eð ∙ Þ, Var ð ∙ Þ, Fi ð ∙ Þ are the
expectation operator, the variance operator, and the indicator operator.
Assuming second-order stationarity, the second term reduces to 0, and the
expression simplifies to Eq. 2:
1 n 2 o
γ Fi ðx, x þ hÞ ¼ E F i ð x Þ F i ð x þ hÞ ð2Þ
2
1 X
γ ⋆ Fi ¼ ½Fi ðxÞ Fi ðx0 Þ2 ð3Þ
2N jxx0 j¼h
The variogram model parameters (type, nugget, range) of the guide variables must
be chosen such that when realizations of the guide variables are converted to facies
realizations as described above, those facies realizations have indicator variograms
consistent with the observed empirical indicator variograms (Eq. 3).
The indicator variogram is defined as
1n
γ Fi ðx, x þ hÞ ¼ P½Fi ðxÞ ¼ 1 þ P½Fi ðx þ hÞ ¼ 1
2
T o ð4Þ
2P½ðFi ðxÞ ¼ 1Þ ðFi ðx þ hÞ ¼ 1Þ
In comparison to the LRM, where Φ(G1), Φ(G2) 2 [0,1], the guide variables G1 ðxÞ,
G2 ðxÞ 2 ½1, 1 in this section. This means Eq. 4 can be reexpressed in terms of
facies regions as
1n
γ Fi ðx, x þ hÞ ¼ P½G1 ðxÞ, G2 ðxÞ 2 Ri þ P½G1 ðx þ hÞ, G2 ðx þ hÞ 2 Ri
2 o ð6Þ
2P½ðG1 ðxÞ, G2 ðxÞ 2 Ri Þ \ ðG1 ðx þ hÞ, G2 ðx þ hÞ 2 Ri Þ
1
fP½G1 ðxÞ, G2 ðxÞ 2 Ri þ P½G1 ðx þ hÞ, G2 ðx þ hÞ 2 Ri g ¼ Ai ð7Þ
2
Ai is the area of region Ri or equivalently the global proportion of facies i. Under the
assumption that the guide variables are Gaussian, the final term of Eq. 6 can be
evaluated by taking the integral of a quadvariate Gaussian distribution centered at μ
Facies Inversion with Plurigaussian Lithotype Rules 631
The guide variables are assumed to have μ ¼ 0. Assuming that the guide variables
are independent and have covariance functions ρ1(h) and ρ2(h), respectively, the
covariance matrix Σ can be expressed as
2 3
1 0 ρ1 ð hÞ 0
6 0 1 0 ρ2 ð hÞ 7
Σ¼4 5 ð9Þ
ρ1 ð hÞ 0 1 0
0 ρ2 ð hÞ 0 1
For demonstrating this process, a study was conducted on a synthetic wedge model
with maximum thickness of 90 m and three facies (clean sand, shale, and sand-shale
mix). The two guide variables were simulated using Sequential Gaussian Simula-
tion and the variogram parameters outlined in Table 1.
The grid thickness was chosen to be 2 m to allow for the generation of thin beds
of sand as seen in Fig. 4. The lithofacies rule used for categorizing the facies is
depicted in Fig. 5. The thresholds were chosen to be 0.4 for thresholds of both guide
variables. A rock model for laminated sand and shale (Allen 1984) was chosen. In
this rock model, the three lithological facies can be defined by net-to-gross ratio
(NTG) of 0 %, 50 %, and 100 %. The thresholding procedure yielded a global
proportion of 40 % shale, 36 % sand, and 24 % sand-shale mix.
Fig. 4 2D profile view of a medial slice of the ground truth. The top two images indicate the guide
variables, while the bottom image is the generated facies map
Fig. 6 2D profile view of seismic scale inversion for NTG in comparison with true NTG (slice 14)
634 L. Li et al.
The fine-scale case assumes that both horizontal and vertical variograms can be
estimated accurately, and the prior facies probabilities are assumed to be equal
throughout the model grid. In this example of three facies, the prior probability for
each facies is 33 %. The inverted result is shown in Fig. 7. The inverted result yields
an overall sand fraction of 41.4 %, which is very close to the ground-truth value of
42.6 %. This illustrates a reduction in the error misfit when compared to the seismic
scale case. The reduction of misfit may be attributed to the finer scale on which the
inversion was performed. Further studies are required to understand the impact of
model scales on the quality of Promise inversion results. Moreover, it should be
noted that the Plurigaussian inversion was able to generate realizations with
multiple contacts between facies (purple is in contact with both red and green
facies).
The algorithm was also tested with a marine survey on a deepwater channel-lobe
system. Internal petrophysical studies indicated that the reservoir consists of three
facies (clean sand, shale, and a mixture of the two), and prior facies probability
maps have been postulated for each facies (Fig. 8).
Lithofacies Rule The lithofacies rule can be generated given the expected facies
proportions and the facies contact proportions. In this study, an initial geological
concept model of three facies has been built for the reservoir. The facies proportions
were obtained from the initial model, while the contacts between facies were
estimated by computing the facies transition probabilities (Table 2) from the initial
model.
The nine facies transition probabilities constitute nine constraints, but for three
rectangular facies, only two thresholds are available for regular rectangular facies
map, resulting in an overdetermined system. Therefore, a least squares approach is
used to fit the thresholds and results in the LRM shown in Fig. 9.
Variogram Fitting The final step before inversion is obtaining estimates for the
variograms. A prior most likely facies map is obtained by identifying the facies with
Facies Inversion with Plurigaussian Lithotype Rules 635
Fig. 7 2D profile view of slice 14 of the ground truth (left) compared to the inverted result (right)
of 3D inversion with correct horizontal and vertical variograms
Fig. 8 Postulated prior facies probability maps for an offshore reservoir in deepwater channel-
lobe system
Fig. 10 Empirical indicator variograms computed for the sand-shale mix on the most probable
facies map shown in green along with the corresponding indicator variograms by modeling two
continuous variograms for each guide variable
the highest probability from the input probability maps. By computing the indicator
variogram on the most likely facies map, and computing the integral in Eq. 8 over
the averaged thresholds, one continuous variogram can be fitted for each guide
variable. In a 3D grid, the variograms must be fitted to six directions in the space for
each facies. Variograms fitted for the sand-shale mix facies are shown in Fig. 10.
The fitted variogram parameters are shown in Table 3. It should be noted that
these parameters are not necessarily unique. Ideally, a single set of parameters
Facies Inversion with Plurigaussian Lithotype Rules 637
Fig. 11 Arithmetic average map of NTG over the P50 of the posterior realizations of the inversion
should be used for each facies; however, in practice this is a difficult task. Conse-
quently, the averaged variogram over all three facies was used.
For comparison, a seismic scale inversion (30 m cells) was performed with the
resulting average NTG map of the P50 realization of the inversion results shown in
Fig. 11a. Using the estimated LRM and variograms, fine-scale inversion (2 m) was
performed, with the resulting maps shown in Fig. 11b. The fine-scale inversion
indicates isolated regions of higher NTG than the coarse-scale result. The overall
trends for both approaches are similar, but the fine facies inversion provides
additional insight on areas of higher NTG which would be useful for drilling
decisions. The top view map and a vertical cross section shown in Fig. 12 demon-
strate results of the P50 realization of the fine-scale facies inversion. Lenticular
shape sand bodies of channels and lobes are identifiable in the vertical cross section.
In the top view, the transition between sand and shale is observed in the top view
map, which demonstrates the effect of the Plurigaussian lithofacies rule.
638 L. Li et al.
Fig. 12 A vertical cross section and a map of the top view of the P50 posterior realization
6 Conclusions
Bibliography
Allen D (1984) Laminated sand analysis. SPWLA 25th Annual Logging Symposium. New
Orleans: Society of Petrophysicists and Well-Log Analysts, 20
Armstrong M et al (2011) Plurigaussian simulations in geosciences. Springer Science & Business
Media, New York
Coléou T, Manuel P, Azbel K (2003) Unsupervised seismic facies classification: a review and
comparison of techniques and implementation. Lead Edge 22(10):942–953
Doyen P (2007) Seismic reservoir characterization: an earth modelling perspective. EAGE
publications, Houten
Doyen PM, Psaila DE, Den Boer LD, Jans D (1997) Reconciling data at seismic and well log
scales in 3-D earth modelling, SPE Annual Technical Conference and Exhibition. Society of
Petroleum Engineers, San Antonio, p 10
Dubois M, Bohling G, Chakrabarti S (2007) Comparison of four approaches to a rock facies
classification problem. Comput Geosci 33(5):599–617
Gelderblom P, Leguijt J (2010) Geological constraints in model-based seismic inversion, SEG
International Exposition and 84th Annual Meeting. Society of Exploration Geophysicists,
Denver, pp 2825–2829
Gómez-Hernández JJ, Rodrigo-Ilarri J, Cassiraga E (2016) Extended abstract example for
GEOSTATS2016, Instructions for GEOSTATS2016 participants. Universitat Politècnica de
València, Valencia, pp 1–4
Grana D, Della Rossa E (2010) Probabilistic petrophysical-properties estimation integrating
statistical rock physics with seismic inversion. Geophysics 75(3):21–37
Gunning JS, Kemper M, Pelham A (2014) Obstacles, challenges and strategies for facies estima-
tion in AVO seismic inversion. 76th EAGE Conference and Exhibition 2014
Hesthammer J, Landrø M, Fossen H (2001) Use and abuse of seismic data in reservoir character-
isation. Mar Pet Geol 18(5):635–655
Larsen AL, Ulvmoen M, Omre H, Buland A (2006) Bayesian lithology/fluid prediction and
simulation on the basis of a Markov-chain prior model. Geophysics 71(5):69–78
Leguijt J (2001) A promising approach to subsurface information integration, 63rd Conference
Extended Abstracts. European Association of Geoscientists & Engineers, Amsterdam, p 4
Leguijt J (2009) Seismically constrained probabilistic reservoir modeling. Lead Edge (Soc Explor
Geophys) 28(12):1478–1484
Rimstad K, Omre H (2010) Impact of rock-physics depth trends and Markov random fields on
hierarchical Bayesian lithology/fluid prediction. Geophysics 75(4):93–108
Ulvmoen M, Omre H (2010) Improved resolution in Bayesian lithology/fluid inversion from
prestack seismic data and well observations: part 1—methodology. Geophysics 75(2):21–35
Combined Use of Object-Based Models,
Multipoint Statistics and Direct Sequential
Simulation for Generation of the
Morphology, Porosity and Permeability of
Turbidite Channel Systems
Abstract This work presents a new methodology for simulating the morphology
and petrophysical properties of hydrocarbon reservoirs in turbidite channel sys-
tems. The simulation of the morphology uses an object-based algorithm that
imposes multipoint statistics of azimuth angle classes and cdfs of width and
thickness as measured from training images of channels. A facies is then assigned
to each block of the reservoir grid according to a conceptual model of facies both
laterally and vertically within the channels. Following this, as each facies has a
specific cdf for both porosity and permeability, simulated images of these
petrophysical properties are generated using direct sequential simulation (DSS)
with local histograms. For illustrative purposes, a case study of a reservoir in the
Lower Congo Basin is presented.
1 Introduction
Turbidite channel systems are one of the most common types of siliciclastic reservoirs
in deep-water settings. Such systems by some estimates account for between 1200 and
1300 oil and gas fields worldwide (Stow and Mayall 2000). Geologically, these
2 Methodology
The methodology comprises five main stages: (i) data preparation, (ii) generation of
the morphology of the complex and the channels through multipoint statistics of
training images, (iii) conversion of the vector morphological model to a grid of
Combined Use of Object-Based Models, Multipoint Statistics and Direct. . . 643
blocks with the assignment of facies to each block, (iv) simulation of the porosity
and the permeability conditional to the facies model, and (v) upscaling of the grid of
blocks in order to be used in a flow simulator.
To apply the proposed methodology, it is first necessary to prepare the data and
information as well as to establish the dimensions of the reservoir and the size of
each reservoir block. Representative training images of the complex pathline and of
the turbidite channel centreline are selected. A set of control points within the
reservoir volume where a channel complex occurrence is confirmed (from seismic
information or well data) are now defined. Each control point has an associated
maximum distance, and the complex pathline should approximate each control
point according to a random distance drawn between zero and the maximum
distance. Two 1D list values (for thickness and width dimensions, respectively) to
assign to the simulated channels are now simulated by DSS and Co-DSS (Soares
2001) imposing a correlation metric between the two variables. A conceptual facies
model is designed that accounts for the lateral and vertical distributions of the
porosity and permeability of each facies.
6. Evaluate the multipoint statistics of the azimuth classes from the training image
using a 1D template with a maximum predefined dimension, and create the
search tree (Marques 2015). The training image is scanned by the template, and
all the azimuth class sequences found in the image are stored in the search tree
along with their conditional probabilities. The structure of the search tree
histogram is quite simple: first, the proportions of each single class are stored,
and then the conditional probability of sequences of two classes is calculated and
stored, followed by the conditional probability of sequences of three classes, and
so on, until the maximum size of the template is reached. Because the probabil-
ities are stored as relative values, the sum of the proportions of all the sequences
composed of the same number of classes is always equal to 1.
7. Simulate the complex pathline through polyline object-based simulation condi-
tional to the multipoint statistics of the azimuth class sequences stored in the search
tree. A starting point is selected within the volume of the reservoir, and several line
segments with the same length are added to the point. The orientation angles of
these segments are computed, first by Monte Carlo simulation of an azimuth class
conditional to the multipoint statistics stored in the search tree and then by adding a
simulated residual value to the central angle of the class previously simulated.
8. Adjust the complex pathline according to the control points previously defined.
Each segment of the polyline is rotated locally until it reaches the drawn
generated distance for each control point.
The simulated pathline of the complex will condition the simulation of the
centrelines of the individual channels, which are also represented as polygonal
lines. The simulation of the individual channels is constrained to the reservoir area
by respecting a defined tolerance distance between the pathline of the channel
complex and the channel centrelines. The tolerance distance at the top of the
reservoir should be greater than that at the bottom of the reservoir to give the
complex the shape of a large channel. The following steps are used to generate the
morphology of the turbidite channels:
9. Using the training image of the individual channel, compute turbidite channel
centrelines at different depths within the reservoir volume, following the same
procedure as that used to generate the complex pathline between steps 1 and
7 above.
10. Assign thickness and width dimensions previously simulated during the data
preparation phase to each vertex of the centrelines. At this stage, each channel
is represented by four polygonal lines: the top of the channel, two lateral
margins and the base of the channel.
Having generated the morphology model, a conceptual facies model of the
section of a channel is applied to the simulated pathlines in order to obtain a global
facies model for the reservoir. In the present study, a model with five different
regions was considered (Fig. 1): Region I, coarse-grained sand; Region II, medium-
grained sand; Region III, fine-grained sand; Region IV, fine-grained sand and silt;
and Region V, clay. Region V corresponds to the outer channel region.
Combined Use of Object-Based Models, Multipoint Statistics and Direct. . . 645
Fig. 1 Proposed conceptual facies model defining five regions within a cross-section of a channel
At this stage, each high-resolution reservoir block has a facies assigned according
to the conceptual model of facies and the simulated pathlines of the channels. The
simulation of porosity and permeability conditional to the facies images is now
performed using an adapted version of DSS (proposed by Soares 2001, with the
algorithm parallelization proposed by Nunes and Almeida 2010) that uses local
histograms of the variables (here, the porosity and permeability conditional to each
facies) instead of a global histogram (Roxo et al. 2016). The stochastic simulation
process allows several equally probable scenarios of both properties to be generated
for the same morphological model.
The upscaling of both petrophysical properties allows the images to be used in a
dynamic flow simulator. The upscaling of porosity is calculated by the arithmetic
mean of the porosity values of the high-resolution blocks; for permeability, the
upscaling follows the classical approach of combining the arithmetic and harmonic
means of the small blocks, generating a tensor of permeability for each upscaled block.
3 Case Study
Fig. 2 (a) Training image of the complex pathline produced by analysing 3D seismic images of
the reservoir. (b) Training image of the channel centreline
Combined Use of Object-Based Models, Multipoint Statistics and Direct. . . 647
for each facies of the conceptual model, with values of mean and standard deviation
as appropriate for the different lithologies. The box plots of porosity and perme-
ability for each facies of the conceptual model are shown in Fig. 3.
For the case study, a template with a maximum dimension of 10 cells and 18 classes
of 20 for the complex pathline and 36 classes of 10 for the channel centrelines was
used. The complex pathline was adjusted for line segments of 100 m, and the channel
Fig. 3 (a) Box plots of porosity for each region of the conceptual facies model. (b) Box plots of
permeability for each region of the conceptual facies model
648 I. Marques et al.
centrelines were adjusted for line segments of 25 m. The tolerance radius between the
complex pathline and the control points was defined as 250 m.
Figure 4 presents a portion of a 3D vector model of a complex comprising
25 channels, in which it is possible to see the centrelines of the channels and the
sections of the channels separated by a distance of 25 m. Figure 5 presents an image
Combined Use of Object-Based Models, Multipoint Statistics and Direct. . . 649
Fig. 6 (a) Map view (Z ¼ 30 m) and cross-section view of the facies distribution of a simulated
channel complex. (b) Map view (Z ¼ 30 m) and cross-section view of the porosity (%) of a
simulated channel complex. (c) Map view (Z ¼ 30 m) and cross-section view
4 Discussion
Figures 4, 5 and 6 show that the algorithm used is very effective when representing
the complex curvilinear geometries that characterize the channels, even though the
training image of the channel is very sinuous. Also, it can be seen that the set of
generated channels follows a pathline (the complex pathline).
The thickness and width values of the channels were generated based on uniform
distribution laws, carrying a high level of uncertainty. However, these values
650 I. Marques et al.
5 Concluding Remarks
Acknowledgements The authors are grateful to Partex Oil and Gas Group for providing infor-
mation and to Midland Valley for providing an academic licence for Move® software. This work
is a contribution to Project UID/GEO/04035/2013 funded by FCT-Fundaç~ao para a Ciência e a
Tecnologia, in Portugal.
Bibliography
Nunes R, Almeida JA (2010) Parallelization of sequential Gaussian, indicator and direct simula-
tion algorithms. Comput Geosci 36(8):1042–1052
Quininha M (2015) Simulaç~ao de estruturas meandriformes por objetos e estatı́sticas multiponto e
avaliaç~ao da porosidade – aplicaç~ao a reservatórios siliciclásticos. Master thesis, FCT
Universidade Nova de Lisboa (in Portuguese)
Quininha M, Almeida JA, Legoinha P (2015) Multipoint statistics of azimuth angle classes: an
application to the simulation of channel structures and the evaluation of porosity in siliciclastic
reservoirs. 17th annual conference of the International Association for Mathematical
Geosciences
Roxo S, Almeida JA, Matias FV, Mata-Lima H, Barbosa S (2016) The use of sensory perception
indicators for improving the characterization and modelling of total petroleum hydrocarbon
(TPH) grade in soils. Environ Monit Assess 188(3):1–19
Soares A (2001) Direct sequential simulation and cosimulation. Math Geol 33(8):911–926
Stow D, Mayall M (2000) Deep-water sedimentary systems: new models for the 21st century. Mar
Pet Geol 17:125–135
Strebelle S (2002) Conditional simulation of complex geological structures using multiple-point
geostatistics. Math Geol 34:1–22
Recent Advancements to Nonparametric
Modeling of Interactions Between Reservoir
Parameters
1 Introduction
Fig. 1 Examples of three different data features where parametric methods often perform poorly,
nonlinear (a), heteroscedasticity, (b) and pinch-out (c)
among the more popular. The parametric methods may run into problems for data
with certain special or distinct features like pinch-outs, nonlinearity, and heterosce-
dasticity (see Fig. 1 above for an illustration). The simple nonparametric methods
also have limitations and may perform poorly if the data either has a high number of
variables or if there are few data points. They typically struggle to adapt pattern that
is easily seen visually; see also (Leuangthong and Deutsch 2003) for additional
discussion.
Therefore, in order to construct a reliable joint model, that can handle data
commonly encountered in reservoir modeling, it is necessary to rely on more
advanced methods.
Here, we will focus on two recent advances in the field of nonparametric density
(and conditional distribution) estimation. The first is a variation of the Cloud
Transform with directional smoothing (CT, all mentions of CT will hereafter
mean with directional smoothing, unless otherwise stated) (Hermansen et al.
2016), and the other is a nonparametric density estimation called Local Gaussian
Density Estimator (LGDE) (Otneim and Tjøstheim 2016a). Both are good and
robust estimators for lower-dimensional problems, with CT having a slight edge
over LGDE; however, the latter is more stable in problems with a high number of
variables where the curse of dimensionality comes into effect. This curse refers to
the problem of the exponential increase in volume when dimensions increase,
together with limited data support (Sammut and Webb 2010). The KS requires
certain amount of observation within a boundary to make a stable and reliable
estimation, but in a high-dimensional volume data tend to become very sparse.
Parametric models do not have the high-dimensional problems, and we will later
see how nonparametric models can use some parametric structures to overcome this
issue.
In Sect. 2 we will illustrate one limitation of the standard kernel methods on a
synthetic data and also outline the basic mechanics underlying the improvements of
CT. The two main methods, CT and LGDE, are properly introduced in Sects. 4 and 5,
Recent Advancements to Nonparametric Modeling of Interactions Between. . . 655
respectively; both will be tested with synthetic and real data. The results are summed
up in Sect. 6.
Fig. 2 Plot (a) shows the scatter plot of two synthetic data, and (b) is the corresponding
conditional cumulative distribution for Cloud Transform, (c) binned estimator, and (d) standard
Kernel Smoother. The colors represent values from 0 (blue) to 1 (red). These 2D cdf plots illustrate
the conditional cdf of permeability (log) given porosity
Sect. 4 for details). Additional illustrations with the traditional KS are shown in
Fig. 5.
For a specific porosity value in Fig. 2, the corresponding vertical slice will give
the one-dimensional cdf. In Fig. 3, we have illustrated an example of the difference
between KS and CT for permeability (log) given porosity value equal to 0.1,
illustrating the potential of including porosity in the model.
To understand the basic mechanics of the CT method, and also exactly how it
improves on the standard methods, some basic knowledge of the traditional meth-
odology is needed. Kernel smoothers refer to a general class of nonparametric
methods, used for function estimation, such as density and regression. In short, such
Recent Advancements to Nonparametric Modeling of Interactions Between. . . 657
0.8
F(x)
0.4
1D permeability
Kernel Smooth
0.0 Cloud Transform
1 2 3 4 5 6
Permeability (log)
Fig. 3 Three 1D cdf of permeability taken from Fig. 2. Red curve is the 1D cdf of permeability
(log) data in Fig. 2a. The blue and green curves are the 1D cdf of permeability (log) given porosity
obtained by setting porosity equal to 0.1 in Fig. 2 for KS (d) and CT (b), respectively. Note that the
level of porosity seen in Fig. 2 alters the estimated cdf
methods use nearby locations (like a moving window) to borrow strength for the
estimation, e.g., density, distribution, or function, at every location; see Eq. 1 and
Fig. 4 below. Given a k-dimensional data set x1, . . ., xk from an unknown distribu-
tion f, the kernel density estimator is defined as
Xn
^f H ðxÞ ¼ 1 K ðx xi Þ; ð1Þ
n i¼1 H
with K H ðzÞ ¼ jHj2 K H1 z . Here, K ðÞ is the kernel function, H is a
1
Figure 4 shows the porosity plotted against permeability (log) with an ellipse
indicating the bandwidth in x and y direction, indicated by h1 and h2, respectively.
This is KH from Eq. 1 for a 2D data set. In this illustration the density is estimated in
the center of the ellipse at point (xi,xj). Only the points inside the ellipse contribute
significantly to the density in this point, and the choice of the kernel function KH
decides how each point contributes.
In addition to choosing the kernel function K ðÞ; an important part of kernel
density estimation is to determine the bandwidths in the matrix:
h1 h12
H¼ ; ð2Þ
h12 h2
where the elements indicate the smoothing along the axes. Different setups of H
will result in different shapes of the ellipse shown in Fig. 4; h1 ¼ h2 and h12 ¼ 0
gives a circle, h1 > h2 and h12 ¼ 0 gives an ellipse without orientation (it follows the
658 H.G. Olsen and G.H. Hermansen
Fig. 4 Porosity vs
permeability (log) with an
illustration of two
bandwidths in a Gaussian
kernel
axis like the example in Fig. 4), and finally h1 > h2 and h12 6¼ 0 gives an ellipse with
an orientation. From Figs. 2 and 4, it seems clear that the last option provides the
best use of the information (direction) observed in the data.
The choice of bandwidths is important, e.g., too small may result in overfitting,
and very large bandwidths make it impossible to detect the finer details. Choosing
the optimal bandwidths is a nontrivial task. Under reasonable restrictive assump-
tions (Gaussian kernel and the underlying data is Gaussian), there exist “optimal”
bandwidth (rates) recommendations, and for the one-dimensional case, this is
h ¼ 1:06σn0:2 , where σ is the standard deviation and n is the number of data
points; see Silverman (1998). Still, most problems usually require some manual
tuning, which becomes hard for high-dimensional problems.
In two- or higher-dimensional problems, it is more difficult to determine and
tune the bandwidths and one must typically rely more on estimation from iterative
methods. A popular choice is cross-validation, where sections of the data will be
left out for testing and the remaining data will be used for estimation. The
bandwidths are then chosen based on the performance of a predetermined criterion,
but choosing a good criterion is a problem on its own and relies on a clear
understanding of the purpose of the model. One advantage of CT (discussed in
Sect. 4) is that it solves some of these difficulties and will often give good results
(as seen in Fig. 2) when used out of the box with almost no tuning required. The
issue with choosing bandwidths is illustrated in Fig. 5, which shows the same
synthetic data set as in Fig. 2, but estimated with two different bandwidths using
the standard Gaussian KS. In Fig. 2, we used the previously mentioned optimal 1D
bandwidth for each dimension, whereas in Fig. 5, this is scaled trying to obtain a
better fit.
Recent Advancements to Nonparametric Modeling of Interactions Between. . . 659
Fig. 5 Densities estimated with KS and two different bandwidths. Both have used optimal
bandwidth based on the standard deviation of the data, but in (a) we have scaled this by a factor
0.25 and in (b) we have scaled this by a factor of 2
We will now present the main components of the CT method. This was developed
to solve some of the challenges of traditional kernel smoothers for density estima-
tion, like bandwidth estimation, and adapting to the main structure of the data. The
CT method presented here is an extension of the one presented in (Kolbjørnsen and
Abrahamsen 2005) and builds on the work described in (Hermansen et al. 2016).
The main idea is to perform the kernel smoothing in a transformed domain (see
Fig. 6b) which makes the method more robust against outliers and better behaved in
areas where there are few observations. The kernel smoother borrows information
of nearby observations, which is problematic in areas with few observations.
Therefore, it performs better in the transformed domain, for most bandwidth
choices, as each point will have more evenly support.
The main focus will be on the implementation of two-dimensional data sources,
but we will also illustrate how the method can be extended to higher dimensions and
show an example of a three-dimensional case.
Let ðyi ; xi Þ, i ¼ 1, . . . , n be independent observations of (Y,X) from an unknown
joint distribution f(y,x). Here, we will construct the nonparametric estimate for the
conditional cdf, F ^ ðyjX ¼ xÞ; based on these observations. In short, the CT method
estimates this conditional cdf by the following four main steps; see (Hermansen
et al. 2016) for more details:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1. Normalize the response variable ^y i ¼ yi E ^ ðY j X ¼ xi Þ = V d ar ðY jX ¼ xi Þ,
where the estimated expectation and variance are obtained by a local linear
regression and a weighted sample variance.
2. Transform x1 to T i ¼ F ^ ðxi Þ where F
^ ðÞ is the estimated cdf.
3. Estimate the joint density and the conditional cumulative distribution in the
transformed domain using the pairs (y˜i, Ti) and a standard Gaussian kernel.
4. Transform data back to the original domain by inverting step 1 and 2.
660 H.G. Olsen and G.H. Hermansen
Fig. 6 The process of CT. Plot (a) is the original data with estimated mean (red solid) and one
standard deviation (blue stippled). The normalized data is shown in (b) with the corresponding
density estimated with a standard Kernel Smoother. Plot (c) is the backtransformed conditional cdf
in the original domain
Figure 6 shows the process of CT. Plot (a) shows the original pinched-out data
with estimated mean (red line) and one standard deviation (blue lines); these are
used to transform the data to the domain seen in plot (b). Here, the density
is estimated with a KS before transforming back to the original domain, shown in
plot (c). An advantage of step 1 and 2 above is that the transformed data will be
somewhat similar across different data sets and makes it easier to set more general
bandwidths.
We have further applied the method on a real data case. The variables available
in this case were P-wave velocity (Vp, ms/ft), density (Rho), gamma ray (GR), and
total organic content (TOC, weight percent). This data set is of a conventional
reservoir and consists of four blocked wells with a total of 315 observations.
Figure 7 shows two 2D examples where we have used GR conditioned on Vp and
GR conditioned on Rho.
The previous examples have dealt with two variables and specifically exempli-
fied by estimating permeability (log) conditioned on porosity. In this section, we
will outline how CT can be extended to higher-dimensional problems, (Y1, X1, . . .,
Xk) with k fixed (Hermansen et al. 2016):
1. (a) Normalize each variable sequentially. First, let x e2 be the normalized
x2 conditioned on x1, and then let x 3 be normalized x3 conditioned on x
e e2
and x1, and so on for all variables xj. The normalization is done by
computing the corresponding conditional means and standard deviations.
(b) Finally, normalize y by computing the conditional mean and sd given
x
ek , . . . , x
e 2 , x1 .
2. Transform all the normalized x variables from step 1 (a) sequentially. First, let
^ 1 ðx1i Þ, where F
T 1i ¼ F ^ 1 ðÞ is the marginal cdf of X1, and then let
^
T 2i ¼ F2j1 ðe ^ ðÞ is the conditional cdf of normalized X2 given
x 2i jx1i Þ, where F 2j1
X1. This is continued for all variables.
Recent Advancements to Nonparametric Modeling of Interactions Between. . . 661
Fig. 7 Estimated conditional cdf with CT on a real data case, overlain by the observations, GR
conditioned on Vp (a) and GR conditioned on Rho (b)
3. The n-dimensional joint density and conditional cdf can be computed in the
transformed domain by a standard kernel density method.
4. Finally, the density can be transformed back by using inverse operations from
step 1 and 2.
There are some general problems with the generalized version of CT. Firstly, the
normalization and transformation steps become quite involved for data that goes
beyond three dimensions, as each transformation and normalization step is condi-
tioned on all the previous normalization and transformations.
Moreover, we are still vulnerable to the curse of dimensionality since there is a
traditional kernel density smoother at the core of the algorithm after the normali-
zation and transformation. However, CT will generally be more robust, compared
to standard KS, since the kernel estimation takes place in a more well-behaved
space, but it will eventually run into the same high-dimensional challenges; see
(Hermansen et al. 2016) for a more detailed discussion.
We will now show CT applied in three-dimensional real data case. In this model
we will condition TOC on Vp and Rho; see Fig. 8 for a plot of the data. There are
15 high TOC values in this data set (orange to red dots). According to Roxar ASA,
who provided the data, these points may be erroneous, so we have chosen to
exclude them. In addition, we have excluded the low Vp value (under 60) as we
suspect that this may also be an error.
The result of 3D cdf estimated with CT is shown in Fig. 9. Plot (a) shows the data
and (b) shows the corresponding 2D cdf of TOC given Rho (Vp values are ignored).
The bottom row ((c) and (d)) shows two slices of the 3D cdf cube. It illustrates the
effect of the third parameter on the estimated density. From this we can see the
distribution of TOC given Rho is different for the two chosen Vp values (71.1 and
89.2). The lower Vp values give a wider cdf (larger variance). The one-dimensional
cdfs corresponding to Rho equal 2.2 is plotted in Fig. 10, where it is easier to see the
differences. From this we can conclude that to include Vp variable will give
additional information to the estimated density. The performance of CT, in relation
to LGDE, is discussed more in Sect. 6.
662 H.G. Olsen and G.H. Hermansen
Fig. 9 Result of the 3D conditional cdf estimated with CT, here TOC is conditioned on Vp and
Rho. Figure (a) shows the 3D data set and (b) is the 2D CT of TOC given Rho (Vp is ignored). The
bottom row shows two slices of the 3D density, for Vp equal to 71.1 in (c) and Vp equal to 89.2 in
(d). Only sections of the 3D cube where there is ample data support are shown, and only closest
20 % data to the chosen Rho value is plotted. The black lines indicate the corresponding 1D cdf in
Fig. 10
Fig. 10 Three 1D cdf
extracted from Fig. 9 of Rho
equal to 2.2 (black lines).
Red line is from the 2D cdf
(Vp ignored), and blue and
green are from the two
slices of the 3D cdf cube, a
low and high Vp value,
respectively
Z
1 Xn
Lðθ; xÞ ¼ K ðXi xÞlogψ ðXi ; θÞ
i¼1 H
K H ðv xÞψ ðv; θÞdv ð3Þ
n
at the locations x from the sample X1, . . ., Xn (Otneim and Tjøstheim 2016a). This
estimates the unknown density by fitting a parametric family of densities ψ ð; θÞ
664 H.G. Olsen and G.H. Hermansen
locally, where ^ θ ðxÞ maximizes the likelihood. The unknown density ^fX ðxÞ is then
^
equal to ψ x, θ ðxÞ . Here, ψ is chosen to be a standardized multivariate normal
p
distribution, that is, ψ ðz; θÞ ¼ ψ ðz; RÞ ¼ ð2π Þ2 jRj2 exp 12z R1 z , where R ¼ R
1 0
ðzÞ ¼ ρij ðzÞ is the correlation matrix. K H ðÞ is chosen to be Gaussian kernel.
The last step is to combine to a final density and transform the data back to the
original domain. The density given by (Otneim and Tjøstheim 2016a) is
Yp
f ðx Þ
f ðxÞ ¼ f Z Φ1 ðF1 ðx1 ÞÞ, . . . , Φn Fp xp 1i i ; ð4Þ
i¼1 ϕ Φ ðF i ðx i ÞÞ
where fi and Fi are the marginal distributions, and Φ1 ðFi ðxi ÞÞ is the transformation
of the original vector xi to the normal domain (by first the marginal F and then the
inverse standard normal Φ1 ).
We have been investigating the conditional density, which previously has been
obtained by using the relation f ðyjxÞ ¼ f ðx, yÞ=f ðxÞ. This can however be unstable
due to the division of small values of the marginal f(x). The LGDE method makes
use of the Gaussian framework to make a stable estimation of the conditional
density directly. For a Gaussian distribution, the conditional distribution has an
explicit formula; if x ¼ ½x1 ; x2 is multinormal distributed with μ ¼ ½μ1 ; μ2 and
Σ ¼ ½Σ11 Σ12 ; Σ21 Σ22 , then x1 jx2 ¼ a will be multinormal with μ ¼ μ1 þ Σ12 Σ1 22
ða μ2 Þ and Σ ¼ Σ11 Σ12 Σ1 22 Σ21 . With this (Otneim and Tjøstheim 2016b)
obtains an estimator similar to Eq. 4 for f ðyjxÞ.
The main advantage of the LGDE method is for high-dimensional problems
where it is shown to overcome some of the problems related to the curse of
dimensionality. This is done by a simplification which is based on letting the
local correlation matrix at each point, ^ρ ij , only depend on their own pair of
variables, that is, ^ρ ij x1 ; . . . ; xp ¼ ^ρ ij xi , xj . This simplification reduces the
estimation of f(x) to a series of bivariate problems. Otneim and Tjøstheim
(2016a) have tested the LGDE method for several data cases and compared it to
standard kernel density. They have generated data from known distributions,
varying between two Z and ten dimensions and evaluate the results through a relative
square error (IRSE) ^f ðxÞ f ðxÞ 2 =f ðxÞdx. The result is that LGDE outper-
forms the KS for most cases and dimensions, chi-squared marginals with Gaussian
copula, lognormal marginals with t-copula, t-marginals with Clayton copula, uni-
form marginals with Clayton copula, and multivariate t-distribution. The case
where LGDE struggled against KS was a mixture of two Gaussian models; here,
KS performed better for the lower dimensions (less than 4). See (Otneim and
Tjøstheim 2016a) for more discussions.
While the main purpose of the LGDE method is for the high-dimensional data
sets, we demonstrate the use of the method in two dimensions. The results are
shown in Fig. 11a for the synthetic data and Fig. 11b, c for the real data, which is the
Recent Advancements to Nonparametric Modeling of Interactions Between. . . 665
Fig. 11 Result of the LGDE method used on the same data as in Fig. 2 shown in (a). Plots (b), (c)
show the estimated cdf on the real data case, GR conditioned on Vp (b) and GR conditioned on
Rho (c)
same data previously used for CT in Figs. 2 and 7, respectively. The LGDE in
Fig. 11a is seen as an improvement of the KS in Fig. 2 as it is able to achieve some
pinch-out effects. Figure 12 shows the LGDE method applied on the real data set
for three dimensions; this is similar to the example in Fig. 9 for CT. We compare
CT and LGDE on this data set in Sect. 6; see Fig. 13 and Table 1. While the
comparisons show some small favor to CT, we see that LGDE also works well.
6 Discussions
Our experience using these methods is that both CT and LGDE work well for two-
and three-dimensional problems, but that CT has the edge both in performance and
stability. To compare the performances of the methods on the real data sets (Figs. 9
and 12), we backtransform using the estimated cdf and look at the residuals in the
normal transformed domain. This means to first transform each data point to the
uniform domain with the estimated cdf and then transform them to the standard
normal domain by using the inverse cdf of the standard normal distribution. A
perfect transformation, where all dependencies to the underlying parameters (which
we condition on) are removed, will give normal distributed residuals with mean and
standard deviation scattered around zero and one, respectively. The reason we base
our test on transforming to standard normal is that this is of particular interest in
geostatistics where methods like Kriging often uses a workflow of first transforming
the data toward Gaussianity; this is called trans-Gaussian Kriging (Cressie 1993).
See (Hermansen et al. 2016) for a more thorough explanation on this transformation
test, where they used it on CT in two and three dimensions.
The sliding average results of the backtransformations can be seen in Fig. 13
where the mean should be around zero. The summed absolute deviations are listed
in Table 1, where we have also included results for standard deviations and a model
666 H.G. Olsen and G.H. Hermansen
Fig. 12 Result of the 3D conditional cdf estimated with LGDE, here TOC is conditioned on Vp
and Rho. Plot (a) shows the 3D data set and (b) is the 2D LGDE of TOC given Rho (Vp is ignored).
The bottom row shows two slices of the 3D density, for Vp equal to 71.1 in (c) and Vp equal to
89.2 in (d). Only sections of the 3D cube where there is ample data support are shown, and only
closest 20 % data to the chosen Rho value is plotted
Fig. 13 Backtransformed data using the estimated cdfs (2D and 3D) for CT (a) and LGDE (b).
For the 2D estimations, only Rho is conditioned on; Vp is ignored. The red and green lines are
result of backtransforming the Vp and Rho data, respectively. The black lines are from the 3D
model, where both Vp and Rho are used as conditioned on. The goal of the plot is to see if the 2D
estimations have systematic bias in the direction of the variable not used in the model, and we see
that the Vp variable trails off. This indicates that Rho should also be included when modeling the
response TOC
with only using the Rho variable. From this table, and this criterion, we see that CT
performs better than the LGDE (lower values) for the 3D models (Vp + Rho). We
also see that in the 2D models that backtransforming the ignored variable gets high
Recent Advancements to Nonparametric Modeling of Interactions Between. . . 667
values; see line 2 and 5 in Table 1 and Fig. 13 where Vp was the only variable
conditioned on (Rho is ignored). This indicates that 2D model does not capture all
the underlying signals and that this variable should be included in the model.
We emphasize that one should not rely too much on the results in Table 1 for a
comparison of the methods, as this is only for one data case and this criterion. To
create test criteria for comparisons, one must consider what the estimated densities
will be used for. Another evaluation of the methods could be to draw data from
several known distributions, and then repeatedly measure the difference between
the true and estimated distributions, like was done for LGDE by (Otneim and
Tjøstheim 2016a).
The kernel density estimation in Eq. 1 was set up with a constant bandwidth matrix.
One disadvantage with constant bandwidths is that the number of data inside the
kernel varies, depending on the clustering of the data. An improvement to this is
letting the bandwidth vary with the data, e.g., use a small bandwidth in areas with
many observations, and increase the bandwidth in areas with few observations. This
fits within the kernel density estimator described in Sect. 3, but with replacing the
bandwidth H with H(x), one that varies along x in each direction. See (Sain 1994)
for a comparison of adaptive kernel estimation methods.
An example of estimating a density in one dimension with a varying bandwidth
is shown in Fig. 14. Both are estimated with a Gaussian Kernel, but the pdf in (a) is
estimated with a constant h and the pdf in (b) with an adaptive h. From this figure it
is clearly seen that the pdf in (b) is better suited to the underlying data.
As all three methods discussed in this paper use kernel smoothing at one point,
this extension could apply to all methods. However, CT and LGDE utilize trans-
formations of the data before applying the KS, which typically helps in regions with
sparse data.
668 H.G. Olsen and G.H. Hermansen
Fig. 14 Porosity data with pdf (red line) estimated with constant bandwidth (a) and with an
adaptive bandwidth (b)
7 Conclusion
Bibliography
Cressie N (1993) Statistics for spatial sata. Wiley series in probability and mathematical statistics
Hermansen G, Kolbjørnsen O, Abrahamsen P (2016) Extensions and applications of the cloud
transform with directional smoothing. Norwegian Computing Center
Kolbjørnsen O, Abrahamsen P (2005) Theory of the cloud transform for applications. Springer
Leuangthong O, Deutsch CV (2003) Stepwise conditional transformation for simulation of
multiple variables. Int Assoc Math Geol 35:155–173
Otneim H, Tjøstheim D (2016a) The local Gaussian density estimatorfor multivariate data. Stat
Comput
Otneim H, Tjøstheim D (2016b) Non-parametric estimation of conditional densities. A new
method. Chapter 3.3 in the PhD thesis of Håkon Otneim, Department of Mathematics,
University of Bergen, Norway
Sain SA (1994) Adaptive kernel density estimation. PhD thesis
Sammut C, Webb GI (2010) Encyclopedia of machine learning. In: Encyclopedia of machine
learning. Springer US, pp 257–258
Silverman B (1998) Density estimation for statistics and data analysis. Chapman & Hall/CRC,
London
Wand M, Jones MC (1993) Comparison of smoothing parameterizations in bivariate kernel
density estimation. Am Stat Assoc 88:520–528
Geostatistical Methods for Unconventional
Reservoir Uncertainty Assessments
1 Introduction
Unconventional plays are often considered statistical plays. With a traditional play,
the local data measures are reasonably accurate and in combination with local
geologic factors allow for the mapping of reservoir properties that are suitable for
forecasting with the application of flow simulation. In contrast, for a statistical play,
well measures are highly uncertain, and the relationships between the measured
values production is weak (see Fig. 1). In the face of this difficulty, Olea et al.
(2011) suggests directly modeling well production measures such as EUR (although
we utilize IP as a production proxy due to availability). This may first require a
standardization to account for well parameters such as completed length. Well
production is then treated as the regionalized variable of interest and summarized
for reservoir uncertainty modeling. Note, while geologic factors naturally enter the
traditional workflows through hierarchal depo- and lithofacies and porosity, per-
meability, and saturation distributions and relations, the challenge is to integrate
geologic factors in this statistical play workflow. Careful sub-setting and selection
Geostatistical Methods for Unconventional Reservoir Uncertainty Assessments 673
Fig. 1 Workflows for traditional and statistical play assessment. In a statistical play, the well-
based petrophysical measures are highly uncertain and poorly related to production. This moti-
vates direct modeling of production as a regionalized variable (Figure from Pyrcz and Deutsch
2012)
of analogs and matching and sub-setting modeling area of interest with consider-
ation of geologic setting and stationarity are essential.
The statistical play concept has been formalized within the SPEE#3 Monograph
that outlines methods to model uncertainty for undeveloped reserves based on the
bootstrap method (SPEE 2010). The bootstrap is a statistical resampling technique
that permits the quantification of uncertainty in statistics by resampling (with
replacement) from the original data (Efron 1982). Importantly, this method does
not account for neither spatial context such as location of samples nor spatial
continuity.
The SPEE#3 Monograph approach proceeds by identifying analogous wells,
modeling a distribution of EUR, determining the number of wells in the well plan,
and repeated bootstrap sampling to evaluate the uncertainty in production results
over a well aggregate measure (e.g., the average or sum of production). This
approach benefits from simplicity and ease of use.
674 M.J. Pyrcz et al.
2 Methodology
Fig. 2 Naı̈ve and cell-based declustered IP distributions for three US domestic shale plays.
Display includes distribution shape change and change in mean
exists any longer; the well provides no information. Beyond the range there may be
some small negative correlation indicative of spatial trends.
The semivariogram is a fundamental input for geostatistical simulation and the
workflows presented later, but there are benefits in directly utilizing insights from
the semivariograms in unconventional reservoirs. A simple application of the
semivariogram is to determine the maximum extent of information from a well in
map view based on the semivariogram range. In general, wells that have a separa-
tion of twice the range are not redundant and carry the maximum information
content. Also, well location maps with range circles (or ellipses in the case of
anisotropy) provide visualization of data coverage and may be applied to assist in
well planning. The circle size may be reduced to reflect an acceptable level of
spatial uncertainty.
Experimental isotropic semivariograms are calculated for the normal score
transform of well IP (positioned at the well collars) with the filtered IP data for
three US domestic shale plays (see Fig. 3). Each of experimental semivariogram
indicates (1) 30–40 % relative nugget effect, (2) long correlation ranges, and
(3) indication of long-range trends. These shale gas plays indicate a high degree
of variability between adjacent wells, but also some degree of correlation or
information over long distances. Large relative nugget effects are not expected to
be due to allocation issues, but may represent strong initial production sensitivity to
artificial fracture sets.
Dispersion variance is a generalized form of variance that accounts for the scales of
the samples, scale of the area of interest, and spatial continuity. In geostatistics,
given the spatial and scale context, all variances are dispersion variances. Volume-
variance relations refer to the general relationship between the variance and scale
(Pyrcz and Deutsch 2014).
The dispersion variance may be calculated from volume integration of the
semivariogram values (known as gamma bar), γ:
Fig. 3 Semivariograms of normal score transformed IP data for three US domestic shale plays.
The sizes of the experimental points are scaled by number of pairs available
It follows that selecting a larger lease block will reduce dispersion variance
between lease blocks within the play, D2 (block, play), but will retain a larger
dispersion variance within the lease block, D2 (well, block). While estimates of
larger lease block scale are easier, individual well results within the larger lease
block will have more variability.
Volume-variance relations based on gamma bar values are formulated under the
assumption that for the variable of interest, the variogram is stationary over the
domain, averaging is performed with nonoverlapping volumes, and variable aver-
ages linearly (Journel and Huijbregts 1978; Frykman and Deutsch 2002). In the
proposed workflow based on global kriging, well EUR (IP when EUR is not
available) is the variable of interest. This is reasonable if it is assumed that:
1. The EUR variogram is stationary over the block, V.
2. Well EUR is represented as the total volume of production at the scale of the
effective drainage of a nominal well, EURv(u).
3. EURv ðuα Þ, 8uα 2V is an exhaustive set nonoverlapping data support production
volumes of the block, V.
4. There is no interaction in EUR between adjacent EURv(u).
678 M.J. Pyrcz et al.
Kriging is general and may directly integrate scale information. For example,
kriging may be applied to calculate models of data scale uncertainty represented
by a mesh of the kriging variances over the lease block. Also, kriging may be
applied with volume-integrated semivariogram models to directly estimate and
assess uncertainty of a production property over other scales, such as lease block
scale, through the kriging estimate and variance, respectively. Deutsch and Deutsch
(2010) propose the use of global kriging to directly estimate block uncertainty (and
declustered block mean).
The kriging variance equation derived in standard geostatistics textbooks is
composed of global dispersion variance, closeness, and redundancy terms (Pyrcz
and Deutsch 2014). (1) The global dispersion variance (the variance of the property
of interest at the correct scale) is the maximum uncertainty possible at the specified
volume support. This accounts for the size of the block being estimated. (2) Close-
ness accounts for the correlation of the data to the volume being estimated. This
component accounts for the respective coverage of the data within this volume
being estimated. For example, if for a specific dataset, the block expands further
away from the data, overall closeness decreases and the kriging variance increases.
(3) Redundancy accounts for the correlation between the data. For example, if the
spatial continuity range relative to well spacing is large, then the well data redun-
dancy is large and the kriging variance increases.
Closeness and redundancy components interact resulting in interesting behavior
in block uncertainty. A simple example is provided with representative histogram
(Haynesville declustered example) and a stationary 36 36 km lease block. The
well data count was varied from 0 to 100 wells (assuming uniform well spacing)
and the spherical variogram range was varied from 0 to 50 km. The resulting
Geostatistical Methods for Unconventional Reservoir Uncertainty Assessments 679
10 Wells
1,000
750
20 Wells
30 Wells
500
450
40 Wells
Haynesville High
400
Haynesville Low
Haynesville Mid
50 Wells
350
60 Wells
300
70 Wells
80 Wells
250
90 Wells
10 km 20 km 30 km 40 km
Spatial Continuity
Fig. 4 Global kriging uncertainty vs. number of wells and spatial continuity. Response surface is
standard deviation of IP (MCFPD)
assumption of histogram and variogram stationarity over the block). For example,
for the Haynesville mid-variogram range case (vertical line drawn on Fig. 4), about
50 regularly spaced wells are required to reduce the block IP uncertainty to a
standard deviation of 400 MCFPD. This builds on the work of Wilde and Deutsch
(2013) to model uncertainty vs. well count.
Fig. 5 Original well data and local uncertainty in IP and three scenarios for proposed wells,
including 3 3 pads in low and high uncertainty locations and drilling for information. The
proposed well locations are outlined in red
production data uncertainty in MCFPD from the kriging variance). Three proposed
nine well scenarios are shown, including 3 3 pads in low and high uncertainty
locations and drilling for information with their updated local uncertainty maps.
Global IP distribution and spatial continuity model is based on a declustered IP
dataset from a modeled analog dataset.
The resulting well aggregate IP distributions are shown in Fig. 6. The spatial
context significantly impacts the resulting well aggregate uncertainty model. The
mean is influenced by previously drilled well IP within the range of spatial
continuity. The uncertainty for the well aggregate result is significantly higher for
the pad in the second scenario due to the limited correlation with previously drilled
wells along with highly redundant proposed wells, and the third scenario has the
lowest uncertainty due to limited correlation between proposed wells resulting in
averaging out of variability in the aggregate summary.
4 Discussion
Fig. 6 Well aggregate uncertainty distributions (average IP) for three scenarios of nine proposed
wells based on representative statistics, spatial continuity, and spatial bootstrap
the block, this may result in underestimation of uncertainty. Without the concept of
volume-variance relations, the scale of the block is assumed to be the scale of the
data, resulting in overestimation of block uncertainty.
5 Conclusions
Acknowledgments The authors are appreciative to Chevron Energy Technology Company for
support of this work and for allowing this publication. Also, the constructive reviews of two
anonymous reviewers are appreciated.
Geostatistical Methods for Unconventional Reservoir Uncertainty Assessments 683
Bibliography
Deutsch JL, Deutsch CV (2010) Some geostatistical software implementation details. CCG Paper
2010-412, Centre for Computational Geostatistics, University of Alberta, Edmonton
Deutsch CV, Journel AG (1998) GSLIB: geostatistical software library and user’s guide, 2nd edn.
Oxford University Press, New York
Efron B (1982) The jackknife, the bootstrap, and other resampling plans. Society for Industrial and
Applied Math, Philadelphia
Frykman P, Deutsch CV (2002) Practical application of geostatistical scaling laws for data
integration. Petrophysics 43(3):153–171
Journel AG (1994) Resampling from stochastic simulations. Environ Ecol Stat 1:63–84
Journel AG, Huijbregts C (1978) Mining geostatistics. Academic, New York City, 600 p
Olea RA, Houseknecht DW, Garrity CP, Cook TA (2011) Formulation of a correlated variables
methodology for assessment of continuous gas resources with an application to the Woodford
play, Arkoma Basin, eastern Oklahoma. Boletı́n Geológico y Minero, 122 (4), pp 483–496.
ISSN: 0366-0176
Pyrcz MJ, Deutsch CV (2012) Uncertainty models for exploration and appraisal of shale gas. CCG
Paper 2012-131, Centre for Computational Geo-statistics, University of Alberta, Edmonton
Pyrcz MJ, Deutsch CV (2014) Geostatistical reservoir modeling, 2nd edn. Oxford University
Press, New York, p 448
Pyrcz MJ, Gringarten E, Frykman P, Deutsch CV (2006) Representative input parameters for
geostatistical simulation. In: Coburn TC, Yarus RJ, Chambers RL (eds) Stochastic modeling
and geostatistics: principles, methods and case studies, vol II and 5. AAPG Computer Appli-
cations in Geology, pp 123–137
Society of Petroleum Evaluation Engineers (2010) Guidelines for the practical evaluation of
undeveloped reserves in resource plays. Monograph 3
Wilde BJ, Deutsch CV (2013) A methodology for quantifying uncertainty versus data spacing
applied to the oil sands. CIM J
Productivity Prediction Using Alternating
Conditional Expectations
Emmanuel T. Schnetzler
1 Introduction
y ¼ β0 þ β1 x1 þ β2 x2 þ þ βp xp ð1Þ
This imposes a strong assumption of linearity that is not appropriate when complex
relationships exist.
In some cases, nonlinear relationships are present and known, in which case
parametric transformations of the variables can be applied to build a model. This is
the case in the Box-Cox family of transformations for continuous variables (Box
and Cox 1964). However, this also imposes a strong preconceived model on the
data that is not valid in complex cases.
3.1 Theory
4.1 Overview
ACE is applied to a real case in an area of 3000 m2 located in the western part of the
Denver-Julesburg Basin in Colorado. Major oil-producing areas shown in Fig. 2 in
the Niobrara Formation include the Hereford in the northwest, the East Pony in the
northeast, and the Wattenberg Field in the center.
The available data include production information and a set of geophysical and
geological layers.
688 E.T. Schnetzler
5 Response Variable
The response variable selected is the early cumulative oil production, calculated as the
sum of the production of the best 3 months in the first year of production. Because
historical production includes wells drilled over several decades using different
technologies, it is necessary to subset the data to a consistent set of wells; in this
case, only horizontal wells were used. This controls for part of the variability between
wells, although some further filtering could control for additional variables, in partic-
ular engineering factors. Exploratory data analysis shows that in this case the subset of
Productivity Prediction Using Alternating Conditional Expectations 689
Fig. 4 Data layers input to the ACE run as listed in the text
6 Predictor Variables
A set of nine predictor variables available for the full area of interest and deemed
related to productivity is compiled. Figure 4 shows an overview of the data layers
listed below.
Top row, left to right:
• Precambrian depth
• Distance to interpreted faults
• Resistivity from wells
• Bouguer gravity anomaly
• RTP (reduced to pole) magnetic field
Bottom row, left to right:
• Niobrara isopach
• TOC (total organic carbon)
• Tmax (temperature at which the maximum rate of hydrocarbon generation occurs
in a kerogen sample during pyrolysis analysis)
• Vitrinite reflectance
692 E.T. Schnetzler
7 ACE Result
The result of the prediction (early oil prediction in BBL) from the Alternating
Conditional Expectations is shown in Fig. 5. The three major oil-producing areas
are highlighted in the prediction, in part driven by the data. Additional areas with
limited data control are predicted as high-producing locations, in particular in the
northeast, south of the East Pony Field, and in the south where only a few training
data points are available.
Correlation between actual and predicted production is 0.38 for multiple linear
regression compared to 0.74 for ACE (Fig. 6).
Predictor variables can be ranked according to their influence on the response
variable, as shown in Table 1, with increasing influence from top to bottom.
The optimal transformation can be plotted and analyzed to understand the effect
of each variable on the response variable. Figure 7 shows the optimal transforma-
tion of the response variable (early oil production, in BBL).
8 Conclusions
Many regression techniques can be used to predict a variable from a set of predictor
variables. When relationships are complex and cannot be modeled parametrically
either through linear model or more complex functions, a nonparametric approach
is more appropriate.
Machine learning algorithms (Support Vector Machines, Neural Networks, etc.)
are powerful but tend to be difficult to interpret. Nonlinear, nonparametric
approaches based on optimal transformations of the variables such as Alternating
Conditional Expectations and some modifications designed to address some limi-
tations, AVAS, for example (Tibshirani 1988), provide the opportunity to inspect
the optimal transformations and put them in a physical context for validation.
A number of potential pitfalls need to be kept in mind when applying any
regression technique. Overfitting the training data is a common problem in predic-
tive modeling that should be closely watched through cross-validation. The predic-
tor variables should be chosen with care, and it should be possible to formulate a
reason why each is likely to be related to the response variable, even if the link
cannot be expressly defined. The method relies on colocated relationships and does
not take into account relations to neighbors such as is the case in co-kriging or
co-simulation.
The example presented shows how ACE can be applied in a spatial setting to
predict a grid of early oil production. The case study illustrates the improvement in
correlation between transformed predictors and transformed response. Other vari-
ables such as petrophysical properties (e.g., porosity) could be considered for this
approach, and while the case study is 2D, it is applicable in 3D.
Productivity Prediction Using Alternating Conditional Expectations 693
50000. 50000.
Linear Regression
40000. 40000.
V9_S2
30000. 30000.
20000. 20000.
10000. 10000.
0. 0.
0. 10000. 20000. 30000. 40000. 50000. 60000. 0. 10000. 20000. 30000. 40000. 50000. 60000.
OIL_3mth_H2_BBL OIL_3mth_H2_BBL
Fig. 6 Correlation between true and predicted production for linear regression and ACE
Fig. 7 Optimal
transformation of the
response variable
Productivity Prediction Using Alternating Conditional Expectations 695
Bibliography
Barnett RM, Deutsch CV (2013) Tutorial and tools for ACE regression and transformation. Centre
for Computational Geostatistics (CCG) Annual Report 15, 401, University of Alberta,
Edmonton
Box EG, Cox RD (1964) An analysis of transformations. J R Stat Soc Ser B 211–252
Breiman L, Friedman JH (1985) Estimating optimal transformations for multiple. Regression and
correlation. J Am Stat Assoc 80:580–598
Kuhn M, Johnson K (2013) Applied predictive modeling. Springer, New York
Tibshirani R (1988) Estimating transformations for regression via additivity and variance stabili-
zation. J Am Stat Assoc 83:394–405
Wang D, Murphy M (2004) Estimating optimal transformations for multiple regression using the
ACE algorithm. J Data Sci 2:329–346
The Adaptive Plurigaussian Simulation
(APS) Model Versus the Truncated
Plurigaussian Simulation (TPS) Model Used
in the Presence of Hard Data
1 Introduction
B. Sebacher (*)
TU Delft, Delft, The Netherlands
e-mail: [email protected]
R. Hanea
Statoil ASA, Stavanger, Norway
e-mail: [email protected]
A. Stordal
IRIS, Stavanger, Norway
e-mail: [email protected]
generate facies fields conditioned to hard data, the TPS truncates Gaussian fields
that are not stationary. Consequently, the simulated facies fields with TPS are no
longer conditioned to expected facies proportions.
The experiment is performed using a synthetic example with three facies types
of which facies fields exhibit a particular topology, characteristic of a real field from
the North Sea.
In order to give an easy description of the APS method, we describe the method-
ology for a reservoir model where three facies types are present (denoted F1, F2, and
F3). In each field, the probability is 1 at the well locations (or any other location
where we have a facies observation) where the associated facies type occurs and
0 where it does not occur. This means that the facies observations are already
incorporated into the probability fields. Important prior information is regarding the
transition (contact) between any two facies type and any topological information.
Suppose that any two facies types can intersect each other. We denote by pk the
probability field associated to facies type k (where k ¼ 1, 2, 3).
The APS algorithm (for three facies types) for simulating a facies field is
summarized as follows:
1. Given the prior probability fields p1, p2, p3 and information of facies connec-
tions, create a layout of the simulation map (or truncation map1). The simulation
map consists of a reliable decomposition of the square (0, 12) in three
sub-domains, each having assigned a facies type. The area of each sub-domain
is equal to the probability of the associated facies type. In Fig. 1 is shown a
layout of the simulation map used in the APS procedure in the next section. The
construction of the simulation map must take into account the information about
the number of the facies, the facies connection, and the topological characteri-
zation of the reservoir geology.
2. Generate samples from two (stationary) Gaussian random fields, Y1 and Y2, with
predefined dependence structure. The geostatistical properties of the Gaussian
fields are correlated with the prior information of the spatial distribution of the
facies types (indicator variogram).
3. Transform the Gaussian random fields to uniform random fields, α1 and α2, using
the integral transform. For each grid cell j of the reservoir domain, α1 ðjÞ ¼ Φ1
ðY 1 ðjÞÞ and α2 ðjÞ ¼ Φ2 ðY 2 ðjÞÞ, where Φ1 and Φ2 are the marginal Gaussian
cumulative distribution functions (CDF) of Y1 and Y2.
1
We prefer to call it the simulation map instead of truncation map because we consider that in the
APS we simulate a facies type from a discrete distribution rather than truncate some real values.
700 B. Sebacher et al.
4. For each grid cell j of the reservoir domain, build its simulation map from the
layout (using the marginal probabilities pj1 , pj2 , pj3 collected at the grid cell j) and
set the facies type in grid cell j to Fk if the point ð/1 ðjÞ, /2 ðjÞÞ is situated in the
sub-domain assigned to facies type k (in the simulation map assigned to grid cell
j).
Consequently, each grid cell has its own simulation map that depends on the set
of probabilities collected at that location. At the location where we have facies
observation, its simulation map consists of the square (0, 1)2, occupied by the facies
type observed there. This means that irrespective of the Gaussian field values, the
observed facies will be always simulated so the hard data is preserved. With this
method, we always simulate facies fields honoring the marginal facies probabilities
and, consequently, the expected facies proportions.
3 APS vs TPS
The traditional plurigaussian simulation (TPS) model has two main ingredients:
• The Gaussian random fields defined on a region of interest. The region of interest
in reservoir engineering is the reservoir domain, but with a discrete structure
(grid cells).
• A truncation map defined on a multidimensional real space. The truncation map
is defined by the intersection of some curves that divide the space into regions,
each having assigned a facies type.
The Adaptive Plurigaussian Simulation (APS) Model Versus the Truncated. . . 701
The number of the Gaussian fields is equal to the dimension of the space where
the truncation map is designed. If the dimension of the space is two, a pair of
Gaussian random fields, Y1 and Y2, is simulated, and, at each grid cell j, a facies
type is assigned depending on where the point of simulated values (Yj1 , Yj2 ) is
situated in the truncation map. Consequently, in the TPS methodology, the trunca-
tion map tailors the facies fields according to the Gaussian field values at each grid
cell. This methodology can be easily rewritten in a similar manner as we have
presented the APS before. Let us consider for simplicity that the Gaussian fields Y1
and Y2 follow at each location a normal distribution of which CDF’s are Φ1 and Φ2.
Then, the function F : R2 ! ð0; 1Þ2 , Fðx1 ; x2 Þ ¼ ðΦ1 ðx1 Þ, Φ2 ðx2 ÞÞ is a bijection and
has increasing components. Consequently, it defines a bijection between the trun-
cation maps designed in the space R2 and the ones designed in the space (0, 1)2.
Thus, assigning the facies type based on the projection of (Yj1 , Yj2 ) in a truncation
map from R2 is equivalent with assigning a facies type based on projection of
(Φ1(Yj1 ), Φ2(Yj2 ) in a truncation map designed in (0, 1)2. Consequently, the APS can
be viewed as the TPS model conditioned to probability fields of the facies type (soft
data) or is the methodology that incorporates the probability fields in the TPS
model. In addition, the APS with uniform probability fields is the TPS.
In the traditional TPS, when facies observations are available at some locations,
the Gaussian field values are generated such that the simulated facies fields satisfy
the observations (interval conditioning). This implies a change in the mean and
variance functions of the Gaussian fields compared with the case where the facies
observations are not present, i.e., the Gaussian fields are no longer stationary. We
will show that this modification produces a bias for the probability field of each
facies type calculated from an ensemble of realizations.
Let us consider a rectangular domain with 50*100 grid cells with three facies
types present. One facies type exhibits a long correlation from west to east (denoted
facies type 2) and a small correlation from north to south. Another facies type
(denoted facies type 3) is characterized by bodies of rocks that occur most likely on
the edge of the other facies, and the last facies type is the medium where all other
facies types are propagated (denoted facies type 1). This topology is characteristic
to a real field from the North Sea (Hanea et al. 2014). We apply the TPS method-
ology for creating realistic facies distributions, using two Gaussian fields. Initially,
we set both Gaussian fields stationary with its marginal having a standard normal
distribution and having a Gaussian covariance type. The first Gaussian field is
anisotropic, modeled with a long correlation range of 50 grid cells and a small
correlation range of 5 grid cells and with the principal direction being the horizontal
direction. The second Gaussian field is isotropic with the correlation range of five
grid cells.
The truncation map used for generation of the facies fields is shown in Fig. 2,
where the parameters of its lines are calculated based on the expected facies
proportions p1 ¼ 0.425, p2 ¼ 0.425, and p3 ¼ 0.15. This truncation map is obtained
applying the inverse of function F to the simulation map from Fig. 1. Consequently,
702 B. Sebacher et al.
the thresholds α1, α2, and β (see Fig. 2) are calculated based on the inverse of
function F:
pffiffiffiffiffi pffiffiffiffiffi
p3 p3 p3 p3 pffiffiffiffiffi
α1 ¼ Φ1
1 p1 þ 1
, α2 ¼ Φ1 p1 þ þ , β ¼ Φ1
2 2 p3 ð1Þ
2 4 2 4
This truncation map, in combination with the Gaussian fields presented before,
generates facies fields with realistic topological and geometrical structure (Hanea
et al. 2014), but is not necessarily conditioned to facies observations. We consider
13 locations from where we observe the facies types that occur at those positions.
The positions and the facies observations are presented in Table 1 and in Fig. 3.
In order to generate conditional facies fields, we follow a conditional simulation
approach (Armstrong et al. 2003). Firstly, at the observation grid cells, we generate
pairs of Gaussian values such that those pairs yield correct facies observations in
accordance with the truncation map. Secondly, the simulated Gaussian values are
used in a conditional sequential Gaussian simulation process in order to populate
with values the remaining grid cells. With this method, we generate an ensemble of
120 samples of pairs of Gaussian fields. These samples generate by truncation of
120 different facies fields with correct facies observations. We note that the
procedure of conditioning of the Gaussian fields destroys their stationarity. This
means that the mean function of the Gaussian fields is not 0 at each location, and
consequently, the simulated facies fields are no longer conditioned to expected
facies proportions. This raises the question if the truncation map is still valid (i.e.,
produces facies fields conditioned to expected facies proportions).
The Adaptive Plurigaussian Simulation (APS) Model Versus the Truncated. . . 703
10 Facies Facies
type 3 type 1
15
20 Facies Facies Facies
type 2 type 3 type 1
25 Facies Facies
type 2 type 2
30
35
40 Facies Facies Facies
type 1 type 1 type 3
45
50
10 20 30 40 50 60 70 80 90 100
Fig. 3 The facies observations into the domain
10 10 10
20 20 20
30 30 30
40 40 40
50 50 50
20 40 60 80 100 20 40 60 80 100 20 40 60 80 100
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
In Fig. 4 is shown the probability fields of the facies types obtained from the
ensemble of facies fields generated with the TPS method and conditioned to the
facies observations presented in Table 1.
One may observe that these probability fields are biased in the sense that the
geostatistical properties of the Gaussian fields drastically influence their spatial
distribution. For instance, if we look at the top of probability field of the facies type
704 B. Sebacher et al.
10 10 10
20 20 20
30 30 30
40 40 40
50 50 50
20 40 60 80 100 20 40 60 80 100 20 40 60 80 100
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
two, one may observe a long correlation in the horizontal direction due to the
placement of three facies observations of the facies type two within the correlation
influence of the first Gaussian field. The distance between the observation grids is
40 grid cells, whereas the long correlation range of the Gaussian fields is 50 grid
cells. This causes the bias for the probability fields. The bias increases with the
number of the facies observations situated within the correlation influence.
We also apply the APS methodology using stationary Gaussian fields having the
same geostatistical setup as before. In order to do that, we need to have prior
probability fields of the facies occurrence that incorporate the facies observations.
In practice the prior probability fields of the facies are developed by a group of
experts (geologists, geophysicists, geo-modelers) in the early phase of the reservoir
exploration, using various data, such as core information, seismic data, well log
data, etc. In the papers of Beucher et al. (1999) and Doligez et al. (2002), the authors
present methodologies to incorporate facies probability maps that comes out from
seismic inversion, in the geostatistical simulation of the facies distribution. Here
using a synthetic example, we have to create ourselves these probability fields, but
we want to avoid the strong bias observed in the probability fields from the
ensemble obtained with traditional TPS. Using an empirical approach, we generate
a set of “prior” probability fields of the facies that incorporate the facies observa-
tions (Fig. 5). We generate these fields with a small correlation around the obser-
vation locations and, in addition, with the property that the mean of each probability
field is equal with the expected facies proportions of the associated facies type. This
last property ensures that the simulated facies fields are conditioned to expected
facies proportions. The layout of the simulation map used in the APS is presented in
Fig. 1. We generate an ensemble of 120 samples of pairs of stationary Gaussian
fields, and, using the APS, the simulation map layout, and the probability fields, we
generate the ensemble of facies fields.
The probability fields for each facies type, calculated from the ensemble of
realizations with APS, are presented in Fig. 6. One clearly observes that the prior
probability fields are very good preserved and the geostatistical properties of the
Gaussian fields have little influence. This was expected because the APS samples
correctly at each grid cell from the prior marginal distribution of the facies types.
The Adaptive Plurigaussian Simulation (APS) Model Versus the Truncated. . . 705
10 10 10
20 20 20
30 30 30
40 40 40
50 50 50
20 40 60 80 100 20 40 60 80 100 20 40 60 80 100
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
10 10 10
20 20 20
30 30 30
40 40 40
50 50 50
20 40 60 80 100 20 40 60 80 100 20 40 60 80 100
10 10 10
20 20 20
30 30 30
40 40 40
50 50 50
20 40 60 80 100 20 40 60 80 100 20 40 60 80 100
Fig. 7 Facies fields generated with TPS (top) and APS (bottom)
The ensembles of facies fields approximate the prior pdf of the facies field
conditioned to the facies observations and the probability fields from Figs. 4 and
6 approximate of the prior marginal distribution of the facies fields. In order to
completely quantify the quality of the simulation, we still have to look at the facies
fields from the topological and geometrical perspective, to see if they represent the
same geological concept. In Fig. 7 we present three ensemble members of facies
fields in both simulations; at the top is shown the simulation with the TPS condi-
tioned to facies observations and at the bottom the facies fields obtained with the
APS (facies type 1 is in blue, facies type 2 is in green, and facies type 3 is in red).
One can observe that there is no difference regarding topology and geometry of the
facies among these fields. Consequently, we have two ensembles of facies fields
with consistent facies observations at the well locations, both representing the same
geological concept, but the ensemble obtained with APS quantifies better the prior
uncertainty of the facies distribution.
706 B. Sebacher et al.
4 Conclusions
Bibliography
Agbalaka C, Oliver D (2008) Application of the enkf and localization to automatic history
matching of facies distribution and production data. Math Geosci 40(4):353–374
Agbalaka CC, Oliver DS (2009) Automatic history matching of production and facies. SPE
reservoir simulation symposium. Society of Petroleum Engineers
Armstrong M, Galli A, Loc’h G (2003) Plurigaussian simulations in geosciences. Springer,
New York
Astrakova A, Oliver DS (2014) Conditioning truncated pluri-Gaussian models to facies observa-
tions in ensemble-kalman-based data assimilation. Math Geosci 47(3):345–367
Beucher H, Fournier F, Doligez B, Rozanski J (1999) Using 3D seismic-derived information in
lithofacies simulations. A case study.SPE 56736. SPE – annual conference of the Soc. of
Petrol. Eng. SPE, Houston, pp 581–592
Deraisme J, Farrow D (2005) Geostatistical simulation techniques applied to kimberlite orebodies
and risk assessment of sampling strategies. Geostat Ban: 429–438
Doligez B, Fournier F, Jolivet G, Gançarski S, Beucher H (2002) Seismic facies map integration in
geostatistical geological model: a field case. EAGE – 64th conf. and exhibition of the European
Assoc. of Geoscientists & Engineers. EAGE, Florence, pp 215–218
Galli A, Beucher H, Le Loc’h G, Doligez B (1994) The pros and cons of the truncated Gaussian
method. In: Geostatistical simulations. Kluwer, Dordrecht, pp 217–233
Hanea R, Ek T, Sebacher B, Saetrom J, Sollien D (2014) Geologically realistic facies updates for a
north sea field. 76th EAGE conference and exhibition. Amsterdam
Lantuejoul C (2002) Geostatistical simulation: models and algorithms. Springer, Berlin
Le Loch G, Galli A (1997) Truncated plurigaussian method: theoretical and practical points of
view. Geostat Wollongong 96(1):211–222
The Adaptive Plurigaussian Simulation (APS) Model Versus the Truncated. . . 707
1 Introduction
Most oil fields in China have entered into the middle and later development stage.
In order to effectively explore the remaining oil, it is necessary to finely character-
ize the reservoir. For example, the grid of the reservoir model should be refined to
10 10 m in a plane and 0.2 m vertically. For some old oil fields, the nodes of the
reservoir model can reach up to ten million, and the greater the volume of the
model, the longer time the simulation needed. Reservoir stochastic modeling
method usually requires the establishment of multiple equal probability models.
These models are used to predict the risk development and increase the simulation
Data patterns (pattern, called Pat) are based on the style of multipoint geostatistics
modeling algorithm in computing the minimum unit constituting a priori geological
concept model “of training image (TrainImage, referred as TI)” local space-related
multipoint structure. The data pattern is a data model (template, abbreviated T)
which is a framework for the training template scan image acquisition. First, define
the data model as follows:
T ¼ hm , m ¼ i þ j I þ k I J i 2 ½0, I 1, j 2 ½0, J 1, k 2 ½0, K 1
I, J, and K are the length and breadth dimensions of the data template T. Figure 1
is a two-dimensional example of data pattern at position u and is obtained through
scanning the training image:
An MPS Algorithm Based on Pattern Scale-Down Cluster 711
Fig. 1 Sketch of pattern and realization. (a) Pattern based on template. (b) Realization used to
scan by template
PatT ðuÞ ¼ fTIðu þ hÞh 2 Tg
The similarity is the index which can evaluate the degree of similarity between
two objects, and it is widely used in pattern recognition, computer vision, and other
fields. Common similarity functions are Minkowski function (Lk norm), Hsim
function, Pearson coefficient, and Jaccard similarity coefficient. Similarity function
takes two objects as input variables to determine a nonnegative real number which
can represent similarity. SIMPAT using Manhattan distance (L1 norm) calculates
the degree of similarity between data patterns or between the data pattern and data
event. It is defined as
Xd
L1 ðX; Y Þ ¼ i¼1
jxi yi j
sðX; Y Þ ¼ 1=ðdðX; Y Þ þ 1Þ
Figure 2a–e is the Manhattan distance of five data patterns. The Manhattan
distance of data patterns in Fig. 2a is equal to 1. X and Y are the most similar data
pattern. The similarity of X and Y in Fig. 2c is minimum.
The core idea of the pattern-based multipoint geostatistics algorithm SIMPAT is
that it adopts the degree of similarity between data patterns and data events in the
training image, getting all the priori geological model through by scanning the
training images and storing it in the computer memory to avoid repeated scanning
of training images. Calculating the similarity between data patterns and data events,
712 Y. Siyu et al.
matching the most similar data events and data pattern in the database, and covering
and freezing data event until the simulation are completed. The time complexity of
SIMPAT is proportional to the size and scale of pattern database, and the size of the
pattern database is closely related to the size of the training image and data pattern.
The larger the size of the training image and data pattern, the more information the
reservoir geological model library contains and the longer the simulation time.
When SIMPAT is applied to an actual oil field, in which grids reach one million to
ten million, the simulation time can be up to several hours or even days. By
analyzing the modeling features of SIMPAT and considering the memory footprint
drawbacks of previously improved methods, the new PSCSIM algorithm is pro-
posed to balance the contradiction between time-consuming and memory footprint.
The core concept of the new algorithm is the thumbnail pattern. The commonly
used methods for the dimension reduction of data pattern are adjacent equally
spaced sampling method, bilinear interpolation and cubic convolution interpolation
method, etc. In this paper, the basic idea of adjacent equally spaced sampling
method is calculating the coordinate mapping relationship between the thumbnail
pattern and data pattern by the reverse transform of coordinates. If the coordinates
mapped to the data pattern are the floating point, then it changes to an integer. The
An MPS Algorithm Based on Pattern Scale-Down Cluster 713
r i ¼ I Pat =I ThumPat r j ¼ J Pat =J ThumPat , r k ¼ K Pat =K ThumPat , I Pat , and IThumPat is the
dimension of data patterns and thumbnail pattern in the direction of I.
Figure 3 is the principle diagram of calculating the thumbnail pattern based on
the adjacent equally spaced sampling method, where the dimension of the data
pattern is 7 7 as shown in Fig. 3a and the dimension of thumbnail pattern is 4 4
as shown in Fig. 3c. The nodes in the thumbnail pattern are arranged with equal
spaces in the data pattern. The probability of similar patterns classified as the same
kind of data pattern after dimension reduction processing is much larger than the
dissimilar data patterns.
The data pattern corresponding to the same thumbnail pattern was deemed as the
same type of pattern (pattern cluster, referred as PatClusters) as it was able to
cluster the pattern database after reducing the dimension of the data pattern.
Figure 4a shows the PatClusters which contain six data patterns, and Fig. 4b is
the thumbnail pattern corresponding to data pattern. Pattern database is clustered
into a lot of PatClusters after the dimension reduction processing, and they are
collectively called pattern clusters.
Only when data pattern in PatClusters has a link between data event and PatCluster
can it participate in the actual modeling, and that link is called represent pattern.
The represent pattern has the same dimensions as the data pattern. This paper
computed represent pattern based on the ensemble average method. The formula
calculates nodes of represent pattern as follows:
X
n
RepPatði; j; kÞ ¼ ðPatin PatCluster ði; j; kÞÞ=n
1
where n is the number of data patterns in pattern cluster, where the value of nodes in
represent pattern is equal to the average value of nodes in the entire data pattern.
Figure 4 shows the represent pattern of pattern cluster (4c). Each pattern cluster has
a unique represent pattern. The difference between represent pattern and data
pattern is that the value of represent pattern is not binary but continuous.
714 Y. Siyu et al.
Fig. 3 Principle sketch of the procedure for scaling down pattern by space sampling method. (a)
Pattern based on template, (b) process of reverse transformation, and (c) thumbnail pattern of
pattern
Fig. 4 The relationship of pattern cluster, thumbnail pattern, and represent pattern. (a) Pattern
cluster, (b) thumbnail pattern, and (c) represent pattern
[b]
Fig. 5 Flow charts of SIMPAT and PSCSIM. (a) Flowchart of SIMPAT and (b) flowchart of
PSCSIM
from the pattern clusters and also find the data pattern which is most similar to the
data event from the most similar pattern cluster.
The difference between PSCSIM and SIMPAT is the introduction of pattern
clusters and the calculation of similarity based on pattern clusters. The flowchart of
SIMPAT and PSCSIM is shown in Fig. 5.
Process of PSCSIM
1. In inputting the training image, define the size of simulate realizations and data
patterns.
2. Using data pattern, scan the training image and construct pattern database.
3. Clustering all data pattern in the database based on the pattern scale-down
cluster strategy, the data pattern corresponding to the same thumbnail pattern is
deemed as the same type of pattern, and construct pattern clusters.
4. Build a simulated realization-based random path.
5. Mesh nodes randomly access the nodes on the path. If there are no simulation
nodes, go to step 6; otherwise, proceed to step 10.
6. Take the data template as a unit, scanning the simulate realizations to obtain the
data event at the current simulation node.
7. The first similarity comparison: compare the similarity of represent pattern in
data events and PatClusters and identify data patterns most similar with the data
events.
8. The second similarity comparison: compare the similarity of data event and all
data patterns in pattern clusters and identify the data pattern most similar to the
data event.
9. Cover part of the estimated regions with the most similar data pattern and
freeze these areas. Then, return to step 5.
10. Complete the simulation. Output the realizations.
716 Y. Siyu et al.
5 Case Study
Taking two binary images as the training image, compare the efficiency of
PSCSIM, SIMPAT, Snesim, Filtersim, DisPat, and other mainstream multipoint
geostatistics algorithms. The first training image is a two-dimensional fluvial digital
model (Fig. 6a) with a 250 250 dimension and a 1 1 m grid cell. The second
training image is a three-dimensional fluvial digital model (Honarkhah 2011)
(Fig. 6d) with a dimension of 69 69 39 and a 1 1 1 m grid cell. Using
SIMPAT and PSCSIM in conducting the non-conditional simulation, the simulation
results are shown in Fig. 6b–c (black is the channel phase, whereas white is the
background phase) and Fig. 6e–f (red is the channel phase, whereas blue is the
hollow background phase). PSCSIM showed better internal priori geological struc-
ture characteristics of training image through the comparison of SIMPAT and
PSCSIM.
The CPU and memory of the testing hardware are 2.0 GHz and 8 GB, respec-
tively. The time consumed by PSCSIM, SIMPAT, Snesim, and Filtersim algorithm
generated 100 realizations as follows:
1. Taking two-dimensional model as training image, grid dimension of the simu-
lated realizations is 250 250, dimension of data template is 9 9, and dimen-
sion of thumbnail pattern is 5 5 using three-grid simulation. Computation time
of SIMPAT is 5245 s, and PSCSIM algorithm is 204 s. The computational
efficiency of PSCSIM algorithm improved by 25 times compared to SIMPAT
algorithm.
2. Taking three-dimensional model as training images, the grid dimensions of the
simulated realizations are 69 69 39, dimensions of data template are
11 11 5, and dimensions of thumbnail pattern are 5 5 3 using the
two-grid simulation. The calculation time of SIMPAT is 96,588 s and PSCSIM
algorithm is 1498 s. The computational efficiency of PSCSIM algorithm
increased by 64-folds compared to SIMPAT algorithm (Table 1).
An MPS Algorithm Based on Pattern Scale-Down Cluster 717
Fig. 6 Training images and realizations in two dimensions and three dimensions. (a) Largetrain,
(b) realization of PSCSIM for Largetrain, (c) realization of SIMPAT for Largetrain, (d) fluvsim,
(e) realization of PSCSIM for fluvsim, and (f) realization of SIMPAT for fluvsim
6 Conclusions
1. This paper found the efficiency bottleneck of SIMPAT through the analysis of
basic principles of multipoint geostatistics modeling algorithm, which includes
the process of similarity calculation by data event and data pattern, putting
forward the concept of data pattern dimension reduction and introducing the
adjacent equally spaced sampling method to geostatistics. Using adjacent
equally spaced resampling method to reduce the dimension of data patterns
reserved multipoint statistical information of space while effectively reducing
the data dimension.
2. A huge number of data patterns are clustered to establish a database library based
on the data pattern of which dimension is reduced. This method used the
ensemble average method of an E-type to calculate the representative data
pattern. With the proposed PSCSIM algorithm, the algorithm improved one
An MPS Algorithm Based on Pattern Scale-Down Cluster 719
References
Abdollahifard MJ, Faez K (2013) Fast direct sampling for multiple-point stochastic simulation
[J]. Arab J Geosci
Arpat GB (2005) Sequential simulation with patterns. Stanford University
Boisvert JB, Pyrcz MJ, Deutsch CV (2007) Multiple-point statistics for training image selection.
Nat Resour Res 16(4):313–321
Caers J (2001) Geostatistical reservoir modelling using statistical pattern recognition. J Pet Sci
Eng 29(3-4):177–188
Comunian A, Renard P, Straubhaar J (2012) 3D multiple-point statistics simulation using 2D
training images. Comput Geosci 40:49–65
Honarkhah M (2011) Stochastic simulation of patterns using distance-based pattern modeling.
Stanford University
Honarkhah M, Caers J (2012) Direct pattern-based simulation of non-stationary geostatistical
models. Math Geosci 44(6):651–672
Huang T, Li X, Zhang T, Lu D-T (2013) GPU-accelerated direct sampling method for multiple-
point statistical simulation. Comput Geosci 57:13–23
Mariethoz G, Renord P, Straubhaar J (2010) The direct sampling method to perform multiple-point
geostatistical simulation. Water Resour Res 46(11)
Ortiz JM, Emery X (2005) Integrating multiple-point statistics into sequential simulation algo-
rithms. Quant Geol Geostat 14:969–978
Straubhaar J, Renard P (2011) An improved parallel multiple-point algorithm using a list
approach. Math Geosci 43(3):305–328
Straubhaar J, Walgenwit A, Renard P (2013) Parallel multiple-point statistics algorithm based on
list and tree structures. Math Geosci 45(2):131–147
720 Y. Siyu et al.
1 Introduction
Most reservoir modeling projects involve updating current models with new infor-
mation (e.g., logs from new wells, new seismic processing, or early production
data). Models are typically rebuilt from scratch, with very little quantitative effort
to check the consistency of the new data with the current models and to estimate the
impact of those same new data on the project forecasts (e.g., oil-in-place or ultimate
recovery). Yet, if the new data are consistent with the current models, i.e., if the new
data values could have been predicted by some of the current models, there may be
no need to rebuild models; a relationship between new data measurements
predicted from current models and corresponding forecasts could be developed
and used to update current forecasts with the actual new data measurements. This
would not only save a considerable amount of time, allowing rapid reservoir
management decisions in response to new information, but it would also reduce
the risk of irrational fluctuations of the model forecast uncertainty range due to
successive subjective reinterpretations of the data, arbitrary changes in modeling
decisions, and/or introduction of new modeling technologies.
The proposed approach identified as “direct forecast updating” is quite similar to
“direct forecasting,” a new reservoir modeling methodology that aims at making
reservoir forecasts by integrating data without performing any complex condition-
ing or inversion (Scheidt et al. 2015a; Satija and Caers 2015); the main difference is
that prior models, which are fully unconstrained in direct forecasting, are replaced
with reservoir models constrained by previously collected data. One particular
focus of this paper is the direct updating of reservoir forecasts using new well
data or, to be more specific, using statistical measures computed from those new
well data, for example, well net-to-gross or well hydrocarbon pore column. When
building models to make global forecasts, modelers only assess and model global
uncertainties, such as reservoir facies proportions or porosity and permeability
distributions. However, to be able to build a relationship between new well data
predictions and global forecasts, local variability at the new well locations also
needs to be captured in the current models. That local variability is derived from
both geostatistical simulation stochasticity (seed number) and local input modeling
parameter uncertainties, for example, local petrophysical property trends. In this
paper a new methodology is proposed to account for such local variability when
updating current forecasts directly with new well data.
Integrating New Data in Reservoir Forecasting Without Building New Models 723
2 Methodology
To explain the methodology proposed in this paper, the following simple case study
is considered: the original oil-in-place (OOIP) forecasts of a reservoir need to be
updated after a new well was drilled and an average net-to-gross value NTG ¼
NTGm was estimated from the logs at the new well location. Statistically speaking,
we want to compute PfOOIPjNTG ¼ NTGm g, which can be rewritten using Bayes’
formulation as:
Let θ be the set of input modeling parameters representing the major global
geological uncertainties identified in the reservoir, for example, the global reservoir
rock volume or the reservoir porosity distribution.
The numerator and denominator of Eq. 1 can be rewritten using integrals over
the whole input modeling parameter uncertainty space:
Z
PfOOIP and NTG ¼ NTGm θgdθ
PfOOIPNTG ¼ NTGm g ¼ Z
ð2Þ
PfNTG ¼ NTGm θgdθ
well as the correlation coefficient between NTG and OOIP, can be modeled using a
design of experiments built for the global input modeling parameters space θ:
1. For each of the N experimental design runs, which correspond to a specific
combination θi ði ¼ 1 . . . N Þ of global input modeling parameters, generate
L stochastic realizations.
2. For each of the L realizations, compute the OOIP and NTG value at the new well
location.
3. Calculate the means and standard deviations of the L values of NTG and OOIP,
as well as the correlation coefficient between the NTG and OOIP values.
4. Using the N experimental design runs, model a response surface for the NTG and
OOIP means and standard deviations, as well as the correlation coefficient
between NTG and OOIP.
Using the previous response surfaces, under the multi-Gaussian assumption,
PfOOIP and NTG ¼ NTGm j θi g and PfNTG ¼ NTGm j θi g can be computed for a
very large number n of combinations θi randomly drawn from Monte Carlo
simulation; this provides a new updated OOIP probability distribution according
to Eq. 3.
The proposed approach has several advantages:
• First, the multi-Gaussian assumption, combined with the use of response sur-
faces to estimate the parameters of the multi-Gaussian model for any combina-
tion θi of input parameters, allows fully determining the bivariate distribution P
{OOIP and NTG}; there is no need to use any arbitrary interpolation technique
such as the traditional Kernel smoothing (Park et al. 2013; Scheidt et al. 2015b).
• Then, the exact NTGm value can be directly plugged into the multi-Gaussian
function PfNTG ¼ NTGm j θi g for any combination θi of input parameters; there
is no need to determine a quite arbitrary bandwidth around the new data
measurements (Scheidt et al. 2015a).
• PfNTG ¼ NTGm j θi g provides the probability that the NTG value measured at
the new well location will be observed for the specific combination θi of global
input modeling parameters. Typically, one would expect the correlation between
NTG and OOIP over multiple stochastic realizations to be quite low. If this is
indeed the case, i.e., if NTG and OOIP are conditionally independent, which can
be tested, Eq. 3 can be rewritten as:
X
1
PfOOIP θi gPfNTG ¼ NTGm θi g
PfOOIP NTG ¼ NTGm g ¼ n
X ð4Þ
n
1
PfNTG ¼ NTGm θi g
In that new Eq. 4, the updated OOIP forecasts can be interpreted as the linear
combination of the OOIP forecasts corresponding to each possible combination θi
of input parameters weighted by the probabilities that the NTG value at the new
well location be observed in the stochastic model realizations generated for θi.
Integrating New Data in Reservoir Forecasting Without Building New Models 725
Fig. 1 Facies proportion curve and three horizontal sections of the reference model
Table 1 Global input modeling parameters, uncertainty ranges, and reference model
Input modeling parameters P10 P50 P90 Reference
Reservoir sand proportion 20 % 30 % 40 % 25 %
Sand geobody shapes Ellipses Variogram-based Channels Ellipses
Vertical trend None Medium High Medium
Horizontal trend None Medium High None
Porosity average 15 % 17 % 19 % 19 %
Permeability average 10 md 50 md 250 md 50 md
Poro/perm correlation 0.5 0.7 0.9 0.9
the actual NTG value observed at the second new well location corresponds to the
64th percentile. Therefore, in both cases, the new NTG measurements can be
considered as consistent with the existing models, and the previously described
forecast updating process can be applied.
Figure 5 provides a bubble graph displaying the bivariate distribution P
{OOIP and NTG} resulting from Eq. 3 for the first new well location.
Table 2 provides the P10, P50, and P90 OOIP values for the two alternative new
well locations.
For the first well, the P50 value of the updated forecasts is closer to the reference
OOIP value (80 M bbl.), while the uncertainty forecast range significantly
decreased: the new P10-P90 difference is 46 M bbl. vs. 67 M bbl. initially. In
contrast, as expected, because the second well is very close to an existing well, its
Integrating New Data in Reservoir Forecasting Without Building New Models 727
Fig. 2 Initial forecasts and horizontal sections of two models generated using the D-optimal
experimental design
Fig. 3 NTG map from reference model, with locations of the three existing wells (black dots) and
two alternative new wells (white dots)
Fig. 4 Histograms of the NTG predictions for both new well locations. The red line corresponds
to the actual observed NTG value
728 S. Strebelle et al.
Fig. 5 Bivariate distribution P{OOIP and NTG} for the first new well location. Each bubble
corresponds to a particular combination θi of global input modeling parameters; it is centered at the
estimated NTG and OOIP mean values, and its size is proportional to the estimated NTG standard
deviation (only 200 bubbles are displayed). The red line corresponds to the actual NTG value
observed at that first new well location
impact on OOIP forecasts is extremely limited; thus the P10, P50, and P90 values of
the updated forecasts (66, 97, and 125 M bbl.) are very close to the initial forecasts
(61, 93, and 128 M bbl.).
The same methodology based on Eq. 3 can be applied to the case where both
wells 1 and 2 are drilled. This requires the computation of an additional response
surface: the correlation between NTG values at wells 1 and 2 for any combination θi
of global input parameters. Figure 6 provides the scatterplot of predicted NTG
values at well location 1 versus predicted NTG values at well location 2 for 10,000
combinations θi of input modeling parameters. The actual NTG values (0.04 for the
first well and 0.2 for the second well) are in the predicted ranges. Thus the
combination of the two new NTG measurements can be considered as consistent
with the existing models, which confirms that the previously described forecast
updating process can be applied. Note that several methods exist to quantitatively
check that consistency between new observed values and predictions, in particular
the Mahalanobis distance (Mahalanobis 1936).
When both new well locations are used, the P10, P50, and P90 values of the
updated forecasts are 65, 86, and 105 M bbl. The new P10-P90 difference is 40 M
Integrating New Data in Reservoir Forecasting Without Building New Models 729
Fig. 6 Scatterplot of predicted NTG values at well location 1 versus predicted NTG values at well
location 2 for 10,000 combinations θi of input modeling parameters. The red dot corresponds to the
actual NTG values observed at the new well location (0.04 for well location 1 and 0.2 for well
location 2)
bbl., which is, as expected but not guaranteed, smaller than the P10-P90 difference
obtained for each well considered individually.
4 Discussion
The results obtained in the case study above show that reservoir forecasts can be
directly updated in the presence of new information without rebuilding any models
provided that the new information is consistent with the existing reservoir models.
It should be noted that there is no guarantee for the P50 value of the updated
forecasts to be closer to the true reservoir value or for the updated uncertainty range
to systematically decrease; it all depends on the new data measured value. How-
ever, getting more accurate and precise forecasts is expected on average as the
number of additional new wells increases.
In most cases, the previous methodology can be simplified by replacing some
response surface with constant values or straightforward functions. For example, in
the previous case study, it can be observed that, as expected, OOIP varies very little
across multiple stochastic realizations for any particular combination θi of global
input parameters. On average over the 99 experimental design runs, the coefficient
of variation (ratio between standard deviation and mean) is only 0.004. This means
that only the response surface for the OOIP mean need be modeled; the OOIP
standard deviation could be directly estimated by multiplying the OOIP mean by
0.004. This simplification provides updated P10, P50, and P90 OOIP values very
close (less than 0.5 % relative difference) to the updated forecasts obtained using a
full response surface for the OOIP standard deviation. Ignoring completely the
730 S. Strebelle et al.
5 Conclusions
Bibliography
Atkinson A, Donev A, Tobias R (2007) Optimum experimental designs, with SAS. UOP Oxford,
Oxford
Mahalanobis P (1936) On the generalised distance in statistics. Proc Natl Inst Sci India 2(1):49–55
Mecklin C, Mundfrom D (2005) A monte carlo comparison of the type I and type II error rates of
tests of multivariate normality. J Stat Comput Simul 75(2):93–107
Park H, Scheidt C, Fenwick DH et al (2013) History matching and uncertainty quantification of
facies models with multiple geological interpretations. Comput Geosci 17:609–621
Satija A, Caers J (2015) Direct forecasting of subsurface flow response from non-linear dynamic
data by linear least-squares in canonical functional principal component space. Advances in
Water Resources 77:69–81
Scheidt C, Renard P, Caers J (2015a) Prediction-focused subsurface modeling: investigating the
need for accuracy in flow-based inverse modeling. Math Geosci 47(2):173–191
Scheidt C, Tahmasebi P, Pontiggia M et al (2015b) Updating joint uncertainty in trend and
depositional scenario for reservoir exploration and early appraisal. Comput Geosci 19:805–820
Strebelle S (2002) Conditional simulation of complex geological structures using multiple-point
statistics. Math Geol 34(1):1–21
Zagayevskiy Y, Deutsch CV (2013) Impact map for assessment of new delineation well locations.
JCPT 2013:441–462
Statistical Scale-Up of Dispersive Transport
in Heterogeneous Reservoir
Abstract Numerical methods are often used to simulate and analyze flow and
transport in heterogeneous reservoirs. However, they are limited by computational
restrictions including small time steps and fine grid size to avoid numerical disper-
sion. The ability to perform efficient coarse-scale simulations that capture the
uncertainties in reservoir attributes and transport parameters introduced by scale-
up remains challenging. A novel method is formulated to properly represent
sub-grid variability in coarse-scale models. First, multiple sub-grid realizations
depicting detailed fine-scale heterogeneities and of the same physical sizes as the
transport modeling grid block are subjected to random walk particle tracking
(RWPT) simulation, which is not prone to numerical dispersion. To capture addi-
tional unresolved heterogeneities occurring below even the fine scale, the transition
time is sampled stochastically in a fashion similar to the continuous time random
walk (CTRW) formulation. Coarse-scale effective dispersivities and transition time
are estimated by matching the corresponding effluent history for each realization
with an equivalent medium consisting of averaged homogeneous rock properties.
Probability distributions of scale-up effective parameters conditional to particular
averaged rock properties are established by aggregating results from all realiza-
tions. Next, to scale-up porosity and permeability, volume variance at the transport
modeling scale is computed corresponding to a given spatial correlation model;
numerous sets of “conditioning data” are sampled from probability distributions
whose mean is the block average of the actual measured values and the variance is
the variance of block mean. Multiple realizations at the transport modeling scale are
subsequently constructed via stochastic simulations. The method is applied to
model the tracer injection process. Results obtained from coarse-scale models
where properties are populated with the proposed approach are in good agreement
with those obtained from detailed fine-scale models. With the advances in
V. Vishal (*)
School of Mining and Petroleum Engineering, University of Alberta, 7-207 Donadeo
Innovation Centre for Engineering, Edmonton, AB T6G 1H9, Canada
e-mail: [email protected]
J.Y. Leung
School of Mining and Petroleum Engineering, University of Alberta, 6-285 Donadeo
Innovation Centre for Engineering, Edmonton, AB T6G 1H9, Canada
e-mail: [email protected]
1 Introduction
Accurate predictions of flow and transport in natural porous media are crucial in the
management of valuable subsurface resources including water aquifers and hydro-
carbons reservoirs. These predictions are usually assessed with uncertainty due to
(1) underlying heterogeneity or spatial variation in rock and transport properties at
multiple scales and (2) conditioning data. This uncertainty, though can be reduced
by improved geophysical (e.g., seismic), core, well log, pressure test, and tracer test
data, cannot be entirely eliminated. Reservoir models are typically constructed and
subjected to flow and transport simulation to capture the aforementioned uncer-
tainties. Although fine-scale models could capture detailed description of the
heterogeneity, simulation with these models can be computationally demanding.
A commonly adopted alternative is to replace these fine-scale models with lower-
resolution coarse-scale (scale-up) models. During this process of coarsening, a
number of transport properties (e.g., dispersivity), along with reservoir properties
(e.g., porosity and absolute permeability), must be scaled up accordingly. Properly
scale-up models should not only honor the conditioning data, but they should also
account for the associated subscale heterogeneities.
Mass transfer mechanisms in single-phase flow involve: (1) advection or con-
vection, (2) diffusion, and (3) mechanical dispersion. These phenomena are gener-
ally described by the parabolic advection-dispersion equation (ADE). Common
solution techniques, such as finite volume and finite element, suffer various com-
putational restrictions including numerical dispersion; poor predictions are
observed if flow is advection-dominated (high Péclet number). An alternative
solution framework is the classical random walk particle tracking (RWPT), which
is free from numerical dispersion and imposes no restriction on grid size.
Another conceptual issue associated with the ADE is its inability to capture
non-Fickian transport behavior. ADE considers dispersion as a sum of diffusion and
convective spreading, ignoring any additional mixing introduced by the interaction
of these two mechanisms. It neglects heterogeneity that is not resolved completely
at the volume support on which the ADE and its associated parameters are defined.
It is true that if heterogeneities at all scales are modeled explicitly, the ADE reflects
non-Fickian behavior of transport (Salamon et al. 2007; John 2008; Rhodes et al.
2009; Li et al. 2011).
Statistical Scale-Up of Dispersive Transport in Heterogeneous Reservoir 735
∂cðX; tÞ
¼ ∇ ðD∇cðX; tÞÞ ∇ ðVcðX; tÞÞ: ð1Þ
∂t
∇V¼0: ð2Þ
k
V ¼ ∇p: ð3Þ
μ
k is the permeability tensor; p and μ are the pressure and viscosity of the fluid,
respectively. Equation 1 can be solved using the particle-tracking method, which
simulates mass transport by discretizing injected mass concentration into a number
of particles (walkers) of equal mass; over a given time step, the movement of each
particle is controlled by advection and dispersion/diffusion. The particle tracking
scheme is summarized by Eq. 4 (Kinzelbach 1986; Tompson and Gelhar 1990;
LaBolle et al. 1996; Delay et al. 2005; Salamon et al. 2006):
pffiffiffiffiffi
Xp ðt þ ΔtÞ ¼ Xp ðtÞ þ A Xp ðtÞ Δt þ B Xp ðtÞ :ξðtÞ Δt:
ð4Þ
A ¼ V þ ∇D, BBT ¼ 2D:
variance. If the transition time is considered to be deterministic and constant for all
particles, the above formulation reduces to the classical RWPT framework and
converges to the ADE. To capture non-Fickian features due to unresolved hetero-
geneities occurring below the fine scale, Srinivasan et al. (2010) proposed a
different particle-tracking model, in which the transition time is considered to be
stochastic and vary among particles:
pffiffiffiffiffiffi
Xp ðτ þ ΔτÞ ¼ Xp ðτÞ þ A Xp ðτÞ Δτ þ B Xp ðτÞ ξðτÞ Δτ:
ð5Þ
tðτ þ ΔτÞ ¼ tðτÞ þ ηðτ, ΔτÞ:
(ξ, η) is a set of random series that is drawn from the joint probability density
function of transition length ξ and time η. In this formulation, time steps are
incorporated implicitly and sampled from time distribution density function.
This formulation of the RWPT with stochastic transition time is equivalent to the
commonly adopted CTRW framework for modeling non-Fickian transport. The
governing equation in CTRW is the Fokker-Planck with memory equation, Eq. 6,
which is derived from the equivalence of the classical generalized master equation
and CTRW (Berkowitz et al. 2002). Solute particle migration is described as a
series of jumps over different transition times. The jump position (X) and transition
times (t) follow a joint probability density function ψ(X, t) that describes the jump
at position and time t:
se e ðsÞðV ∇e
c ðX; sÞ co ðXÞ ¼ M c ðX; sÞ ∇ ðD∇e
c ðX; sÞÞÞ: ð6Þ
where ec denotes the Laplace transform of c and co(X) is the initial condition. The
particle transport velocity V and the dispersion tensor D are defined by the first and
second moments of the pdf of the transition length, p(X):
Z
1 1X
vi ¼ pðXÞXi dX ¼ Xi pðXÞ:
t1 ℜ t1 X
Z ð7Þ
1 1X
Dij ¼ pðXÞXi Xj dX ¼ Xi Xj pðXÞ:
2t1 ℜ 2t1 X
e ðsÞ that
t1 is the median transition time or lower cutoff time. The memory function M
accounts for the unknown heterogeneities below the model resolution is defined as
e e ðsÞ
st1 φ
MðsÞ ¼ : ð8Þ
1φ e ðsÞ
ψe ðX, sÞ ¼ pðXÞe
φ ðsÞ: ð9Þ
Statistical Scale-Up of Dispersive Transport in Heterogeneous Reservoir 737
Initial conditions:
cðX1 , t ¼ 0Þ ¼ 0: ð13Þ
The corresponding solution in the Laplace space for the flux-average concen-
tration (cf) is determined by
" (sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi )#
vX1 sDL
e
c f ðX1 ; sÞ ¼ co exp 1þ4 1 : ð14Þ
2DL Me ðsÞv2
3 Method
Apart from the volume variance, the spatial correlation length also changes with
volume support. Averaged semi-variogram γ is estimated numerically by averaging
the point-scale variogram γ on two volume supports V and V0 (Journel and
Huijbregts 1978):
Z Z 0 0
0 1 0 1 Xn X n
γ V; V ¼ 0 γ v; v dvdv 00 γ hij : ð16Þ
VV nn i¼1 j¼1
V V0
A procedure, which was described in Leung and Srinivasan (2011) and was also
implemented in Vishal and Leung (2015), is adopted to construct coarse-scale
models of reservoir properties that capture the subscale variability:
1. Compute the variance of mean (VarðZÞ) according to Eq. 15.
2. Estimate averaged variogram γ according to Eq. 16.
3. Draw multiple sets of conditioning data via parametric bootstrapping of a
Gaussian likelihood function, whose mean and variance are the block average
of the fine-scale conditioning data and VarðZÞ, respectively.
4. Perform conditional simulation at the coarse scale (i.e., transport modeling
scale) using histogram (mean ¼ fine-scale global mean; variance ¼ VarðZÞ)
and γ (which accounts for the spatial correlation) for a particular set of
conditioning data from step #3.
5. Repeat step #4 for the remaining sets of conditioning data from step #3.
In this work, porosity (ϕ) is modeled according to above mentioned steps, while
absolute permeability (k) is assumed to be correlated with porosity as k ¼ a ϕb,
740 V. Vishal and J.Y. Leung
where a and b are empirical constants. This assumption would imply that linearly
averaging is also valid in the transformed space of (k/a)1/b; therefore k is estimated
from ϕ using the same empirical relation, assuming the transform relationship is
invariant with scale (Leung and Srinivasan 2011). In principle, permeability can be
scaled up in the same fashion as for the transport parameters, as explained in the
next section, particularly if a simple correlation with porosity is unattainable.
4 Case Study
A B
Fine-scale porosity distribution Coarse-scale porosity distribution
500 500
0.35 0.35
0.30
0.25
0.25
Y-direction
Y-direction
0.20
0.15
0.15
0.10
0.05
0.05 0.00
0 0
0 500
0 X-direction 500 X-direction
Fig. 3 Distribution of porosity: (A) true fine-scale model and (B) a randomly selected realization
of the coarse-scale model
742 V. Vishal and J.Y. Leung
Fig. 4 Normalized cumulative mass flux profile for: (A) Fickian model and (B) non-Fickian
model
profiles for all 100 realizations are compared to the true fine-scale model in Fig. 4B.
Also shown in Fig. 4A is a case where there is no unresolved heterogeneity
occurring below the fine scale; constant transition time is assumed; therefore, αL
and αT , instead of β*, are modeled (Vishal and Leung 2015). It is clear that
significant non-Fickian behavior is observed when there is unresolved heterogene-
ity. In both cases, the true fine-scale response is captured by the variability
exhibited among all 100 coarse-scale realizations.
Bibliography
Abstract This paper discusses methods for interpolating the well data that
describes a hydrocarbon reservoir. The main difficulty lies in the requirement that
the interpolation result should correspond not only to borehole data but also to some
additional criteria, such as variograms and histograms. In a broader sense, it should
reflect our knowledge of the geological environment. The authors are interested in
what capabilities the different stochastic methods, sequential Gaussian simulation
(SGS), multiple-point statistics (MPS) and fuzzy model simulation, can provide.
1 Introduction
Let us have a look at the problem in more detail. An interpolation of well data
requires taking into account the following:
1. A geological environment is of a categorical nature, and it is thus impossible for
the quantitative properties to be interpolated among different categories. The
quality of property interpolation in such an environment is controlled by com-
paring the histograms of initial well data with the histograms computed for
values in the interwell space.
2. The interpolated values should represent the actual variability of a geological
environment, where the term “actual” means meeting a certain criterion. But in
all cases, the variability of interpolated values should not depend on the spacing
of well data points. This requirement can be achieved only by means of multiple
stochastic realizations. Where a variogram is used as a mentioned criterion, the
quality of interpolation is controlled by comparing the variograms of initial well
data with the variograms computed for values in the interwell space.
3. It is geostatistical techniques that are mostly applied to compute stochastic
realizations. All of these methods are based on the assumption of stationarity,
A part of an existing large field in Western Siberia is shown below in Fig. 1. The
shown part was used as a trial field.
Figure 1 shows not only a trial field but also a three-dimensional model calcu-
lated based on the well data. It includes a geometrical framework (as a pile of
stratigraphic surfaces) resulting from the correlation of the well log data and the
subsequent interpolation of the stratigraphic marks. The frame is used as a base for
constructing a stratigraphic grid of 100 100 170 (170 cells on Z). The model
also includes an ASP cube, which is the result of a deterministic quasi 3D (layer-by-
layer) interpolation of the ASP well-logging curves calculated with a 1/R2
weighting. The stratigraphic surfaces and the ASP cube are shown in the paleo-
reconstruction where the surfaces are transformed to the horizontal planes. The
initial ASP curves are presented on the paths of the wells in the window for the
A Comparative Analysis of Geostatistical Methods for a Field with a Large. . . 747
Fig. 1 Deterministic interpolation of the ASP well data. The red and green colours approximately
demonstrate sands and clays, respectively. On the top right is the stratigraphic frame of the model.
In the figures presented below, the colour legend is the same
vertical slice. Both (vertical and horizontal) slices show the intersection line.
Hereinafter, the stratigraphic grid of the model will remain unchanged. What will
change are the ASP cube calculation methods.
The ASP cube, as illustrated in Fig. 1, is considered as a starting point of our
investigation. Why only as a starting point? Because the quality of such an
interpolation is very low. First, it does not take into account the categorical nature
of the environment, being therefore incapable of reproducing a well data histogram.
The biggest error is seen for the AV4 horizon (see Fig. 2, right).
Second, this interpolation type incorrectly displays the variability of the envi-
ronment. If the range of the horizontal variogram for borehole data is about
1,000 m, its range for the interpolated values is about 2,000 m (see Fig. 3, left).
The horizontal variability of the deterministic ASP cube depends on the well
spacing. At the same time, the vertical variogram is displayed well (see Fig. 3,
right). The latter is the result of a layered interpolation. The other differences in the
variograms are not of great importance. The difference in the sills of the variograms
for the borehole and cube data follows from the difference in their histograms and
748 M. Volkova et al.
Fig. 2 Deterministic interpolation, histograms of the ASP values (with the red curves calculated
based on the initial well data; and with the grey curves, on the interpolated values; and with the
data for the AV1–AV5 horizons shown on the left and separately for the AV4 horizon on the right)
Fig. 3 Deterministic interpolation, variograms for the ASP values (the red and grey curves
representing the calculated initial well data and the interpolated values, respectively). Shown on
the left are the horizontal variograms, and on the right are the vertical variograms
the difference in the sills of the horizontal and vertical variograms, from a greater
variability of the environment in the vertical direction.
Based on Figs. 2 and 3, we can conclude that the deterministic interpolation (see
Fig. 1) gives a false view of the environment. However, the cube shown in Fig. 1
has one advantage: it excellently displays the deterministic features contained in the
initial well data. It is immediately seen that the AV1 horizon is a heterogeneous
reservoir of a poor quality, that there is a massive buried channel in the AV3
horizon and that the AV5 horizon is a solid reservoir.
A Comparative Analysis of Geostatistical Methods for a Field with a Large. . . 749
Fig. 4 Interpolation of the ASP well data by the SGS method, using the normal score
(NS) transformation (one realization). A horizontal section within the AV1 horizon is shown on
the top right
750 M. Volkova et al.
Fig. 5 Interpolation of the ASP well data by the SGS method with NS transformation. Histograms
and variograms of the ASP values (with the red curves indicating values calculated based on the
initial well data and the blue curves on the interpolated values). In the centre are the horizontal
variograms, and the vertical variograms are on the right
However, it should be kept in mind that the NS method allows fitting of the
histograms of initial and interpolated values only for the entire cube. To prove that
this method can lead to errors, it is quite enough to look at the “local” results it can
deliver. Figure 6 displays the histograms for the well data and for the interpolated
values for the AV1 horizon. What we see is that they have essential differences.
Figure 4 (top right) shows the horizontal section of the ASP cube within the AV1
horizon. It is noteworthy that all the red and green “spots” (high and low ASP
values) are located exactly between the wells. The AV1 horizon well data has no
such values.
Of course, the above problem of the vertical nonstationarity has been well
addressed. Each layer with different statistics (e.g. AV1) should be calculated
separately as well as each zone with different statistics on the XY plane. Let’s
A Comparative Analysis of Geostatistical Methods for a Field with a Large. . . 751
have a look at the AV3 horizon once more (Fig. 4 top left). The presence of the
reservoir zones outside the channel predicted by NS is very unlikely. These
problems were mentioned in the Introduction section.
The second of the three techniques used to represent the actual variability of the
environment is multiple-point statistics (Strebelle 2002). Appropriate tools have
been developed and tested for this use (Volkova 2015). We have made such a
calculation as well. We used a fragment of the deterministic model shown above in
Fig. 1 as a training image (TI). It is a small cube of 13 % of the total field with the
highest density of wells. The quality of the TI can be argued, but it is the best TI we
can obtain. The calculation was carried out using the direct sampling
(DS) algorithm (Mariethoz et al. 2010). Parameters of the DS algorithm are as
follows: the number of neighbours, 40; their weights, 1/R2, where R is the distance
in the 3D space measured in the cells; the threshold value, 0.10; and the maximum
fraction for TI scanning, 1 %. The result of the interpolation (as well as the contour
of the fragment taken as the TI) is illustrated in Fig. 7, and the histograms and
variograms are presented in Fig. 8.
The direct sampling algorithm distinguishes areas having distinct statistical
properties. For example, the AV1 horizon is modelled almost correctly (see AV1
in Figs. 9 and 10). At the same time, the method leads to some errors. The well data
does not indicate that the AV5 horizon reservoir has a gap (see AV5 in Fig. 9).
The third of the above-named methods employed to represent the actual variability
of the environment is the fuzzy model (Kovalevskiy 2015). This method does not
rely on statistical generalizations. Computations are performed on a 3D strati-
graphic grid. In the most recent edition (aimed at simplification), the calculation
algorithm (for one realization) consists of the following steps:
1. We randomly choose an empty cell (as the forecast cell). Then we select an n
number of the non-empty cells nearest to that empty cell. (Those cells initially
contain borehole data.) For example, n can be equal to 6. The selection should
provide azimuthal control to ensure selection of a single nearest cell from each
azimuth sector (see Fig. 11, left). The selection of cells from the upper or lower
layer of the grid can be performed as an option.
752 M. Volkova et al.
Fig. 7 Interpolation of the ASP well data by the MPS method, using the direct sampling algorithm
(one realization). A horizontal section of the deterministic cube is added on the top right, with an
indication of the fragment taken as a training image
Fig. 8 Interpolation of the ASP well data by the MPS method, using the direct sampling
algorithm. Histograms and variograms of the ASP values (with the red curves calculated based
on the initial well data and the pink curves on the interpolated values). In the centre are the
horizontal variograms, and on the right are the vertical variograms
A Comparative Analysis of Geostatistical Methods for a Field with a Large. . . 753
Fig. 9 Interpolation of the ASP well data by the MPS method, using the direct sampling
algorithm. On the top right is the horizontal section of the AV1 horizon. On the top left is the
horizontal section of the AV5 horizon
Fig. 11 Selection of the data affecting the forecast cell (on the left). On the right is the weight of
values calculated as 1/R2. The one cell’s value is presented with its estimated accuracy
Fig. 12 On the left is the fuzzy presentation of the parameter value in the forecast cell. On the
right is the random selection from the integral distribution
2. Each selected cell is represented by two numbers: the value in the cell and its
“weight” (see Fig. 11, right). The weight is calculated as 1/R2, where R is the
distance from the cell with data to the forecast cell. Each cell value is presented
with its estimated accuracy (set by expert evaluation).
3. On the basis of the selected values and taking into account the accuracy of each
value, we compute the fuzzy value in the forecast cell (see Fig. 12, left). We do it
by a simple summation of the inputs. After that, we normalize the membership
function so that its integral is equal to 1.
4. Next, we compute the integral distribution of the membership function (which is
an analogue of integral probability distribution – see Fig. 12, right). After that,
we select a random number from the uniform distribution on the interval [0, 1].
Using the integral distribution of the membership function, we transform the
selected random number to another, also random number, but one distributed
A Comparative Analysis of Geostatistical Methods for a Field with a Large. . . 755
Fig. 13 Interpolation of the ASP well data as a realization of the fuzzy model. A horizontal
section of the AV1 horizon is added on the top right
according to the membership function. And it is this second random number that
we insert into the forecast cell and that is assigned the status of a data point. Then
we select the next forecast cell and repeat the same process. Thus, the described
algorithm is very similar to a standard sequential simulation.
The so calculated stochastic realization of the fuzzy model is illustrated in
Fig. 13 and the histograms and variograms in Figs. 14 and 15.
As regards selection of non-empty cells from the upper or lower layer of the grid,
we have not selected such cells in this calculation. However, because the initial well
data was presented as the vertical (or almost vertical) columns of cells, each of the
simulated predictions is also made for the entire vertical column, with individual
random numbers used for each forecast cell.
756 M. Volkova et al.
Fig. 14 Interpolation of the ASP well data as a realization of the fuzzy model. Histograms and
variograms of the ASP values (with the red curves calculated based on the initial well data and the
green curves on the interpolated values). In the centre are the horizontal variograms, and on the
right are the vertical variograms
6 Conclusions
In conclusion, it is worth paying attention to the following. As was said above, the
deterministic interpolation technique (reproducing neither a histogram nor a
variogram) has an important advantage: it excellently displays the deterministic
features of the data. Indeed, it is possible to see in the AV3 horizon very specific
meandering facies (see Fig. 16(1)). This means that the well data indicates that
meandering facies are present.
On the other hand, no meandering facies can be seen as the result of the
interpolation by the SGS method (see Fig. 16(2)), while, just to repeat, the well
data indicates their presence. The interpolated values obtained using this technique
A Comparative Analysis of Geostatistical Methods for a Field with a Large. . . 757
is “sterile” with respect to those deterministic features, which have not been
specially marked.
This is that very “sterility” which was mentioned in the Introduction section. The
histograms and variograms of the initial data are displayed well, with the caveat that
SGS can lead to errors where the local histograms differ from the overall histogram.
The result of interpolation by the MPS method (see Fig. 16(3)) displays the
meandering facies well. On the whole, it can be said that it works, with the
variograms being displayed well and the histograms being displayed satisfactorily.
758 M. Volkova et al.
It is important that MPS allows dealing with nonstationary data and a nonstationary
training image. It can be assumed that the result may be improved by improving the
training image.
The fuzzy model realization as regards showing the meandering facies is also
good (see Fig. 16(4)). It clearly shows that the facies (meandering facies, channel
facies, etc.) cannot be considered as statically uniform. What we can see is a gradual
transition from one into the other. The well data histograms and variograms are
displayed well, even locally. So it can be stated that this approach seems to have
great potential.
Bibliography
Deutsch CV, Journel AG (1998) GSLIB: geostatistical software library and user’s guide. Oxford
University Press, Oxford
Kovalevskiy EV (2015) Fuzzy geological model: stochastic realizations preserving deterministic
features of data. EAGE Conference “Petroleum Geostatistics 2015”
Mariethoz G, Renard Ph, Straubhaar J (2010) The direct sampling method to perform multiple-
point geostatistical simulations. Water Resour Res 46(W11536) doi:10.1029/2008WR007621
Strebelle S (2002) Conditional simulation of complex geological structures using multiple‐point
statistics. Math Geol 34:1–22
Volkova M (2015) Stochastic modeling on the base of multiple-point statistics with seismic
attribute maps as training images. Geophysics 6:68–73 [in Russian]
Part V
Hydro(geo)logy
Building Piezometric Maps: Contribution
of Geostatistical Tools
1 Introduction
B. Bourgine (*)
BRGM (French Geological Survey), 36009, 45060 Orléans Cedex 2, France
e-mail: [email protected]
M. Saltel • N. Pedron
BRGM, Direction Régionale Aquitaine, 33600 Pessac, France
e-mail: [email protected]; [email protected]
E. Lavie
Conseil Régional Aquitaine, 33077 Bordeaux, France
e-mail: [email protected]
Kriging hydraulic head has been used for a long time to build reference maps either
to characterize high or low water levels or to calibrate the parameters of flow
simulations (Renard and Jeannée 2008). As usually hydraulic heads are not station-
ary, kriging with a drift should be used. An example can be found in (Aboufirassi
and Marino 1983), where the drift depends on the coordinates. In unconfined
aquifers, the hydraulic head is generally well correlated with topography. Thus
hydraulic heads can be interpolated by co-kriging with topography (Hoeksema
1989) or by using a variable derived from topography as an external drift (Desbarats
et al. 2002). In this last case the authors compare two external drift variables: the
topography itself and another variable “TOPMODEL” which combines several
parameters such as the slope and the upslope area draining to a given point. This
second method is supposed to take into account the fact that the water table
elevation is much smoother than topography (Wolock and Price 1994) but authors
conclude that the results “are not always physically plausible.” Renard and Jeannée
(2008) use a smoothed topography as external drift and obtain smoothing by a
standard moving average algorithm. The smoothing radius is set empirically.
Boundary conditions are important constraints to take into account. The two
main ones are the Dirichlet condition (known and constant head) and the Neumann
condition (known and constant flow). The first one is easily satisfied by adding data
points with given hydraulic head along the boundary (e.g., on the banks of a lake).
The second one is generally used to take into account no flow boundary conditions.
A first application proposed by Delhomme (1979) was described by Chilès and
Delfiner (1999). This work was continued by Le Cointe (2006) during a joint work
with J.P. Delhomme as tutor. In this work, a co-kriging of heads and their deriva-
tives is implemented using a finite difference approach. At last, Kuhlman and Pardo
Building Piezometric Maps: Contribution of Geostatistical Tools 763
3.1 Methodology
where f l(x) are known drift functions that can depend on the coordinates or can be
also external drift and where al are unknown coefficients.
The residual Y(x) can be non-stationary, with a covariance σ(x, x0 ). In case of a
IRF-k, the polynomial components with degree lower or equal to k are part of m(x).
Data is denoted Za and can belong to two groups:
• The first group represents known hydraulic heads at point xα:
Z α ¼ Z ðxα Þ ð3Þ
764 B. Bourgine et al.
• The second group corresponds to the value of the component of the gradient at
point xα along a unit vector uα and is used to represent non-flow boundary
conditions or fixed gradients:
∂Z
Zα ¼ ðxα Þ ð4Þ
∂uα
In this paper, we use a finite increment formulation. The value of the component
of the gradient at point xα is represented by two dummy points xα1 and xα2 separated
by a distance 2bα along the unit vector uα (the macroscopic gradient is calculated on
length 2bα). Thus a data of the second group has the following form:
Z ðxα þ bα uα Þ Zðxα bα uα Þ
Zα ¼ ð5Þ
2bα
Kriging can be expressed in the classical way. The estimator at point x0 is:
X
Z * ðx 0 Þ ¼ λ Z
α α α
ð7Þ
Note that the Zα are of one of the two forms shown in (3) and (4). In the particular
case of no flow boundary condition, the Zα of the second group are equal to zero
(partial derivative equal to zero perpendicularly to the boundary) and don’t con-
tribute directly to the estimator in Eq. 7. But in fact they modify the weights and
consequently contribute indirectly to the estimation.
The kriging system can be written in a matrix form (Chilès and Delfiner 1999):
Σ F λ σ0
0 ¼ ð8Þ
F 0 μ f0
where Σ is the matrix of the covariances σ αβ between data points, F the matrix of
drift functions fl(xα) at data points xα, λ the vector of weights λα, μ the vector of
Lagrange parameters μα, σ 0 the vector of covariances σ α0 between data points and
estimated point, and f0 the vector of drift functions at estimated point.
The covariances σ αβ have different expressions according to the type of data. If
both α and β are indices
of points of type “hydraulic heads,” the covariance is
simply σ αβ ¼ σ xα ; xβ where σ is the covariance of Y(x). If both α and β are indices
of points of type “component of a gradient,” the covariance σ αβ can be written as
(using the finite increment formulation)
Building Piezometric Maps: Contribution of Geostatistical Tools 765
σ αβ ¼ wα wβ σ xα2 ; xβ2 σ xα1 ; xβ2 σ xα2 ; xβ1 þ σ xα1 ; xβ1 ð9Þ
The expressions above were introduced in GDM software (Geological Data Man-
agement, http://gdm.brgm.fr) and were applied in a large aquifer of the Aquitaine
Region (France): the Middle Eocene aquifer. This aquifer shows many lateral facies
variations, from continental fluviatile deposits in the eastern part to marine deposit
to the west. It covers a 17,290 km2 area around Bordeaux city (Fig. 1) and is
intensively pumped. First measurements of hydraulic heads started in 1873.
The oldest available piezometric map represents the initial state of hydraulic
head corresponding to the 1870–1900 period. The aquifer is now monitored by a
network of piezometers that has been set up since 1959.
Nowadays a reference piezometric map of the average level of the last quarter of
the year is built every year. For the year 2007, that will be illustrated here,
119 piezometers are available. The aquifer is connected to the river network in
the northern and eastern parts of the studied area. Therefore points along the river
can be added as constant heads, the hydraulic head being equal to the surface
topography at these points. The aquifer is confined in almost the whole area, except
in the few square kilometers near the outcrops where it is unconfined.
The variogram of the hydraulic head (computed from piezometers only) is
non-stationary (Fig. 2) and can be fitted, for example, by a power model. Several
kriging options were tested and are presented in (Bourgine and Pédron 2011),
including co-kriging using data from different years and kriging with an external
drift based on a hydrodynamic flow model, but we will focus here on kriging with
boundary conditions.
Since 1993, a hydrodynamic flow model has been developed in order to improve
groundwater management and to find solutions to reduce impact due to pumping
(Saltel et al. 2012). The hydraulic head obtained from this model for year 2007 is
displayed in Fig. 3. The red ellipse indicates an area where the aquifer becomes
impervious, along its south boundary, just to the north of Villeneuve sur Lot City.
This boundary condition was integrated in the flow model, and consequently
contour lines are perpendicular to this boundary.
766 B. Bourgine et al.
Fig. 1 Available data for the Middle Eocene aquifer around Bordeaux City (France)
200
150
Variogram (m2)
100
50
0
0 10000 20000 30000 40000
Distance (m)
Fig. 2 Variogram of hydraulic head for year 2007 (piezometers data only)
hydraulic head derived from the hydrodynamic flow model for year 2007. Note that
additional constant heads have been added in the ocean (green point along the
western boundary) in order to calibrate the flow model.
The same data are used again with ordinary kriging (map b): the obtained
groundwater ridge is less continuous and the map is unstable in this area. Boundary
conditions are then introduced as no flow data perpendicularly to the groundwater
ridge. The new computed map (Fig. 6 – (map c)) is now much more continuous and
satisfactory. The comparison of the southwestern part of map (a) with the same part
of maps (b) or (c) shows that the high gradient area seen on map (a) near the
southern aquifer boundary is not reproduced on maps (b) and (c). This difference is
due to a supplementary point (named 08512X0001), not measured in 2007, that was
introduced in the flow model to calibrate it. After adding it to the 2007 piezometer
data with the same hydraulic head value, we get the map (d) that is now very
satisfactory. This shows that even if introducing boundary conditions in the kriging
improves the results, a sufficient number of measurements of hydraulic head is
essential to calibrate the correct value.
Building Piezometric Maps: Contribution of Geostatistical Tools 769
As recalled in the review of previous work, kriging hydraulic heads with external
drift based on smoothed topography can be used in unconfined aquifers, but the
remaining question lies in the way to smooth the topography. We propose to smooth
the topography using a standard moving average and to set the smoothing radius
using the following method:
• Several smoothing radius are tested. In each cell of the initial topography grid,
we compute a new field equal to the initial topography averaged within the
specified radius.
• For each smoothing radius, we compute the variogram of the residual:
770 B. Bourgine et al.
Fig. 6 Use of boundary conditions given by a groundwater ridge; (a) hydraulic head from the flow
model; (b) from ordinary kriging; (c) from kriging with boundary conditions; (d) idem c with a
supplementary data
where R(x) is the residual at point x, HH(x) the measured hydraulic head, and ZS
(x) the smoothed topography at this point.
• The scatter diagram of both variables and the linear regression of hydraulic head
against the smoothed topography are also computed.
The smoothing radius is then chosen close to the value that minimizes the
variance of the regression error and that gives the most correlated variogram as
well as a nearly linear scatter diagram.
Building Piezometric Maps: Contribution of Geostatistical Tools 771
4.2 Application
This method has been applied successfully for several unconfined aquifers (Saltel
and Bourgine 2015). We report here the example of an aquifer located in the
Turonian chalk located near Perigueux City (Dordogne department, southwest of
France) (Bourgine and Pédron 2011). In this aquifer, 67 piezometers are monitored
in an area of 38 40 km. A manual contouring is available for the high waters of
year 2002 (Fig. 7). This figure shows that the manual piezometric map is much
smoother than topography but follows more or less the topography. As the ground-
waters are converging toward the La Dronne river, the contour lines were drawn
perpendicularly to the river axis.
To quantify the correlation between hydraulic head and topography (initial or
smoothed), the topography was smoothed using a moving average. The radius of
smoothing was increased from 0 (no averaging compared to initial 50 m resolution
digital elevation model) to 3,000 m. For each window size the standard deviation of
the error of the linear regression of the measured hydraulic head as a function of
smoothed topography was computed, as well as the variogram of the residual R(x)
(Fig. 8).
Figure 8 (left) shows that the standard deviation of the regression error reaches a
minimum for a window radius of 1,000 m. This radius is chosen as the optimal one
and the variogram of the residual R(x) is computed (Fig. 8 – right). It can be fitted
with an exponential model with a practical range of 5,300 m.
Figure 9 shows the linear regression between the measured hydraulic head and
(left) the initial topography (50 m resolution) or (right) the topography smoothed
with a radius of 1,000 m. The correlation is much better in the second case. In this
case it is clear that the window radius of 1,000 m gives the best results. We have
772 B. Bourgine et al.
Fig. 8 Residual standard deviation versus radius of the moving average (left) and variogram of
residual for the optimum size (right)
encountered other situations where the minimum in Fig. 8 can be chosen between
two or three window radius. In this case we compare the variograms of the residuals
and chose the radius that gives the best variogram continuity. It is also possible to
compare the linear regressions obtained for various radii and select the radius that
gives the best results. At last it is also possible to perform a cross validation for each
window radius and select the radius that gives the best results in the key areas of the
aquifer to be studied.
At last kriging with the smoothed topography as external drift was performed in
the whole aquifer, and results were compared to other kriging options: ordinary
kriging and universal kriging with polynomial drift, as well as with manual
contouring. Kriging was performed on a 500 m cell-size grid.
Table 1 shows a comparison between manual contouring and three types of
kriging. The second column gives the number of meshes in the estimated grid for
which the estimated hydraulic head is above topography. Considering that the
aquifer is not confined, this should not be observed. Results show that kriging
with external drift is far ahead from other methods and gives the best results. In
addition, we indicate in the table the mean and maximum value of the difference
between interpolated value and topography, when interpolated value is above
topography. Ordinary kriging and universal kriging give results of equivalent
quality and are not significantly better than manual contouring, whereas kriging
with external drift gives clearly the best results.
Another way to quantify the quality of the three kriging is to use a subset of data
as validation data set. For this we selected an area south of the Mareuil City where
the water table draws a dome (Fig. 10). The 19 piezometers located in the dome (red
points on the figure) were discarded, and only the other piezometers located outside
this dome (blue triangles) were used as input data for kriging.
Results of this validation test are shown Fig. 11. The upper part of the figure
shows the interpolations obtained using all data (blue triangles + red points). All the
interpolations are consistent and close to manual contouring. The lower part of the
Building Piezometric Maps: Contribution of Geostatistical Tools 773
Fig. 9 Linear regression between hydraulic head and topography (initial and smoothed)
figure shows the result obtained when only the points represented by blue triangles
on Fig. 10 are used. As expected, ordinary kriging and kriging with polynomial drift
are not able to reproduce the dome at the center of the figure. Conversely kriging
with external drift gives very satisfactory results, as it tends to mimic the shape of
the smoothed topography when interpolation is performed far from the data.
4.3 Discussion
Similar tests were performed in other areas of the aquifer, and similar results were
obtained.
As a conclusion of this example, kriging with external drift based on smoothed
topography gives consistent (and the best) results in this unconfined aquifer. The
influence of external drift becomes dominant far (compared to variogram range)
from the data, which is one of the well-known properties of this method. In areas
774 B. Bourgine et al.
Fig. 10 Test zone in the “Mareuil dome” area. Blue triangles are points kept as input for
interpolation, and red points are data used to validate the interpolation result. AB cross section
shown next figure. Coordinates in Lambert II projection, meters
with sparse data, it can be a drawback if the correlation is not linear or weaker, but
that was not observed here.
One of the limitations of this method is that the window size used to smooth the
topography is constant over the whole studied area. Moreover it does not take into
account hydrodynamic properties of the aquifer. It is likely that results could be
improved by letting the size of the smoothing window vary, as a function of rock
hydrodynamic properties, or according to other topographical of hydrogeological
information (draining valley, dry valley, plateau, terrain slope or change of slope,
etc.), but that was not investigated in this work.
5 Conclusions
This work illustrates the use of two types of kriging in order to take into account
hydrogeological constraints when building piezometric maps. These two types of
kriging (kriging with external drift based on smoothed topography and kriging with
boundary conditions) are not new but not very well known by hydrogeologists. For
both methods we obtain much better results than with ordinary kriging or universal
kriging that are often proposed “by default” in many software. For kriging with
external drift, we propose a method that facilitates the choice of the smoothing
radius when smoothing the topography. Both kriging methods were implemented in
Building Piezometric Maps: Contribution of Geostatistical Tools 775
Fig. 11 Comparison of kriging results on a cross section (a) kriging from all data, (b) kriging from
data represented by the blue triangles. OK ordinary kriging, UK universal kriging, KED kriging
with external drift
GDM software. This software, developed by the BRGM, is used for many appli-
cations (geological modeling, geochemistry, environment, hydrogeology,
geostatistical analysis). A report was also written and training organized, in order
to describe the methodology to be followed by hydrogeologists. In this way a better
standardization of procedures is obtained that facilitates the work.
Of course the expertise of the hydrogeologist remains fundamental in order to
define, for example, the appropriate boundary conditions that should be taken into
account to assess the quality of the resulting map and to introduce additional
constraints when necessary.
Bibliography
Aboufirassi M, Marino M (1983) Kriging of water levels in the Souss Aquifer, Morocco. Math
Geol 15:537–551
776 B. Bourgine et al.
Abstract Inverse modeling for subsurface flow and transport in porous media is
expected to improve the reliability of predictions in that the realizations generated
are consistent with the observations of states. A gradient-based blocking Markov
chain Monte Carlo (McMC) method is presented for stochastic inverse modeling.
The method proposed effectively takes advantage of gradient information for
tuning each realization to create a new “candidate” proposal, and hence it is capable
of improving the performance of McMC. The gradients are efficiently computed by
an adjoint method. The proposal mechanism is based on the optimization of a
random seed field (or probability field), and thus it is able to preserve the prior
model statistics. The method proposed has better performances than the single-
component McMC and also avoids directly solving a difficult large-scale
ill-conditioned optimization problem simply by turning it into a sampling proce-
dure plus a sequence of well-conditioned optimization subproblems. A synthetic
example demonstrates the method proposed.
1 Introduction
J. Fu (*)
Institute for Water and Environmental Engineering (Currently with Chevron),
Universitat Politècnica de València, Valencia, Spain
e-mail: [email protected]
J.J. Gómez-Hernández
Institute for Water and Environmental Engineering, Universitat Politècnica de València,
Valencia, Spain
e-mail: [email protected]
S. Du
Petroleum Engineering Department (Currently with Chevron), Texas A&M University,
College Station, TX, USA
e-mail: [email protected]
2 Blocking McMC
Consider a random function (RF) discretized at n grid nodes and assume that there
are m hard data (i.e., direct observations) and k linear or nonlinear state data where
the term “nonlinear” means that the dependent state data are a nonlinear function
(e.g., through flow and transport partial differential equations) of model parameters.
Specifically, let x denote the RF, x1 ¼ xobs denote the m hard data, and y ¼ yobs
denote the dependent state data. The objective of a stochastic modeling is to
generate realizations of x conditioned to x1 and inverse conditioned to y, i.e., x|
x1,y. If we adopt a multi-Gaussian distribution to characterize the RF, then
e Nðμ, Cx Þ, where μ is the prior mean of the RF and Cx is the two-point
x
covariance. After conditioning to the hard data, the distribution remains multi-
Gaussian, xjx1 Nðμxjx1 , Cxjx1 Þ, where μx|x1 is the conditional mean of the RF and
Cx|x1 is the conditional covariance, which, respectively, correspond to the simple
kriging (SK) estimate and the SK covariance commonly used in geostatistics.
The joint prior probability density function (pdf) of the multi-Gaussian random
field x conditioned to the hard data is
1 T 1
πðxjx1 , θÞ / exp ðx μxjx1 Þ Cxjx1 ðx μxjx1 Þ , ð1Þ
2
where the hyperparameter θ represents the prior information about the RF given by
the prior mean μ and the prior covariance Cx, which are necessary to compute the
SK estimate and the SK covariance. These prior values have to be either adopted as
a priori subjective estimates or modeled after some measurements. Essentially, this
pdf (1) measures the dissimilarity (or distance) of a realization from the prior; the
closer to the prior, the higher the probability. Usually the covariance reflects our
limited knowledge on the subsurface reality. The realizations generated from this
distribution by Monte Carlo are all considered as “equally likely.”
We assume a multi-Gaussian error for the discrepancy between the observed
state y and the state resulting
fromthe approximate solution of the state equations
ysim ¼ gðxÞ, ysim je
x N gðxÞ, Cy , where Cy describes the degree of discrepancy
between the transfer function g(x) and the true but error-prone observation y.
The joint pdf of y for a realization of the parameters x is given by
1 T 1
πðyjxÞ / exp ðy gðxÞÞ Cy ðy gðxÞÞ : ð2Þ
2
The error covariance matrix Cy is generally assumed diagonal. For the deterministic
optimization problem, the objective function JML is often defined as lnπ(y|x),
780 J. Fu et al.
1
J ML ¼ ðy gðxÞÞT C1
y ðy gðxÞÞ; ð3Þ
2
πðxjx1
, y, θÞ
1 1 ð4Þ
/ exp ðx μxjx1 ÞT C1
xjx1 ðx μ xjx1 Þ ðy gðxÞÞT 1
C y ðy gðxÞÞ :
2 2
1 1
J MAP ¼ ðx μxjx1 ÞT C1 T 1
xjx1 ðx μxjx1 Þ þ ðy gðxÞÞ Cy ðy gðxÞÞ, ð5Þ
2 2
3 Gradient-Based BMcMC
∂J ∂J ∂x ∂J ∂^
x
¼ ; ð8Þ
∂^z ∂x ∂^z ∂^x ∂^z
∂J
where ∂^x
is simply the computed gradient of the objective function with respect to
model parameters using the adjoint method (Fu et al. 2010, 2011). Note that ∂^ x
∂^z
is
simply the lower triangle matrix L22 in (6).
In this way, the structure information contained in the prior covariance is
explicitly applied to the realization generated, hence ensuring model consistence.
782 J. Fu et al.
More details can be found in Fu and Gómez-Hernández (2008). The accuracy of the
approximation in (8) largely depends on the correlation length of the field and the
flow scenario. However, the quality of the approximation does not damage the
quality of the solution because it only serves to find a candidate for the construction
of the Markov chains.
Sensitivity analysis shows that the size of the computational template associated
with each updating block is generally not necessarily very big (Fu and Gómez-
Hernández 2009a); a common multidimensional gradient-based optimizer will
work fine in finding the “optimal” ^z opt for an updating block. The BFGS method
(Broyden-Fletcher-Goldfarb-Shanno; also called the quasi-Newton method) is one
of the most efficient gradient-based optimization methods and is applied here for
the local optimization problem. Its key feature is to iteratively build up a good
approximation to the inverse Hessian matrix (Press et al. 1992). Within this method,
the deviate at the iteration l is updated by
^ 1 ð^g l ^g l1 Þ;
^z l ¼ ^z l1 þ H ð9Þ
l
1
^
where H l is the inverse Hessian matrix.
Several additional comments are pertinent:
1. Although one can start gBMcMC for stochastic inverse modeling with a history
matched model, which can be obtained by some deterministic approaches
(Fu et al. 2010, 2011), we assume that this large-scale model is not available
in this work but one can initialize gBMcMC with an unconditional guess model
that was generated by some geostatistical tool (Gómez-Hernández and Journel
1993).
2. Because updating one block may not necessarily reduce the objective function
too much (especially if the block is located at an insensitive zone), a general
recommendation is to perform only a couple of iterations for solving the local
optimization problem.
3. Only the relatively sensitive locations (e.g., above a certain sensitivity threshold)
should be subject to the gradient-based model update, while the insensitive
locations enter the standard McMC update without prior gradient optimization.
By doing so, many local optimization problems can be avoided.
4. In the case that the convergence is problematic for the local optimization, an
escape mechanism is required to ensure the continuous evolution of the chain
without break, for example, the chain can be advanced by the best result at hand
even though the convergent result is not available, or the chain keeps the current
state and finds another place to tune, i.e., the so-called “run-and-hit” scheme;
note that a “hit-and-run” scheme may not work fine with a gradient-based
proposal since if an “optimal” proposal is rejected, it is usually hopeless to
tune the same block at the current state to match the data.
An adaptive version of the probability perturbation gBMcMC can be formulated
when the gradient-based update is always located at the most sensitive places, i.e.,
A Gradient-Based Blocking Markov Chain Monte Carlo Method for Stochastic. . . 783
the block with the highest gradients. Again, a “run-and-hit” scheme may be invoked
if the objective function does not decrease too much after several iterations by
tuning the most sensitive places.
4 An Illustrative Example
Fig. 1 The reference lnK field and the well configuration: Four empty circles denote injection
wells and five solid dots are extraction wells
784 J. Fu et al.
of simulation such that the simulated transient pressure head distribution is to the
least degree influenced by the time discretization. Nine wells are drilled throughout
this confined aquifer (Fig. 1): four of them are injection wells with a constant flow
rate and the other five are production wells with a constant pressure. As the state,
the hydraulic head at the four injection wells is continuously collected for the first
50 time steps (approximately until after 40 time units).
The stochastic inverse modeling problem is to infer the hydraulic conductivity
field according to the observed hydraulic head data at the four injection well bores.
As a consequence of the method used for the generation of the conductivity fields,
lnK will preserve the prior distribution statistics, i.e., lnK ~ N(0,1), and the
exponential covariance with the range equal to 25. Measurement errors of state
(i.e., the pressure head) are assumed to have a normalized error variance of 5 % at
all locations, and the error covariance matrix is diagonal.
5 Results
We first want to compare the performance of the BMcMC algorithms, i.e., how the
convergence rates and acceptance ratios of various algorithms look like. Figure 2
plots the traces of three BMcMC algorithms with different block sizes. The x axis
shows the chain index, scaled in log; the y axis shows the log normalized mismatch,
that is, the mismatch (i.e., the objective function) is first normalized to one by
dividing by the original initial value and then a log value of the normalized
mismatch is taken as the y value. An obvious observation is that, for the relatively
large block size (i.e., of 25 25), the two gradient-based algorithms (i.e., the
gBMcMC and adaptive gBMcMC) do improve the convergence rate of BMcMC
(Fig. 2b); moreover, as expected, the adaptive scheme that is able to locate the most
sensitive places for model updating (i.e., the adaptive gBMcMC) seemingly has a
faster convergence rate than the others (both gBMcMC and BMcMC). However,
another observation is that for a small block size (i.e., of 8 8), such improvement
is not very evident (Fig. 2a). This is mainly because the objective function can be
brought down much more quickly by the gradient-based methods if the block size is
larger; on the other hand, the difficulty of the underlying optimization problem is
also increased and may become very challenging for a large-scale problem. In other
words, updating the entire field may have the fastest convergence rate, but the
optimization problem for finding the updates is too challenging. In addition, from
Fig. 2, one can find that the adaptive algorithm has the highest rate of convergence,
yet the quality of inverse solution seemingly is the worst. By quality, we mean the
objective function value to measure the mismatch. Note that the adaptive algorithm
has the largest average objective function; yet, since the y axis is taken as log of the
normalized mismatch, the real difference of the objective functions between them is
quite small in reality. These observations are supported by several parallel chains.
Next, we want to see the quality of the history matching, i.e., how the historical
data are reproduced and how reliable the predictions are. Figure 3 plots the histories
A Gradient-Based Blocking Markov Chain Monte Carlo Method for Stochastic. . . 785
1 1
(A) Convergence rate (B) Convergence rate
0 0
Normalized mismatch (in log)
-2 -2
-3 -3
-4 -4
0 1 2 3 4 5 6 0 1 2 3 4 5
Chain # (in log) Chain # (in log)
Fig. 2 Convergence rate of McMC inverse modelings: gBMcMC1 shows the gBMcMC method
and gBMcMC2 represents the adaptive gBMcMC method
of the pressure head at the four injection wells from the three inverse modeling
algorithms. For comparison, the unconditional cases that were generated by
GCOSIM3D are also plotted. The conditioning state data are plotted in circles,
while the gray lines plotted 200 independent realizations from different methods.
Note that the inverse modelings only constrain the models to the first 40-day’s state
data, the rest of states is treated as the prediction. Clearly, the unconditional cases
cannot reproduce the historical data but stochastic inverse-conditional realizations
match well the given data; also, stochastic inverse modelings are able to offer much
better predictions than the unconditional cases. Moreover, the data are efficiently
enveloped by the inverse-conditional realizations, meaning that the realizations
generated by the stochastic inverse modeling algorithms (BMcMC, gBMCMC, and
adaptive gBMcMC) well represent a model of uncertainty consistent with the data.
The failure of unconditional realizations in representing the uncertainty is probably
because the well performances are very sensitive to the heterogeneity of lnK around
the well bores. In addition, the adaptative gBMcMC has a relatively larger uncer-
tainty than the others simply because it has larger objective function values (Fig. 2
also shows the mismatches).
Finally, we point out that the main spatial pattern of the reference field has been
reasonably reproduced by the proposed method using various block sizes. The
variograms follow the prior specifications and are consistent with the reference
field (Fig. 4). Note that the prior model represents the models before history
matching; BMcMC plots history matched models using BMcMC method; while
gBMcMC represents the models after history matching using the method proposed.
Therefore, the method proposed can effectively preserve the spatial statistics and
structure for stochastic inverse modeling.
786 J. Fu et al.
(a) (b)
5 Well #2 (nonconditioning) 5 Well #2 (gBMcMC1 13x13)
4 4
Pressure head
Pressure head
3 3
2 2
1 1
0 0
0 100 200 300 400 500 0 100 200 300 400 500
TIME TIME
Fig. 3 Reproduction of the historical state data at well #2: (a) before history matching and (b)
after history matching
1.4 1.4
(A) x direction (B) x direction
1.2
1 1
0.8 0.8
γ
0 0
0 10 20 30 40 50 0 10 20 30 40 50
distance distance
Fig. 4 Ensemble mean of semivariogram of MCMC inverse modeling: (a) x direction and (b)
y direction
Bibliography
V. Gavinhos (*)
Faculty of Engineering, University of Beira Interior, DECA/GEOBIOTEC, Covilh~a, Portugal
e-mail: [email protected]
J. Carvalho
Faculty of Engineering, University of Porto, DEM/CERENA, Porto, Portugal
e-mail: [email protected]
1 Introduction
Dams are among the most complex and sensitive manmade structures both in terms
of project design and construction. To prevent water passing under the main
structure and particularly to avoid the effects of under pressure and foundation
erosion, safety standards are observed very closely. This makes the foundation
treatment design and construction a crucial phase in any dam project. With the
increasingly overall quality and safety awareness regarding the foundation condi-
tions, most designers adopt an intensive construction stage dedicated to ground
improvement and quality control. The overall criteria and general strategies are
established prior to construction stage, but the real amount of work and the
adaptations to real conditions are made with close monitoring of some geological–
geotechnical parameters (like fracture density, weathering), permeability, and grout
intake. A very common parameter used in decision making for depth and density of
drilling is permeability. In this case study the project design team (EDP 2011)
established the foundation permeability as the parameter to be monitored, used in
decision making, and, finally, optimized to values under 1 Lu, all during the
construction stage.
The foundation treatment consists basically of a consolidation stage and a
sealing stage where a deep curtain of grouted drill holes is built. Curtain grouting
is a series of parallel and vertical drill holes along the dam foundation which, after
grouting and sealing the surrounding fractured rock, work as an underground
barrier to water percolation. The usual construction methodology consists in apply-
ing a sequentially phased drilling strategy in which drill holes are grouted by stages
(upward or downward according to the quality of the rock). The curtain is materi-
alized by successively drilling and grouting a first line of primary boreholes
followed by a secondary collinear line of in-between boreholes. This implies that
the permeability of the secondary drilling locations may be affected by the first
phase grouting.
The general approach is to drill all primary-stage grout positions by sectors to a
target depth and evaluate the permeability of every 5 m in each drill hole and
immediately decide to go deeper and/or prepare for second-stage treatment if
permeability above 1 Lu is found.
Geostatistical models are used to identify the areas to be targeted with further
treatment at an early stage of construction development and enable decision making
for technical specification of further treatment and, in addition, allow production
management and planning to be optimized.
Geostatistical Modelling and Simulation Scenarios as Optimizing Tools for. . . 791
3 Geological Setting
The studied site is located at Sabor river mouth, northeastern part of Portugal. This
river crosses two main geological settings along its course: schist and granitic
complexes. At the site the dam foundation rock mass is described as a medium to
good-quality schist alternating with quartz-rich wakes with a permeable fracture
network developed in a very compact rock matrix. Regionally there is an important
tectonic active fault – Vilariça Fault – crossing the center of the valley graben
located to the west side of the site (Fig. 1). The river flows from northwest to
southeast.
Groundwater flow in this fractured medium is governed by three main planar
orientations, with various fillings and close aperture. Due to fracture network and
foliation/schistosity pattern, the dam foundation has low permeability although
some occurrences of small faults and altered veins work as preferred pathways.
Regional foliation/schistosity and some important fracture systems occur with
dips of about 40 –50 along average bearings of N 70 E (EDP 2007). A cross
section of the rock mass shows the interpreted model from the early stages of the
project, Fig. 2.
4 Technical Features
The Feiticeiro Dam is the first of two dams built at the end of the Sabor river basin.
Along with the Laranjeiras Dam (4 km upriver), it comprehends the hydroelectric
complex of Baixo Sabor project from Energias de Portugal (EDP). Specifically the
Feiticeiro Dam is a conventional gravity concrete dam with 22 blocks up to 45 m
high with a 290 m development along a straight axis approximately perpendicular
to the river current.
792 V. Gavinhos and J. Carvalho
Fig. 1 Geological setting of the Feiticeiro Dam (Adapted from IGM 1989)
Fig. 2 Cross section of the study area. Geological–geotechnical profile from early studies of the
project. View from upstream to downstream (Adapted from EDP 2007)
Fig. 3 Location of the sampled area: primary drilling campaign, spaced 5 m. Axis convention for
data georeferencing (Adapted from EDP 2007)
5 Data
The data chosen for this procedure was acquired at the construction site at a location
of production where there was enough information for a probabilistic approach. The
first sector to be drilled and monitored for permeability was the center of the river
valley (riverbed), from inside the already built foundation drainage gallery (FDG)
of the Feiticeiro Dam (Fig. 3).
The sampling occurs within the cycle of wireline core drilling where a pressure
packer test (Lugeon test) is performed every 5 m in descending mode until at least
30 m depth. Then, for the 17 primary drill holes inside the bottom rectangle in
Fig. 3, 102 Lugeon values were collected. The interpretation was made according to
the (Houlsby 1976) guidelines, assuming the values to represent the point at the
bottom of each 5 m chamber. Also these drill holes are separated by 5 m from each
other along the FDG. From this description the resulting sampling grid is regular
(5 5 m2) from 0 to 80 m along X and from 5 to 30 m along Y, Fig. 4.
This 2D grid was obtained by shifting the depths up toward the positive Y axis
by a vertical translation of 30 m. So the value 30 in Fig. 4 actually corresponds to a
depth of 5 m and the value 10 corresponds to the depth of 20 m.
6 Descriptive Statistics
Fig. 4 Spatial representation of the data grid (Surfer12), with measured (interpreted) Lugeon
values. Assumed punctual at the bottom of each 5 m chamber
There is a total of seven permeability classes. The predominant classes are low
permeability (0 Lu and 1 Lu) with some scarce high values (4 Lu and 6 Lu). Table 2
summarizes the descriptive statistics of the permeability 102 values sample set.
As expected, data distribution is very positively skewed. High skewness (2,43)
and kurtosis (6,94) indicate asymmetry and departure from normality. The high
variation coefficient indicates influence from high values which can become prob-
lematic in the estimation process.
Next charts on Figs. 5 and 6 graphically complement the information on Table 2
concerning the statistical characteristics of the primary permeability data set. As
previously observed the data distribution follows a clear lognormal tendency. Data
stationarity was evaluated using moving windows statistics for the mean and
variance and polynomial regression analysis using the function polytool (Matlab
13, Mathworks) (Fig. 6).
No significant trend was identified along horizontal and vertical directions. This
is an expected behavior for permeability in the context of dam foundation rock
masses. Stationarity and homoscedasticity were assumed and ordinary kriging was
used in estimation.
Geostatistical Modelling and Simulation Scenarios as Optimizing Tools for. . . 795
Fig. 6 Data regression models along horizontal and vertical directions (polytool-Matlab13)
7 Analysis of Continuity
(Fig. 8).
796 V. Gavinhos and J. Carvalho
Fig. 7 Omnidirectional
variogram with spherical
fitted model
Direction 45 presented the major range. The 3D model of the variogram surface
(Fig. 9) confirms the higher continuity along direction 45 .
This evidence is corroborated by the real geological setting where there are
important structures oriented according to the identified ranges of anisotropic
behavior.
The process of variogram model adjustment was affected by an overall random-
ness tendency of the data. Nevertheless cross-validation was systematically used to
check error variance during variogram parameter adjustment. The adopted values
are summarized in Table 3.
From the refinement of the adjustments, the nugget effect was reduced to 0,75,
and a final geometric spherical variogram model was assumed, modeled with an
anisotropic ratio of 1,7 along direction 45 .
8 Estimation
A contour map of the ordinary kriging estimation, based on the referred variogram
model, is shown in Fig. 10. The underlying regular grid was defined as 100 32
nodes with a spacing of about 0,8 m covering all data points in the sampled area of
the foundation (Figs. 3 and 4).
The error study by cross-validation (Fig. 11) and residual analysis showed an
error variance of 0,6 and a maximum error of 3,6.
Geostatistical Modelling and Simulation Scenarios as Optimizing Tools for. . . 797
Fig. 8 Directional variograms for the four selected directions: 0 , 45 , 90 , 135 , and fitted
spherical models
This map shows that error is concentrated in the areas that most likely will be
treated with secondary drilling and grouting.
The observation of the geological–geotechnical profile of the foundation rock
mass (Fig. 2), when compared with the estimated permeability map obtained by
ordinary kriging (Fig. 10), shows clear correlation in particular in which concerns
the referred regional foliation/schistosity pattern as well as some important fracture
systems occurring with dips of about 40 –50 along average bearings of N 70 E
(EDP 2007). Although the good correlation between the geological models with the
estimation map (Fig. 12) is a positive sign of the adequacy of the analysis and of the
reliability of the permeability spatial variability model, the main objective is to
798 V. Gavinhos and J. Carvalho
Fig. 11 Error map (residuals) from the kriging estimation. Values above error variance (0,6)
(Surfer12)
Fig. 12 Geological–geotechnical model compared with OK estimation map (From Figs. 2 and 10)
For this purpose, in addition to Surfer 12, GsLib software package was used.
The binary indicator transform of the data consisted in defining the cutoff
permeability (for P1 data), following (Soares 2006), as
1, if P1 1
I ð1; pÞ ¼ ð1Þ
0, if P1 > 1
Given the short amount of data available, only two threshold data sets were
considered by grouping the classes [0–1] and [2–6].
Although the P1 data set gave fairly adjustable experimental variograms, the
second phase, P2, data set showed randomness, and it was more difficult to model
also due to the number of pairs being even smaller. For indicator kriging the main
anisotropy angle, 45 , identified previously in the variogram study of the P1 data,
was also consistent with class [0–1] but for class [2–6] an omnidirectional expo-
nential model was adopted (Fig. 13 and Table 4).
The major geometric anisotropy range is 45 . Zonal anisotropy is residual and
discarded of the analysis.
The output indicator probability values were mapped on a grid identical to the
one used on estimation (Fig. 10) with 3,200 nodes, providing at this time a least
square estimate of the ccdf (Deutch and Journel 1998). The probability map
providing information on the areas within the threshold 1 Lu, e.g., low permeabil-
ity, with no need for further grouting treatment is shown in Fig. 14.
From Fig. 14 the predicted areas to be treated with further grouting of the
foundation become clear. Also the areas of very low permeability become very
well defined. It is observed that the areas that were more affected by error in the
estimation by ordinary kriging are the ones that bear more certainty to be treated
with a secondary stage of grouting.
For a general risk analysis, it is possible to use the E-type estimate. This is
simply the average permeability derived by weighting the permeability in each
cutoff class by the probability for that class, e.g., it is an optimal estimation of the
ccdf value at each grid node location (Goovaerts 1997).
When assessing the permeability in terms of the E-type estimates, it is assumed
that each grid cell has a single value. In this case the task is simple in terms of
computation and visualization because we only have two classes. So, the map in
Fig. 15 is consistent with the cutoff map in Fig. 14 because high mean values are
associated with high original data values, class [2–6], and complementary low
mean values are associated with class [0–1].
Geostatistical Modelling and Simulation Scenarios as Optimizing Tools for. . . 801
Fig. 13 Indicator variograms of the threshold data sets: class [0–1] – example of the 45 fit; class
[2–6] omnidirectional fit
Fig. 15 E-type estimation of the mean of the conditional cumulative distribution function (ccdf)
10 Conclusion
Fig. 16 Indicator sequential simulation realization map obtained with sisim, from GsLib
initial very promising results is intended to be developed and applied in the project
of a new dam in river T^amega, northern part of Portugal.
After a preliminary statistical, spatial continuity and error analysis study, the
obtained estimation, probability, and simulation maps were compared. All estima-
tion and simulation results appear to be very consistent with each other and with the
known geology and geotechnical information of the site. Ordinary kriging estima-
tion, indicator kriging, and sequential indicator simulation allowed defining prob-
abilistically quantified foundation zones needing treatment. Indicator kriging, being
a nonlinear interpolator, was useful for highly positively skewed data like the data
set used in this investigation.
Obtained models were successfully correlated with the initial information of the
project base study, namely, the geological–geotechnical model, indicating that the
described approach has an interesting potential as an optimizing tool for curtain
grouting design at the dam foundation during the construction stage.
A crucial further step in this investigation will be reducing model uncertainty by
including more data and to use secondary variables to complement information.
Nevertheless any accepted and validated method of predicting the permeability in
unsampled areas will be very useful in construction stages that last up to 3 years
long, which represent a significant percentage of cost and planning. In terms of
project and production management, a simple exercise of discarding secondary
drilling and grouting labor from the identified impermeable zones would give a
clear improvement in planning, cost control, and also environment and safety gains,
indicating this methodology as having an interesting potential for support in
efficiency decisions.
804 V. Gavinhos and J. Carvalho
Acknowledgments The authors thank the support of GEOBIOTEC and the contributions of EDP.
This work is financed by national funds through FCT-Fundaç~ao para a Ciência e Tecnologia, I.
P., within the project GEOBIOTEC – UID/GEO/04035/2013.
Bibliography
Deutch CV, Journel AG (1998) Geostatistical software library and user’s guide. Oxford University
Press, New York
EDP (2007) Relatório Geológico Geotécnico Escal~ao Jusante: Empreitada geral de construç~ao do
Aproveitamento Hidroeléctrico do Baixo Sabor
EDP (2011) Memória de Projecto de Execuç~ao: Tratamento de Fundaç~ao do Escal~ao Jusante.
Empreitada geral de construç~ao do Aproveitamento Hidroeléctrico do Baixo Sabor
Goovaerts P (1997) Geostatistics for natural resources evaluation. Oxford University Press,
Oxford
Houlsby AC (1976) Routine iinterpretation of the Lugeon water test. Q J Eng 9:303–313
IGM (1989) Geological chart 11C – Torre de Moncorvo. Serviços Geológicos de Portugal, Lisboa
Soares A (2006) Geoestatı́stica para as Ciências da Terra e do Ambiente. IST – Instituto Superior
Técnico, Lisboa
Inverse Modeling Aided by the Classification
and Regression Tree (CART) Algorithm
1 Introduction
In the last decade, ensemble Kalman filtering (EnKF) has been used to characterize
aquifer heterogeneity by assimilating observations such as piezometric heads and
concentrations. The filter has shown very good results when reproducing piezo-
metric heads (Zhou et al. 2011), but not as good when reproducing concentrations
(Xu et al. 2013; Liu et al. 2008). It has been shown that matching concentrations is
easier when assimilating both piezometric heads and solute concentrations than if
only piezometric heads are used (Li et al. 2012), because piezometric heads do not
contain enough information about the transport process. Furthermore, for the
purpose of matching concentrations, results are improved if both porosity and
hydraulic conductivity are updated simultaneously.
One problem we can face is that when the aquifer is large and the density of
observation small, the filter update may be restricted to small areas around the
observations, producing nonrealistic anomalies in the spatial distribution of the
parameters.
Another problem we have found is related to observations that are too close in
space and that are much more different between them than what could be attributed
to measurement errors. In these cases, some observations must be treated as outliers
and removed from the dataset, otherwise, the filter will introduce artifacts trying to
match these dissimilar observations.
When the filtering is not able to provide good results in a specific area of the
aquifer, it could be a sign that the conditions imposed for the simulation are not
correct, but this is hard to know a priori. A typical problem is the treatment of
recharge as deterministic, since there is always large uncertainty in its estimation.
The main objective of this manuscript is to address these issues by using a
decision tree so the parameter or parameters that are causing troubles in the filtering
process can be modified when needed. To do so, we use the classification and
regression tree (CART) algorithm (Breiman et al. 1984); this algorithm creates a
decision tree by a binary recursive partitioning procedure, up until a previously
imposed stopping rule is met; at this point the tree is “pruned” to obtain a clearer
and more understandable decision tree.
In this context, we present the CART-aided EnKF as an alternative to the
standard EnKF to avoid the problems mentioned above. The CART algorithm
discriminates which of the updated parameter fields should be replaced by another
field in order to avoid convergence issues or if the parameter field is incongruent
with the case being analyzed. Section 2 explains the EnKF methodology as well as
the CART algorithm. Section 3 explains the synthetic example used throughout this
paper. Results and discussions are found in Sect. 4. Finally, conclusions are in
Sect. 5.
Inverse Modeling Aided by the Classification and Regression Tree (CART). . . 807
2 Methodology
Kalman filtering methods have been used for many years to estimate both system
states and parameters in an optimal manner (Haykin 2001). The EnKF differs from
the original Kalman filter in the use of an ensemble of realizations in order to obtain
a better estimation of the augmented state covariance matrix P, necessary for
optimal updating of parameters and system state variables.
In the EnKF, the augmented state vector x is a collection of parameters and state
variables; in our case x consists of logarithm of hydraulic conductivity (LnK) and
porosity (Φ)—the parameters—, and hydraulic heads (h) and concentrations (c)—
the state variables. Vector x is updated with the aim of reducing the difference
between model predictions and observations, through the assimilation of state
observations in time.
For a number of realizations, Nre in the ensemble x ¼ ðx1 ; x2 ; . . . ; xNre Þ, any
realization i, xi includes parameters and state variables:
T !
A a1 ; a2 ; . . . ; aNe Np
xi ¼ ¼ ð1Þ
B i ðb1 ; b2 ; . . . ; bNn Ns ÞT i
where vector A is the vector of parameters and vector B is the vector of system state
variables, Ne is the number of elements in which the model is discretized (in our
example over a finite element mesh), Np is the number of parameters (in our case,
2, loghydraulic conductivity and porosity), Nn is the number of nodes in which
heads and concentrations are calculated (in our example, at the vertices of the finite
element mesh), and Ns is the number of state variables (in our case, 2, heads and
concentrations). For the parameters and state variables considered in our example,
Eq. (1) becomes
ðLnK 1 , Φ1 , LnK 2 , Φ2 , . . . , LnK Ne , ΦNe ÞT
xi ¼ ð2Þ
ð h1 ; c 1 ; h2 ; c 2 ; . . . ; hN n ; hN n Þ T i
The filtering process then consists of two steps; the first one is called the forecast
step and the second one the update step. In the forecast step, an initial ensemble of
parameters is used as input for the transient groundwater flow and solute transport
model:
xt ¼ Fðxt1 Þ ð3Þ
where F(∙) is the transfer function (numerical model) used to calculate state vari-
ables, and it would correspond to the solution of Eq. (8) for heads and of Eq. (9) for
concentrations shown below. The transfer function only modifies the state vari-
ables, leaving parameters unchanged.
808 J.C. Gutiérrez-Esparza and J.J. Gómez-Hernández
After the forecast is done, the update step starts by collecting observations at
observation locations. The discrepancy between observations and predictions is
used to update the augmented state vector according to the following equation:
xtu ¼ xtf þ Gt zt þ ε Hxtf ð4Þ
where xut is the updated augmented vector; xft is the forecasted vector obtained with
the flow and transport model; zt is the observation vector at time t; ε is the
observation error following a normal distribution of zero mean and diagonal
covariance R; H is the observation matrix, it is composed of 1s and 0s when
observations are located at mesh nodes; and Gt is known as the Kalman gain and
is given by the following expression:
1
Gt ¼ Ptf H T HPtf H T þ Rt ð5Þ
where Pft is the covariance matrix of the augmented state, computed from the
ensemble of realizations.
In order to reduce spurious correlations, we have used covariance localization as
a way to constrain correlation up to a certain range. Using the distance-dependent
localization function used by Xu (Xu et al. 2013):
8
>
> 1 d 5 1 d 4 5 d 3 5 d 2
>
> þ þ þ 1, 0 d a;
>
< 4 a 2 a 8 a 3 a
ρe ðdÞ ¼ ρYY ðdÞ¼ 1 d 5 1 d 4 5 d 3 5 d 2
> d 2 d 1
XY >
> þ þ 5 þ4 , a d 2a;
>
>
: 12 a 2 a 8 a 3 a a 3 a
0 d > 2a:
ð6Þ
In Eq. (7) denotes the Schur product between correlation coefficients and
covariance terms.
Inverse Modeling Aided by the Classification and Regression Tree (CART). . . 809
The groundwater flow equation in saturated porous media can be written in the
form:
∂h
∇ ðK∇hÞ þ w ¼ Ss ð8Þ
∂t
∂c
Φ ¼ ∇ ðqcÞ þ ∇ ðΦD∇CÞ ð9Þ
∂t
3 CART Algorithm
Here Xj is one of the possible values of variable j in our training dataset, while XRj
is the threshold or attribute for variable j that will discriminate the data. This is
repeated, until all the dataset is analyzed and terminal nodes contain information
only for one class. For every split, the information is homogenous, so that the left
split and the right split separate the information in an optimum way; to do so the
threshold XR is searched within all the available values of attributes in the training
dataset with the CART algorithm. For variables that are continuous, as in our case,
the CART algorithm then solves the next minimization problem:
where i(t) is the impurity measure for node t, yk is the response for training set k, and
μk is the mean of responses in the node.
After all nodes in the tree have been created, pruning is the next step, and one
way to do it is to set a threshold for the reduction in the impurity measure, below
which no split will be made. A preferred approach is to grow a very large tree and
then prune the tree back to an optimal size (Moisen 2008). For our case we used the
cost-complexity function used by Timofeev (2004) with no parameters:
! min T
CðTÞ ¼ RðTÞ þ ðTÞ ð13Þ
where R(T ) is the misclassification error observed in a cross validation dataset and
T is a complexity measure of the tree T which is equal to the sum of terminal nodes
in the tree.
The flowchart in Fig. 1 shows how the three parts are coupled. First, we start with
the flow and transport model and the initial ensemble of parameters for LnK and Φ,
where part of the ensemble is used as training dataset for the regression tree, and the
other part is used in the CART algorithm to prune a tree. Finally, the EnKF is used
to update the fields.
Inverse Modeling Aided by the Classification and Regression Tree (CART). . . 811
Initial K Initial Φ
Auxiliary K
Flow and transport model t0
Auxiliary Φ
CART Algorithm
ENKF
4 Synthetic Example
12. 09
9. 765
7. 442
5. 119
2. 796
0. 4727
-1.85
After 50 time steps and using ensembles of 100 realizations per parameter (together
with a reserve of 100 auxiliary realizations for each parameter), some fields of
hydraulic conductivity and porosity after updating are shown in the next set of
figures. Figures 4, 5, 6, 7, 8, and 9 show LnK fields before and after the first time
step updating. In general, conductivities increase after the first update because the
initial hydraulic head field was not very consistent with the observations at that time
step.
Porosity fields are also shown from Figs. 10, 11, 12, and 13; here four fields are
shown at the initial time step prior to any modification and after 50 time steps.
LnK after 50 time steps is shown for the same three fields in Figs. 14, 15, and 16.
It can be seen that an increase in LnK has occurred in the western-central part of the
aquifer. That is due to a drastic difference between the recharge needed in order to
obtain a good estimation and the one actually imposed to the model.
Hydraulic head is shown at four observation points in Figs. 17, 18, 19, and 20. In
general, the adjustment of hydraulic heads is not as good as we expected, and we
attribute it to the uncertainty on the recharge. At this moment we are working on
introducing modifications in the recharge field to improve the reproduction of the
hydraulic heads with realistic hydraulic conductivity patterns. To show the
Inverse Modeling Aided by the Classification and Regression Tree (CART). . . 813
Fig. 4 LnK field for realization 3 before the first time step updating
Fig. 5 LnK field for realization 3 after the first time step updating
Fig. 6 LnK field for realization 26 before the first time step updating
importance of recharge in the matching of the model, we have run a test with-
out modifying the recharge patterns, and the filter produces unrealistic conductivity
fields such as the one in Fig. 21.
Concentrations at observation points are shown in Figs. 22 and 23; the large
spread is similar as in the hydraulic head calculations, since the recharge is being
modified without any optimization criteria. Posterior optimization of recharge
814 J.C. Gutiérrez-Esparza and J.J. Gómez-Hernández
Fig. 7 LnK field for realization 26 after the first time step updating
Fig. 8 LnK field for realization 64 before the first time step updating
Fig. 9 LnK field for realization 64 after the first time step updating
6 Conclusions
The application of data assimilation algorithms to a realistic case faces the problem
that when some of the components defining the conceptual model is improperly
characterized, the filtering algorithm may result in very unrealistic heterogeneity
patters; patterns that are built to compensate the conceptual model errors. As it is
the case of our example with the initial estimation of the recharge, the hydraulic
816 J.C. Gutiérrez-Esparza and J.J. Gómez-Hernández
head predictions do not match the observations, so the EnKF modifies drastically
the parameters. Since it was clear that modifying conductivities alone was not
enough to match observations, we have implemented a decision tree to force
modifications in the normal procedure of the filtering algorithm. For instance,
when an updated conductivity field is not plausible, the CART algorithm enters
into play proposing a modification in the workflow, either by replacing the
Inverse Modeling Aided by the Classification and Regression Tree (CART). . . 817
Fig. 21 LnK field for realization 3 after 50 time steps if no recharge modification is implemented
Acknowledgment Financial support to carry out this work was received from the Spanish
Ministry of Economy and Competitiveness through project CGL2014-59841-P and through the
Ministry of Culture, Education, and Sports FPU program.
Inverse Modeling Aided by the Classification and Regression Tree (CART). . . 819
Bibliography
Breiman L, Friedman J, Oshen R, Stone CJ (1984) Classification and regression trees. Taylor &
Francis, Oxon
Haykin S (2001) Kalman filtering and neural networks. Wiley, New York
Li L, Zhou H, Gómez-Hernández JJ, Hendricks Franssen H-J (2012) Jointly mapping hydraulic
conductivity and porosity by assimilating concentration data via ensemble Kalman filter. J
Hydrol 428:152–169
Liu G, Chen Y, Zhang D (2008) Investigation of flow and transport processes at the MADE site
using ensemble Kalman filter. Adv Water Resour 975:986
Moisen GG (2008) Classification and regression trees. Elsevier, Oxford
Remi N (2004) Geostatistical earth modelling software: user’s manual. Blackwell, Palo Alto
Springer-Verlag (2007) Top 10 algorithms in data mining. Knowl Inf Syst 14(1):1–37
Timofeev R (2004) Classification and Regression Trees (CART) Theory and Application Msc.
Thesis. Humboldt University, Berlin
Xu T, Gómez-Hernández JJ, Zhou H, Li L (2013) The power of transient piezometric head data in
inverse modeling: an application of the localized normal-score EnKF with covariance inflation
in a heterogenous bimodal hydraulic conductivity field. Adv Water Resour 54:100–118
Zhou H, Gómez-Hernández JJ, Hendricks Franssen H-J, Li L (2011) An approach to handling
non-Gaussianity of parameters and state variables in ensemble Kalman filtering. Adv Water
Resour 54:844–864
Numerical Simulation of Solute Transport
in Groundwater Flow System Using Random
Walk Method
Abstract This paper presents the new random walk solute transport model
(RWSOLUTE) for solute transport simulation in groundwater flow system. This
model is novel in using an efficient particle tracking algorithm. The proposed model
is validated against analytical and other reported numerical solutions for chosen test
case. The accuracy and stability of the RWSOLUTE model solutions are verified
through mass balance error checks and Courant stability criteria. Further the
sensitivity of the model solutions is analyzed for varying values of time step size
and particle mass.
1 Introduction
Groundwater quality degradation in many parts of the world due to poorly planned
municipal, agricultural, and industrial waste disposal practices has continuously
drawn the attention of research worker to modify methods of predicting and
analyzing the impact of the migration of the dissolved solutes reliably in the
aquifers. The pollution of this vital water resource resulted into a serious environ-
mental problem which may damage human health and destroy the ecosystem. Thus
it has become essential to assess the severity of groundwater pollution and chalk out
the strategies of aquifer remediation, which are made possible by the use of the
statistical numerical models.
2 Mathematical Model
yield [dimensionless]; h is the hydraulic head averaged over vertical [L]; t is the
time [T]; x and y are spatial coordinates [L]; Qi is the pumping rate when (Qi < 0)
and injection rate when (Qi > 0) at ith pumping and/or injection well [L3/T]; nw is
the number of pumping and/or injection wells in the domain; np is the number of
nodes in the domain with distributed discharge and/or recharge; δðxo xi , yo yi Þ
is the Dirac delta function; xo and yo are the Cartesian coordinates of the origin [L];
xi and yi are the coordinates of ith pumping and/or injection well, [L]; and qj is the
distributed discharge rate when (qj < 0) and recharge rate when (qj > 0) at jth nodes
[L/T].
The hydrodynamic dispersion coefficients in the tensor form can be given as
2
Dxx Dxy αL v2x vx vy αT vy
vx vy
¼ þ ð2Þ
Dyx Dyy jvj vx vy v2y jvj v v
x y v2x
where α L and α T are longitudinal and transverse dispersivities [L] and jvj is the
magnitude of the average linear groundwater velocity [L/T].
An initial concentration of the solute is prescribed in the entire aquifer domain Ω
by
where vc0 is the specified advective solute flux across the boundary Γ2 [M/L3/T] and
[D] ∇c is the dispersive solute flux across the boundary Γ2 [M/L3/T].
Although only numerical solutions of the Eq. (1) are sought, it is very important
to get head distribution from the following mass balance equation for the ground-
water flow which is given as
∂h
∇:v ¼ Sy þq ð6Þ
∂t
vx ¼ T xx ∇h , vy ¼ T yy ∇h ð7Þ
where Txx and Tyy are components of the transmissivity tensor [L2/T] which are
approximated as Txx Kxx h and Tyy Kyy h, provided the change in the head in
unconfined aquifer is negligible as compared to its saturated thickness
(Illangasekare and D€oll 1989); Kxx and Kyy are components of the hydraulic
conductivity tensor [L/T].
The Random Walk Solute Transport Model (RWSOLUTE) (Kulkarni 2008) uses
real particles generated on the basis of total solute mass in the aquifer. Unlike
conventional random walk models, this model tracks real particles instead of
imaginary particles, which minimizes book keeping efforts. It also employs an
efficient particle tracking algorithm by taking weighted average of particle veloc-
ities computed at four time instances in a given time step.
This model does not directly solve the governing advection–dispersion equation.
It decouples that equation into advection and dispersion parts. The advective and
dispersive transport of the solute mass is simulated with the help of particles. Unlike
the other Eulerian–Lagrangian methods, only tracer particles are moved in the flow
field. Each particle represents the fraction of the total solute mass in the aquifer
system. The number of particles in a computational cell is worked out as
NPit, j
0
cit, j ðθ Δx Δy bÞ þ ci, j Qi, j δ x0 xi , y0 yj þ qi, j ðθ Δx Δy bÞ
¼
PM
ð8Þ
where NPti;j is the total number of the particles to be generated in the computational
cell (i, j) at a given point of time t and PM is the fixed particle mass, [M].
The advective transport is solved by tracking the movement of the particles
along flow lines. The advective displacement components of the particle in x- and
y-direction during a given time step due to average linear groundwater flow velocity
can be given as
V xp 2
dxpa ¼ V xp Δt þ /L Δt
jV j
V yp 2
dypa ¼ V yp Δt þ /T Δt ð9Þ
jV j
Numerical Simulation of Solute Transport in Groundwater Flow System Using. . . 827
where dxap , dyap are the advective displacement components of p the particle during a
r ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2 2ffi
time step Δt, [L]; dlp ¼ a
dxpa þ dypa is a resultant displacement of the
particle, [L]; and V xp and V yp are the particle velocities, [L/T].
The particle velocity components in Eq. 9 are obtained from the nodal velocity
components using bilinear interpolation method. The time weighted value of the
particle velocity is calculated by using fourth-order Runge–Kutta method. In the
bilinear interpolation method, the computational cell in which the particle has
moved after advection during a given time step is identified. The groundwater
velocity components at the representative node of the identified cell along with the
velocities at the three closest nodes that are obtained from the velocity field are used
to interpolate the particle velocity. For the location of the particle in the first
quadrant, the x-component of the particle velocity is interpolated using the follow-
ing equations:
t t
xp xi, j xp xi, j
V x 1p ¼ 1 V xi , j þ V xiþ1, j
Δx Δx
t t
xp xi, j xp xi, j
V x 2p ¼ 1 V xi, jþ1 þ V xiþ1, jþ1 ð10Þ
Δx Δx
t t
yp yi, j yp yi, j
V xp ¼ V x 1p þ 1 V x 2p
Δy Δy
where Δx and Δy are the dimensions of the computational cell, [L], and (i, j), (i + 1,
j), (i, j + 1), and (i + 1, j + 1) are the indices of the four closest nodes to the particle.
The above components of the particle velocity are weighted for taking into
account the variation of the velocity during a time step by fourth-order Runge–
Kutta method. The particle velocity components are obtained by the following
equations:
828 N.H. Kulkarni and R. Gupta
1
1
V tþΔt
xp ¼ V xp
t
þ V xp þ 2 2 V xp þ 2 3 V xp þ 4 V xp V tþΔt
yp
6 ð12Þ
1
¼ V yp
t
þ 1
V yp þ 2 2 V yp þ 2 3 V yp þ 4 V yp
6
where the superscripts 1, 2, 3, and 4 stand for the velocity components at t, t + Δt/2,
t + Δt/2, and t + Δt time levels. There are two velocity components (corresponding
to superscripts 2 and 3) at time instant t + Δt/2.
The dispersive transport of the particle due to hydrodynamic dispersion is
calculated from the random movement of particles following the Gaussian normal
distribution. The dispersive transport of a particle in x-and y-direction during a
given time step can be given as
dxpa dypa
dxpd ¼ ZLP sin φ þ ZT P cos φ ¼ Z LP þ Z TP
dlla dlla
dyp a
dxpa ð13Þ
dypd ¼ ZLP cos φ Z T P sin φ ¼ Z LP a ZT P a
dll dll
where dxdp , dydp are the dispersive transport components of pth particle during Δt,
[L]; Z Lp ¼ Nð0, σ 2L Þ&ZT p ¼ Nð0, σ 2T Þ are random numbers; N(0, σ 2L ) & N(0, σ 2T ) are
normally distributed random numbers with zero mean and one standard deviation
and which range from 6 to +6; φ is the angle between the direction of advective
qffiffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffi
transport dlap and the y axis; and σ Lp ¼ 2αL dlpa and σ T p ¼ 2αT dlpa are standard
deviations. Thus, the particle under consideration moves during a given time step
first by advective transport in proportion to the velocity of groundwater flow, and its
location after this movement is determined. Then the change in location during the
same time step due to dispersive movement is superimposed over this location.
Thus, the total displacement of a particle in x- and y-direction during a given time
step can be given as
xtþΔt
p ¼ xpt þ dxpa þ dxpd
ð14Þ
ytþΔt
p ¼ ypt þ dypa þ dypd
where xtþΔt
p , ytþΔt
p is the location of the pth particle at t + Δt time level and xtp , ytp is
the initial location of pth particle. This procedure is repeated for all the particles
generated in the cell. The change in nodal solute concentration during a time step is
given as
ctþΔt
i, j ¼ cit, j þ NPtþΔt
i, j PM ð15Þ
where NPtþΔt
i, j is the updated number of the particles residing in the computational
cell after both advective and dispersive transport. The model takes care of
Numerical Simulation of Solute Transport in Groundwater Flow System Using. . . 829
The stability of the transport model solutions can be ensured by selecting proper
time step size which will satisfy the Courant number criteria that can be given as
V x Δt
Cx ¼ 1
Δx ð16Þ
V y Δt
Cy ¼ 1
Δy
where Cx and Cy are the Courant numbers calculated for the maximum magnitudes
of the x- and y-components of the velocity vectors, respectively. It relates the
distance a particle travels during one time step to the spatial increment. The Courant
number controls the oscillations of the numerical solution arising from the temporal
discretization.
The accuracy of the transport model is verified with respect to the average mass
balance error in the solutions of Eq. (1) which is given as
100 Δ Mf cT Δ MscT
Ec ¼ ð17Þ
M0c
where Ec is the average mass balance error in the solute transport solutions
[percent], ΔMfTc is the net solute flux in the aquifer system during the simulation
period [ppm], ΔMsTc is the change in solute mass in the aquifer system [ppm], and
M0c is the initial solute mass in the aquifer system [ppm].
830 N.H. Kulkarni and R. Gupta
Figure 1 shows the schematic of aquifer modeled in test case (Sun 1996), which is
aimed at validating the RWSOLUTE model. The aquifer is 80 m long and 70 m
wide. The left and right sides of the aquifer are subject to no-flow boundary
conditions, while the top and bottom sides are subject to prescribed head boundary
conditions with the constant head of 100 m. The groundwater is flowing in from the
bottom to the top side at the uniform velocity of 1.0 m/day.
The boundary conditions for solute transport simulation involve zero solute flux
boundaries across the left and right sides and a concentration of 1 ppm on the
bottom side and zero concentration at the top side of the aquifer. Starting from an
initial state of zero solute concentration, the aquifer is gradually contaminated due
to transport of solute from a line source of pollution at the bottom side due to both
advection and dispersion.
For solute transport simulations by the proposed model, for advective transport
simulation, the aquifer is discretized by using a mesh-centered finite difference grid
with uniform nodal spacing in x- and y-direction as 5 m resulting into total
255 nodes involving 60 boundary nodes. For dispersive solute transport simulation,
50
Distance (m)
40
30
20
10
0 10 20 30 40 50 60 70 80
Distance (m)
Fig. 1 Schematic of aquifer modeled in test case for the simulation of solute transport in uniform
one-dimensional groundwater flow in unconfined aquifer with line source of pollution
Numerical Simulation of Solute Transport in Groundwater Flow System Using. . . 831
a triangular finite element mesh is used with isosceles triangular elements of size
5 m resulting into 448 finite elements. The aquifer parameters used in this test
simulation are aquifer thickness (b ¼ 30 m), effective porosity (θ ¼ 0.30), transmis-
sivity (T ¼ 10 m2/d), specific yield (Sy ¼ 0.10), and longitudinal dispersivity
(αL ¼ 10 m).
Figures 2 and 3 show comparison of transverse concentration profiles obtained at
the bottom side from the proposed model with the reported analytical solutions (Sun
1996) and FESOLUTE and RWSOLUTE solutions (Kulkarni 2008). It is found
from the comparison of the concentration profiles obtained from solute transport
simulation by proposed model that the initial error in numerical solutions dampens
out as the solution progresses through the time. The RWSOLUTE concentration
profile shows numerical overshoots at early times and deviates leftward from the
analytical concentration profile at later times. At distance of 10 m from the line
source of pollution, the percentages of the deviation between MMOCSOLUTE and
RWSOLUTE and analytical solutions are 5 % and 30 %, respectively. However,
RWSOLUTE model is justified as it simulates the randomness in actual solute
transport behavior.
The rise in mass balance curve is very sharp in the early stages of the simulation up
to 3 days. Thereafter the mass balance error curves increase gradually till the end of
the simulation period of 30 days. It is found from the results that the mass balance
error curve of RWSOLUTE simulations deviates from other reported model sim-
ulations by an order of 34 %. The RWSOLUTE simulations are more erroneous
because of the randomness in generation of the particles to simulate the solute
transport process. As the simulation progresses, the numerical oscillations get
dampened, and consequently the errors in the later stages of the simulation are
comparatively less than the error in early stages of the simulation.
It is seen from Fig. 4 that the RWSOLUTE simulated concentration profiles for the
3.75-day time step size are severely affected by numerical oscillations and
dispersions.
The solutions are found to be reasonably accurate for the time step size of 1.25
days. Even for the choice of time step size of 0.25 day, an overshoot in the
numerical solutions of around 30 % is observed at 10 m downstream from the
line source of pollution. The unacceptable overshoot of 100 % is noticed at the
832 N.H. Kulkarni and R. Gupta
1.6
Analytical
1.4 MMOCSOLUTE
FESOLUTE
RWSOLUTE
1.2
Relative Concentration (C/Co)
0.8
0.6
0.4
0.2
0
0 10 20 30 40 50 60 70
Distance (m)
Fig. 2 Validation of RWSOLUTE model with analytical solutions and reported model solutions
* FESOLUTE
MMOCSOLUTE
RWSOLUTE
1
* * * * *
0.8
*
PercentError: (%)
* *
* *
*
0.6
*
0.4
0.2
0
**0 5 10 15 20 25 30
Time: (days)
Fig. 3 Comparison of mass balance errors in RWSOLUTE and other reported models
distance of 25 m downstream of the source for 3.75-day time step simulation. Thus,
the RWSOLUTE simulations are found to be severely restricted by the choice of
time step size.
Numerical Simulation of Solute Transport in Groundwater Flow System Using. . . 833
2 RWSOLUTE
Time Step Size (Days)
1.8 0.125
1.25
1.6 3.75
Relative Concentration (C/Co)
1.4
1.2
0.8
0.6
0.4
0.2
0
0 10 20 30 40 50 60 70
Distance (m)
Figure 5 shows the effect of particle mass variation on the shape of the concentra-
tion profiles simulated by RWSOLUTE model. The fixed particle mass signifies the
fraction of the total solute mass residing in the unit volume of groundwater at a
given point of time. The particle mass is a fixed quantity while the number of
particles is only varying due to the transport of the solute mass in a given time step.
The particle mass is varied from 1.0 to 10.0 g. The numerical solutions experienced
oscillations up to the distance of 20 m from the source of the pollution for all the
chosen values of the particle mass. The maximum overshoot is observed in the
concentration profile simulated for the 10.0-g particle mass to the order of 50 %,
and the maximum undershoot is found in the concentration profile simulated for
5.0-g particle mass which is of the order of 30 %.
4 Conclusion
1. Validation of RWSOLUTE model for chosen test case shows that the numerical
dispersion in solute concentration profiles is 24 %.
2. The investigations pertaining to the effect of time step size (Courant number) on
proposed transport models reveal that the RWSOLUTE solutions suffer from
834 N.H. Kulkarni and R. Gupta
1.6
RWSOLUTE
Particle Mass (mg)
1.4 1.0
5.0
10.0
Relative Concentration (C/Co)
1.2
0.8
0.6
0.4
0.2
0
0 10 20 30 40 50 60 70
Distance (m)
Bibliography
Banton O, Delay F, Porel G (1997) A new time domain random walk method for solute transport in
1 – D heterogeneous media. Ground Water 35(6):1008–1013
Bellin A, Rubin Y, Rinaldo A (1994) Eulerian-Lagrangian approach for modeling of flow and
transport in heterogeneous geological formations. Water Resour Res 30(11):2913–2924
Numerical Simulation of Solute Transport in Groundwater Flow System Using. . . 835
Bentley LR, Pinder GF (1992) Eulerian-Lagrangian solution of the vertically averaged ground-
water transport equation. Water Resour Res 28(11):3011–3020
Delay F, Ackerer P, Danquigny C (2005) Simulating solute transport in porous or fractured
formations using random walk particle tracking. Vadose Zone J 4(2):360–379
Dentz M, Cortis A, Scher H, Berkowitz B (2004) Time behavior of solute transport in heteroge-
neous media: transition from anomalous to normal transport. Adv Water Resour 27
(2):155–173
Freeze RA, Cherry JA (1979) Groundwater. PrenticeHall Inc, Englewood Cliffs
Illangasekare TH, D€ oll P (1989) A discrete kernel method of characteristics model of solute
transport in water table aquifers. Water Resour Res 25(5):857–867
Kulkarni NH (2008) Numerical experiments on the solute transport in groundwater flow systems.
IIT Bombay, Mumbai
Lu N (1994) A semianalytical method of path line computation for transient finite-difference
groundwater flow models. Water Resour Res 30(8):2449–2459. http://doi.org/10.1029/
94WR01219
Neuman SP (1984) Adaptive Eulerian – Lagrangian finite element method for advection –
dispersion. Int J Numer Methods Eng 20(2):321–337
Prickett TA, Naymik TG, Lonnquist CG (1981) A “random-walk” solute transport model for
selected groundwater quality evaluations, vol 65. Illinois State Water Survey Champaign,
Champaign
Sorek S (1988) Eulerian-Lagrangian method for solving transport in aquifers. In: Groundwater
flow and quality modelling. Springer, Dordrecht, pp 201–214
Sun N-Z (1996) Mathematical modelling of groundwater pollution. Springer, New York
Uffink GJM (1988) Modeling of solute transport with the random walk method. In: Groundwater
flow and quality modelling. Springer, Dordrecht, pp 247–265
Wang HF, Anderson MP (1995) Introduction to groundwater modeling: finite difference and finite
element methods. Academic, San Diego
Wen X-H, Gomez-Hernandez JJ (1996) The constant displacement scheme for tracking particles
in heterogeneous aquifers. Ground Water 34(1):135
Zhang R, Huang K, van Genuchten MT (1993) An efficient Eulerian-Lagrangian method for
solving solute transport problems in steady and transient flow fields. Water Resour Res 29
(12):4131–4138
A Comparison of EnKF and EnPAT Inverse
Methods: Non-Gaussianity
Abstract The EnKF has been extensively used for real-time data assimilation in
areas such as reservoir/groundwater modeling. One of the big challenges of the
EnKF is how to handle the non-Gaussianity of aquifer properties, particularly for
channelized aquifers where preferred flow conduits are encountered. EnPAT is a
pattern-based inverse method and was developed to deal with the non-Gaussianity
of model updating. In this work, we compared the EnKF and EnPAT on a bench-
mark example. The results show that EnPAT can better reproduce the curvilinear
geological features and thus has a better transport prediction.
1 Introduction
For developing and managing production of oil fields, it is important to predict the
locations of hydrocarbon-bearing areas accurately and thus determine optimum
production well locations. This is critical because of the significant cost of drilling
wells. However, one of most challenging tasks is to characterize the geological
structure and identify reservoir properties such as permeability and porosity, which
define the locations of hydrocarbons. The challenge mainly stems from inherent
strong heterogeneity of the deposits and from scarcity of data.
L. Li (*) • H. Zhou
Geology and Geological Engineering, South Dakota School of Mines and Technology, 501
E. Saint Joseph St., Rapid City, SD 57701, USA
e-mail: [email protected]
J.J. Gómez-Hernández
Institute for Water and Environmental Engineering, Universitat Politècnica de València,
Valencia, Spain
e-mail: [email protected]
S. Srinivasan
Petroleum and Natural Gas Engineering, PennState University, 110 Hosler Building|University
Park, State College, PA 16802, USA
e-mail: [email protected]
2 Example
3 Results
Figure 2 shows two individual realizations before and after conditioning to head
data using the EnKF and the EnPAT. The updated realization using the EnKF tends
to lose the channel features of the training image. Partly, this is expected because in
the EnKF each updated realization can be interpreted as a linear combination of the
ensemble of prior models. For the EnPAT, the individual conductivity realizations
retain the channel features of the training image and end with similar channel
structures as the ones in the reference field.
A Comparison of EnKF and EnPAT Inverse Methods: Non-Gaussianity 839
Figure 3 shows the concentration distribution, after 30 days, for the reference
field, and the ensemble mean and variance of the predicted concentrations for the
prior models and for the updated models using the EnKF and the EnPAT. The
results show that the concentration profile is heavily impacted by the connectivity
of the conductivity channel structures.
4 Conclusion
The EnKF and the EnPAT are compared for inverse modeling of a conductivity
field characterized by curvilinear channel features. The EnKF has been widely used
in petroleum engineering and hydrogeology over the past decade. The remarkable
advantages of the EnKF are the capability to handle multiple parameters, compu-
tational efficiency, and real-time data assimilation. However, one significant draw-
back of the EnKF is that, because it is based on two-point statistics, it is optimal
only for linear state functions and parameters following multi-Gaussian distribu-
tion. As an alternative to the EnKF, the EnPAT method was proposed to condition
to dynamic data without the limitation that state variables should follow a multi-
Gaussian distribution. The application of the EnKF and the EnPAT to a synthetic
case shows that the EnPAT yields a more accurate characterization of conductivity
than EnKF, and more importantly, the updated models obtained using the EnPAT
honor the prior geologic features exhibited in the training image.
840 L. Li et al.
2 2
1 1
North
North
0 0
-1 -1
-2 -2
-3 -3
.0 .0
.0 East 50 .0 East 50
2 2
1 1
North
North
0 0
-1 -1
-2 -2
-3 -3
.0 .0
.0 East 50 .0 East 50
2 2
1 1
North
North
0 0
-1 -1
-2 -2
-3 -3
.0 .0
.0 East 50 .0 East 50
Fig. 2 Randomly selected two individual realizations before and after head data conditioning
using the EnKF and the EnPAT
A Comparison of EnKF and EnPAT Inverse Methods: Non-Gaussianity 841
Fig. 3 Ensemble mean and variance of concentration before and after head conditioning using the
EnKF and the EnPAT
Bibliography
Evensen G (2003) The ensemble Kalman filter: theoretical formulation and practical implemen-
tation. Ocean Dyn 53(4):343–367
Zhou H, Gomez-Hernandez JJ, Li L (2011) A pattern-search-based inverse method. Water Resour
Res 48(3):W03505
Calibration of Land Subsidence Model Using
the EnKF
Abstract Land subsidence modeling has been developed for reliable modeling and
prediction in the last several decades. Calibration of hydraulic properties such as
transmissivity and elastic and inelastic specific storages using observation data is a
challenge because of the strong nonlinearity of groundwater flow equation espe-
cially when it accounted for the interbed drainage process. The ensemble Kalman
filter is applied to calibrate hydraulic properties in a synthetic land subsidence
model. The characterization of transmissivity and specific storages and prediction
of land subsidence are improved after the drawdown and subsidence observation
data are conditioned.
1 Introduction
As a global scale problem, land subsidence has been studied by many researchers
(Galloway et al. 1999; Bell et al. 2002, 2008; Zhang and Burbey 2015). Due to
declining water levels, decreasing pore water pressures within the aquifer system
have led to significant increases in effective stress, which accounted for large-scale
compaction of sediments (Terzaghi 1925; Poland and Davis 1969; Holzer 2010).
Subsidence data, when combined with groundwater drawdown data, can be used to
improve groundwater model calibration of the hydrologic parameters such as
elastic and inelastic skeletal specific storage, transmissivity, the compaction time
constant, and others (Burbey 2001; Hoffmann et al. 2001, 2003b; Zhang et al. 2013;
Zhang and Burbey 2015). Hoffman et al. (2003a, b) used a regional groundwater
flow and subsidence model in conjunction with UCODE (Poeter et al. 2005) to
L. Li (*) • H. Zhou
Geology and Geological Engineering, South Dakota School of Mines and Technology, 501
E. Saint Joseph St., Rapid City, SD 57701, USA
e-mail: [email protected]
M. Zhang
Department of Agricultural and Biological Engineering, University of Florida, Gainesville, FL
32611, USA
estimate spatially varying compaction time constants and inelastic specific skeletal
storage coefficients in Antelope Valley, California (Hoffmann et al. 2003a). Zhang
et al. (2013) presented a discrete adjoint algorithm for identifying suitable zona-
tions of elastic and inelastic skeletal specific storage coefficients and hydraulic
transmissivity from hydraulic head and subsidence measurements (Zhang et al.
2013). The ensemble Kalman filter (EnKF) based on the sequential Bayesian
updating rule can be used to obtain results similar to those obtained by Monte
Carlo (MC)-type inverse methods but with reduced CPU time (Burgers et al. 1998;
Evensen 2003). The EnKF method is able to jointly map the hydraulic conductivity
and porosity fields accurately and efficiently by assimilating dynamic piezometric
head and multiple concentration data (Li et al. 2012). However, no research has
been done to joint estimation of spatially distributed hydraulic transmissivity and
elastic and inelastic specific skeletal storage coefficients using land subsidence and
groundwater level measurements. We will demonstrate the capability of the EnKF
to jointly map hydraulic transmissivity and elastic and inelastic specific skeletal
storage coefficients in a synthetic land subsidence model with MODFLOW 2005
(Harbaugh 2005).
2 Example
A synthetic aquifer has 38 58 grid blocks with block size of 1 1 km. Aquifer
thickness is 200 m. The aquifer is assumed to be confined and simulated as transient
flow. The total simulation time of flow and land subsidence is 15 years, with each
year being divided into two 6-month periods. A poorly permeable but highly
compressible clay interbed of constant thickness of 70 m is distributed within the
permeable aquifer. All the boundaries are assumed to be no-flow boundary, and the
initial hydraulic head is 800 m over the domain. The preconsolidation head is set as
795 m. Aquifer storage coefficient is set as 0.002. Five wells are pumped at a
constant rate in 6-month intervals (6 months on during the summer and 6 months off
during the winter) (see Fig. 1). The vertical hydraulic conductivity of interbed is
0.00006 m/day.
Three cases are considered. For the first case, no conditioning data are consid-
ered. Like the “true” fields, the initial models are generated using GCOSIM3D,
without conditioning any data. For the second case, only the measured static data at
well locations will be considered. For the case three, besides the static data,
dynamic data (drawdown and subsidence data) will be included as well; the initial
models of case 2 will be used in this case, and a further conditioning on dynamic
data is conducted using the EnKF such that a best characterization of both trans-
missivity and specific storages is achieved and a smallest uncertainty would be
observed since all the data are integrated into the models.
Calibration of Land Subsidence Model Using the EnKF 845
Fig. 1 Boundary
conditions of hypothetic
example
3 Results
4 Conclusion
0.4
Case 1
Well #1
0.3
Subsidence [m]
0.2
0.1
0
1000 2000 3000 4000 5000
Time [days]
0.4
Case 2
Well #1
0.3
Subsidence [m]
0.2
0.1
0
1000 2000 3000 4000 5000
Time [days]
Fig. 2 Simulated subsidence for Well #1 for three cases. The red line indicates the observation
data and the gray lines denote the simulation results using calibrated properties
Calibration of Land Subsidence Model Using the EnKF 847
0.4
Case 3
Well #1
0.3
Subsidence [m]
0.2
0.1
0
1000 2000 3000 4000 5000
Time [days]
Fig. 2 (continued)
Bibliography
Bell JW, Amelung F, Ramelli AR, Blewitt G (2002) Land subsidence in Las Vegas, Nevada,
1935–2000: new geodetic data show evolution, revised spatial patterns, and reduced rates.
Environ Eng Geosci 8:155–174
Bell JW, Amelung F, Ferretti A, Bianchi M, Novali F (2008) Permanent scatterer InSAR reveals
seasonal and long‐term aquifer‐system response to groundwater pumping and artificial
recharge. Water Resour Res 44:W02407
Burbey TJ (2001) Stress‐strain analyses for aquifer‐system characterization. Groundwater
39:128–136
Burgers G, Jan van Leeuwen P, Evensen G (1998) Analysis scheme in the ensemble Kalman filter.
Mon Weather Rev 126:1719–1724
Evensen G (2003) The ensemble Kalman filter: theoretical formulation and practical implemen-
tation. Ocean Dyn 53:343–367
Galloway D, Jones DR, Ingebritsen SE (1999) Land subsidence in the United States. US Geolog-
ical Survey, Reston
Harbaugh AW (2005) MODFLOW-2005, the US Geological Survey modular ground-water
model: the ground-water flow process. US Department of the Interior, US Geological Survey
Hoffmann J, Zebker HA, Galloway DL, Amelung F (2001) Seasonal subsidence and rebound in
Las Vegas Valley, Nevada, observed by synthetic aperture radar interferometry. Water Resour
Res 37:1551–1566
Hoffmann J, Galloway DL, Zebker HA (2003a) Inverse modeling of interbed storage parameters
using land subsidence observations, Antelope Valley, California. Water Resour Res 39:5
Hoffmann J, Leake S, Galloway D, Wilson AM (2003b) MODFLOW-2000 ground-water model –
user guide to the subsidence and aquifer-System compaction (SUB) package. DTIC Document
848 L. Li et al.
1 Introduction
Shallow geothermal system uses the energy stored in the first approximately 400 m
under the earth surface (Llopis Trillo and López Jimeno 2009). From about
10–20 m in depth, temperature is considered to be constant during the year. Deeper
below the surface, temperatures increase according to the geothermal gradient (3 C
for each 100 m of depth on average) (Sanner 2001).
Due to low temperatures (10 C to max. 30 C) in the shallow zone, the so-called
low enthalpy energy is obtained. Low enthalpy energy cannot be used directly and
geothermal systems have to be applied to make use of it.
The most common system to extract heat from the underground is the ground
source heat pump (GSHP) system. A GSHP system extracts thermal energy from a
cold zone to transport it to a warmer zone. The natural form of heat transport
would be in the opposite direction (from warm to cold) according to the second
law of thermodynamics. To invert the natural heat flow, it is necessary to supply
the system with energy, normally with a compressor. In these systems for each
kWh of electric energy used for the compressor, up to 4.5 kWh of thermal energy
can be provided (Conde Lázaro and Ramos Millán 2009). Another advantage of
GSHPs is the reversibility which allows to obtain heating and cooling with the
same system.
Natural groundwater or collectors installed in the underground in which a fluid
circulates are used as heat sources. In the first case, the natural groundwater is used
directly; it is pumped up with a well and transported to the heat exchanger. After
extracting energy it is reinjected to the ground. These systems are called open-loop
systems. In the second case, a fluid circulates through the collectors which are
installed in the underground. The fluid is heated up on this way in the collectors and
transports the energy to the GSHP system. This type of systems is called closed-
loop systems (Llopis Trillo and López Jimeno 2009).
Different investigations on heat transport in the subsurface have been made so far,
most of them assuming homogenous aquifer conditions. Kupfersberger (2009)
developed a 2D numerical groundwater model to simulate the impact of ground-
water heat pumps on groundwater temperature in the Leibnitzer Feld aquifer,
Austria. He validated the simulated results comparing them to field site measure-
ments. A 3D density-dependent groundwater flow and thermal transport model was
developed and validated using the results of the thermal injection experiment by
Molson (1992).
The effect of heterogeneity on heat transport simulation was the object of several
investigations over the last few years. Ferguson (2007) presented a study on the
topic, using stochastic modeling on two aquifers with low and high degrees of
heterogeneity. He concluded that there is considerable uncertainty in the distribu-
tion of heat associated with injection of warm water into an aquifer. Bridger and
Allen (2010) developed a model to evaluate the influence of aquifer heterogeneity
as a result of geologic layering on heat transport and storage in an aquifer used for
thermal energy storage. Bridger and Allen (2010) used FEFLOW to create a three-
dimensional groundwater flow and heat transport model. All these investigations
considered only the heterogeneity of the permeability, porosity, and thermal con-
ductivities that were assumed to be constant.
The present work has been made in order to get more information about heat
transport modeling in aquifer systems. Based on the results obtained by Shuang
(2009), further investigation on how heterogeneity affects heat transport simulation
Influence of Heterogeneity on Heat Transport Simulations in Shallow. . . 851
has been made. Synthetic aquifers with different grades of heterogeneity were
created using the Stanford Geostatistical Modeling Software (SGeMS) (Remy
et al. 2009).
A set of heat transport simulations were performed using MT3DMS (Zheng and
Wang 1999) as heat transport code.
To evaluate the importance of heterogeneity in permeability as well as hetero-
geneity in porosity and thermal conductivity, different simulations with homoge-
neous and heterogeneous parameters were made and compared to each other.
Therefore, the main objectives of this work are:
• To analyze the influence of heterogeneous distributions of permeability on heat
transport simulations in shallow geothermal systems
• To analyze the influence of heterogeneous distributions of porosity, bulk density,
and thermal conductivity on heat transport simulations in shallow geothermal
systems
Mean values of permeability and porosity, injection rates, initial temperature
distributions, and the model dimensions and well layout were taken from
Shuang (2009).
MT3DMS code (Zheng and Wang 1999) was originally written to simulate solute
transport. The comparison of the solute transport equation (Eq. 1) and the heat
transport equation (Eq. 2) shows the similarities of these two processes. Table 1
shows the nomenclature used hereafter:
ρ K d ∂C
1þ b n ¼ ∇½nðDm þ αs νa Þ∇C ∇ðνa nCÞ þ qs Cs λnC ð1Þ
n ∂t
! " ! #
ðρcÞe ∂T λe
n ¼∇ þ αh νa ∇T ∇ðn νa T Þ
n ρf c f ∂t n ρf c f
qh λu 0
þ ΔT ð2Þ
ρf cf ρf cf HF
MT3DMS was verified for heat transport by Mendez Hecht (2008) and Molina
Giraldo (2008). Therefore, MT3DMS can be used for the simulation of heat
transport with some adaptations on the equation coefficients. Further information
about heat transport modeling with MT3DMS can be found in other studies such as
Molina (2009). In order to perform heat transport modeling, the following adapta-
tions of the original mass transport parameters were performed:
Heat exchange between solid and liquid phase:
852 J. Rodrigo-Ilarri et al.
Table 1 Nomenclature
Parameter Unit Description
C [kg m3] Dissolved mass concentration
cs [J kg1 K1] Specific heat capacity solid
Cs [kg m3] Concentration sources and sinks
Dm [m2 s1] Coefficient of molecular diffusion
F [m] Depth of the water table
H [m] Thickness of the aquifer
k [m s1] Hydraulic conductivity
Kd m3 kg1 Distribution coefficient
n [–] Effective porosity
Q [m3 s1] Water injection rate
qh [W m3] Heat injection or extraction
qs [m3 s1 m3] Flow rate of sources and sinks solute transport
R [–] Retardation factor
t [s] Simulated time period
T0 [K] Initial temperature
Tf [K] Temperature of the water
Tin [K] Temperature of the injected water
Ts [K] Temperature of the solid
va [m s1] Seepage velocity
αh [m] Horizontal transverse dispersivity coefficient
αv [m] Vertical transverse dispersivity coefficient
αL [m] Longitudinal dispersivity coefficient
αs [m] Dispersivity coefficient
γ [–] Unit weight of the fluid
λ [W m1 K1] Thermal conductivity
λe [W m1 K1] Overall thermal conductivity of the saturated aquifer
λf [W m1 K1] Thermal conductivity of the fluid
λs [W m1 K1] Thermal conductivity of the solid
λu [W m1 K1] Thermal conductivity of the unsaturated soil
μ [kg s1 m1] Viscosity of the fluid
νs [m s1] Velocity
ρb [kg m3] Bulk density
ρece [J m3K1] Volumetric heat capacity of the saturated aquifer
ρf [kg m3] Density of water
ρfcf [J m3K1] Volumetric heat capacity of the fluid
ρs [kg m3] Density of the solid
ρscs [J m3K1] Volumetric heat capacity of the solid
σ2 [–] Variance
Influence of Heterogeneity on Heat Transport Simulations in Shallow. . . 853
!
ρ Kd ðρcÞe
1þ b ¼ ð3Þ
n n ρf c f
ρb ¼ ð 1 nÞ ρs ð5Þ
λe
Dm ¼ ð6Þ
n ρf c f
The model layout and dimensions were assumed considering the field site data
of the Esseling site (Shuang 2009). It consists of a grid with 100 100 cells (1 m
1 m x 1 m) and 40 layers. The same layout as Shuang (2009) was chosen to compare
results. Following Rasouli (2008), the aquifer is assumed to be confined. The left
and the right boundaries of the model are considered as constant head boundaries.
The upper and lower boundary and the bottom of the model are no-flow boundaries.
All other cells are assigned as active flow cells.
Flow simulations were performed assuming a hydraulic gradient i ¼ 0.02. Pre-
scribed constant heads are 45 m on the left boundary and 43 m on the right
boundary. Horizontal and the vertical hydraulic conductivities were set equal to
make the aquifer isotropic.
The injection well is located in the cell (30, 50) from layer 10 to layer 15 of the
model. The total injection rate was divided and assigned to each layer. The recharge
was assumed to be constant over 360 days. The total simulation time was divided
into 12 stress periods and a steady-state simulation was performed. As the definition
of the boundary conditions is required, a constant temperature value was assigned
to the left-hand boundary of the model. All other cells are active temperature cells.
The advection term of the heat transport equation was solved with the hybrid
MOC/MMOC (HMOC) solution scheme as it runs faster than the ultimate TVD
scheme and it is free of numerical dispersion (Zheng and Wang 1999). The type of
sorption in the chemical reaction package was set to linear isotherm sorption. No
first-order reaction was simulated.
Table 2 summarizes the flow and heat transport model input parameters that
were considered to be known throughout the simulation process.
The stochastic simulation of the hydraulic conductivity fields was done with the
Stanford Geostatistical Modeling Software (SGeMS). The sequential Gaussian
simulation algorithm was used. Based on the set of hard data and a target histogram,
ten synthetic conductivity fields were created for the scenarios shown in Table 3.
Scenarios 1–3 correspond to heterogeneous permeability distributions and constant
n and λ values, and Scenarios 4–6 correspond to heterogeneous permeability
distributions and heterogeneous n and λ values.
Figure 1 shows the logK histogram and theconductivity field view of the first
simulated field (Simulation #0) for Scenario 3 σ2 logk ¼ 1 .
Figure 2 shows the heat transport model results and the position of the heat
plume after 360 days of injection for realizations #0 of Scenarios 0, 1, 2, and 3.
Influence of Heterogeneity on Heat Transport Simulations in Shallow. . . 855
Results show that heterogeneity has an important effect on the shape of the
temperature plume, which decreases both in length and width when heterogeneity
increases.
To demonstrate the influence of heterogeneous hydraulic conductivity on the
cold plume development, simulated temperatures along the observation line (from
the injection well downstream, 12 m depth) were plotted.
Figure 3 shows the temperatures along the observation line of Simulation #0 for
each scenario. The deviation of the cold plume in this model is mainly caused by
predominating advective transport.
To visualize the differences between each of the simulations for a specific
scenario, Fig. 4 shows the results obtained for the ten simulations of Scenario 1
856 J. Rodrigo-Ilarri et al.
Fig. 1 Histogram and logK field – Simulation #0 of Scenario 3 σ2 logk ¼ 1
σ2 logk ¼ 0:1 and Scenario 3 σ2 logk ¼ 0:1 (n, λ constant) together with their
correspondent mean values.
In order to estimate the influence of heterogeneity of all parameters over the heat
transport simulation results, a set of ten new realizations for each one of the
Scenarios 4–6 were obtained. These scenarios consider heterogeneous distributions
of permeability, porosity, and thermal conductivity.
There are some known empirical relationships between hydraulic conductivity
and porosity such as those proposed by Kozeny (1927), Carman (1937), Carrier
(2003), Schneider (2003), Regalado and Carpena (2004), and Mohnke (2008). In
this work, the relation of Busch and Luckner (1993) was used, as it provides a
simple linear relationship between hydraulic conductivity and porosity.
Following Eqs. 5 and 6, these new porosity fields were afterward used to
compute every other needed parameter, like bulk density (ρb) and thermal conduc-
tivity (λ) and in a further step the effective molecular diffusion coefficient (Dm).
In order to summarize the results, Fig. 5 shows the simulated temperatures along
the observation line for ten realizations of Scenarios 1, 3, 4, and 6.
To quantify the uncertainty in the prediction of temperature distribution caused
by heterogeneity of the parameters, the variance of the simulated temperature
plumes for the ten simulations was computed for all the scenarios. Figure 6
shows the temperature variances of the ten simulations of Scenarios 3 and 6.
When comparing results obtained between simulations of corresponding scenar-
ios (Scenarios 1 and 4 or Scenarios 3 and 6), it has been found that the temperature
differences are higher as heterogeneity increases. Some simulations made for
Scenario 3 σ2 logk ¼ 1 show temperature differences up to 1 K.
Results show that an increasing variance of permeability causes increasing
variance in the expected temperature distribution. Scenario 3 σ2 logk ¼ 1 shows
2
considerably higher variances than Scenario 1 σ logk ¼ 0:1 . The highest variances
have been found in the first 20 m from the injection well. The changes in flow
direction generate an uncertainty in the prediction of the temperature plume.
Influence of Heterogeneity on Heat Transport Simulations in Shallow. . . 857
280.71
279.50
278.28
283.15
281.90
280.65
279.41
278.16
283.16
281.91
280.66
279.41
278.15
283.
282.
281.
279.
278.
Fig. 3 Simulated temperatures along the observation line for Simulation #0 of Scenarios 0, 1,
2, and 3 (n, λ constant) (top) and mean value of ten realizations for each scenario (bottom)
If hydraulic gradients were lower and consequently flow velocities decrease, the
influence of dispersion would be higher, and changes in thermal conductivity and
porosity could have a major effect on the temperature distribution. Further simula-
tions should be made using different hydraulic gradients to investigate the influence
of heterogeneity in combination with the hydraulic gradient.
This shows that the effect of the heterogeneity of the porosity n and thermal
conductivity λ can be important for heat transport simulation in highly heteroge-
neous systems. It is possible that, in combination with a low hydraulic gradient,
high heterogeneity of the porosity n and thermal conductivity λ could cause a
considerable increase in the uncertainty of the predictions of the temperature
plume.
Influence of Heterogeneity on Heat Transport Simulations in Shallow. . . 859
Fig. 4 Temperatures on observation line for ten realizations of Scenarios 1 (top) and 3 (bottom)
6 Conclusions
This work has been performed in order to analyze the influence of parameter
heterogeneity distributions on heat transport in shallow geothermal systems. Het-
erogeneous distributions of permeability, porosity, bulk density, and thermal con-
ductivity were considered on a set of simulations obtained by sequential Gaussian
simulation techniques. The following conclusions can be taken:
• Heterogeneity in the hydraulic conductivity field causes significant changes in
the hydraulic head distribution. This affects the flow velocity field which is used
for heat transport simulation.
860 J. Rodrigo-Ilarri et al.
Fig. 5 Temperatures on observation line for ten realizations of Scenarios 1–4 (top) and 3–6
(bottom)
2
1.667
1.333
1
0.6667
0.3333
1e-07
2
1.667
1.333
1
0.6667
0.3333
1e-07
Fig. 6 Temperature variances (layer 13/row 50) of Scenarios 1–3 (left) and 4–6 (right)
Acknowledgments This work has been partially funded by the Spanish Ministerio de Economía
y Competitividad through research project ¿QUIÉN HA SIDO? – Ref: CGL2014-59841-P.
862 J. Rodrigo-Ilarri et al.
Bibliography
Bridger D, Allen D (2010) Heat transport simulations in a heterogeneous aquifer used for aquifer
thermal energy storage (ATES). Can Geotechnik J 47:96–115
Busch K-F, Luckner L (1993) Geohydraulik Band 3 von Lehrbuch der Hydrogeologie. Gebr.
Borntraeger, Berlin
Carman P (1937) Fluid flow through a granular bed. Trans Inst Chem Eng 15:150–167
Carrier DW (2003) Good by Hazen; Hello Kozeny-Carman. J Geotech Geoenviron
129:1054–1056
Conde Lázaro E, Ramos Millán A (2009) Guı́a Técnica de Bombas de Calor Geotérmicas. Gráficas
Arias Montano, S.A, Madrid
Ferguson G (2007) Heterogeneity and thermal modeling of ground water. Ground Water
45:485–490
Kozeny J (1927) Über kappilare Leitung des Wassers im Boden. Wien: H€ older-Pichler-Tempsky,
A.-G. [Abt.:] Akad. d. Wiss
Kupfersberger H (2009) Heat transfer modelling of the Leibnitzer Feld aquifer, Austria. Environ
Earth Sci 59:561–571
Llopis Trillo G, López Gimeno C (2009) Guı́a Técnica de Sondeos Geotérmicos Superficiales.
Gráficas Arias Montano S.A, Madrid
Mendez Hecht J (2008) Implementation and verification of the USGS solute transport code
MT3DMS for groundwater heat transport modelling. Eberhard Karls Universität Tübingen,
Tübingen
Mohnke O (2008) Pore size distributions and conductivities of rocks derived from magnetic
resonance sounding relaxation data using multi-exponential delay time inversion. J Appl
Geophys 66:73–81
Molina Giraldo NA (2008) Verification of MT3DMS as heat transport code using analytical
solutions. Eberhard Karls Universität Tübingen, Tübingen
Molson JW (1992) Thermal energy storage in an unconfined aquifer: 1. Field injection experiment.
Water Resour Res 28–10:2845–2856
Rasouli P (2008) Numerical verification of shallow geothermal models using FEFLOW. Eberhard
Karls Universität Tübingen, Tübingen
Regalado C, Carpena R (2004) Estimating the saturated hydraulic conductivity in a spatially
variable soil with different permeameters: a stochastic Kozeny-Carman relation. Soil Tillage
Res 77:189–202
Remy N, Boucher A, Wu J (2009) Applied geostatistics with SGeMS. A user’s guide. Cambridge
University Press, Cambridge
Sanner B (2001) Shallow geothermal energy. Justus-Liebig University, Giessen
Schneider JH (2003) New least squares model used for development of permeability-porosity
correlation. Poteet Texas
Shuang J (2009) Geostatistical modeling of shallow open geothermal systems. Eberhard Karls
Universität Tübingen, Tübingen
Zheng C, Wang PP (1999) MT3DMS documentation and user guide. U.S. Army Corps of
Engineers, Washington
Part VI
Environmental Engineering and Sciences
Building a Geological Reference Platform
Using Sequence Stratigraphy Combined
with Geostatistical Tools
Abstract This paper presents a methodology that is currently tested at the French
geological survey in order to validate drill holes interpretation. Validated drill holes
are intended to be included in the future French geological reference platform
which is under construction. To validate drill holes, a first subset of high-quality
holes is selected. This data is interpreted in terms of geology and a geostatistical
analysis is performed. A 3D geological model is built to assess the overall geolog-
ical consistency. Then the rest of the drill holes is progressively and iteratively
validated by geostatistical cross validation. As several thousands of drill holes are
to be validated, specific software and workflows have been developed and are
presented here.
The objective of this work is the setup of a methodology for the validation of drill
holes that will be included in the French geological reference platform.
The French geological reference platform (Fig. 1) is one of the major scientific
programs of the BRGM (BRGM is the French geological survey). It aims at
delivering 3D validated geological data to public institutes, to the scientific com-
munity or to private companies, as well as to every citizen.
The data to be validated is of various types: geological maps, well logs, field
data, seismic profiles, etc.
Drill holes are one of the most important subsurface data to be validated in the
RGF (french geological reference platform). These drill holes are currently stored
in the French national drill holes database (BSS), which actually includes more than
800,000 drill holes. For most of the drill holes, the only provided information is the
Building a Geological Reference Platform Using Sequence Stratigraphy. . . 867
Fig. 2 First litho-stratigraphic interpretation from initial lithology description of drill holes
(D-NF marine dunes, Fy-z quaternary alluviums, SUB bedrock)
The methodology of validation of drill holes will be now presented in more details
and illustrated on two real datasets, one in the Paris basin and the other in the North
Aquitanian basin (France).
The first example shown in Fig. 4 is a 120 * 80 km area around Paris. In this area,
approximately 7000 drill holes are to be validated.
Building a Geological Reference Platform Using Sequence Stratigraphy. . . 869
Fig. 3 Overall
methodology
2460000
N
2440000
2440000
2420000
2420000
2400000
2400000
2380000
0 5 10 20
km Système de coordonnées: NTF France II degrees
Among them, 168 reference drill holes with a gamma ray are chosen for the
reference set (2.5 %).
These reference drill holes are grouped along vertical transects (Fig. 4) that are
aligned along the major geological axes. Drill holes and well logging are displayed
along these transects (e.g., Fig. 5 in the Aquitanian basin). The geologist can
interpret well logs using sequence stratigraphy concepts. Basically, the geologist
870 B. Bourgine et al.
Fig. 5 Interpretation of drill holes using gamma ray logging and sequence stratigraphy concepts.
Example in the Aquitanian basin
has to identify surfaces or time lines that correspond to depositional context change.
This allows restoring correctly the geometry of sedimentary bodies. This will also
make possible facies simulation in each geological body.
Once the well logs are interpreted in terms of sequence stratigraphy, it is possible
to define a litho-stratigraphic pile along the wells (Fig. 5). The litho-stratigraphic
pile interpretation defined here (shown at the left of each gamma ray (Fig. 5)) is a
simplification of sequence stratigraphy interpretation (shown at the right of each
gamma ray on Fig. 5), where several different facies are grouped together between
two “major” time lines.
A geostatistical analysis of reference drill holes is then performed. The studied
variable is the elevation of the top or base of each formation. It can also be the
formation thickness.
For each formation, a “formation status map” is drawn. It shows if the formation
is present or absent and if the drill hole has intersected only the top of formation, the
base of formation, or both top and base (Fig. 6).
For example, in Fig. 6 the green “#” symbols correspond to a formation gap.
When isolated and surrounded by holes intersecting the formation, it can indicate an
error of interpretation or a coordinate error. Otherwise, these “#” symbols show
where the formation has not deposited and help the geologist to check his
interpretation.
From a geostatistical point of view, the variograms of tops and bottoms are
computed and fitted. Generally, a polynomial drift (linear or quadratic) is consid-
ered, and the fitted variogram is that of the residual after removal of a global drift
(fitted by least squares to all data – which is an approximation of the true residual).
Figure 7 shows an example of variogram of the residual computed from high-
quality data. The error in the interpretation of logging is of the order of few
Building a Geological Reference Platform Using Sequence Stratigraphy. . . 871
Fig. 6 Formation status map. Example for Champigny limestone. Paris basin
240
180
Variogramme (m2)
120
Distance (m)
Fig. 7 Example of the variogram of the top of a formation (residual after removing a polynomial
drift)
centimeters, and the nugget effect can be neglected. It would not be the case with
destructive holes for which the identification of formation change, based on cut-
tings, can be affected by an error of 1–2 m.
Then a cross validation (based on kriging) is then performed. At this step, the
coefficient of the drift function is computed locally by kriging using data of the
neighborhood, so the local drift (and then the local residual) may differ from the
global one used to fit the variogram of the residual. Ideally, the new residual should
be recomputed and the process should be iterated as shown by Hengl et al. (2007).
This is not done in our case. Consequently, there is a risk of error due to a
nonoptimal geostatistical model. However, in practice the cross validation of the
reference data gives very satisfactory results in terms of error and of normalized
error, and outliers found in data can generally be attributed to errors in the data. In
our case the approximation seems acceptable and the geostatistical model is
validated.
872 B. Bourgine et al.
During this cross validation, inequality constraints given by the drill holes that
have not crossed or reached a given formation are taken into account. For example,
if a drill hole does not reach a given formation, this gives an upper bound value for
the elevation of the top of this formation. When the top of this formation is
interpolated from other drill holes where it has been observed, the interpolated
value at the location of the drill hole is compared to the upper bound. A warning is
sent if the interpolated value is above the upper bound and data can be verified.
Note that at this step a standard kriging is used instead of a “kriging taking into
account inequalities” (Freulon and De Fouquet 1993; Abrahamsen and Benth 2001)
because the objective is first to check data consistency and detect potential error
(even in the inequality constraint themselves).
Other information like the DTM and the geological map are also taken into
account. For example, the geological map, if correct, virtually gives infinity of
inequality constraints: if at a given point the formation A is outcropping that gives
constraints on the elevation of other geological formations.
The whole litho-stratigraphic pile has to be processed. If the litho-stratigraphic
pile in the area includes 20 formations, this procedure has to be repeated for all the
formations.
To make this work possible in a consistent and largely automated way, specific
software has been developed by the BRGM: GDM MultiLayer (Geological Data
Management, http://gdm.brgm.fr) (Bourgine et al. 2008; Bourgine 2015). This
software and associated algorithms include the management of gaps due to erosion
or formation pinching.
At last a preliminary 3D geological model is built using the reference dataset.
This helps checking the overall interpretation and possibly to correct it.
Once the reference dataset is validated for all the formations, other drill holes
(dataset B) are compared to the reference dataset.
For this, we again apply a cross validation technique. For example, if we
consider the top of a formation, and if we denote TA as the top of the formation
known from reference drill holes (set A) and TB the top of the same formation
measured on drill holes dataset B, we estimate TB from drill holes belonging to
dataset A and compare estimated value TB* to the true value TB. Thus, we can
compute the estimation error (TB*-TB) and the normalized error if we divide by the
kriging standard deviation.
The B holes, where we do not know the true value TB, but an inequality
constraint on TB, can also be tested.
At last the software we have developed displays automatic maps highlighting
outliers and potential errors. By clicking on the map, the geologist can get vertical
cross sections showing the potential anomaly.
Building a Geological Reference Platform Using Sequence Stratigraphy. . . 873
Fig. 8 Validation of B holes along a cross section. Example for the top of “TOA” formation
(dark-blue formation along holes)
For example, on the cross section in Fig. 8 (data from the Aquitanian basin), the
interpolation of the base of the dark-blue formation (named TOA) using reference
holes A is not consistent with B hole where the base of dark-blue formation is
intersected much lower. Either drill hole B is correct (e.g., in the case of a fault or a
fold between the two holes) or drill hole B is not located here (not GPS-tagged drill
hole, coordinate input error, database error) or has not been correctly interpreted.
Typically, this hole has to be validated or corrected by manual check.
In fact all the B holes are not checked systematically, but only when necessary.
In a first step, B holes which are consistent with reference holes and with the
geological model are validated automatically. These new validated holes are then
added to the reference set A, and the process is reiterated with nonvalidated holes
remaining in set B. The iterative process ends when there are no more automatically
validated holes in set B. The holes remaining in set B can then be checked and
corrected manually. At each step, the variogram can (on demand, if necessary) be
recomputed and remodeled, especially to improve fitting for the short-range
components.
For the selection of automatically validated holes, we use the table shown in
Fig. 9. Drill holes for which the estimation error is low and the normalized error is
lower than 2 (green cell in Fig. 9) are validated automatically and added to the
reference set A. The threshold between “low” and “high” estimation error is fixed
arbitrarily and depends on geological context, as well as on the lithology contrast
between successive formations, which can make the drill holes interpretation easy
or difficult.
The threshold value of 2 for the normalized error corresponds to a 95 %
confidence interval if we assume that the kriging error distribution is Gaussian. It
is well known that kriging error is generally not Gaussian. However, we have
verified that it is a reasonable approximation in our case.
874 B. Bourgine et al.
All the holes that are not automatically validated are kept for further checking,
and the process is iterated using reference A holes plus the validated B holes, as
soon as there is no new auto-validated B hole.
The B holes that are not automatically validated can be checked manually one by
one. As it is not always possible to check all drill holes, we check preferentially drill
holes that are located in critical zones, for example, that intersect important
aquifers. Outliers of the cross validation are also checked because often they reveal
a major error in the interpretation or in the data.
Another criterion is to select the next candidate for validation in areas where the
kriging error map obtained from reference dataset shows large errors and where
present B holes are that have not yet been validated.
For example, on the kriging standard deviation map in Fig. 10, reference holes
A are in black and B holes waiting for validation in red. We will try to validate first
B holes located in the yellow and orange areas in order to get a maximum standard
deviation of 10–12 m.
At each time new holes are validated, it is possible to rebuild the whole
geological model automatically.
Cross sections and isopach maps can be derived from the model and are used for
verification and for the final validation.
4.3 Automatization
Fig. 10 Use of the kriging standard deviation map to define the next B holes to be validated
managing gaps in case of erosion or formation pinching. However, it does not yet
handle variogram automatic fitting nor drift degree identification. This work is still
left to the geologist and requires a short training.
This software is able to handle several thousands of drill holes and work on large
areas. But in practice, it is preferable to work on a limited area, not for performance
reasons, but rather to consider relatively homogeneous areas where the drift degree
is constant or where the geological context does not change much.
4.4 Results
Examples of results are given for two areas: the North Aquitanian basin and the
Paris basin.
In the North Aquitanian basin (Fig. 11), the geological context is a carbonate
ramp, and we looked at Jurassic formations. In this area, 117 reference holes were
selected (set A), and 10 formations of the Jurassic were considered.
For the Toarcian, which is one of those ten formations, 60 drill holes were to be
validated (set B).
Two thirds of these drill holes could be automatically validated by the proce-
dure, so only one third was left for manual verification.
For control purposes, they were all checked: four were discarded, seven were
validated, and nine were erroneous but could be reinterpreted.
876 B. Bourgine et al.
Fig. 11 Result in North Aquitanian basin and location of the study area in France
5 Conclusion
A methodology has been set up in order to enable drill holes validation. It is based
on geological concepts mixed with basic geostatistical tools.
This methodology helps finding quickly consistent drill holes and discarding
erroneous ones.
As we are considering large datasets and many geological formations, automatic
and semiautomatic tools have been developed to save time and ensure repeatability.
Building a Geological Reference Platform Using Sequence Stratigraphy. . . 877
These tools have been made accessible to the geologists in charge of validation
and who are not geostatisticians.
Further work to be done concerns (1) improvement in the automation (e.g.,
automatic looping on formations and automatic database management when adding
validated B holes to dataset A), (2) automatic variogram fitting and drift identifica-
tion, and (3) finding a way to manage drill holes that have not been automatically
validated nor discarded by the present procedure. These holes are kept as “not yet
validated” but are not rejected. They are candidate for a further validation, but it
would be useful to assign to these drill holes some kind of probability index of
being consistent with present knowledge and develop other methods to assess their
quality.
References
Abrahamsen P, Benth FE (2001) Kriging with inequality constraints. Math Geol 33(6):719–744
Bourgine B (2015) MultiLayer, Manuel de référence, version 2014. BRGM Report RP64115-FR
Bourgine BLM, Lembezat C, Thierry P, Luquet C, Robelin C (2008) Tools and methods for
constructing 3D geological models in the urban environment. The Paris case. In: Ortiz JM,
Emery X. GEOSTATS 2008. Santiago, pp 951–960
Freulon X, De Fouquet C (1993) Conditioning a Gaussian model with inequalities. In: Soares A
(ed) Geostatistics Troia 92, vol 1. Kluwer Academic, Dordrecht, pp 33–51
Hengl T, Heuvelink G, Rossiter D (2007) About regression-kriging: from equations to case
studies. Comput Geosci 33:1301–1315
Homewood P, Mauriaud P, Lafont F (1999) Best practices in sequence stratigraphy for
explorationists and reservoir engineers. Bull centre Rech Elf Explor Prod Mem 25:81, pp 50
Constrained Spatial Clustering of Climate
Variables for Geostatistical Reconstruction
of Optimal Time Series and Spatial Fields
Abstract The purpose of this work is to present a new methodology for identifying
geographical regions within which the climatic behaviour of a meteorological
variable is coherent. We have chosen temperature as the variable of interest, and
thermal coherence is defined here in the sense of having a strong (negative)
correlation between terrain altitude and temperature. An improved method of
constrained spatial cluster analysis is described in the form of a new constrained
clustering algorithm. The methodology includes spatial bootstrap statistical tests to
provide a more realistic measure of the uncertainty of the coefficient of correlation
together with a spatial test of the correlation of residuals. The results are used as
optimal estimates of areal temperature averages. The methodology is illustrated by
applying it to the annual mean temperature measured at 1220 temperature stations
across Spain.
1 Introduction
Changes in temperatures are perhaps the most common and reliable indicator of
climate change or global warming (Morice et al. 2012; Robeson 1994). Tempera-
ture is usually measured at a finite number of sampling locations over regional or
national networks of weather stations. Temperature measurements are affected by
P. Dowd (*)
University of Adelaide, Adelaide, Australia
e-mail: [email protected]
H. Wang • Y. Yang
School of Resources and Geosciences, China University of Mining and Technology, Xuzhou,
Jiangsu Province 221116, China
e-mail: [email protected]; [email protected]
E. Pardo-Igúzquiza
Geological Survey of Spain, Rı́os Rosas 23, 28003 Madrid, Spain
e-mail: [email protected]
2 Methodology
Classical cluster analysis identifies groups of objects that are similar. It does so by
maximising the similarity of objects (for the purposes of this paper, temperature
measurements from weather stations) within a group and maximising the dissimi-
larity of different groups of objects (Gordon 1981). There are two broad types of
clustering methods: hierarchical clustering and non-hierarchical clustering. Among
the non-hierarchical clustering algorithms, the most widely used is the k-mean
algorithm. The similarity of objects is usually defined in terms of a distance
Constrained Spatial Clustering of Climate Variables for Geostatistical. . . 881
where ni is the number of objects that belong to the ith group and ρ*i is the value of
the modified Pearson product-moment correlation coefficient:
where ρi is the estimated Pearson correlation coefficient of the ith group and SE(ρi)
is the associated standard error. The correlation between altitude and temperature is
negative, and, thus, from Eq. 2, ρi < ρ*i . The modified coefficient, ρ*i , can be used
instead of ρi as an experimental measure of correlation between altitude and
temperature that accounts for the size of the group, i.e. the uncertainty of the
estimated value of the correlation coefficient as represented by the standard error
in Eq. 2. The value 1.96 * SE(ρi) is the lower bound of the 95 % confidence interval
882 P. Dowd et al.
and has been chosen as a conservative value for including cluster size in the
comparison of the correlation coefficient of two different clusters. The estimated
Pearson product-moment correlation coefficient, ρi, for the i th group is
Xni rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Xni
ffirffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2 Xni
ρi ¼ T ij
T i H ij
H i = T ij T i Hij H i 2
j¼1 j¼1 j¼1
ð3Þ
where Tij is the temperature at the j th station of the ith group, Hij is the altitude of the
j th station of the i th group and Ti and H
i are the respective means of the i th group.
Weather stations are thus defined by their pair (T, H ) of temperature and altitude.
Stations are also defined by their geographical co-ordinates (X, Y), which are
implicitly included in the proposed methodology by the contiguity constraint.
It is obvious that
OFðG1 ; . . . ; GM Þ N ð4Þ
with the minimum value N being unattainable in practice because it would imply
the unlikely case of a perfect (negative) correlation between altitude and temper-
ature for a given partition (G1, . . ., GM).
The standard error, SE(ρi), of the estimated Pearson correlation coefficient may
be calculated by a parametric method such as Student’s t-distribution or by using a
non-parametric method such as the bootstrap. The advantage of the latter is that it
works when the sampling distribution of the correlation coefficient is asymmetrical
and the data are (spatially) correlated, as is the case in the application
described here.
The clustering process is applied to each set of annual temperatures. In the
clustering process, there are two permitted operations: coalescence of two groups
and moving an element from one group to another. Both operations use the
definitions of contiguity given above.
Two groups Gi and Gj will coalesce to form a new group, Gk, if:
1. The two groups, Gi and Gj, are contiguous.
2. The value of the objective function improves.
An object, ok, that belongs to group Gi can move to group Gj if:
1. The object ok is contiguous with the group Gj.
2. The value of the objective function improves.
Finally, the constrained clustering algorithm comprises the following steps:
(i) Start with a partition of M clusters, where M is greater than the expected
optimal number of clusters. For example, M ¼ 100 is used in the case study.
Select at random M stations from the total N stations (N > M ). Call these
locations the seed stations. Next, each of the N stations is assigned to the
Constrained Spatial Clustering of Climate Variables for Geostatistical. . . 883
nearest seed station, and then M groups that comply with the contiguity
constraint are formed. This is the initial random partition with M groups:
fG1, G2 , . . . , GM g ð5Þ
where ρ*ij is the modified Pearson correlation coefficient, defined in Eq. 2, for
the merged group {Gij} ¼ {Gi} [ {Gi}. Note that maximum operator is used in
Eq. 6 because the correlations are negative, and groups with the largest
possible negative correlation are required.
(iii) If the operation of merging the two groups fails (because Eq. 6 is not satisfied),
then the operation of moving an element to the closest group is tried. If the pair
of elements {oik, ojl} is the two closest elements between groups {Gi} and
{Gj}, such that {oik}2 {Gi} and {ojl} 2 {Gj}, there are two possibilities to try:
(1) the station {oik} leaves group {Gi} and joins group {Gj}, and (2) the station
{ojl} leaves group {Gj} and joins group {Gi}. Note that the possibility of both
element swapping groups is not allowed because it violates the contiguity
constraint. Thus, for possibility (1), let ρ*i and ρ*j be the correlation
coefficients of {Gi} and {Gj}, respectively, and let ρ*i and ρ*jþ be the
correlation coefficients of {Gi} and {Gj+} where {Gi} ¼ {Gi} {ok},
i.e. group {Gi} without object {ok}, and {Gj+} ¼ {Gj} + {ok} is group {Gj}
with object {ok} added. The proposal to move an object from group i to
group j is accepted if
884 P. Dowd et al.
max ρ*i ; ρ*jþ < max ρ*i ; ρ*j ð7Þ
Combining (7) and (8), and noting that both conditions cannot hold simul-
taneously, gives
n o
min max ρ∗ , ρ
i jþ
∗
, max ρ ∗
, ρ
iþ j
∗
< max ρ ∗ ∗
i , ρ j ð9Þ
(iv) Go to (ii) to operate on the next group until all N groups have been visited
in turn.
3 Case Study
The study area is mainland Spain as shown in Fig. 1. It is bordered to the north and
northeast by France, to the south and east by the Mediterranean Sea and to the west
and northwest by Portugal and the Atlantic Ocean. A DEM of mainland Spain with
a resolution of 751 728 m is used in this study. We used the Universal Transverse
Mercator (UTM) projection, and we selected UTM-30N as the reference system.
The DEM is a re-projection from the original resolution of 752 752 m. The
experimental data are 1220 mean annual temperatures for year 2010, the locations
of which are also shown in Fig. 1.
Scatterplots of altitude and mean temperature for the year 2010 in Fig. 2 show a
clear linear relationship but with significant dispersion within a broad band. In
Fig. 2 the dispersion of temperatures is approximately 7 C, which is most evident
for altitudes close to zero. Stations at zero altitude are at, or near, the coast and thus
belong to different climate regions than those at higher altitudes. The total corre-
lation coefficient is 0.729 with a 95 % confidence interval of 0.755, 0.702.
Student’s t-statistic was used to calculate this interval as, for such a large sample,
the non-parametric bootstrap evaluation gives virtually the same value. The clusters
retained by the spatial clustering algorithm are shown in Fig. 3.
Constrained Spatial Clustering of Climate Variables for Geostatistical. . . 885
The scatterplots between altitude and mean annual temperature for each spatially
constrained cluster are shown in Fig. 4 in which stations with different symbols and
colours represent the different clusters.
The 13 clusters are clearly distinguishable in these plots, and the stations in each
cluster display a strong (negative) correlation between altitude and mean annual
temperature. In particular, the weakest correlation coefficient among all clusters is
0.848, which is significantly higher than the correlation coefficient (0.729) for
all temperatures taken as a single group, and the strongest correlation coefficient is
0.959. The value of the objective function (formula (1)) for the 13 retained
clusters is 1087.55, which is higher than the value of 889.38 when all stations
are considered as one group. These results demonstrate the ability of the
constrained spatial clustering approach to identify geographical regions that are
thermally coherent.
In this section, we assess the ability of the spatial clustering method to retain
clusters that explain most of the structural variance in temperature. If the retained
clusters are thermally coherent, then the regressions of temperature on altitude
within each cluster should explain almost all of the structural variance in temper-
ature within that cluster. If this is so, then the semi-variogram of the residuals
should indicate more or less spatial randomness, i.e. a pure nugget effect. The semi-
variogram of the residuals from all 13 retrained clusters taken together displays
some spatial variability up to a range of ~115,000 m, with a nugget variance of 0.37
and a structural variance of 0.24 as shown by the semi-variogram model fitted in
886 P. Dowd et al.
Fig. 2 Linear relationship between altitude and mean annual temperature for Spain, for all 1220
weather stations in the year 2010
Fig. 5. Thus, 61 % of the total variability is nugget (or random) variance, and only
39 % is spatially structured variance. By way of comparison, the model fitted to the
semi-variogram of the original temperatures also has a range of ~115,000 m, but the
Constrained Spatial Clustering of Climate Variables for Geostatistical. . . 887
Fig. 4 Scatterplots of altitude and mean annual temperature based on constrained spatial clusters
for year 2010
nugget variance of 0.60 accounts for only 8 % of the total variance with spatially
structured variance accounting for the remaining 92 %. This confirms that the bulk
of the structural variance has been removed by spatial clustering. Some residual
structural variance might be expected from border effects of the clusters; even so,
the structured variance of the residuals (0.24) represents only 3 % of the total
variance of temperature (7.5).
In this section we test the significance of the estimated correlation coefficients using
a bootstrap statistical test and the Fisher transformation. The estimated correlation
coefficients for all other clusters are inside their confidence intervals. Thus, all
888 P. Dowd et al.
Fig. 5 Omnidirectional experimental variogram (red crosses) of the regression residual for the
year 2010 and the fitted model (black line): a spherical model with a nugget variance of 0.37( C)2,
a structural variance of 0.24( C)2 and a range of 115,000 m
estimated correlation coefficients are significant. For the sake of brevity, we show
in Fig. 5 only two bootstrap histograms (clusters 57 and 18). With respect to the
estimated confidence limits, when the distribution is symmetrical, such as in B in
Fig. 5, both estimators give very similar 95 % confidence intervals. Differences
between the two estimators increase as the skewness of the sampling distribution of
the correlation coefficient increases as can be seen in A in Fig. 6.
4 Discussion
Fig. 6 Bootstrap histogram of two clusters: 95 % confidence limits of correlation coefficient from
Student’s t-test (green dashed lines) and bootstrap percentile estimates (red dashed lines); blue
dashed lines are estimated correlation coefficients
a b
0.85
18
16 0.8
14
12 0.75
10
0.7
8
6 0.65
4
2 0.6
Fig. 7 (a) Estimated annual mean temperature; (b) estimated standard error
1. Firstly, the spatial clustering algorithm delivers a stable solution. Provided steps
(i) to (iv) in the methodology section are repeated at least 1000 times, all results
will have stabilised, in particular, the number of groups, the values of the
correlation coefficients and the value of the objective function in formula (1).
In addition, the objective function will have converged to an acceptable approx-
imation of its minimum value.
2. Secondly, the role of the modified correlation coefficient. In the case study, the
modified correlation coefficient outperforms the Pearson correlation coefficient
because the inclusion of the standard error reduces the effects of the varying
cluster sizes. In formula (2) the modified correlation coefficient is defined as the
estimated correlation coefficient plus 1.96 times the corresponding standard
error. This modified value is equal to the upper limit of the 95 % confidence
interval using the Fisher transformation. Although the coefficient 1.96 is
required for a 95 % confidence interval, it could be set to any other value
depending on the requirements of the application.
890 P. Dowd et al.
3. Thirdly, the method used to quantify the uncertainty of the correlation coeffi-
cient. Provided there are sufficient data, the bootstrap will give the same result as
the Fisher transformation. However, we prefer to use the bootstrap procedure
because there may be significant differences when the sampling distribution of
the correlation coefficient is asymmetrical as, for example, in A in Fig. 6.
5 Conclusions
The improved constrained spatial clustering analysis presented in this paper has
proved to work well when applied to our experimental data. The spatial clusters that
it generates are reasonable and stable, and the correlation between annual mean
temperature and altitude in each cluster is more significant than the equivalent
correlation for all sample stations taken as a single group. In addition, by using our
modified correlation coefficient, the constrained spatial clustering approach reduces
the effect of the uncertainty caused by different cluster sizes.
In testing the spatial residual of mean annual temperature, the semi-variograms
of the residuals display near-random spatial variability for the complete data set.
This indicates that, within each cluster, altitude explains almost all of the structural
variation in mean annual temperature. For all stations taken together, it is possible
that cluster (region) border effects may introduce a small amount of structural
spatial variability in the residuals.
In testing the significance of the correlation coefficient, the bootstrap procedure
is preferred to the Fisher transformation. The reason is that for asymmetrical
distributions, which are common in practice, the bootstrap performs better.
Finally, the areal temperature average, estimated from the cluster results, pro-
vides an effective means of identifying climate zones across Spain and of detecting
climate patterns.
These results may assist in identifying effects of climate change and detecting
evidence of global warming in Spain especially when our approaches are applied to
more extensive data sets, such as a series of annual average temperature data over a
long sequence of years. We intend to investigate these possibilities in future work.
References
Carrega P (1995) A method for reconstruction of mountain air temperatures with automatic
cartographic application. Theor Appl Climatol 52:69–84
DeGaetano AT (2001) Spatial grouping of United States climate stations using a hybrid clustering
approach. Int J Climatol 21:791–807
Fereday DR, Knight JR, Scaife AA, Folland CK (2008) Cluster analysis of North
Atlantic-European circulation types and links with tropical Pacific sea surface temperatures.
J Clim 21:3687–3703
Constrained Spatial Clustering of Climate Variables for Geostatistical. . . 891
Gerstengarbe F-W, Werner PC, Fraedrich K (1999) Applying non-hierarchical cluster analysis
algorithms to climate classification: some problems and their solution. Theor Appl Climatol
64:143–150
Gordon AD (1981) Classification: methods for the exploratory analysis of multivariate data.
Chapman and Hall, London. ISBN 0412228505
Gordon AD (1996) A survey of constrained classification. Comput Stat Data Anal 21:17–29
Hoffman FM, Hargrove JR, Erickson DJ, Oglesby RJ (2005) Using clustered climate regimes to
analyse and compare predictions from fully coupled general circulation models. Earth Interact
9:1–27
Mahlstein I, Knutti R (2010) Regional climate change patterns identified by cluster analysis. Clim
Dyn 35:587–600
Morice CP, Kennedy JJ, Rayner NA, Jones PD (2012) Quantifying uncertainties in global and
regional temperature change using an ensemble of observational estimates: the HadCRUT4
data set. J Geophys Res 117:1–22
Robeson SM (1994) Influence of spatial sampling and interpolation on estimates of air temperature
change. Clim Res 4:119–126
Stooksbury DE, Michaels PJ (1991) Cluster analysis of South-Eastern U.S. climate stations. Theor
Appl Climatol 44(3):143–150
Tang L, Su X, Shao G, Zhang H, Zhao J (2012) Clustering-assisted regression(CAR) approach for
developing spatial climate data sets in China. Environ Model Softw 38:122–128
Thacker WC, Lewandowicz R (1997) Partitioning the North Atlantic into regions of similar
seasonal sea-surface temperature anomalies. Int J Climatol 17:3–23
Unal Y, Kindap T, Karaca M (2003) Redefining the climate zones of Turkey using cluster analysis.
Int J Climatol 23:1045–1055
Viers G (1975) Éléments de climatologie. Fernand Nathan, Paris, 309p
Zhou D, Khan S, Abbas A, Rana T, Zhang H, Chen Y (2009) Climatic regionalisation mapping of
the Murrumbidgee irrigation area, Australia. Proc Natl Acad Sci U S A 19:1773–1779
Zscheischler J, Mahecha MD, Harmeling S (2012) Climate classification: the value of
unsupervised clustering. Procedia Comp Sci 9:897–906
Constraining Geostatistical Simulations
of Delta Hydrofacies by Using Machine
Correlation
P. Dowd (*)
University of Adelaide, Adelaide, Australia
e-mail: [email protected]
E. Pardo-Igúzquiza
Geological Survey of Spain, Rı́os Rosas 23, 28003 Madrid, Spain
e-mail: [email protected]
S. Jorreto • A. Pulido-Bosch • F. Sánchez-Martos
University of Almerı́a, Almerı́a, Spain
e-mail: [email protected]; [email protected]
1 Introduction
2 Methodology
where i is the depth index, zi is the centre of an interval in the reference well, k is the
offset between the centres of the intervals in the two wells being compared and n is
the number of data either side of the central interval value; the number of data in the
interval is thus 2n + 1.
However, while in CORRELATOR the similarity index is based on the concept
of shale content as estimated from gamma logs, we have defined a similarity index
for hydrofacies:
X K Xiþn Xiþn
s¼1
j¼in
I s ð jÞ j¼in
I s ð j þ k Þ
α1, 2 ði; k; nÞ ¼ 1 ð2Þ
2ð2n þ 1Þ
where Is ( j) is the indicator of the sth hydrofacies at the j th location in the interval in
well 1, Is ( j + k) is the indicator of the sth hydrofacies at the ( j + k)th location in the
interval in well 2 and K is the total number of hydrofacies. For example, if i ¼ 5,
k ¼13, n ¼ 3 and K ¼ 4, then the total number of each of the four hydrofacies in well
1 in the interval from depth 2 to depth 8 is compared with the total number in the
interval from depth 15 to depth 21 in well 2. The value of α1,2(i, k; n) ranges
between 0 and 1 in the same way as the similarity function based on the shale
content in CORRELATOR.
The Pearson correlation coefficient is defined in the standard way (Olea 2004):
cov12 ði; j; nÞ
r 1, 2 ði; k; nÞ ¼ ð3Þ
s1 ði; nÞs2 ði; k; nÞ
where cov12(i, j; n) is the covariance between the two intervals (one in each
borehole) and s1(i; n)and s2(i, k; n) are the respective standard deviations. The
covariance term is
where E{.} is the mathematical expectation operator and C(i) is the category
number at ith location CðiÞ 2 f1, 2, . . . , K g.
The final correlation function in Eq. (1) is used to establish the likelihood of
correlations between different parts of the boreholes, which, in turn, can be used to
assess the plausibility of different geostatistically simulated realisations of the
spatial distribution of hydrofacies.
3 Case Study
The study area (Fig. 1) is the detrital aquifer of the Andarax river delta in the
province of Almerı́a (Southern Spain). The aquifer comprises deltaic deposits from
the Pleistocene overlain by fluvial and deltaic deposits from the Quaternary
(Sanchez-Martos et al. 1999). The Andarax river is ephemeral, with flow usually
resulting from big storms, and is typical of rivers in the semiarid coastal regions of
the Mediterranean. Within the study area, there are 19 boreholes and three clusters
of four piezometers each (Jorreto-Zaguirre et al. 2005), the locations of which are
shown in plan view in Fig. 1. The borehole cores were classified into five types of
hydrofacies according to their permeability: very permeable (category 1), perme-
able (category 2), low permeability (category 3), impermeable (category 4) and
very impermeable (category 5).
Figure 2 shows the original 15 boreholes that were drilled along the coast. Their
absolute locations are shown in Fig. 1, whereas in Fig. 2 relative spatial locations
are used so as to show the distribution of the facies along the boreholes as well as
the relationships between them. A visual inspection shows that the most abundant
of the hydrofacies is number 2, and the least abundant is number 1. The proportions
of the hydrofacies are 2.65 %, 43.70 %, 27.10 %, 18.50 % and 7.80 % for
hydrofacies 1 to 5, respectively.
It is also apparent from Fig. 2 that there are no clear patterns of spatial
distribution of the hydrofacies across the boreholes. This could, however, be a
subjective evaluation, whereas an evaluation by machine correlation could provide
a quantitative assessment that would be the same for everyone independent of
individual experience. The three functions given in Eq. 1, i.e. similarity, Pearson
correlation and weighted final correlation, are shown for boreholes 1 and 3 in
Figs. 3, 4 and 5, respectively. In these graphs if the maximum correlations are on
the 1:1 line, the implication is that the maximum correlation is along the horizontal
between the boreholes, while deviations from the 1:1 line give the likely dip of the
layer according to machine correlation. The similarity function has the effect of
screening large correlations between intervals with different hydrofacies but with
the same correlation pattern. Negative correlations are of no interest for hydrofacies
correlation, so the final image of interest is the weighted correlation using positive
correlations (Fig. 5).
Figure 5 shows that there is a relatively high probability (>0.6) that the
hydrofacies are correlated in the interval between depths of 36 m and 62 m in
Constraining Geostatistical Simulations of Delta Hydrofacies by Using. . . 897
Fig. 1 Location of the study area (yellow square) in the Andarax river delta. 1 Quaternary detrital
material. 2 Pliocene detrital material
Fig. 2 Distribution of categories (from 1 to 5) along the 15 boreholes (from 1 to 15, Fig. 1) that are
located parallel to the coast
simulations of the hydrofacies. Although the machine correlation between all the
boreholes should be considered, for illustration purposes we concentrate on bore-
holes 10 and 11 for which the machine correlation is given in Fig. 7. These two
boreholes are shown in 3D in Fig. 9.
Figures 10, 11 and 12 show panel 74, from the 3D simulations in Fig. 8b–d,
respectively, and which passes close to the line that joins boreholes 10 and 11. The
idea is to rank the plausibility of the three simulations of the panels between
boreholes 10 and 11 represented by the yellow square in Figs. 10, 11 and 12 by
using the machine correlation in Fig. 7.
This task could be accomplished in several ways using the fact that the condi-
tional simulation is a 3D grid of voxels so that each column of voxels (keeping x
and y constant) can be assimilated to a simulated or synthetic borehole. Thus, the
equidistant “synthetic borehole” (Fig. 13) could be correlated with boreholes 10 and
11 in turn (Fig. 14), and the weighted correlation function could be compared with
Fig. 7. The procedure is repeated for each simulation; it would then be possible to
rank the simulations in Figs. 10 and 11 according to this rank.
Constraining Geostatistical Simulations of Delta Hydrofacies by Using. . . 899
Fig. 6 Left random permutation of hydrofacies of boreholes 1 and 3. Right raw correlation
(upper), similarity function (middle) and weighted correlation (bottom)
Fig. 8 (a) 3D view of boreholes and (b, c and d) three different realisations of conditional
simulations of the geologic delta mediums by simulating the hydrofacies
Constraining Geostatistical Simulations of Delta Hydrofacies by Using. . . 903
Fig. 9 3D location of
boreholes 10 and 11
Fig. 13 Synthetic borehole equidistant from experimental boreholes 10 and 11 created by the
column of voxels at the intersection of two panels from the 3D simulation
In practice, however, data are often so sparse that the range of generated
realisations is unrealistically wide. In the absence of sufficient data or other
information, machine correlation offers one possibility for constraining the
range of realisations by determining the possible continuity of the horizons of
the hydrofacies. In this work a method of machine correlation for hydrofacies has
been proposed as an extension of the CORRELATOR approach (Olea 2004) and
which does not require any additional information. The procedure has been
validated with real data, and we have shown how it could be used in constraining
the 3D realisations of hydrofacies such as those given in Dowd et al. (2015). The
work presented here is just a first step, and further work is needed to rank
realisations by a plausibility index, based on machine correlation, which could
then be used for screening the realisations.
In future work we will also explore a variation to the approach described in this
paper by conducting the well-to-well correlation before doing the geostatistical
simulations. The well-to-well correlation would define the stratigraphic grid on
which the simulations would be generated. As a reviewer pointed out, this would
avoid mixing correlation uncertainty and geostatistical uncertainty. Nevertheless,
for the particular application described in this paper, correlation after simulation is a
useful approach because underground geology data are sparse and there are no 3D
geophysics data available; this lack of data makes correlation before simulation far
too restrictive.
906 P. Dowd et al.
Fig. 14 Machine correlation between the synthetic borehole in Fig. 13 for the three simulations in
Fig. 8b–d and boreholes 10 and 11
Acknowledgement This work was supported by research project CGL2015-71510-R from the
Ministerio de Economía y Competitividad of Spain.
We are grateful to an anonymous reviewer for the comments on the alternative approach of
conducting the well-to-well correlation before doing the geostatistical simulations.
References
Barakat MKA (2010) Modern geophysical techniques for constructing a 3D geological model on
the Nile Delta, Egypt. Thesis dissertation. Technische Universität Berlin, 158 p
Cabello P, Cuevas JL, Ramos E (2007) 3D modelling of grain size distribution in quaternary
deltaic deposits (Llobregat Delta, NE Spain). Geol Acta 5(3):231–244
dell’Arciprete D, Bersezio R, Felleti F, Giudici M, Comunian A, Renard P (2011) Comparison of
three geostatistical methods for hydrofacies simulation: a test on alluvial sediments. Hydrogeol
J 20:299–311
Constraining Geostatistical Simulations of Delta Hydrofacies by Using. . . 907
Abstract Nowadays, climate data series are used in so many different studies that
their importance implies the essential need of good data quality. For this reason, the
process of homogenisation became a hot topic in the last decades, and many
researchers have focused on developing efficient methods for the detection and
correction of inhomogeneities in climate data series. This study evaluates the
efficiency of the gsimcli homogenisation method, which is based on a geostatistical
simulation approach. For each instant in time, gsimcli uses the direct sequential
simulation algorithm to generate several equally probable realisations of the cli-
mate variable at the candidate station’s location, disregarding its values. The
probability density function estimated at the candidate station’s location (local
probability density functions (PDF)), for each instant in time, is then used to verify
the existence of inhomogeneities in the candidate time series. When an inhomoge-
neity is detected, that value is replaced by a statistical value (correction parameter)
derived from the estimated local PDF. In order to assess the gsimcli efficiency with
different implementation strategies, we homogenised monthly precipitation data
from an Austrian network of the COST-HOME benchmark data set (COST Action
ES0601, Advances in homogenization methods of climate series: an integrated
approach – HOME). The following parameters were tested: grid cell size, candidate
order in the homogenisation process, local radius parameter, detection parameter
and correction parameter. Performance metrics were computed to assess the effi-
ciency of gsimcli. The results show the high influence of the grid cell size and of the
correction parameter in the method’s performance.
1 Introduction
surrogate and synthetic. The first group contains real inhomogeneous data, while
the other two enclose simulated data with inserted inhomogeneities, outliers,
missing data periods, local station trends and a global trend, per network (Venema
et al. 2012). Fifteen simulated networks were prepared, and they are located in
different places within Europe. The networks comprise 5, 9 and 15 stations.
The submitted methods were evaluated by the calculation of performance
metrics. Based on the performance metrics, the best homogenisation contributions
were ACMANT (Domonkos et al. 2011), MASH (Szentimrey 1999, 2007, 2008),
PRODIGE (Caussinus and Mestre 1996, 2004) and USHCN (Menne and Williams
2009; Menne et al. 2009). Recently, some of the homogenisation methods were
transformed into software packages, in order to become fully automatic procedures,
and they are available in (http://www.climatol.eu/tt-hom/).
This study assesses the efficiency of the gsimcli homogenisation method, which
is based on a geostatistical simulation approach. To do so, we homogenised
monthly precipitation data from an Austrian network of the HOME benchmark
data set. The following parameters were tested: grid cell size, candidate order in the
homogenisation process, local radius parameter, detection parameter and correction
parameter. Performance metrics were computed to assess the efficiency of gsimcli.
Precipitation is the focus on this study, since it is one of the most important
variables for climate and hydrometeorology studies. Changes in precipitation
pattern may lead to floods, to droughts and consequentially to the loss of biodiver-
sity and agricultural productivity (Sayemuzzaman and Jha 2014).
This work is organised as follows. The following section depicts the study
domain and data. The methodological framework includes the description of the
gsimcli method and the set of performed homogeneity tests. In the results and
discussion section, the performance metrics are scrutinised. Finally, the conclusion
section brings a summary of the lessons learned and recommendations for
future work.
This study analyses monthly surrogate precipitation data that are part of the HOME
data set, namely, the network 16. This network comprises 15 stations and it is
located in Austria (Fig. 1). The data series include 100 years of precipitation values,
between 1900 and 1999. It covers a rectangular area of approximately 24,640 km2
(220 112 km). Considering the statistics for the annual series (Table 1), the lowest
values were recorded in stations 4313302 (northeast corner) and 4315421 (west
area), with 374.3 mm and 384.6 mm, respectively. The maximum value was
recorded in station 4319710 (southwest corner). The variability of both annual
and monthly series is very high. For example, the standard deviation of the annual
series varies between 131.4 mm (station 4315515) and 287.7 mm (station
4320123).
912 S. Ribeiro et al.
purpose of the variography study. Seven semivariograms were modelled for each of
the monthly series, in a total of 84.
3 Methodological Framework
This study evaluates the gsimcli homogenisation method, which is based on the
direct sequential simulation (DSS) algorithm (Soares 2001). The gsimcli method
uses the DSS in the calculation of the local probability density functions (PDF)
(Costa and Soares 2009) at the location of the candidate station. Such calculation is
prepared solely with the temporal and spatial observations of nearby reference
stations. A breakpoint is identified whenever the interval of a specified probability
p, centred in the local PDF, does not include the real observation of the candidate
station. The detected irregular value is then replaced by a statistic value of the local
PDF formerly computed (e.g. mean, median or a given percentile).
This method turned into a software package, which allowed the homogenisation
process to become direct and quasi-automatic (Caineta et al. 2015). Two subsets of
parameters must be defined before starting the homogenisation procedure: the
simulation parameters and the homogenisation parameters. The former define the
number of simulations, the kriging type of the geostatistical method, the maximum
number of nodes to be found, the number of CPU cores, the simulation grid size and
the semivariogram parameters. The latter depict the candidate order, the detection
parameter, the local radius and the correction parameter. The simulation grid
describing the area where the stations are located must also be analysed, in order
to ascertain the cell size, the number of columns and the number of rows. These
values take into account that the bordering stations of the network must be
surrounded by a number of cells at least equal to the value of the local radius
parameter. The gsimcli method is freely available at http://iled.github.io/gsimcli/.
Table 2 Different homogenisation strategies (grid cell size, detection parameter, correction
parameter and local radius parameter)
Test Grid cell size Detection Correction parameter Local radius
# (metres) parameter (p) (percentile p) parameter (r)
1 1,000 0.95 0.975 0
2 1,000 0.95 0.95 0
3 5,000 0.95 0.90 1
4 5,000 0.95 0.90 2
5 5,000 0.95 0.90 3
6 5,000 0.95 0.975 0
7 5,000 0.95 0.975 1
8 5,000 0.95 0.975 2
9 10,000 0.95 0.975 0
10 10,000 0.95 0.975 1
11 10,000 0.95 0.975 2
12* 10,000 0.95 0.90 0
13* 10,000 0.95 0.90 0
14 10,000 0.95 0.90 1
15 10,000 0.95 0.90 2
16 10,000 0.975 0.975 0
In all Tests the candidate order was based on the stations’ data variance, except in Test #13 (*) that
was based on the network deviation
parameter (the statistic value used for the inhomogeneities correction: the inhomo-
geneities, outliers or missing values can be replaced by the mean, median, skewness
and percentile), the local radius (sets the radius of a circle centred at the candidate
station location where the simulated values of the nodes located within the circle are
considered in the calculation of the local PDF) and the candidate order (the order by
which the candidate stations are homogenised).
The analysed grid cell sizes are 1000 m, 5000 m and 10,000 m, which corre-
spond to grids of 27,709 cells (229 121 cells), 2088 (58 36 cells) and 792 cells
(36 22 cells), respectively. The values of the detection parameter analysed are
0.95 and 0.975. The values considered for the correction parameter are the percen-
tiles of 0.90, 0.95 and 0.975. The investigated local radii are 0, 1, 2 and 3. It is
noteworthy to mention that the area of the circle centred in the candidate station
depends on the grid cell size and the local radius. Regarding the candidate order,
Tests #1 to #12 and #14 to #16 used the descending value of the stations’ data
variance, while Test # 13 used the network deviation (difference between the
station and the network average values) to define the sequence of the candidate
stations to be homogenised (Table 2).
The values defined for the remaining simulation parameters are common to the
16 strategies and correspond to default values proposed by Ribeiro et al. (2016):
• Number of simulations: 500
• Kriging type: ordinary kriging
Assessing the Performance of the Gsimcli Homogenisation Method with. . . 915
5 Conclusion
Acknowledgements The authors gratefully acknowledge the financial support of Fundaç~ ao para
a Ciência e Tecnologia (FCT), Portugal, through the research project PTDC/GEO-MET/4026/
2012 (“GSIMCLI – Geostatistical simulation with local distributions for the homogenization and
interpolation of climate data”).
Assessing the Performance of the Gsimcli Homogenisation Method with. . . 917
References
Aguilar E, Auer I, Brunet M, Peterson T, Wieringa J (2003) Guidelines on climate metadata and
homogenization. In Llansó P (ed) WMO/TD No. 1186, WCDMP No. 53. World Meteorolog-
ical Organization, Geneva
Bernstein L, Bosch P, Canziani O, Chen Z, Christ R, Davidson O et al (2007) Climate change
2007: synthesis report. Cambridge University Press, Cambridge
Bližňák V, Valente M, Bethke J (2014). Homogenization of time series from Portugal and its
former colonies for the period from the late 19th to the early 21st century. Int J Climatol
2400–2418
Caineta J, Ribeiro S, Soares A, Costa AC (2015) Workflow for the homogenisation of climate data
using geostatistical simulation, vol 1, Informatics, geoinformatics and remote sensing. Inter-
national Multidisciplinary Scientific GeoConference – SGEM, Albena, pp 921–929
Caussinus H, Mestre O (1996) New mathematical tools and methodologies for relative homoge-
neity testing. In: Service HM (ed) Proceedings of the first seminar for homogenization of
surface climatological data. World Meteorological Organization, Budapest, pp 63–82
Caussinus H, Mestre O (2004) Detection and correction of artificial shifts in climate series. Appl
Stat 53:405–425
Chair of the Management Committee of the Action (2011) Monitoring progress report 03/05/2007-
01/06/2011. HOME – advances in homogenisation methods of climate series: an integrated
approach (COST Action ES0601)
Conrad V, Pollack LW (1962) Methods in climatology. Havard University Press, Cambridge, MA
Costa AC, Soares A (2009) Homogenization of climate data: review and new perspectives using
geostatistics. Math Geosci 117(1):91–112
Domonkos P (2013) Measuring performances of homogenization methods. Időjárás Q J Hung
Meteorol 117(1):91–112
Domonkos P, Poza R, Efthymiadis D (2011) Newest developments of ACMANT. Adv Sci Res
6:7–11
Hannart A, Mestre O, Naveau P (2014) An automatized homogenization procedure via pairwise
comparisons with application to Argentinean temperature series. Int J Climatol 34:3528–3545
Menne MJ, Williams CN Jr (2009) Homogenization of temperature series via pairwise compar-
isons. J Clim 22(7):1700–1717
Menne MJ, Williams CN Jr, Vose RS (2009) The U. S. historical climatology network monthly
temperature data, version 2. Bull Am Meteorol Soc 90:993–1007
Ribeiro S, Caineta J, Costa AC (2015) Review and discussion of homogenisation methods. Phys
Chem Earth, (in press)
Ribeiro S, Caineta J, Costa A, Henriques R, Soares A (2016) Detection of inhomogeneities in
precipitation time series in Portugal using direct sequential simulation. Atmos Res
171:147–158
Sayemuzzaman M, Jha MK (2014) Seasonal and annual precipitation time series trend analysis of
North Carolina, United States. Atmos Res 137:183–194
Soares A (2001) Direct sequential simulation and cosimulation. Math Geol 33(8):911–926
Szentimrey T (1999) Multiple analysis of series for homogenization (MASH). In: Proceedings of
the second seminar for homogenization of surface climatological data. WCDMP-No.
41, WMO-TD No. 1962. World Meteorological Organization, Budapest, pp 27–46
Szentimrey T (2007).Manual of homogenization software MASH v3.02. Hungarian Meteorolog-
ical Service
Szentimrey T (2008) Development of MASH homogenization procedure for daily data. In:
Lakatos M, Szentimrey T, Bihari Z, Szalai S (eds) Proceedings of the fifth seminar for
homogenization and quality control in climatological databases. WCDMP-No. 71. World
Meteorological Organization, Budapest, pp 123–130
Venema V, Mestre O, Aguilar E, Auer I, Guijarro J, Domonkos P et al (2012) Benchmarking
homogenization algorithms for monthly data. Clim Past 8(1):89–115
918 S. Ribeiro et al.
Yingjun Sun
Abstract Soil pollution caused by heavy metals was studied. The potential eco-
logical risks posed by seven heavy metals (Cu, Pb, Cd, Cr, Ni, Zn, and As) in the
agricultural soils of Yanggu, Shandong Province, China, were analyzed. The spatial
variation of the seven heavy metals was used to develop a probability map of the
heavy metals based on sequential simulation methods. The ecological risks to the
region from the heavy metals were then assessed using the Hakanson potential
danger index. The result showed that Cd was the main problem in the Yanggu area.
The potential ecological risk from Cd was nearly 80 and the risk classification was
medium. The total potential ecological risk from all seven heavy metals was
114.656. This indicated that heavy metal pollution in the entire study area posed
only a slight ecological risk.
1 Introduction
Y. Sun (*)
College of Geomatics, Shandong University of Science and Technology, Qingdao 266510,
China
Department of Civil Engineering, Shandong Jianzhu University, Jinan 250014, China
e-mail: [email protected]
2 Experiment
Yanggu (35 550 –36 190 N, 115 390 –116 060 E) is located in the west of Shandong
Province (Fig. 1). It covers an area of 1,064 km2 and the population size is 794,800.
It has a typical continental monsoon climate, with warm temperatures, semi-humid
conditions, and well-defined seasons. The mean annual temperature is 13.9 C, and
the mean precipitation is 523.9 mm. The soil of Yanggu is typically moist; the soil
types are mainly sandy soil, loam soil, and clay soil. Yanggu is an important cotton-
producing area for both the Shandong Province and China.
2.2 Sampling
A digital map of Yanggu was used to establish uniform 2 2 km grids for the entire
study area. Each grid center represented a sampling point. Topsoil samples
(0–20 cm) from Yanggu were collected from 226 different grids in 2008 (Fig. 1).
Some sampling point locations varied slightly from the grid center. These changes
were based on specific topography, land use, and soil type provided by 1:50,000
scale topographic maps, 1:100,000 scale soil maps, and 1:10,000 scale aerial
photos. Soil samples were obtained by mixing five subsamples from each site
within a 20 20 m area and were recorded for the central point position using
Ecological Risk Evaluation of Heavy Metal Pollution in Soil in Yanggu 921
All soil samples were air-dried, ground, and sieved through a 2-mm nylon sieve to
remove coarse materials and debris. Then a portion of each sample was ground with
a mortar and pestle until all particles passed through a 0.15-mm nylon sieve. For
922 Y. Sun
analysis of total heavy metal concentrations, 1 g of each dry soil sample was
digested in a Teflon tube using a mixture of perchloric acid (HClO4), nitric acid
(HNO3), and hydrogen fluoride (HF). The solution of each digested sample was
analyzed by inductively coupled plasma atomic absorption spectrometry
(ICP/AAS) for the following heavy metals: As, Cd, Cr, Cu, Ni, Pb, and Zn (Meng
2014). Standard reference materials were obtained from the Center of National
Standard Reference Materials of China, and blank samples were used with each
batch of samples (one blank and one standard for each ten samples) for quality
assurance and quality control. The analytical precision values for the tests, mea-
sured as relative standard deviations, were all less than 10 %. All samples were
analyzed in duplicate and results were accepted when the relative standard devia-
tion was within 5 %. The results met the accuracy demand of the Technical
Specification for Soil Environmental Monitoring HJ/T 166–2004 ((SEPAC) 2004).
We used the Hakanson potential ecological risk equations. The equations used to
compute the potential ecological risk of each heavy metal and the integrated
potential ecological risk were the following:
where Tir is the toxic response factor of the individual heavy metal. The
corresponding values of Cd, As, Pb, Ni, Cu, Cr, and Zn were 30, 10, 5, 5, 5, 2,
and 1, respectively. Cfi ¼ Ci =Cri is the pollution coefficient of each heavy metal; Ci
is the concentration of each heavy metal; Cir is the recommended values of heavy
metal concentration in soils. The paper took the recommended values of Shandong
Province (Table 1).
A combination of different heavy metals can result in a relatively greater total
potential ecological risk. The Hakanson classification is shown in Table 2.
To obtain the spatial distribution of a heavy metal using the simulation method, we
initially tested for satisfaction of the assumption of multivariate normality. The
histograms of each heavy metal were generated (Figs. 2 and 3). Cd and Cr had
Ecological Risk Evaluation of Heavy Metal Pollution in Soil in Yanggu 923
almost normal distributions, with low skewness and kurtosis values close to 3. The
other five heavy metal values were log transformed to satisfy requirements for the
normal distribution.
YG-AS-Variogram YG-Cd-Variogram
3.38 7.618E-04
Semivariance
Semivariance
2.53 5.714E-04
1.69 3.809E-04
0.84 1.905E-04
0.00 0.000E+00
0.00 16666.67 33333.33 50000.00 0.00 16666.67 33333.33 50000.00
YG-Cr-Variogram YG-Zn-Variogram
60.2
42.4
Semivariance
45.1
Semivariance
31.8
30.1
21.2
15.0
10.6
0.0
0.0
0.00 16666.67 33333.33 50000.00
0.00 16666.67 33333.33 50000.00
Separation Distance (h)
Separation Distance (h)
YG-Ni-Variogram YG-Pb-Variogram
7.57
Semivariance
12.4
5.68
Semivariance
9.3
3.79
6.2
1.89
3.1
0.00
0.0 0.00 16666.67 33333.33 50000.00
0.00 16666.67 33333.33 50000.00 Separation Distance (h)
Separation Distance (h)
YG-Cu-Variogram
12.6
Semivariance
9.4
6.3
3.1
0.0
0.00 11568.68 23137.35 34706.03
Separation Distance (h)
The spatial distributions of the seven heavy metals were based on 1,000 simulations
(shown in Fig. 5). We used the upper limit of the background value of heavy metals
in soil established by the local government as the standard (Table 1). To present
spatial distribution details, the legend of each heavy metal had ten or more intervals.
We did not find an obvious concentration point of Pb, Cr, and Cu in the
simulation results. The areas contaminated with Zn, As, and Ni represented
25–50 % of the entire area. Cd pollution was widely spread throughout the area
with higher accumulations in eastern and southern locations.
The probability map of each heavy metal was based on the 1,000 realizations, and
the corresponding recommended threshold values listed in Table 1. Cr, Pb, and Cu
had little probability of exceeding the threshold value (Fig. 6). Some areas for As,
Zn, and Ni had probabilities >50 % for exceeding the threshold value. For Cd,
almost 80 % of the total area had a probability >90 % of exceeding the threshold
value. These results were the same as those from the spatial distribution analysis in
Sect. 3.3.
An evaluation of the Hakanson potential ecological risk was completed for each
heavy metal and the total risk using the contamination sum. The results indicated
that Cd was the main heavy metal contaminant problem in Chiping. The potential
ecological risk from Cd was nearly 80 and the risk classification was medium
(Fig. 7).
The total potential ecological risk from all seven heavy metals was 114.656, and
this level of heavy metal pollution poses a slight potential ecological risk (Fig. 8).
Ecological Risk Evaluation of Heavy Metal Pollution in Soil in Yanggu 927
3.6 Discussion
Analysis of the seven heavy metals showed that the sampling schedule was impor-
tant. The goal of simulation is to get the attribution of unsampled locations. But the
uniform 1-km grid sampling was limited. No information was available for dis-
tances smaller than 2 km. This meant that short-distance spatial variation was
poorly estimated. For example, the nugget of Pb is forced to zero in Fig. 4, where
this was not supported by the data.
In addition, cross-correlations between heavy metals should be taken into
account. We found a strong positive relationship between Cu and Zn with the
Pearson coefficient 0.714. Thus, a co-kriging approach could be used which
estimates values of cross-variograms. This might provide an improved linear
unbiased prediction of intermediate values.
4 Conclusions
Levels of seven heavy metals (Cu, Pb, Cd, Cr, Ni, Zn, and As) were studied in
agricultural soils. Spatial variation of these heavy metals was determined based on
sequential simulation methods. The ecological risk of the heavy metals was
assessed using the Hakanson potential danger index. The results indicated that Cd
was the main problem in Yanggu. The potential ecological risk from Cd contam-
ination had a value of 80 and a medium risk classification. The total potential
Ecological Risk Evaluation of Heavy Metal Pollution in Soil in Yanggu 931
ecological risk in the study area from all seven heavy metals had a value of 114.656,
and this total constitutes a slight ecological risk.
Acknowledgments This work was supported by the following projects: (1) the National Natural
Science Foundation of China (No. 41,301,509, 41,271,413), (2) China Postdoctoral Science
Foundation funded project (No. 2014 M561950), and (3) Qingdao Municipal Bureau of Human
Resource and Social Security (No. 201,409). The authors thank LetPub (www.letpub.com) for its
linguistic assistance during the preparation of this manuscript.
References
(SEPAC), S. E. (2004) The technical specification for soil environmental monitoring HJ/T
166–2004. China Environmental Press, Beijing
Goovaerts P (1997) Geostatistics for natural resource evaluation. Oxford University Press,
New York
Goovaerts P (1998) Geostatistics in soil science: state-of-the-art and perspectives. Geodema
89:1–45
Li SQ, Yang JL et al (2014) Atmospheric deposition of heavy metals and their impacts on soil
environment in typical urban areas of Nanjing. China Environ Sci 34(1):22–29
Liu HL, Yin CQ et al (2010) Distribution and speciation of heavy metals in sediments at a littoral
zone of Meiliang Bay of Taihu Lake. China Environ Sci 30(3):389–394
Lu JS, Zl Z et al (2012) Sources identification and hazardous risk delineation of heavy metals
contamination in Rizhao city. Acta Geograph Sin 67(7):971–984
Meng F, Ding N et al (2014) Assessment of heavy metal pollution in Chinese suburban farmland.
Pol J Environ Stud 23(6):447–454
Muyessar Trudi JA (2013) Distribution characteristics of soil heavy metal content in northern
slope of Tianshan Mountains and its source explanation. Chin J Eco-Agric 21(7):883–890
Sun YJ, DN (2008) Spatial analysis of heavy metals in surface soils based on GeoStatistics.
Geoinformatics 2008 and joint conference on GIS and built environment: monitoring and
assessment of natural resources and environments, pp 714514-714514-8
Zheng J, Jl Z (2012) Assessment of farmland soil heavy metal pollution in Urumqi Midong sewage
irrigation region. Arid Environ Monit 26(1):17–21
Comparison of Trend Detection Approaches
in Time Series and Their Application
to Identify Temperature Changes
in the Valencia Region (Eastern Spain)
1 Introduction
conducted using serially independent simulation data and did not address confi-
dence levels and power for serially dependent data. In Yue et al. (2002b) the authors
proved that the TFPW-MK test has a larger power than that of PW-MK test with AR
(1) simulated time series; performances were not assessed for correlated time series
with a skewed distribution.
In spite of the very large volume of literature comparing the performances of
various tests, there has been very little exploration of these six statistical tests and
especially of the assumptions under different models. In this paper, we use Monte
Carlo experiments to conduct a comprehensive comparison of the performances of
rank-based, LR and modified MK tests with respect to the confidence level and
power in order to assess the ability to detect trend in the presence and absence of
serial correlation. To do so we use two groups of simulated time series – one with
and one without serial correlation.
The remainder of this paper proceeds as follows. A brief description of
approaches employed in this study is given in Sect. 2. Section 3 compares the
performance of statistical tests using a simulation study. A real case study is
presented in Sect. 4. Section 5 provides conclusions from the study.
2 Methodologies
The statistical tests employed in this study are introduced in Sects. 2.1 and 2.2.
The MK test (Kendall 1975; Mann 1945) is a rank-based non-parametric test. The
MK statistic S is calculated as:
Xn1 X n
S¼ i¼1 j¼iþ1
sgn xj xi ð1Þ
where j>i, xj and xi are the values at i and j, respectively; n is the length of the time
series, and
8
< 1 if θ > 0
sgnðθÞ ¼ 0 if θ ¼ 0 ð2Þ
:
1 if θ < 0
h Xm i
VarðSÞ ¼ 1=18 nðn 1Þð2n þ 5Þ k¼1
t k ð tk 1 Þ ð2tk þ 5 Þ ð3Þ
where m is the number of tied values and tk is the number of observations in the kth
value.
The standardized test statistic Z is computed by
8 pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffiffiffiffiffiffiffi
< ðS 1Þ= VarðSÞ if S > 0
Z¼ 0 pffiffiffiffiffiffiffiffiffiffiffiffiffiffi if S ¼ 0 ð4Þ
:
ðS þ 1Þ= VarðSÞ if S < 0
where di ¼ rxi – ryi; rxi and ryi are the rank numbers of xi and yi, respectively; and
n is the length of the time series. When there are ties in rxi and ryi the average rank
is used. Positive (negative) ts indicates an increasing (decreasing) trend. At an α
significance level the null hypothesis will be rejected if jts j > tð1α=2Þ, n2 .
The PW-MK test is proposed by Von Storch (1995) and implemented in the
following steps.
Comparison of Trend Detection Approaches in Time Series and Their. . . 937
i. Compute the lag-1 correlation coefficient (r1) (Salas et al. 1980) of the time
series (x1, x2, . . . . . ., xn) with a significance level of α.
ii. If the lag-1 correlation coefficient is significantly different from zero at the
specified significance level of α, the original time series will first be
pre-whitened as ðx2 r 1 x1 , x3 r 1 x, . . . . . . , xn r 1 xn1 Þ, then MK test is
applied to the pre-whitened time series. Otherwise, the MK test is directly
applied to the original time series.
The TFPW-MK procedure (Yue et al. 2002b) is applied in the following manner
to identify a trend in a time series.
i. Estimate the slope of the original time series using the approaches of Sen (1968)
and Theil (1950).
ii. Compute the lag-1 correlation coefficient (Yue and Pilon 2004) of the time
series with a significance level of α. If the lag-1 correlation coefficient is not
significant, the time series is treated as serially independent and the MK test is
applied to the original time series. Otherwise, it is treated as serially correlated,
and the pre-whitening technique is applied to the de-trended time series.
iii. The MK test is applied to the blended series obtained by combining the
de-trended series and the estimated slope of the trend.
In Yue and Wang (2004) a procedure is proposed for correcting the variance of
the MK test statistic S by using an effective sample size (ESS) that takes into
account the impact of serial correlation on the variance of S. The variance V(S)*
modified by ESS is
where n is the actual sample size (ASS) of the actual sample data; n* is the ESS; Var
(S) is the variance of S which can be calculated using Eq. 3. The formula for
computing n* for the lag-1 autoregressive process is given in (Matalas and
Langbein 1962) as:
n* ¼ n= 1 þ 2 ρnþ1
1 nρ21 þ ðn 1Þρ1 = nðρ1 1Þ2 ð9Þ
3 Simulation Study
At the significance level of α ¼ 0.05 for a two-tailed test, the six statistical
approaches were applied to each time series with and without drift. The confidence
level and the power were then computed. For the sake of brevity, as the test results
are insensitive to different sample sizes, we show in this study the experimental
results for sample size 60.
Comparison of Trend Detection Approaches in Time Series and Their. . . 939
Fig. 1 Four realizations of simulated data used in the simulation study. (a) A realization of
Gaussian1; (b) A realization of Gaussian3 with drift specified by σL/σH which is equal to 0.3; (c) A
realization of Chi-square1; (d) A realization of Chi-square1 with drift specified by σL/σH which is
equal to 0.3
3.2.1 Confidence Level of Statistic Tests for Six Types of Time Series
Without Drift
In this subsection, we assess the influence of different time series in the trend
detection analysis. For this purpose, we employ the simulated time series data
without drift. The confidence interval is calculated for measuring purposes. Results
are summarized in Fig. 2, which shows the estimated confidence levels of the six
statistical tests for Gaussian1, Gaussian2, Gaussian3, Chi-square1, Chi-square2 and
Chi-square3, respectively, in the first, second, third, fourth, fifth and sixth columns.
In particular, Gaussian1 and Chi-square1 are simulated time series without serial
correlation while the other four are time series with serial correlation. In this Monte
Carlo simulation, the nominal confidence level is 95 % as the nominal significance
level is set to 0.05. Using these experimental results, we now compare the perfor-
mance of the six tests with respect to the presence or absence of time series
correlation.
Comparing the first (Gaussian1) and fourth (Chi-square1) columns in Fig. 2, it
can be seen that the six statistical tests obtain almost the same confidence level
(around 95 %) and achieve the nominal confidence level for independent time series
(Guassian1 and Chi-square1). Thus, for time series without serial correlation,
irrespective of the type of distribution (Gaussian or Chi-square), the performances
of the six tests for trend detection are very similar.
For the simulated data for correlated time series, shown in the second, third, fifth
and sixth columns in Fig. 2, the performances of all the six tests are similar. In
particular, (1) two of the modified MK tests, PW-MK and TFPW-MK, have almost
the same confidence level, which is slightly lower than that of the VC test; (2) the
confidence level of the VC test is similar to each of the other four correlated time
series and achieves the nominal value (95 %); (3) the confidence level of the three
modified MK tests is obviously higher than that of the rank-based tests and the LR
test when the time series is serially correlated.
940 H. Wang et al.
Fig. 2 Confidence levels of the six statistical tests for time series without drift at the significance
level of α ¼ 0.05
On the other hand, the statistical tests perform differently for the correlated time
series. With the exception of the VC test, the second and third columns show that
the confidence levels of the six statistical tests for time series with short correlated
range (Gaussian2) are higher than the corresponding confidence levels for time
series with long correlated range (Gaussian3). This is also the case for the fifth
(Chi-square2) and sixth (Chi-square3) columns. In particular, the confidence levels
of the PW-MK and TFPW-MK tests for time series with long correlated range is
slightly lower than the corresponding values of these two tests for time series with
short correlated range. The differences are small and can be accepted. Similar
results were obtained in Yue et al. (2002b).
Thus, the three modified MK tests demonstrate almost the same ability to detect
no trend in a correlated time series without drift.
3.2.2 Powers of Statistical Tests for Independent Time Series with Drift
In this section, we summarize trend detection with the six tests in order to evaluate
their performances on time series without serial correlation and with drift. The
power criterion is calculated for measuring purposes. Rows 1–5 of Table 1 present
the power of the six tests for uncorrelated normally distributed time series (Gauss-
ian1) with different drifts specified by σL/σH. It can be seen that the six statistical
tests have almost the same power. Thus, for independent normally time series with
drift, the six tests demonstrate similar abilities to detect trend.
Rows 6–10 of Table 1 show the power for uncorrelated skewed distributed time
series (Chi-square1) with different drifts specified by σL/σH. It can be seen that the
Comparison of Trend Detection Approaches in Time Series and Their. . . 941
Table 1 Power of six statistic tests for Gaussian1 and Chi-square1 with different drifts specified
by ratios (σL/σH) at the significant level of α ¼ 0.05
Data type Ratio MK PW-MK TFPW-MK SRC LR VC
Gaussian1 0.1 0.14 0.14 0.14 0.14 0.14 0.14
0.2 0.37 0.37 0.37 0.37 0.39 0.37
0.3 0.65 0.65 0.65 0.65 0.67 0.65
0.4 0.85 0.84 0.84 0.84 0.89 0.85
0.5 0.94 0.94 0.94 0.94 0.97 0.94
Chi-square1 0.1 0.58 0.58 0.58 0.58 0.15 0.59
0.2 0.93 0.92 0.93 0.93 0.43 0.93
0.3 0.99 0.99 0.99 0.99 0.73 0.99
0.4 1.00 1.00 1.00 1.00 0.89 1.00
0.5 1.00 1.00 1.00 1.00 0.96 1.00
rank-based tests and the modified MK tests have almost identical powers. However,
the LR test has the lowest power among the three groups of tests when drift is
present in time series. Thus, for uncorrelated skewed time series with drift, LR has
the lowest ability to detect trend.
Table 1 indicate that the LR test is sensitive to the distribution of uncorrelated
time series with drift. The other five statistical tests have almost the same ability to
detect trend irrespective of whether the time series is Gaussian or Chi-square.
3.2.3 Powers of Statistical Tests for Correlated Time Series with Drift
We conducted the trend detection with the six statistical tests to investigate their
performance on correlated simulated time series and with drift. The powers were
computed for four types of correlated time series (Gaussian2, Gaussian3,
Chi-square2, Chi-square3) with drift and are shown in Table 2.
For the four types of correlated time series with drift in Table 2, common
experimental results for the six statistical tests are (1) The LR test and rank-based
tests have significantly higher powers than those of the modified MK tests. This is
because positive serial correlation within a time series can increase the possibility
of a trend (Von Storch 1995). Thus, although the LR and rank-based tests have
more power than the modified MK tests, they are not our choices when correlated
time series data with drift; (2) the power of the TFPW-MK test is higher than that of
other two modified-MK tests (PW-MK and VC). This agrees with the studies in
Von Storch (1995) and Yue and Wang (2004) that show that the PW-MK and VC
tests underestimated the drift when both trend and positive serial correlation are
present in a time series. However, in our case, the power of PW-MK is higher than
that of VC for four correlated time series. This implies that, for correlated time
series with drift, the VC test is more likely to underestimate the drift than the
PW-MK test; (3) powers of the PW-MK and VC tests for time series with short
serial correlation (Gaussian2 and Chi-square2) are greater than the corresponding
942 H. Wang et al.
Table 2 Power of six statistic tests for Gaussian2, Gaussian3, Chi-square2 and Chi-square3 with
different drifts specified by ratios (σL/σH) at the significant level of α ¼ 0.05
Date type Ratio MK PW-MK TFPW-MK SRC LR VC
Gaussian2 0.1 0.37 0.08 0.08 0.38 0.40 0.07
0.2 0.47 0.12 0.14 0.48 0.50 0.11
0.3 0.63 0.18 0.20 0.63 0.65 0.17
0.4 0.75 0.25 0.29 0.76 0.77 0.22
0.5 0.85 0.32 0.39 0.85 0.87 0.29
Gaussian3 0.1 0.66 0.09 0.10 0.67 0.69 0.04
0.2 0.71 0.10 0.12 0.72 0.75 0.04
0.3 0.74 0.11 0.21 0.74 0.77 0.05
0.4 0.78 0.13 0.30 0.79 0.80 0.06
0.5 0.82 0.16 0.40 0.83 0.86 0.07
Chi-square2 0.1 0.60 0.07 0.09 0.62 0.63 0.04
0.2 0.68 0.09 0.14 0.69 0.72 0.05
0.3 0.79 0.13 0.21 0.79 0.79 0.09
0.4 0.84 0.17 0.30 0.85 0.86 0.12
0.5 0.91 0.21 0.42 0.91 0.92 0.17
Chi-square3 0.1 0.73 0.14 0.17 0.74 0.76 0.03
0.2 0.76 0.16 0.21 0.77 0.79 0.04
0.3 0.82 0.17 0.28 0.83 0.84 0.06
0.4 0.86 0.20 0.37 0.87 0.88 0.08
0.5 0.91 0.23 0.48 0.91 0.92 0.09
values for time series with long serial correlation (Gaussian3 and Chi-square3).
This is because the estimated lag1 in a time series with long range is greater than
that in a time series with short range. A clear linear positive relationship between
lag1 and drift is reported in Yue and Wang (2002).
On the other hand, different performances are observed for the statistical tests for
correlated time series with drift. The results show that the power of TFPW-MK for a
time series with short range (Gaussian2) is similar to that of a time series with long
range (Gaussian3). This is also the case for Chi-square2 and Chi-square3. These
results for a Gaussian time series are similar to those reported in Yue et al. (2002b).
The results show that the TFPW-MK test can effectively eliminate the serial
correlation in the Chi-square long range time series. This can be considered as a
supplement to the results in Yue et al. (2002b) as the authors did not consider
Chi-square time series.
Thus, we can conclude that the TFPW-MK test has a greater ability than the
PW-MK and VC tests to detect trend in correlated time series.
Comparison of Trend Detection Approaches in Time Series and Their. . . 943
4 Case Study
In this study the six statistical tests were used to assess the significance of trend in
mean annual temperatures measured at 13 weather stations located in the Valencia
region in Eastern Spain. The spatial distribution of the stations is illustrated in
Fig. 3. Of 13 time series, four mean annual temperature time series for stations 1, 5,
8 and 11 are shown in Fig. 4. Weather stations identified by number, X (east-west)
coordinate, Y (north-south) coordinate, altitude, time series period, skewness and
kurtosis are shown in rows (1–7) of Table 3 (Note that actual Y co-ordinate is
Y þ 4,000,000). For a normally distributed series, the skewness and kurtosis should
be approximately 0 and 3, respectively. From Table 3 it can be seen that the
experimental data are negatively skewed and are poorly described by a normal
distribution.
The lag-1 correlation coefficient and its lower and upper confidence interval
limits at the 0.05 significance level are shown in rows (8–10) of Table 3. All mean
annual temperature time series are serially correlated except that for station 2. The
magnitude of the slope, computed using Theil-Sen (Sen 1968; Theil 1950), is
shown in column (11) of Table 3. The P-values (Yue and Pilon 2004) of MK,
SRC, LR, PW-MK, TFPW-MK, VC tests are shown in column (14–17). The
Fig. 3 Locations of 13 weather stations in the Valencia region (Eastern Spain). In the figure,
triangles are station locations and the red circle is the city of Valencia
944 H. Wang et al.
P-values of the modified MK tests are obviously higher than those of the rank-based
and LR tests for all stations except station 2. In addition, the VC test has the highest
P-values for almost all stations. We conclude that the results of this real case study
are entirely consistent with the simulation results.
5 Conclusions
This study investigated the performances of six commonly used trend detection
approaches (two rank-based, LR and three modified MK tests) on two groups of
simulated time series with and without serial correlation. The theoretical compo-
nent of the study used Gaussian and Chi-square distributions as representative
symmetrical and skewed distributions; used semi-variogram models with different
ranges to quantify correlation in time series; and used linear forms of drift to asses
trends. We used confidence level and power to quantify the ability of the six tests to
detect trend in serially correlated, and in uncorrelated, time series. The Monte Carlo
simulation experiments document the performances of the different statistical tests
performances on different simulated time series. As an outcome of this work we
recommend the following strategy for choosing a suitable statistical test: (a) assess
the distribution of the time series (Gaussian or otherwise); (b) determine whether
Table 3 Basic information and P-values of mean annual temperature at 13 stations
Station (1) 1 2 3 4 5 6 7 8 9 10 11 12 13
X Co-ord. 680807 684514 712468 718904 701067 700857 629220 707758 726991 717204 726514 714909 750113
(2)
Y Co-ord. 283759 248681 239983 250120 205830 212364 377457 300798 340568 373677 373419 414204 427117
(3)
Altitude 515 500 43 81 1 2 652 350 12 69 11 364 35
(4)
Period (5) 1972–2011 1977–2011 1967–2011 1946–2011 1927–2006 1947–2006 1966–2010 1941–2011 1987–2011 1966–2011 1938–2010 1945–2001 1976–2009
Skewness 0.06 0.56 0.11 0.29 0.62 0.11 0.26 0.24 1.4 0.05 0.23 0.32 0.2
(6)
Kurtosis 2.75 2.7 1.81 2.45 3.55 2.4 2.85 2.44 4.89 2.19 2.28 2.9 2.05
(7)
Lag-1 (8) 0.37 0.21 0.71 0.49 0.72 0.65 0.58 0.47 0.51 0.76 0.73 0.39 0.75
Upper 0.28 0.3 0.27 0.23 0.21 0.24 0.27 0.22 0.35 0.27 0.21 0.28 0.31
limit (9)
Lower 0.34 0.36 0.31 0.26 0.23 0.27 0.31 0.25 0.43 0.31 0.24 0.33 0.37
lim. (10)
Slope (11) 0.036 0 0.027 0.007 0.021 0.021 0.016 0.009 0.03 0.04 0.03 0.023 0.069
MK (12) 0 0.455 0 0.039 0 0 0.062 0.031 0.126 0 0 0.003 0
SRC (13) 0 0.401 0 0.039 0 0 0.029 0.022 0.081 0 0 0.05 0
LR (14) 0 0.292 0 0.067 0 0 0.072 0.020 0.042 0 0 0.015 0
TFPW- 0.026 0.455 0.282 0.301 0.013 0.022 0.396 0.351 0.491 0.167 0.017 0.03 0.229
MK (15)
PW-MK 0.028 0.455 0.292 0.315 0.018 0.025 0.388 0.335 0.491 0.182 0.023 0.032 0.234
(16)
VC (17) 0.001 0.463 0.079 0.148 0.008 0.025 0.209 0.129 0.251 0.042 0.001 0.032 0.010
946 H. Wang et al.
the time series is correlated; and (c) choose a test from the analysis in Sect. 3 based
on (a) and (b). We also remark here that our case study could form the basis of a
feasible way of exploring proper test techniques for detecting trend.
References
Dahmen ER, Hall MJ (1990) Screening of hydrological data. International institute for land
reclamation and improvement (ILRI), Netherlands
Gilbert RO (1987) Statistical methods for environmental pollution monitoring. Van Nostrand
Reinhold Company, New York
Kendall MG (1975) Rank correlation methods. Griffin, London
Khaliq MN, Ouarda TBMJ, Gachon P, Sushama L, St-Hilaire A (2009) Identification of hydro-
logical trends in the presence of serial and cross correlation: a review of selected method and
their application to annual flow regimes of Canadian rivers. J Hydrol 368:117–130
Khaliq MN, Ouarda TBMJ, Ondo JC, Boee B (2006) Frequency analysis of a sequence of
dependent and/or non-stationary hydrometeologyical observations: a review. J Hydrol
329:534–552
Lins HF, Slack JR (1999) Streamflow trends in the United States. Geophys Res Lett 26:227–230
Mann HB (1945) Nonparametric tests against trend. Econometrica 13:245–259
Matalas NC, Langbein WB (1962) Information content of the mean. J Geophys Res 67:3441–3448
Salas JD, Delleur JW, Yevjevich V, Lane WL (1980) Applied modelling of hydrologic time series.
Water Resources Publications, Littleton
Sen PK (1968) Estimates of the regression coefficient based on Kendall’s tau. J Am Stat Assoc
63:1379–1389
Sonali P, Kumar DN (2013) Review of trend detection methods and their application to detect
temperature changes in India. J Hydrol 476:212–227
Theil H (1950) A rank-invariant method of linear and polynomial regression analysis. Proc K Ned
Acad Wet 53:1397–1412
Von Storch VH (1995) Misuses of statistical analysis in climate research. In: von Storch H,
Navarra A (eds) Analysis of climate variability: applications of statistical techniques. Springer,
Berlin, pp 11–26
Yu YS, Zou S, Whittemore D (1993) Non-parametreic trend analysis of water quality data of river
in Kansas. J Hydrol 150:61–80
Yue S, Pilon P (2004) A comparation of the power of the t test, Mann-Kendall and bootstrap tests
for trend detection. Hydrol Sci J Sci Hydrol 49:21–37
Yue S, Pilon P, Cavadias G (2002a) Power of the Mann-Kendall and Spearman’s rho tests for
detecting monotonic trends in hydrological series. J Hydrol 259:254–271
Yue S, Plion P, Phinney B, Csvadias G (2002b) The influence of autocorrelation on the ability to
detect trend in hydrological series. Hydrol Process 16:1807–1829
Yue Y, Wang CY (2002) Applicability of prewhitening to eliminate the influence of serial
correlation on the Mann-Kendall test. Water Resour Res 38
Yue S, Wang CY (2004) The Mann-Kendall test modified by effective sample size to detect trend
in serially correlated hydrological series. Water Resour Manag 18:201–218
Zhang X, Harvey KD, Hogg WD, Yunzy TR (2001) Trend in Canadian streamflow. Water Resour
Res 37:987–998
Part VII
Big Data
Urban Dynamic Estimation Using Mobile
Phone Logs and Locally Varying Anisotropy
1 Introduction
The usage and ubiquity of mobile phones in all countries and social groups are
generating an unprecedented amount of behavioral data. According to the Interna-
tional Telecommunication Union (ITC 2015), by the end of 2015, there are more
Call detail records are the essential data stored for billing purposes by all telecom-
munication companies. In Table 1 we can observe some of the fields stored in this
database, with sample data attached. Telefónica Movistar stores these records using
a large number of column fields (more than 60) and keeps each day of records in a
long-term storage (more than 1 year of data). In this study only voice and data
records are collected. The voice CDR contains the duration, origin antenna, and
destination antenna, storing the antenna IDs registered at the beginning of the call.
The data CDR contains the size in kilobytes of the transaction and the origin
antenna, but in this case the timestamp of the event can be registered with a delay
fluctuating between 0 and 60 min on average. Although this delay can affect the
final results, theoretical studies (Song et al. 2010) have shown that even using
events with timestamp errors of 60 min approximately, the patterns of movement of
each mobile phone can be predicted with 93 % of accuracy. The CDR dataset, in
voice and data, has inherently low spatial (at antenna/tower scale) and temporal
resolution (not enough to estimate velocities or other short-range properties), but
contains the mobility patterns of all active cellular subscribers without support of
additional data, like GPS or network monitoring tools. This feature makes the CDR
dataset very attractive to perform studies on mobility for urban planning and
transportation policies, among many other use cases.
Regarding the antennas and towers, each antenna is attached to a specific tower,
colloquially known as base transceiver station (BTS). In Fig. 1, top, we can observe
several towers located in a central neighborhood of Santiago, Chile. Each tower can
have attached several antennas of different technologies (2G, 3G, or 4G), as shown
in Fig. 1, bottom. As mentioned in the previous section, there are algorithms that
can decouple the location of each antenna from the tower, increasing the number of
geographical positions with valuable data in the map. However, in this work only
the tower-based approach is used, where all antennas attached to the same tower
have the same geographical positions. Future versions of the method will explore
estimations increasing the number of locations for each antenna in the map,
including the antenna orientation and other geometrical and physical parameters.
With both data sources, the CDRs and the tower/antenna positions, we can count
the number of events registered in each tower/antenna in each second of the day.
For a fixed timestamp, we can define Z(x) as the random field that represents the
number of events in the position x. If {xa} is the set of positions where the towers/
antennas are located (without repetition, in case of erroneous tabulation of the
tower/antenna positions) and are registering CDR events, we can estimate the
number of events in the geographical zone of study (the complete city or some
specific regions) using geostatistical tools like inverse distance weighting (IDW)
and simple or ordinary kriging (Chilès and Delfiner 2012). In the related literature,
typically inverse distance weighting (IDW) is applied for the estimation (Horanont
and Shibasaki 2008; Horanont 2012), since no variographic analysis is involved in
the workflow. As a drawback, the estimation is highly smooth and does not reflect
the landscape features, such as empty lots or large open spaces, where the proba-
bility of finding a mobile device is relatively low compared to a street or a highway
(Fig. 2).
As the populations of interest are devices in movement around the city, only
outdoor antennas are used and a mobility filter is applied, where a minimum
different tower indicator is calculated for each device (if the device does not register
events in at least N different antennas on different towers, it is not included in the
mobility analysis). The indoor antennas capture with higher probability the events
of devices that are not in movement in the streets; for this reason they are not
included in the study (the inverse is not true, since an outdoor antenna can capture
events from indoor devices, but there are no methods to infer this using CDRs).
Since the average distance between two towers fluctuates between 400 and 2,500 m
approximately, the filter value is set to N ¼ 2, meaning that each device must
register a movement of at least 400 m between two different towers (approximately
20 % of the devices in this study can be dropped). Although the mobility filter works
on the majority of cases, there are some special behaviors of the cellular network
where a device can be “jumping” between different antennas located in the same or
different towers, even if the device is not in movement. This behavior is triggered
by network congestion and antenna misconfiguration, and according to recent
Urban Dynamic Estimation Using Mobile Phone Logs and Locally Varying Anisotropy 953
Fig. 1 Top: towers of antennas located in a neighborhood of Santiago, Chile. Bottom: zoom in a
tower with several antennas attached
954 O.F. Peredo et al.
studies (Li et al. 2016), it represents less than 3 % of the observed records in
network data and an order of magnitude less in the aggregated CDR dataset.
Fig. 4 Zoom in to streets extracted from the OpenStreetMap platform. The grid matches with the
orange region of interest in Fig. 3
volume of traffic are motorway, primary and secondary. Using these representative
categories, the idea is to calculate the azimuth angle of each street segment that
intersects with a user-defined grid. In Figs. 3 and 4 the user-defined grid is
represented by orange regular squares. In Fig. 5 we can observe all the intersections
with their corresponding azimuth values (arrow orientation w.r.t. north) and also a
zoom to a particular region of the same grid, where the intersection of the street
segments and the grid square sides can be observed.
A standard IDW interpolation is applied to the azimuth values of the sparse 2D
field from the previous section. After that, a full 2D LVA field is reconstructed as
shown in Fig. 6. We can observe that the streets with wide azimuth are colored in
white, and the streets with narrow azimuth in black. With the full 2D LVA field, we
956 O.F. Peredo et al.
Fig. 5 Top: primary and secondary streets in the zone of interest (red arrows). Bottom: zoom in to
observe the intersections between the street segments and the grid square sides (arrows with
orientation)
can use LVA-based geostatistical codes (Boisvert and Deutsch 2011) to obtain the
estimation, adding the ratio between the X and Y axes.
4 Technical Aspects
Up to this point, the CDR data have been assumed to exist in an aggregate form that
can be used to extract the event number Z(xa) in each sampled position xa
corresponding to tower/antenna locations. However, it is not straightforward to
obtain these values using standard tools for data management, like text-plain or
relational databases. For example, on October 6th of 2014, Telefónica Movistar
generated approximately 91,088,956 CDR registers only in data traffic (24 GBytes
of plain text) for all contract users located in Chile. Just in Santiago, the capital of
Chile, there are approximately one million mobile devices with Movistar contracts,
which are producing 38 % of the total amount of data traffic CDRs. If several days
must be aggregated, a large storage facility must be used to handle the data. For
instance, each period of 6 months of data recording can increase an estimated of
4 TBytes the total space in the storage devices. Using data replication, in order to
backup the data in a different storage, at least 8 TBytes of space are needed to
handle the new 6-month CDR dataset. To solve the previous problem, we use
Hadoop technologies (White 2009), mainly because of the cheap pricing and the
on-demand features delivered by cloud computing providers, like Amazon Web
Services (Amazon 2008) or Microsoft Azure (Microsoft 2010). The CDRs are first
collected with an Extract-Transform-Load automatic process, which selects only
the columns of interest, as shown in Table 1. Those trimmed registers are stored in a
Hadoop distributed filesystem (HDFS) for long-term persistency using data redun-
dancy in different geographical locations. Finally, a Hadoop-based query is
launched periodically over the last days stored in the HDFS in order to collect the
event traces (timestamp and antenna ID) of each device, ordered by day and time.
The aggregated data can occupy space in the storage that is one or two order of
magnitude less than the raw CDRs (from 24 GBytes to 100 Mbytes in 1 day of
CDRs geo-located in antennas based in Santiago, Chile).
In terms of computing capacity, the query that aggregates the CDRs and the
estimation using the LVA process are the most challenging. The query is
implemented in Pig (Gates 2011), a high-level data processing language that allows
958 O.F. Peredo et al.
defining complex data flow patterns using syntax similar to SQL. The advantage of
this processing language is that it can use different back end distributed/parallel
computing technologies without changing the data flow specification or significant
portions of the code. Among the available back end technologies, we can mention
MapReduce (Dean and Ghemawat 2008), Tez (Saha et al. 2015) and Spark (Zaharia
et al. 2010). By using HDFS, the Hadoop distributed file system as storage tech-
nology, combined with a data processing language like Pig, we can launch instances
of clusters of computers using on-demand services in order to execute the query and
store the aggregated results in a storage container. For instance, Microsoft Azure
provides the service HDInsight (Microsoft 2013), which consists in a cluster of
computers, masters, and workers, all of them ready to use with an implementation
of Hadoop and the data processing tools. To allow the HDInsight instance to query
the raw CDR dataset, the access credentials to the corresponding HDFS storage
account must be previously loaded in the cluster management services.
Regarding the LVA process, we use a modified version of the LVA kriging code
developed by Jeff Boisvert (Boisvert and Deutsch 2011). This version uses multi-
core support (Chandra et al. 2001) allowing multiple threads to execute algebraic
operations related with the L-ISOMAP multidimensional scaling. Although the
multi-core support accelerates the computation of the distance matrix in the
multidimensional space, it is only useful the first time the code is executed. For
large estimation grids, the parallelization of the estimation loop as shown in Peredo
et al. (2015) is not yet developed. The distributed capacities of the on-demand
cluster can be exploited further to accelerate the estimation of different aggregated
dataset corresponding to different times of the day. We use MPI to distribute the
data preparation and execution of several LVA estimations of different timestamps
for each day. Using a time granularity of 1 min, a maximum of 1,440 LVA
estimations can be executed efficiently in parallel up to the total number of CPU
cores and depending on the free memory space in each node of the cluster. With this
schema, we can use an on-demand service to query and aggregate the raw CDRs
and with the same computational resources execute all the LVA estimations using
the CPU cores of the nodes in parallel. The final results, expressed as GEO-EAS or
CSV files containing the LVA estimations for each timestamp of the day, can be
stored in the HDFS persistent storage for further usage.
The proposed methodology is applied to the city of Santiago, Chile. In the first
scenario, the region of interest is described in Fig. 4, which corresponds to a zone
with 83 towers of antennas in a total area of 10 km2. The total number of event
traces registered on 12 August of 2015 in this zone is 129,598.
The northeast part in this zone is a business area with high density of tower/
antennas; it includes many office buildings, shops, and retail stores. The west part
corresponds to a residential area with low density of tower/antennas. The middle
Urban Dynamic Estimation Using Mobile Phone Logs and Locally Varying Anisotropy 959
Fig. 7 Comparison of estimation contours between the IDW (top) and ordinary kriging using the
LVA field with ratio r1 ¼ 0.01 (bottom). The small triangles represent the towers of antennas with
event numbers in the range [0,125] in each case. The time of these events is 6:08 PM
and southeast parts are mixed areas where the population can be traveling to other
zones of the city. Figure 7 shows two estimations, inverse distance weighting (top)
and ordinary kriging using the LVA field (bottom) generated with the primary and
secondary streets. The kriging parameters are depicted in Table 2.
960 O.F. Peredo et al.
We can observe that the local anisotropy of the underlying urban landscape is
qualitatively more accurate using the LVA method. In Fig. 8, we can observe
different zones of the LVA estimation at different times. At 3:11 pm, more events
are registered in the commercial zone, and at 8:55 pm more events are registered in
the residential zone.
In order to compare quantitatively both estimations, IDW and LVA, a ground-
truth data must be necessary, for example, GPS or network monitoring data. With
the ground-truth data, we can locate with high precision each device at the specific
time of study. According to Calabrese et al. (2011) and Leontiadis et al. (2014),
extensive GPS campaigns and large volumes of network data are needed to build
the ground-truth model data. In our case, we have developed a special Android
application which collects the GPS position of the mobile phone each 60 s, together
with the local area code (LAC) and cell ID (CI) which are the basic network
parameters to identify the antenna sector of influence. Our idea is to measure a
specific day in the region of interest, using several devices at our disposal, tracking
their positions and antennas where the connections are made. With that ground-
truth data, we can approximately pair the sparse CDR events with some of the GPS
Urban Dynamic Estimation Using Mobile Phone Logs and Locally Varying Anisotropy 961
Fig. 9 Azimuth angles in Santiago metropolitan area. Using primary and secondary objects (left)
and using the motorway object (right)
and transportation, the most common estimation method is the inverse distance
weighting, without including secondary information from the streets in the urban
landscape. To the best of the author’s knowledge, this work is the first application of
LVA-based geostatistical estimation in the context of mobile data geo-located in
towers of antennas. In our scenario, the mobile data corresponds to billing infor-
mation, denoted call detail records (voice and data traffic), and the network of
towers with antennas which are provided by Telefónica Movistar in the city of
Santiago, Chile.
A case study is presented in a specific zone of Santiago, with total area of 10 km2.
The estimations using inverse distance weighting and ordinary kriging with LVA are
compared qualitatively. An extended quantitative comparison is left as future work
due to its complexity and the amount of ground-truth data necessary to build a pivot
comparative model. The LVA estimation represents in an accurate way the flow of
mobile devices around the city streets. In order to obtain the orientations of the
streets, we have used the OpenStreetMap platform to extract objects from the city
map, like primary and secondary streets. After that, a gridded sampling of the
orientations (azimuths) is extracted and posteriorly interpolated using standard
methods. With the full 2D LVA field estimated, standard LVA-based geostatistical
tools can be used to obtain the final results in each timestamp of the day. The process
can be repeated each day, and further analysis can be done with the urban dynamic
estimation provided by the current methodology. LVA estimation parameters, such as
the landmark grid, offset number, or variographic parameters, must be inferred
previously using data from similar past days (working days, weekends, or holidays).
The presented work is the first step in a long-term effort with the final objective
of highly accurate dynamic position estimation of mobile devices in the city using
indirect geo-localization data, such as the CDR dataset. Although there are other
technologies that can provide directly this information, such as GPS or network
monitoring systems, the advantage of the CDRs is that no extra applications need to
be active in the client (possible battery exhaustion of the mobile device) or server
side (possible saturation of the network). The proposed methodology is flexible
enough to allow other types of data sources, by just changing the geo-location and
dataset passed to the LVA-based codes. The main feature is the inclusion of the
streets into the geostatistical estimation. If stochastic simulations must be obtained
to quantify uncertainty, sequential Gaussian simulation can be executed following
the same steps described in this work, using the LVA-based corresponding code.
Nonconventional sources of information, like antenna radiation and power, or
signal propagation models for wireless communications, can be added to the
models in the form of variographic parameters or local definitions of the ratio and
azimuth values of the LVA field (small ratios for small influence zones and large
ratios for large influence zones). This topic will be studied in future versions of the
proposed methodology.
Acknowledgments The authors would like to acknowledge the project CORFO 13CEE2-21592
(2013-21592-1-INNOVA_PRODUCCION2013-21592-1), Telef onica Investigaci
on y Desarrollo
SPA and Telef
onica Chile SA for the support and data provided in this study.
Urban Dynamic Estimation Using Mobile Phone Logs and Locally Varying Anisotropy 963
References
Oliver N, Matic A, Frias-Martinez E (2015) Mobile network data for public health: opportunities
and challenges. Front Public Health 3:189
Peredo O, Ortiz JM, Herrero JR (2015) Acceleration of the Geostatistical Software Library
(GSLIB) by code optimization and hybrid parallel programming. Comput Geosci
85:210–233, PA (December 2015)
Saha B, Shah H, Seth S, Vijayaraghavan G, Murthy A, Curino C (2015) Apache Tez: a unifying
framework for modeling and building data processing applications. In: Proceedings of the 2015
ACM SIGMOD International Conference on Management of Data (SIGMOD ‘15). ACM,
New York, pp 1357–1369
Saravanan M, Pravinth S, and Holla P (2011) Route detection and mobility based clustering. In:
Proceedings of IEEE 5th international conference on Internet Multimedia Systems Architec-
ture and Application (IMSAA), pp 1–7
Song C, Qu Z, Blumm N, Barabási A-L (2010) Limits of predictability in human mobility. Science
327(5968):1018–1021
Soto V, Frias-Martinez V, Virseda J, Frias-Martinez E (2011) Prediction of socioeconomic levels
using cell phone records. Proceedings of User Modeling, Adaption and Personalization: 19th
International Conference, UMAP 2011, Girona, Spain, July 11–15, 2011. Ed. Springer, Berlin,
pp 377–388
Traag VA, Browet A, Calabrese F, Morlot F (2011) Social event detection in massive mobile
phone data using probabilistic location inference. Privacy, Security, Risk and Trust (PASSAT)
and 2011 I.E. Third International Conference on Social Computing (SocialCom), 2011, pp
625–628
Toole JL, Colak S, Sturt B, Alexander LP, Evsukoff A, González MC (2015) The path most
traveled: travel demand estimation using big data resources. Transp Res Part C: Emerg Technol
58(Part B):162–177
White T (2009) Hadoop: the definitive guide, 1st edn. O’Reilly Media, Sebastopol
Zaharia M, Chowdhury M, Franklin MJ, Shenker S, Stoica I (2010) Spark: cluster computing with
working sets. In: Proceedings of the 2nd USENIX conference on Hot topics in cloud computing
(HotCloud’10). USENIX Association, Berkeley, pp 10–10
Part VIII
Health
Using Classical Geostatistics to Quantify
the Spatiotemporal Dynamics of a
Neurodegenerative Disease from Brain MRI
R. Marschallinger (*)
Interfaculty Department of Geoinformatics Z_GIS, Universität Salzburg, Schillerstr. 30, 5020,
Salzburg, Austria
Department of Neurology, Christian Doppler Medical Centre, Paracelsus Medical University,
Ignaz Harrer-Straße 79, 5020 Salzburg, Austria
e-mail: [email protected]
M. Mühlau
Department of Neurology, Klinikum rechts der Isar, TU München, Munich, Germany
P. Schmidt
Department of Statistics, Ludwig-Maximilians-University München, Munich, Germany
P. Atkinson
Faculty of Science and Technology, Engineering Building Technology, Lancaster University,
Lancaster LA14YR, UK
E. Trinka • S. Golaszewski • J. Sellner
Department of Neurology, Christian Doppler Medical Centre, Paracelsus Medical University,
Ignaz Harrer-Straße 79, 5020 Salzburg, Austria
1 Introduction
2 MRI Data
Fig. 2 Geometrical transformation of individual brain geometry (left) to MNB geometry (right).
Axial slice of MS-affected brain at top of ventricle level
WML presents rather heterogeneously across patients not only with regard to the
number and overall volume but also with regard to spatial pattern, predilection
sites, and shape of single lesions. Routine radiological evaluation reports the
number of WML, the total lesion load (volume), and the associated changes as
from follow-up investigations. Our goal is to describe the spatial structure of a
WML pattern in MNB geometry with standard geostatistical parameters
(Marschallinger et al. 2014, 2016).
1 X
n
Variogram function γ ðhÞ ¼ ððzðxi Þ zðxi þ hÞÞ2 ð1Þ
2nðhÞ i¼1
970 R. Marschallinger et al.
eð3jhjÞ
Exponential variogram model γ ðhÞ ¼ c 1 ð2Þ
a
Since the MNB can be considered a Euclidean space which is then dissected into
equally sized voxels, the constant extent and support of the MNB enable rapid, grid-
based variography (Eq. 1), variogram modeling (Eq. 2), and the sensitive compar-
ison of longitudinal and cross-sectional studies.
For efficient characterization of WML from MRI, we developed the following,
fully automatic processing pipeline that builds on a pair of T1 and FLAIR MRI
images, acquired on a 3-tesla scanner:
1. From FLAIR and T1 sequences of a single MRI investigation, automatically
derive a binary 3D model of the WML pattern in MNB geometry (e.g., Fig. 3
right). Software used is Lesion Segmentation Tool (LST) (Schmidt et al. 2012).
2. Compute directional empirical variograms (Eq. 1) in x,y,z directions (dextral-
sinistral, caudal-rostral, dorsal-ventral orientations). Variograms are confined to
lag distances from 0 to 15 mm, because this area holds the most relevant correlation
information and a variogram model can be fitted straightforwardly. Software used
is Geostatistical Software Library (GSLIB) (Deutsch and Journel 1997).
3. Fit an exponential variogram model function (Eq. 2) to each directional empir-
ical variogram to yield associated range and sill parameters. These characterize
overall surface smoothness, total lesion volume (TLL, total lesion load), and
preferred spatial continuity of a white matter lesion (WML) pattern. Software
used is R (R development core team 2008).
4. Derive scatterplots of ln(range) vs. ln(sill) to portray WML patterns in a space
defined by total pattern surface smoothness vs. total lesion volume. For the
current purpose, “surface smoothness” is defined as the ratio of total WML
pattern volume/total WML pattern surface. This space, making up the so-called
Fig. 3 Left: 3D model of white matter of MS-affected brain, normalized to MNB geometry (white
matter visualized in translucent style to show some MS lesions for orientation in 3D). Right:
associated WML model, white matter striped off, and axial slicing plane at ventricle level
indicated for orientation. MNB geometry
Using Classical Geostatistics to Quantify the Spatiotemporal Dynamics of a. . . 971
Fig. 4 Combining geostatistical Range ln(a) and geostatistical Sill ln(C) of 259 MS-affected,
geometrically normalized brains in the MS-Lesion Pattern Discrimination Plot (LDP). Four WML
pattern 3D models and their associated positions in the LDP are indicated. nLesions: number of
individual lesion objects per WML pattern, coded by symbols. See text for details (After
Marschallinger et al. 2016)
972 R. Marschallinger et al.
Fig. 5 WML evolution in three MS patients, three time steps each chronology from left to right,
time interval 6 months each. (See text for details, compare with Fig. 6)
WML anisotropy. Working through Fig. 4 shows that complex patterns with many
lesions or patterns with a “rough”/“complex” surface generally are positioned at the
left fringe of the point cloud while patterns with few, big, and “smooth” lesions are
placed toward the right border. Patterns around the long axis of the elliptic cloud
mediate between rough and smooth extremes.
More clinically relevant, the LDP can be used to portray the evolution of WML
during single-patient follow-up. Figure 5 shows WML patterns from MRI follow-
ups of three patients. Each patient was scanned three times at 6-month intervals,
yielding a total observation period of 1.5 years per patient. Patients are arranged in
rows, and columns represent MRI investigations (time series advancing from left to
right, also indicated by WML colors red-yellow-green). Using the abovementioned
processing pipeline, the evolution of the three individual WML patterns can be
Using Classical Geostatistics to Quantify the Spatiotemporal Dynamics of a. . . 973
Fig. 6 WML pattern evolution paths of three MS patients in the LDP. Arrows indicate evolution
direction. Compare Fig. 5 (See text for details)
visualized and quantitatively described by means of the LDP (Fig. 6). In Fig. 5,
Patient 1 shows a decrease in MS lesion sizes and concurrent decrease in surface
smoothness due to decomposition of big MS lesions into smaller ones; accordingly,
in the LDP, the evolution path arrows point to lower volumes and a less smooth
overall surface (Fig. 6). About the same is due for Patient 3; the WML pattern of
which shows volume loss at the cost of lesion pattern smoothness. This is more
pronounced between scans two and three where the biggest lesion continues to
decrease but new small lesions show up, yielding approximately constant volume.
Patient 2 shows increasing lesion volume at increasing WML pattern surface
smoothness caused by confluence of small lesions into bigger ones.
4 Conclusions
A workflow and processing pipeline has been established that enables the charac-
terization of MS-related white matter lesion (WML) patterns from MRI data by
estimation of geostatistical parameters’ range and sill. These are the bases for
representing a WML pattern in the MS-Lesion Pattern Discrimination Plot
974 R. Marschallinger et al.
(LDP). The LDP is a versatile framework that combines WML pattern volume,
WML pattern surface smoothness, and geometrical anisotropy information in a
single, well-arranged plot. Major changes as well as subtle fluctuations in
MS-lesion pattern geometry can be visualized straightforwardly. The LDP provides
precise insight into the spatial development of WML patterns (i.e., selective
growth/shrink in specific directions) without requiring object-based characteriza-
tion. The LDP is considered an EDA tool that informs on the spatial/spatiotemporal
properties of WML patterns in cross-sectional and longitudinal studies and in
monitoring medication efficacy.
References
Collins DL, Zijdenbos AP, Kollokian V et al (1998) Design and construction of a realistic digital
brain phantom. IEEE Trans Med Imag 17(3):463–468
Compston A, Coles A (2008) Multiple sclerosis. Lancet 372:1502–1517
De Smith M, Goodchild M, Longley PA (2007) Geospatial analysis. Matador, Leicester, 491 pp
Deutsch CV, Journel AG (1997) GSLIB Geostatistical software library and user’s guide. Oxford
University Press, 380pp
Filippi M, Rocca MA (2011) MR imaging of multiple sclerosis. Radiology 259(3):659–681
Garcia-Lorenzo D, Francis S, Narayanan S, Arnold DL, Collins L (2012) Review of automatic
segmentation methods of multiple sclerosis white matter lesions on conventional magnetic
resonance imaging. Med Image Anal 17:1–18
Marschallinger R, Golaszewski S, Kunz A et al (2014) Usability and potential of geostatistics for
spatial discrimination of multiple sclerosis lesion patterns. J Neuroimaging 4(3):278–286.
doi:10.1111/jon.12000
Marschallinger R, Schmidt P, Hofmann P, Zimmer C, Atkinson P, Sellner J, Trinka E, Mühlau M
(2016) A MS-lesion pattern discrimination plot based on geostatistics. Brain Behav 6(3), doi:
10.1002/brb3.430
Penny WD, Friston JK, Ashburner JT, Kiebel SJ, Nichols TE (2007) Statistical parametric
mapping: the analysis of functional brain images. Academic Press, London, 625pp
Polmann CH, Reingold SC, Banwell B et al (2011) Diagnostic criteria for multiple sclerosis: 2010
revisions to the McDonald criteria. Ann. Neurology 69(2):292–302
R Development Core Team (2008) R: a language and environment for statistical computing. R
Foundation for Statistical Computing, Vienna, ISBN 3-900051-07-0
Schmidt P, Gaser C, Arsic M, Buck D, Forschler A (2012) An automated tool for detection of
FLAIR-hyperintense white-matter lesions in multiple sclerosis. Neuroimage 59:3774–3783