Scribd Logo
Upload

Welcome to Scribd!

  • 12 views

    Geostatiscs Applied To Groundwater Contamination I-Metodology

    Uploaded by

    Marcio Alberto
    Sobre geoestatística aplicada a contaminação

    Copyright:

    © All Rights Reserved
    Download as PDF or read online from Scribd

    Geostatiscs Applied To Groundwater Contamination I-Metodology

    Uploaded by

    Marcio Alberto
    12 views9 pages
    Sobre geoestatística aplicada a contaminação

    Original Title

    Geostatiscs Applied to Groundwater Contamination I-Metodology

    Copyright

    © © All Rights Reserved

    Available Formats

    PDF or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    Sobre geoestatística aplicada a contaminação

    Copyright:

    © All Rights Reserved
    Download as PDF or read online from Scribd
    Download as pdf
    12 views9 pages

    Geostatiscs Applied To Groundwater Contamination I-Metodology

    Uploaded by

    Marcio Alberto
    Sobre geoestatística aplicada a contaminação

    Copyright:

    © All Rights Reserved
    Download as PDF or read online from Scribd
    Download as pdf
    of 9
    ‘Ao FICA, an IDE COMUTS BIBLIOGR BTIDO ATRAVES DO PROGRAMA ESTE DOCUMENTO FOI Ot (CAPESISESUIFINEPIIBICT Grostatistics APPLIED TO GROUNDWATER Contamination. I: METHODOLOGY snd Jonathan D. Istok,! Associate Members, ASCE By Richard M. Cooper! ‘Assrnact:_ This paper presents an introduction for the nonspecialst to the use of geostatistics {0 estimate and map contaminant concentrations land estimation errors in a groundwater plume from a set of measured ‘contaminant concentrations. The paper begins with a bref review of the ‘essential elements of geostatistical theory. The remainder of the paper ‘describes the four steps ofa geostatistical analysis with special emphasis ‘on the interpolation technique known as point kriging, This procedure ‘can be used to obtain the best (.e., minimum estimation error, linear (Gce., the estimated concentration at an unmeasured point is given by a linear combination of nearby measured concentrations), unbiased esti- mate, and the estimation error for any point within the plume. These point values can then be mapped (L.e., contours of equal values of Expected contaminant concentration and estimation error can. be drawn), e.g, to display the extent and severity of groundwater contam- IwrrooucTion Geostatistics refers to a set of statistical procedures for describing the correlation of spatially distributed random ‘variables and for performing interpolation and areal estimation for these variables, The fundamentals of seostatistcs were developed by Krige, Sichel, and de Wij i the 1950's to estimate ore reserves inthe god fields of South Aftica (Cark 1979; David 1977: and Henley 1981). Matheron (1963) gave this largely empirical work its theoretical basis in the theory of regionalized variables ‘As shown im the second and third papers in this series (Cooper and Istok 1988a, 19886), geostatistics isa potentially powerful tool forthe analysis of contaminant Concentration data from contaminant plumes in groundwater. The principal reason is that estimation techniques based on the theory of regionalized variables can be used to obtain the best (in the sense that the éstimation error is minimized), linear (.e., estimated concentrations are fiven by linear combinations of measured concentrations), Unbiased estimates of ‘contaminant concentrations at unmeasured points in the plume based om a limited number of groundwater samples. These estimates Can be used tc» map contaminant concentrations within the plume. Further, because the estimation errors at each point are also computed, it is potentially possible 1 evaluate the "information content” ofa groundwa- ter sampling scheme, ic. to determine the value of individual measure Water Resour. Engr, Water Resour. Div., South Florida Water Mgmt. Dist P.0. Box V. West Pain Beach, FL 33403; formely, Grad. Res. Ast, Dept of Age. Enuins Orcyom Suate Unies Corvalin, OR 97330 "Aus. rok, Dept of Age. Engrs, Oregon State Univ. Corvali, OR 97331 Note Discussion open nil September 188. Separate discussions should be submited forthe individual papers in this symposium. To extend the closing date ne month, 4 wntten request ust be led withthe ASCE Manager of Journal. Ths ftunuserpi for this paper wes submited for view and posible publication on Jy Baise This paper part the Journal of Environmental Engineering, Vol 14, No. 2 Api, 1588: ASCE, ISSN 0753 9372788002 02708100 > 3.15 per page Paper No” 23349 270 ments in defining the areal extent of groundwater contamination. The estimation errors also may be useful in developing procedures for optimiz ing the design of future groundwater sampling and monitoring programs at ‘site to obtain the most additional information at the least cost, and to design efficient groundwater treatment and contaminant recovery systems {6,10 identify the optimum locations for contaminant recovery wells. ‘Several authors have reported applications of geostatistics to character- ization of the spatial distribution of contaminants in the soil. Barnes (1978) used the estimation technique known as kriging to map the distribution of four radioactive isotopes deposited on the Enewatak Atoll during nuclear ‘weapons tests conducted there. Bares et al. (1980) conducted a similar analysis for six radioactive isotopes ata nuclear weapons test site in Nevada. Myers and Bryan (1984) used kriging to map lead concentrations in soils near two lead smelters. They were able to use these maps to locate sources of lead contamination. Zirschky and Harris (1986) examined a dioxin spill along a highway. They used kriging to map dioxin concentra- tions and to identify the extent of areas requiring remedial action, Zirschky etal. (1985) used kriging to analyze the distribution of dioxin-contaminated sediments in a stream. Only a few papers have been reported in the literature concerning application of geostatistics to characterization ofthe spatial distribution of groundwater contaminants beneath waste disposal’ sites. Moore and ‘McLaughlin (1980) mapped concentrations of ruthenium-106 in a portion of 4 groundwater contaminant plume using kriging. However, the geosta tistical methodology used in this application was not complete. The type of variable chosen for their analysis, concentration in units of mass per unit volume of groundwater, is not additive, i.e., linear combinations of measured values of this ‘variable do not have the same meaning as the variable itself Gournel and Huijbregts 1978), Only additive variables can be Studied vit the tue of geostatistical methoodlegy Waed by those tions Also, the number of groundwater samples used in their analysis Was very sali cight samples), and such a small number of samples is insufficient to accurately compute the experimental semivariogram Journ! and Hui}. bregts 1978). (Additional discussion ofthe requirement of additivity and of ‘minimum sample size js given in step 2 of the Methodology section). The experimental semivariogram is used to describe the pattern of spatial correlation displayed by the data. A mathematical model is fitted to the experimental semivariogram, and this model is used in kriging to obtain estimates of the variable of interest at unmeasured points. Because the sample size was small, the experimental semivariogram computed from the data presented by Moore and McLaughlin did not fit the type of model they selected very well, and none of the available validation procedures were used to assess the goodness-oftit. A poor fit may cause the estimates obtained by kriging to be unreliable Because so few case histories of geostatistics applied to problems of groundwater contamination have been reported inthe technical literature, comprehensive methodology for this application has not been developed. ‘Although the theoretical basis for such a methodology has been thoroughly’ described in the geostatistical literature, tis literature is often inaccessible to the nonspecialist. The objective of the first paper in this series is to provide the required methodology in a form that is usable by the practicing an engineer. We begin with a brief review of the essential elements of ‘geostatistical theory and then provide a step-by-step procedure for apply~ ing this theory to problems of groundwater contamination. The second and third papers in this series (Cooper and Istok 1988a, 1988b) presents a case history that illustrates the application of the developed methodology to the ‘analysis of groundwater contaminant data from a hazardous waste site Basic Concerts Consider a groundwater contaminant plume V for which set of n values [c(e))---2(t,)] of a contaminant have been measured at the points Xy.coty, where, in general, x, = a point in three-dimensional space. Because the set of measured values is distributed in space, z(x) is called a regionalized variable (ReV). In a geostatistical analysis the ReV z(x) is Considered to be a single realization (j.c., a single set of measurements) of, the random function (RF) Z(x) for contaminant values within the plume. This terminology simply reflects the fact that 2(x) is only known exactly at the measurement points (Actually all experimental measurements are subject to measurement error. If the magnitude of this error is known, it an be included in the estimation procedure). Everywhere else within the plume z(x) is considered to be a random variable. The collection of random Variables for all points in the plume is the RF Z(x). One objective of geostatistical analysis of groundwater contaminant data frequently is to map the contaminant concentrations within the plume, i.e., to draw contours of equal contaminant concentration. To accurately map contam: inant concentrations when the number of measured values is small (the usual case) requires estimates for Z(x) at unmeasured points in the plume. For any unmeasured point x» in the plume, the best possible estimator of the random variable (RV) Z(x9) is given by the conditional expectation ELZU oN Zee ho ZO] oa o tq = the measurement points. Eq. 1 is read “the expected where x» Value at the unmeasured point x given the measured values at the set of points x), . %,-" To compute the conditional expectation requires knowledge of the joint probability distribution of the RF Z(x), i.e., the probability distributions of the n + I random variables Z(), Z(ty),-.-.Z() Gournel and Huijbregts 1978). Since the inference of the joint probability distribution is not possible when only one realization of Z(x) exists, an estimator with fewer requirements is needed. Itis usually adequate to limit the consideration of estimators to the class of linear estimators (Delfiner 1976; Olea 1975). Linear estimators require only knowledge of the first two statistical moments (the mean and variance) of the probability distribution of Z(x). It is still impossible to estimate even these moments unless this probability distribution can be ‘assumed to be stationary over the sample domain, .e., one distribution characterizes all the random variables in Z(x), and the mean of the distribution is constant at each point in the plume. In this case, measure- ments at two different points, z(x) and z(x),,), can be considered to be separate realizations of Z(x), and the ReV (2) can be used to estimate the required first two moments of Z(x) The procedures that use linear estimators and the assumptions of stationarity are referred to as linear geostatistics, and in this paper, the are term geosfatstcs-Tefers.to linear geostatistics unless specifically s (secs Is most arondvatr “plats there Waly are inst srensuficient") ‘of data to justify the use of nonlinear estimators (which require) \ kpowlesiae.of additional moment 0) Hower a) the second paper of this series (Cooper and Istok 1988a), linear estimators can pero ery wel this appation wough not requirements of linear geostatistics, for the discussions that follow two further assumptions will be made about the eaperimental data that constitute the ReV. The first assumption is that the data were collected from samples of the Same-sive ail theme ae ee See obtained using the same measurement technique [ie.. x) has constant ‘t]. The second assumption is that the samples are small relative to the domain of estimation (.., 24) Bas point support). For most sampling schemes used (o characterize groundwater contaminant plumes, these assumptions are very good. Usually the same sampling techniguc is used to obiain all samples in a particular study, and the samples are much salle than the domain oetimaton (Le, the plime and surrounding area). When samples are not similar in size or have been obtained diferent methods, or i they are not smal Compared othe domain of estimation, the sample values must first be adjusted using a procedure known as regularization (David 1977; Journel and Hujbregts 1978). ‘An example of changing support would be the ease where some groundwater -samples were obtained by bailing groundwaterjrom the well casing and others were obtained 6y pumping-vThe effective sample volume: would probably be much smal To the aed spies than forthe pumped Samples. By using regularization, the dliferenee in sample sizes car eliminated from te aanysis; ” “n" “imerence in samp a MerHopoLocy A geostatistical analysis of contaminant concentrati yntaminant concentrations from a contam- inant plume in groundwater, f four steps. The first step is to etermine ithe measured contami Sncenteaions aie aldniGe isibuted and io make appropriate iansormatianes thoy oe ms i th ‘Bat (¢g., a natural fog transformation Sten improves the heal theta “a normal distribution). The second step is to estimate th is to estimate the spatial correla. sion of pairs of measured concentrations as a function ofthe distance find sigection oftheir separation, This scaled the calculation OFT mental semivariogram. Inthe third step, called structural analyse, theoretical model is filted to the experimental semivariogram, and 7h ame ihe fourth step, the fitted model i used in an interpolation procedure kno kriging to. estimate expected values of Contaminant concentrations Tey tnmettred porte wis he plume. Some Tor ‘Step 1. Additivity and Normal Distribution nabs use of linear geostatistics requires thatthe ReV be additive, i that linear combinations of the values of the ReV z(x) have the same ‘meaning (Journel and Juijbregts 1978) ¢1n groundwater appli ions, this gneans that only goncentations expresed in untsof mass (or weight or olume) of slut per unit var should Be used. Concentra- ions expressed if units of mass (or weight Or Volunie) Of solute per unit 273 Volume of groundwater do not represent true accumulations and are not additive, ccontat concentrations obtained from groundwater sees must be-multiplied by the porosity of the aquifer af Te potnt Sie sae wag eke Ye ree 0 conver cotaiant one tions to additive variables means that iter atso ast be mi ted) for each sample location.. sis the analysis is two-dimensional, the contaminant concentrations must also be averaged over the thickness of the plume at each sample mple was taken point. Therefore, in a two-di rensional analysis, it will also be required to reasure (oF estimate) the thickness of the contaminant plume. It is also important to remember that although we express the contaminant concen- rations in units of mass per unit volume of aquifer for purposes of a geostatistical analysis, the measured concentrations represent only that Portion of the contaminant dissolved in groundwater and do not include the Portion of the contaminant sorbed to the porous media (the sorbed portion Could be computed from the contaminant concentrations and measured or timated values of the distribution coefficients if the governing chemical reactions are at equilibrium). hhasize the requitement of additivity in this paper contaminart concentrations that have been converted to_an_adaltive-Re be feferred to as contaminant densities eecond requirement of geostatistics is that the ReV be normally distributed. Although geostatistical techniques have been proposed for ‘cases where the probability distribution of the ReV deviates from normal- ity (Cressie and Hawkins 1980), they are not usually necessary, &s deviations from normality often’ can be corrected by an appropriate S {ransformation or by the elimination of erroneous measurements in the sample data (Armstrong 1984). The probability that the ReV is normally S distributed may be determined from any of several statistics, e-g., the chi-squared statistic (Henley 1981), the Kolmogorov and Smirnov statistic (Henley 1981), or the Shapiro-Wilk statistic (Delhomme 1976). Tin many cases, a log-transformation will improve the fit of the ReV toa JB normal distrbution. This transformation is given by SVs) = log (21x) + A] @ & viet ~ he natal ogee conn ents: 29 ‘constant added to Z(x) to improve the the contaminant densities; and A fit (Re 78). fF the RF of contaminant densities Z(x) has a multivariate lognormal distribution, then the RF Yx) = log Z(x) has a multivariate normal distribution. The parameters of these distributions are equated by ss (@8) o? = m[exp (B?) — 1]. where m and o? = the mean and variance of the RF Z(x); and a and the mean and variance of the RF YG) (Rendu 1978) - 4 i ‘Since the RF Y(x) is multivariate normally distributed, the best possiple estimator of the conditional variable Y(xo)|Y(x),....¥(&,) at point Xo is given by the conditional expectation E1Y()|¥(x)),(). 214 Vee) The (+ 1)-variate distributions of ¥(x) are also normal. The conditional expecta- tionof a normally distributed stationary RF is identical to the linear estimator with the smallest variance of estimation (Journel and Huijbrests 1978). Since the transformed RF Y{x) is second-order stationary and normally distributed [if Z(x) is second-order stationary], an estimate ¥*(x) Of ¥(1) at the point x) may be obtained by a linear combination of the data Values in the same way as for normally distributed data. In the equations that follow, Z(x) may be replaced by Y(x) (i.e., contaminant densities may be replaced by logransformed contaminant densities) in cases where the contaminant densities are determined to be log-normally distributed, : ae Step 2. Calculation of Experimental Semivariogram Recall that estimates of the first two moments of Z(x) are required for a linear geostatistical analysis. The first-order moment is simply the expec: tation (or mean) of Z(x). There are three second-order moments defined in geostatistics: the variance, the covariance, and the variogram. They are, in general, functions of position x. However, under the assumptions of stationarity, the covariance and the semivariogram do not depend on the positions of a pair of measurement points x, and x,.., but only on the vector h separating them (Le. = x|~ x01). These moments are defined as follows: expectation: E[Z(x}] = m © variance: Var{Z(x)] = E{LZ(x) — mP?} = C(0) -. 6 covariance: Ch) = E{{Zex + ITZ} — aac) variogram: — 2y(h) = E{{Zix +b) ~ Z(x)}*} = C10) — Cth). ® “Uh is called the semivariogram and is the form of the variogram most ‘tien used in practice Tn most situations, itis not possible to define an priori variance and covariance for the RF Z(x) as is required forthe assumption of second- border stationarity (Cooper and Istok 1988a; Zrschky and Harris 1986). In this ease, a less strict hypothesis of stationarity, the intrinsic hypothesis, can be used. The expectation and semivariogram are defined as inthe case of assumed second-order stationarity (.e., Eqs. and 8), but the variance land covariance are not used. Because the semivariogram may be applied under the less restrictive assumptions of the intrinsic hypothesis, while the variance and covariance anno, its the principal tol used in geostatistics to describe the pattern of spatial correlation displayed by the data. ‘The semivariogram defined in Eq. 8s also called the true semivariogram of the ReV. Since only one realization of the RF of contaminant densities is available inthe sample data 2), itis only possible to estimate the true semivariogram. This estimate is called the experimental semivariogram: ro) [atm] E emo 200)P oo © where N(h) the number of sample pairs separated by the vector h. 275 There are two sources of error associated with the estimate of the true semivariogram given ‘by Eq. 9 (Journel and Huijbregts 1978). Since a realization of the RF is not known for every point in the plume, the first ‘Source of error is from estimating the value of the true semivariogram from the limited number of measurements that are available. This error is called the variance of estimation, and it is a function of the number of sample pairs N(h) used to estimate (h) for each vector h. The amount of error decreases as N(h) increases. The second source of error is from fluctua- tions of the local mean (defined from point-to-point in the plume) about the ‘assumed mean for the entire plume (or for the sliding neighborhood, described later). These fluctuations cause estimates of the semivariogram to also vary from point-to-point in the plume. As a result, there will be an ‘error associated with estimating the semivariogram for the whole domain of the plume by combining the local estimates obtained for each set of sample pairs. This source of error is called the fluctuation variance and is a function of the distribution of the separations of the sample pairs. A theoretical analysis of these two sources of estimation error for the experimental semivariogram can be used to determine the minimum number of sample pairs and the maximum magnitude of the vector h for which the true semivariogram can be estimated from a set of sample values. The analysis requires that the fourth-order moments of Z(x) be known (Journel and Huijbregts 1978), and for a small number of samples, these moments cannot be computed accurately. However, practical rules for estimating the true semivariogram from a set of sample values are given by Journel and Huijbregts (1978) rule 1: (hy > 30-50 (10a) rule 2: L tny<% (100) 2 Where [h| = the magnitude of the separation vector h; and L = the longest dimension (in the direction of h) of the contaminant plume. ‘The restriction on [hj (Eq. 105) has important implications for the estimation procedure (kriging) to be described in step 4. Since ‘he experimental semivariogram cannot be computed for [hj greater than L/2, any calculations made with the fitted semivariogram model are restricied to [hl less than Z/2. One important consequence of this restriction is that kriging should not be used to estimate contaminant densities for points that are ata distance greater than £2 away from the nearest sample. ‘The restriction on the number of sample pairs (Eq. 10a) also affects the estimation of the true semivariogram. In the general case, an adequate ‘number of sample pairs can usually only be obtained by grouping sample pairs into distance and angle classes for each specified vector hy . Since in two-dimensional problems, h may be specified in polar coordinates as (hi, 6), sample pairs can be grouped by defining tolerance intervals forthe it and the 0p , i-e., Aly) and A@, (Fig. 1) (David 1977). If the magnitude and 276 x FIG. 1. Grouping Sample Pairs into Distance and Angle Classes angle describing a vector h separating the two points of a sample pair fall within the respective tolerance intervals for a specified vector hp , that ‘sample pair is retained and is used in the calculation of 7*(h). Sample pairs that do not meet these eriteria are discarded. For the example in Fig. 1, the vectors h', h?, and h* defining the separations between sample 0 and Samples 1, 2, and 3, respectively, are within the tolerances specified for hy and are used to calculate +*(h). "The sizes of the tolerance intervals forthe [hg| and the 0 are selected to account for suspected anisotropies in the spatial structure of the RF and to provide adequate sample pairs for the calculation of the various *(h). If the regionalization is believed to be isotropic, the tolerance intervals for the Oy are selected to include all angles of the ho i.e, 9 * 180°, so that the hy depend only on the respective magnitudes jh), not on the directions of the hy When the number of samples is small, itis difficult to select tolerance intervals that provide an adequate number of sample pairs for calculation of all the +/"(h) for the experimental semivariogram. In this case, the algorithm may be modified so that sample pairs are selected using only the tolerance intervals for 8. The selected sample pairs are then sorted using the value of [hi for each pair. The sample pairs are grouped into classes that contain the same number of sample pairs (specified to be =30). For each of these groups of sample pairs the average [hl [h|, is determined and the value +({hl) calculated. This procedure ensures that values of the exper- imental semivariogram are based on adequate numbers of sample pais. an Step 3. Structural Analysis: Fitting Semivariogram Model ‘Sisuctural analysis is the fitting of a mathematical model to the experi- mental semivariogram. Any mathematical function may be used to model fan experimental semivariogram as long as it is conditional positive-definite Gncreasing of constant with increasing h and nonnegative) (Olea 1975). In practice only a few types of functions are typically used. Models are of two types depending on whether or not the experimental semivariogram exhibits a sill (ve., an a priori variance). The most commonly used models are shown in Fig. 2 patial structure in a contaminant plume may be present at any scele from a few micrometers to hundreds of kilometers. However, information con the spatial correlation of z(x) will be limited to distances greater than the Smallest separation between a pair of samples, Spatial structure with a Tange of influence less than the smallest experimental value of h will be ‘Observed on the experimental semivariogram as a discontinuity at the origin, (Sampling and analysis errors also contribute to this discontinuity.) The discontinuity is termed a nugget Co , and its presence is termed the nugget effect (Fig. 3). When there is no spatial structure evident over the range of h for which the semivariogram is defined, i.e., y(h) = C(O) forall hy, the semivariogram is modeled as a pure nugget effect (Fig. 2). Recall that his a vector defined as the separation between two measured values of the ReV cia). Since there is no reason to expect the spatial structure of contaminant densities within a plume to be the same in every direction, the Semivariograms can be expected to be different for various directions of b. [fall of the semivariograms are identical regardless of the direction of hy 2{x) is said to be isotropic. The spatial correlation then depends only on the distance of separation jh), not on the direction 0. When the semivariogram js a function of the direction of h, the spatial structure of the ReV is said to be anisotropic. The isotropic structures representing the various direc tions of anisotropy are added together or nested to give the spatial risen sise0 mth) ees vt) h h Fovsoteshense [yee f mt 7H) FIG. 2. Common Semivariogram Models 278 C-voriance of the tronstion mode! €(0)=Cor C FIG. 3. Simple Nested Models and Nugget Effect: (a) Linear Model and Nugget; (b) ‘Spherical Model and Nugget, correlation for all directions [see Clark (1979) or Journel and Huijbregts (1978) for methods to reduce anisotropies to their isotropic cases). The concept of nested structures may be generalized to include the various anisotropies and the nugget effect: wnt) = YA) . ay where = the number of nested structures; and h, = the vector h transformed to make the nested structure represented by 4(hy) isotropic. Examples of two simple nested models are shown in Fig. 3. A nugget effect is nested with a linear model in Fig. 3(a) and with a spherical model in Fig. 3(6). For the latter case, the a priori variance C(0) is equal to the sum of the Value of the nugget effect Cand the value of the sill of the spherical model Ci ices, CO) = Cy + Cy The procedure for fitting a mathematical model to the experimental semivariogram consists of several steps. Preliminary estimates of model parameters (e.g., 4 and C(0)] can be obtained by eye or by the method of ‘weighted least squares (weighted because the number of sample pairs used to compute each point on the semivariogram is, in general, different). At this stage, the fitting procedure can be guided by knowledge of the physical 279 1 AKE (2 5, (Zt = ax) = 0 where Z*(x) = the estimated value at x;, 2) = the measured value ats, And nis the numberof samples. For the kriging errors to be consistent wth the kriging variances, the standardized errors should be normally distrb- uted with a mean close to zero and a variance close to one. Several different models may be compared by this method to determine the one that best represents the experimental semivariogram. The choice of model is optimized by finding the model with the minimum mean squared error (MSE) MSE =~ ¥) [2*(x) ~ 2x)? = Minimum 13) As a practical rule, the MSE also should be less than the variance of the sample values: MSE ‘The best linear unbiased estimate (and thus the conditional expectation) of Z(%9) is obtained by using Lagrangian techniques to minimize the estimation variance and then optimizing the solution of the resulting system of equations when constrained by the nonbias condition of Eq. 18. The following system of equations results from the optimization: E Ayrlvin x) + w= eo. 3s eee (194) z% where 4 = the Lagrange multiplier. This is the kriging system, and the procedure based on Eq. 19s referred to as point kriging. It isa (symmetric and full) system of n + | linear equations in n + 1 unknowns which can casily be solved using Gauss elimination. The solution yields the n weights of Eq. 16 plus the Lagrange multiplier, 4. This particular derivation also is sometimes called point kriging because it gives estimates at points in the domain of estimation, ‘The minimum estimation variance, or kriging variance, is written as Leaibess ues) of = F(Z) — 2%} Hes Fo) + EX) +H 20) For the case where Z(x) is lognormally distributed, the inverse transfor~ mation of the estimate of ¥(x9), ie., exp{¥x)], is not the best linear 281 ‘estimate of the untransformed conditional variable Z(x)|Z(x)),..-.20,) Sohere Z(4,).-..Z0t,) are the n untransformed data (Henley 1981). Instead, the estimate of the expectation and the kriging variance of Z(xq) are Obtained from the kriging estimates for Y(ro) using Eqs. 21 and 22 Journel and Huijbregts 1978): ELZ0

    You might also like