Physics and Chemistry of the Earth 28 (2003) 103–109

www.elsevier.com/locate/pce

Assessment of climate impact on vegetation dynamics

by using remote sensing
G.J. Roerink *, M. Menenti, W. Soepboer, Z. Su
Wageningen University and Research Centre, Alterra Green World Research, P.O. Box 47, 6700 AA Wageningen, Netherlands

Abstract
Climate variability has a large impact on the vegetation dynamics. To quantify this impact a study is carried out with Normalized
Difference Vegetation Index (NDVI) satellite images and meteorological data over part of Sahelian Africa and Europe over several
years. The vegetation dynamics are quantified as the total amount of vegetation (mean NDVI) and the seasonal difference (annual
NDVI amplitude) by a time series analysis of NDVI satellite images with the Harmonic ANalysis of Time Series algorithm. A
climate indicator (CI) is created from meteorological data (precipitation over net radiation). The relationships between the vege-
tation dynamics and the CI are determined spatially and temporally. The driest areas prove to be the most sensitive to climate
impact. The spatial and temporal patterns of the mean NDVI are the same, while they are partially different for the seasonal
difference. The question whether climate impact on vegetation dynamics is the same everywhere on earth in the time and space
domain cannot be satisfactorily answered with these limited datasets.

2003 Elsevier Science Ltd. All rights reserved.

Keywords: Climate; Impact; Remote sensing; Vegetation; Dynamics; Ecosystem; Fourier; HANTS

1. Introduction relation between annual climate parameters and sea-

sonality of NDVI. Seasonality is the difference between
Climate variability has a large impact on the ecosys- the NDVI at the beginning of the growing season and
tem, but the impact is not the same for every region; it the maximum NDVI. NDVI data are corrected by
will be stronger for regions with a delicate balance be- Fourier smoothing algorithms (FA) and solar zenith
tween climate and ecosystem, like the Sahelian or parts angle (S) corrections to further remove extreme ano-
of the Mediterranean region. Vegetation is a very sen- malous signals from global 1 NDVI datasets. They
sitive part of the ecosystem for climate change. Both the conclude that the best predictors of FAS-NDVI sea-
growing season and the total amount of vegetation, sonal extremes are growing degree days (GDD) and
together called the vegetation dynamics, are strongly annual precipitation total (PPT), both cumulative indi-
affected by climatic changes. ces. They also support the hypothesis that the capacity
Since the use of satellite data has become a common for plant communities to produce relatively high bio-
process in the quantitative description of vegetation mass amounts, depends more on the total availability of
growth (Menenti et al., 1993; Justice and Hiernaux, heat and moisture over the same year, than on the
1986), satellite data are also increasingly used for relat- conditions during one month of that year, or the
ing vegetation indices to climatic parameters. Examples anomalous conditions of the previous year.
of such studies are: (i) Richards and Poccard (1998), These previous studies make clear that climate impact
who conclude that mean seasonal Normalized Differ- on vegetation growth has to be analysed on a yearly
ence Vegetation Index (NDVI) is in accordance with the basis. To do this, a study is carried out with NDVI and
mean seasonal rainfall over the majority of southern meteorological data over part of Africa and Europe over
Africa, except over deserts (<300 mm/year) and (ii) several years. In particular, vegetation performance,
Potter and Brooks (1998), who analysed the empirical vegetation dynamics, climate variability and spatial and
temporal sensitivity of vegetation dynamics to climate
*
Corresponding author. Tel.: +31-317-474300; fax: +31-317-
variability will be analysed. In the framework of this
419000. study, vegetation performance is synonymous to vege-
E-mail address: [email protected] (G.J. Roerink). tation growth averaged over a particular region, while
1474-7065/03/$ - see front matter
2003 Elsevier Science Ltd. All rights reserved.
doi:10.1016/S1474-7065(03)00011-1
104 G.J. Roerink et al. / Physics and Chemistry of the Earth 28 (2003) 103–109

vegetation dynamics refers to the temporal behaviour of quantified by applying a Fourier analysis to the time
the vegetation performance in monthly, seasonal and series of NDVI images (Verhoef, 1996). If the resulting
annual time scales. Climate variability is the ensemble NDVI Fourier components (FC) are coupled to climate
of net radiation, precipitation, wind and temperature parameters, the relationship between vegetation dy-
characteristic for a regions in certain time scale (e.g. namics and climate can be established (Roerink et al.,
monthly, seasonal annual). The temporal and/or spatial 1999).
sensitivity of vegetation dynamics to climate variability The response of vegetation dynamics to climate
will be used to characterise the quantitative relationship variability can be determined simultaneously in spatial
between these two quantities in temporal and/or spatial and temporal scales. This method allows us to try to
scales. The objective of the study is to quantify the answer the question whether the earth is an ergodic
spatial and temporal relationships between climate system, i.e. whether the response of vegetation dynamics
variability and vegetation dynamics. has the same spatial and temporal patterns in relation to
climate variability?

2. Method
3. Materials
The theoretical framework and research method are
schematised in Fig. 1. Vegetation performance is deter- The case study is carried out with two datasets,
mined by two major impact sources, climate and human. covering Europe and part of the Sahel. The data char-
Besides these two, several other minor impact sources acteristics are given in Table 1. NDVI data, generated
can be determined, such as soil type and elevation. The from NOAA/AVHRR satellite images are used to
earth surface can be monitored on a daily basis by sat- monitor the vegetation dynamics. Meteorological data,
ellite sensors, such as NOAA/AVHRR and METERO- precipitation and net radiation are used to describe the
SAT. Several of these sensors provide the information climate. In Europe, data of 422 meteorological weather
for the establishment of vegetation indices, like the stations are available from the World Meteorological
commonly used Normalized Difference Vegetation Index Organisation. For the Sahel, precipitation was measured
(NDVI), which give an indication of the vegetation by 107 rain gauges. Net radiation was measured by only
performance (Goward et al., 1985). A time series of one meteorological station (Lebel et al., 1997), which is
NDVI images show the temporal behaviour of the assumed to representative for the 1
1 area, where all
vegetation performance, the vegetation dynamics rain gauges are located. The satellite imagery and loca-
(Spanner et al., 1990). The vegetation dynamics can be tion of the meteorological weather station is shown in
Fig. 3.

Climate Impact Human Impact Other Impact

4. Climate indicator
Measurables: Measurables: Measurables:
- precipitation - crop type - land cover
- net radiation - irrigation - soil type
- temperature - fertilisation - altitude Evapotranspiration is the key process in plant de-
- etc. - etc. - etc. velopment. It is mainly controlled by two meteorologi-
cal parameters, net radiation and precipitation; either
one of them is the limiting factor. The Budyko aridity
Vegetation
Performance
Table 1
Characteristics of the input datasets
Time Series of
NDVI images Characteristics Europe Sahel
Spatial coverage 29N–55N, 11N–16N,
10W–20E 0E–5E
Time span 1995–1997 1992–1993
Climate Indicator HANTS analysis Satellite
- Frequency Sensor AVHRR/NOAA-14 AVHRR/NOAA-11
CI = LP/Rn - Amplitude Resolution 0.01 0.01
- Phase
Parameters 10-days-maximum- Daily NDVI
NDVI
Meteorological
Quantified relationship
between climate variability Data source 422 weather stations 107 rain gauges, 1
and vegetation dynamics meteo station
Parameters Annual averaged Annual averaged
P ; Rn P ; Rn
Fig. 1. Theoretical framework and research method of this study.
 G.J. Roerink et al. / Physics and Chemistry of the Earth 28 (2003) 103–109 105

index is the ratio between net radiation over precipita- peated until the maximum error is acceptable or the
tion times the latent heat of vaporation (Budyko, 1974; number of remaining points has become too small. For
Henning and Flohn, 1977). It is developed for dry areas, a detailed description of the HANTS algorithm one is
like deserts, which are water controlled. For wetter areas referred to Roerink and Menenti (2000) and Roerink
(radiation controlled), like parts of Europe it is better to et al. (2000).
take the inverse of the Budyko index as (Roerink et al., An example of the HANTS procedure is given in
1999) Fig. 2. Fig. 2a represents an annual temporal profile of
LP an arbitrary pixel of 10-days-max-NDVI-composites.
CI ¼ ð1Þ Fourteen cloud affected NDVI points were filtered out
Rn
during the iterative HANTS procedure, using only three
where CI is the so-called climate indicator, L is the latent frequencies (frequency 0 ¼ mean NDVI, frequency
heat of vaporisation, P is the precipitation and Rn is the 1 ¼ annual NDVI cycle, frequency 2 ¼ 6 months NDVI
net radiation. In other words, the LP term is the amount cycle). The rest of the points have a maximum deviation
of energy necessary to evaporate the available precipi- from the fitted curve of, in this case, 0.05 NDVI units.
tation P . Fig. 2b shows the FC (cosine functions) of the three
Between CI ¼ 0 and CI ¼ 1 there is shortage of water. individual HANTS frequencies; the arrows represent the
Theoretically potential evapotranspiration is reached at amplitude and phase values of the annual NDVI cycle
CI ¼ 1, where all precipitation can be evaporated by the (frequency ¼ 1).
net available radiation. When CI is larger than 1 radi-
ation is the limiting factor and water can even become
an impediment in vegetation development. 6. Results

The first step is the HANTS analysis of the NDVI

5. HANTS time series of Europe and Sahel. The number of fre-
quencies used by HANTS was set at 3 for both cases: the
In order to extract the characteristics of the vegeta- mean NDVI (frequency ¼ 0) and the yearly amplitude
tion dynamics, a Fourier analysis is performed on the (frequency ¼ 1), which represent the basic dynamics of
NDVI images. The FC reflect the start, length and the vegetation and the amplitude of 6 months (fre-
magnitude of the vegetation dynamics during the year. quency ¼ 2), which is necessary to smoothen the fit.
For this reason the Harmonic ANalysis of Time Series Fig. 3 shows the results of the HANTS procedure, which
(HANTS) algorithm is developed (Verhoef, 1996). This is a colour composite of the mean NDVI in red, the
algorithm considers only the most significant frequen- amplitude of 1 year in green and the 6 months amplitude
cies expected to be present in the time profiles, and ap- in blue. Within the covered areas a wide variety in
plies a least squares curve fitting procedure based on combinations of FC can be found, due to differences in
harmonic components (sines and cosines). For each land use and climatological conditions. Areas without
frequency the amplitude and phase of the cosine func- vegetation cover during the year are represented in
tion is determined during an iterative procedure. Input black, like the Saharan desert, the peaks of the Alps and
data points, which have a large positive or negative urban areas like Paris. Arable farmlands and savannah
deviation from the current curve (like cloudy and grasslands have strong 12 months components, which
missing pixels), are removed by assigning a weight of results in greenish colours. A high vegetation cover
zero to them. After recalculation of the coefficients on throughout the year results in reddish colours (central
the basis of the remaining points, the procedure is re- European grasslands, tropical forests). Blue colours may

0.8 0.4
measured
rejected phase
0.6 HANTS 0.2
amplitude

0.4 0

0.2 -0.2 frequency = 0 (average)

frequency = 1 (annual cycle)
frequency = 2 (6 months cycle)
0 -0.4
1 6 11 16 21 26 31 36 0 9 18 27 36
(a) (b)

Fig. 2. Principle of HANTS algorithm.

106 G.J. Roerink et al. / Physics and Chemistry of the Earth 28 (2003) 103–109

0.6

0.5

0.4

0.3

0.2

0.1

0
0 1 2 3 4 5 6

Fig. 4. Relation of CI against HANTS amplitude of frequency 0

( ¼ mean NDVI) for all weather station locations in Europe, 1995.

average of the NDVI FC components, which are plotted

in Fig. 5.

6.1. Spatial relationship

Fig. 5 shows the relation between CI and NDVI FC.

A strong correlation is found between them. The spatial
relationship between CI and NDVI FC is found similar
for Europe and Africa. From CI ¼ 0 till 1 the mean
vegetation components (frequency ¼ 0) increases stee-
ply, after which it stabilises around CI ¼ 2, where po-
tential evapotranspiration is reached. Due to irregular
precipitation occurrence and water losses to other terms
of the water balance (seepage, run-off), the actual CI
value where maximum evapotranspiration (corre-
sponding to maximum NDVI FC) is reached is much
higher than the equilibrium point (CI ¼ 1). When CI
further increases radiation becomes a limiting factor for
Fig. 3. Colour composites of the NDVI Fourier components of the vegetation development and the mean vegetation
the Sahel and Europe (red ¼ mean NDVI, green ¼ amplitude 1 year, component decreases again.
blue ¼ amplitude 6 months). The locations of the meteorological sta-
The trends are the same for the mean vegetation
tions are plotted as well.
component (frequency ¼ 0) and the seasonal difference,
which is the FC amplitude of 1 year (frequency ¼ 1),
represent strong bi-annual vegetation peaks (irrigated
areas), but in most cases it represents sharp temporal
changes in vegetation performance, like a small growing 0.5
season in savannah areas, or harvesting of arable lands. mean vegetation Europe
seasonal difference Europe
This indicates that NDVI FC can be used as a quanti- 0.4
mean vegetation Sahel
seasonal difference Sahel
tative measure for vegetation dynamics as discussed in spatial fit avg. vegetation
spatial fit seasonal diff.
the following. 0.3
The second step is to couple the calculated CI to the
NDVI FC for all available locations. Fig. 4 shows an 0.2
example for Europe, 1995. For each weather station the
CI value is plotted against the corresponding mean 0.1
NDVI, derived by HANTS. The cloud of data points
can be explained by Fig. 1; NDVI FC are not only in- 0
fluenced by climate parameters (P ; Rn ), but also by 0 1 2 3 4 5 6

human and other parameters. In this case we assume

that all these effects can be averaged out, which leads to Fig. 5. Spatial relationship between climate variability (CI) and veg-
the next step. By increments of 0.1 CI units we take the etation dynamics (NDVI FC).
 G.J. Roerink et al. / Physics and Chemistry of the Earth 28 (2003) 103–109 107

with the exception that the maximum of the seasonal 0.8

Temporal sensitivity mean vegetation
difference is found at CI ¼ 3. However, since there are Temporal sensitivity seasonal diff.
only a few data points with CI values higher than 2.5, 0.6 Spatial sensitivity mean vegetation

the found maximum is not very reliable. The resulting Spatial sensitivity seasonal diff.

HANTS values of the 6-months FC components are not 0.4

used here, because their primary use was to create a

0.2
better fit and not as indicator of vegetation dynamics.
The relationship between CI and NDVI FC is fitted
0
according to the following equation:
NDVI FC ¼ a þ b  CI þ c  d CI ð2Þ -0.2
0 1 2 3 4 5 6
Four coefficients (a, b, c and d) are used to create the fit.
Eq. (2) is composed by three parts: a represents the
Fig. 6. Spatial and temporal relationships between climate variability
fictive value on the y-axis from where the graph starts to (CI) and vegetation dynamics (NDVI FC).
increase or decrease with the value of the second part
b  CI. The last part, c  d CI , in the formula represents
the strong increase of the vegetation index from no Sahelian case, because DðNDVI FCÞ should be at least
precipitation to potential evaporation. The derived re- 0.10, to avoid the disturbing influence of measurement
gression coefficients of Eq. (2) can be found in Table 2. errors. Due to the threshold value almost all Sahelian
The spatial sensitivity of NDVI FC for CI changes and a lot of European data points are rejected. The re-
can be calculated as the first derivative of Eq. (2): maining data points are averaged in CI classes of 0.3 and
 plotted in Fig. 6, together with the derived spatial sen-
oðNDVI FCÞ  sitivity.
 ¼ b þ c  lnðdÞ  d CI ð3Þ
oðCIÞ CIðx;yÞ The mean NDVI FC sensitivity shows similar spatial
and temporal patterns. Only the first temporal (driest)
The fitted curves are plotted in Fig. 6. It is clear that the point is much higher, which means that vegetation in
driest areas (like Spain, Sahel and Sahara) have the very dry areas reacts very strongly to small changes
highest sensitivity. in CI, which is mainly caused in dry areas by changes in
precipitation.
6.2. Temporal relationship The seasonal NDVI FC sensitivity shows different
patterns. The spatial sensitivity is as expected; from a
Data is available for several years. This allows us to high sensitivity at low CI it gradually decreases to zero
analyse not only the spatial sensitivity, but also the at high CI values (Fig. 6). The temporal sensitivity
temporal sensitivity of the vegetation dynamics (NDVI shows a steep decreasing trend from CI ¼ 0 till
FC) for climate variability as CI ¼ 0.75. At CI ¼ 0.75 the sensitivity is slightly nega-
 tive, which is considered an anomaly caused by limited
oðNDVI FCÞ  DðNDVI FCÞyearX yearY
 ¼ ð4Þ data points. From CI ¼ 0.75 till CI ¼ 2.55 the temporal
oðCIÞ CIðtÞ DðCIÞyearX yearY
sensitivity increases slowly from slightly negative values
where yearX and yearY represent 1995, 1996 and 1997 in to slightly positive values, where after it slowly decreases
the case of Europe. Eq. (4) could not be applied to the again.

Table 2
Derived coefficients a, b, c and d for Eq. (2)
NDVI FC a b c d Standard error
Mean vegetation
1995 0.682 )0.121 )0.599 0.369 0.03840
1996 1.258 )0.221 )1.151 0.626 0.06561
1997 1.178 )0.162 )1.011 0.683 0.01753
3 years (class) 0.625 )0.093 )0.525 0.386 0.01183

Seasonal difference
1995 0.271 )0.247 )2.684 0.868 0.03624
1996 2.705 )0.302 )2.672 0.839 0.04425
1997 0.496 )0.041 )0.455 0.665 0.03513
3 years (class) 0.784 )0.094 )0.742 0.729 0.00831
108 G.J. Roerink et al. / Physics and Chemistry of the Earth 28 (2003) 103–109

If we translate the found temporal sensitivity patterns amount of vegetation (mean NDVI FC) shows similar
in common words, we can conclude the following: spatial and temporal patterns. This indicates that cli-
mate impact on vegetation dynamics is an ergodic phe-
• In dry areas (deserts, parts of Spain) where net radi- nomenon for the examined areas in the space and time
ation is usually abundant, the vegetation sensitivity is domain.
very high to yearly precipitation changes; both the However, the sensitivity of the seasonal vegetation
total amount of vegetation and the seasonal differ- difference (annual NDVI FC) has different spatial and
ence between dry and wet season increase with in- temporal patterns. Both have the highest sensitivity in
creasing precipitation. A vegetation cycle starts to dry areas, but where the spatial pattern gradually de-
develop with the appearance of the first rainfall. creases to zero in very humid areas, the temporal pattern
• In wetter areas (most of Europe) increasing precipita- has after the high sensitivity in dry areas, almost no
tion still results in higher amounts of vegetation, sensitivity in moderately humid areas and a small neg-
however the seasonal difference remains almost con- ative sensitivity in the wettest areas.
stant; the possible increase or decrease of the amount In this case study the results give an ambiguous an-
of vegetation in winter and summer is the same. This swer to the question if climate impact on vegetation
indicates that in such areas, combined effects of pre- dynamics is an ergodic phenomenon over the whole
cipitation and net radiation are the limiting factors earth. Larger datasets in time and space are needed to
for vegetation development. give a more profound answer to this question.
• In very humid areas (mountains, Northern Europe)
where precipitation is excessive, the major limiting
factor to vegetation development is net radiation.
Acknowledgements
Both the total amount of vegetation and the seasonal
difference becomes smaller.
The authors would like to thank the Free University
of Berlin (Germany) for the provision of the NOAA/
AVHRR 10-days-max-NDVI composites of Europe.
7. Conclusion
They are also grateful to the reviewers for their con-
structive comments.
From this study it can be concluded that the HANTS
algorithm is an appropriate tool to describe the tem-
poral behaviour of remote sensing parameters, like the
NDVI. The major advantage of HANTS is that the FC References
reflect the cloud and error free properties of the vege-
tation cycle. Budyko, M.I., 1974. Climate and Life. Academic Press, New York,
508 pp.
The HANTS algorithm has been applied to time se- Goward, S.N., Tucker, C.J., Dye, D.G., 1985. North American
ries of NDVI images of Europe (1995–1997) and Sahel vegetation patterns observed with the NOAA-7 advanced very high
(1992–1993). A climate indicator (CI) has been formu- resolution radiometer. Vegetatio 64, 31–40.
lated from meteorological data (precipitation over net Henning, D., Flohn H., 1977. Climate aridity index map. In:
radiation). A strong correlation has been found between Explanatory note of United Nations Conference on Deser-
tification. 29 August–9 September, 1977, vol. A/CONF 74/31,
these FC of the vegetation dynamics (NDVI FC) and pp. 7–9.
the CI. Of the considered FC the mean NDVI and the Justice, C.O., Hiernaux, P.H.Y., 1986. Monitoring the grasslands of
yearly amplitude prove to be sensitive indicators of cli- the Sahel using NOAA AVHRR data: Niger 1983. Int. J. Remote
mate variability. The 6 months amplitude is less suitable Sensing 7, 1475–1497.
as a climate indicator, since it was chosen to smoothen Lebel, T., Taupin, J.D., DÕAmato, N., 1997. Rainfall monitoring
during HAPEX-Sahel. 1. General rainfall conditions and clima-
the fit and not to reflect the vegetation dynamics. tology. J. Hydr. 188–189, 74–96.
The spatial analysis over Europe revealed a clear re- Menenti, M., Azzali, S., de Vries, A., Fuller, D., Prince, S., 1993.
lationship between the climate indicator and the NDVI Vegetation monitoring in Southern Africa using temporal Fourier
FC. From no precipitation to the point where potential analysis of AVHRR/NDVI observations. In: Proceedings, Int.
evapotranspiration is reached the NDVI FC increase Symp. on Remote Sensing in Arid and Semi-Arid Regions. LIGG,
Lanzhou, China, pp. 287–294.
strongly; if CI increases even further, the net radiation Potter, C.S., Brooks, V., 1998. Global analysis of empirical relations
becomes a limiting factor and the FC decrease again. In between annual climate and seasonality of NDVI. Int J. Remote
the case of Sahel the relationship between CI and NDVI Sensing 19, 2921–2948.
FC could only be established for very dry areas (low CI Richards, Y., Poccard, I., 1998. A statistical study of NDVI sensitivity
values). The found relationship is remarkably similar to to seasonal and interannual rainfall variations in Southern Africa.
Int J. Remote Sensing 19, 2907–2920.
the European case. The sensitivity of the vegetation Roerink, G.J., Menenti, M., Su, Z., 1999. A method for assessment of
dynamics for climate variability is spatially and also interannual climate variability by using Fourier components. Paper
temporally examined. The sensitivity of the total presented at C.9, International Geoscience and Remote Sensing
 G.J. Roerink et al. / Physics and Chemistry of the Earth 28 (2003) 103–109 109

Symposium Hamburg, 28 June–2 July 1999, IGARSSÕ99, pp. 681– Spanner, M.A., Pierce, L.L., Running, S.W., Peterson, D.L., 1990. The
693. seasonality of AVHRR data of temperate coniferous forests: rela-
Roerink, G.J., Menenti, M., Verhoef, W., 2000. Reconstructing tionship with leaf area index. Remote Sensing Environ. 33, 97–112.
cloudfree NDVI composites using Fourier analysis of time series. Verhoef, W., 1996. Application of Harmonic Analysis of NDVI Time
Int J. Remote Sensing 21 (9), 1911–1917. Series (HANTS). In: Azzali, S., Menenti, M. (Eds.), Fourier
Roerink, G.J., Menenti, M., 2000. Time series of satellite data: analysis of temporal NDVI in the Southern African and American
development of new products. BCRS-report NRSP-2 99-33, Delft, continents. Wageningen, The Netherlands, DLO Winand Staring
The Netherlands, 101 p. Centre, Report 108, pp. 19–24.