ESTIMATION OF LEAF AREA INDEX AND CHLOROPHYLL FOR A
MEDITERRANEAN GRASSLAND USING HYPERSPECTRAL DATA
R. Darvishzadeh a,b*, A. Skidmore a, M. Schlerf a, C. Atzberger c, F. Corsi d, M. Cho e
a
International Institute for Geo-information Science and Earth Observation (ITC), Hengelosestraat 99, P.O. Box 6,
7500 AA Enschede, The Netherlands - (darvish, skidmore, schlerf)@itc.nl
b
Department of Remote Sensing and GIS, Faculty of Earth Sciences, Shahid Beheshti University, Tehran, Iran
abpjii@r.postjobfree.com
c
Joint Research Centre of the European Commission, TP 266, Via Enrico Fermi 1, 21027 Ispra (VA), Italy -
abpjii@r.postjobfree.com
d
712 Locksley Rd, Yorktown Heights, NY, 10598, USA - abpjii@r.postjobfree.com
e
Ecological Remote Sensing, Natural Resources and the Environment, Earth Observation Group, Building 33 CSIR, Pretoria, SA -
abpjii@r.postjobfree.com
Commission VII, WG VII/3
KEY WORDS: Hyperspectral, Vegetation, Estimation, Leaf Area Index, Chlorophyll
ABSTRACT:
The study shows that leaf area index (LAI) and canopy chlorophyll content can be mapped in a heterogeneous Mediterranean
grassland from canopy spectral reflectance measurements. Canopy spectral measurements were made in the field using a GER 3700
spectroradiometer, along with concomitant in situ measurements of LAI and chlorophyll content. We tested the utility of univariate
techniques, involving narrow band vegetation indices and the red edge inflection point, as well as multivariate calibration techniques,
such as partial least squares regression. Among the various investigated models, canopy chlorophyll content was estimated with the
highest accuracy (R2cv = 0.74, relative RMSEcv = 0.35) and LAI was estimated with intermediate accuracy (R2cv = 0.67). Compared
with narrow band indices and red edge inflection point, partial least squares regression generally improved the estimation accuracies.
The results of the study highlight the significance of using multivariate techniques such as partial least squares regression rather than
univariate methods such as vegetation indices for providing enhanced estimates of heterogeneous grass canopy characteristics. To
date, partial least squares regression has seldom been applied for studying heterogeneous grassland canopies. However, it can
provide a useful exploratory and predictive tool for mapping and monitoring heterogeneous grasslands.
single species was investigated. Therefore, investigation is
1. INTRODUCTION
required to assess the capability of remote sensing models when
Owing to its fast, non-destructive and relatively cheap it comes to natural heterogeneous canopies with a combination
characterization of land surfaces, remote sensing has been of different plant species in varying proportions. Mediterranean
recognized as a reliable method for estimating various grasslands are characterized by highly heterogeneous canopies,
biophysical and biochemical vegetation variables (Curran et al., and present a challenge for remote sensing applications because
2001; Hansen and Schjoerring, 2003; Weiss and Baret, 1999). the reflectance is often a mixture of different surface materials
Hyperspectral remote sensing with narrow and continuous (Fisher, 1997; Roder et al., 2007).
spectral bands that provide an almost continuous spectrum is
considered more sensitive to specific vegetation variables such The aim of this study was to examine the utility of
as leaf area index (LAI) (Hansen and Schjoerring, 2003). hyperspectral remote sensing in predicting canopy
Because of the role of green leaves in controlling many characteristics such as LAI and canopy chlorophyll content in a
biological and physical processes of plant canopies, LAI (the heterogeneous Mediterranean grassland by means of different
total one-sided leaf area per ground surface area) is a key univariate and multivariate methods. We compared narrow
structural characteristic of vegetation and thus widely used as band vegetation indices, including red edge inflection point
an indicator of vegetation status. (REIP), with partial least squares regression. The suitability of
these different methods will be analyzed in terms of their
LAI has been estimated in numerous studies by using remote prediction accuracy. Naturally, the significance of the results is
sensing in either statistical approaches or physically based valid only for Mediterranean grasslands and the biophysical
(canopy reflectance) models. Many of the previous studies, variables considered. The study is based on canopy spectral
however, are based on simulated data (Atzberger, 2004; Broge reflectance measured in a heterogeneous grassland during a
and Leblanc, 2001; Haboudane et al., 2004), on agricultural field campaign in the summer of 2005 in Majella National Park,
crops (Atzberger, 1995; Atzberger, 1997; Baret et al., 1987; Italy.
Broge and Mortensen, 2002; Jacquemoud et al., 2000; Walter-
Shea et al., 1997; Weiss et al., 2001) or on forest (Chen et al.,
1997; Fang et al., 2003; Kalacska et al., 2004; Running et al.,
1986; Schlerf and Atzberger, 2006; White et al., 1997), where
471
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
were randomly selected in each subplot, and their SPAD
2. METHODS
readings were recorded. From the 30 individual SPAD
measurements, the average was calculated (Table 1). These
2.1 Study area and sampling
averaged SPAD readings were converted into leaf chlorophyll
The study site is located in Majella National Park, Italy (latitude content (units: g cm-2) by means of an empirical calibration
41 52' to 42 14' N, longitude 13 14' to 13 50'E). The park function provided by Markwell et al. (1995). The total canopy
covers an area of 74,095 ha and extends into the southern part chlorophyll content (CCC; units: g m-2) for each subplot was
of Abruzzo, at a distance of 40 km from the Adriatic Sea. The obtained by multiplying the leaf chlorophyll content by the
region is situated in the massifs of the Apennines. The park is corresponding LAI.
characterized by several mountain peaks, the highest being
Mount Amaro (2794 m). Coordinates (x y) were randomly Measured Min Mean Max StDev Range
generated in a grassland stratum to select plots. A total of 45 variables
plots (30 m x 30 m) were generated and a GPS (Global LAI (m2 m-2) 0.39 2.76 7.34 1.50 6.95
Positioning System) was used to locate them in the field. To CCC (g m-2) 0.1 0.87 2.7 0.55 2.56
increase the number of samples in the time available, four to
five randomly selected subplots were clustered within each plot.
Table 1. Summary statistics of the measured biophysical and
This resulted in a total of 191 subplots being sampled. The 1 m
biochemical variables of grassland sample subplots (n=191);
x 1 m subplots differed in species composition and relative
CCC is the canopy chlorophyll content.
abundance while the within-subplot variability was small.
2.5 Data analysis
2.2 Canopy spectral measurements
We selected the normalized difference vegetation index (NDVI)
Fifteen replicates of canopy spectral measurements were taken
(Rouse et al., 1974) as a representative of ratio indices, and the
from each subplot, using a GER 3700 spectroradiometer
second soil-adjusted vegetation index (SAVI2) (Major et al.,
(Geophysical and Environmental Research Corporation,
1990) as a representative of soil-based indices, for the analysis
Buffalo, New York).
in this study. The narrow band NDVI and SAVI2 indices were
systematically calculated for all possible (584 x 584 = 341,056)
The fiber optic, with a field view of 25, was handheld
band combinations between 400 nm and 2400 nm. The soil line
approximately 1 m above the ground at nadir position. The
parameters were calculated from soil spectral measurement of
ground area observed by the sensor of GER had a diameter of
bare soils which were acquired from few subplots with no
45 cm and was large enough to cover the center of the subplots
vegetation. We assumed that the measured soil optical
without being influenced by the surroundings. The 15 replicate
properties were representative for the study area. Consequently,
spectral measurements taken from each subplot enabled to
the soil line parameters were considered constant for all 191
suppress much of the measurement noise by averaging the
subplots.
replicate measurements. Prior to each reflectance measurement,
the radiance of a white standard panel coated with BaSO4 and
For this study, we used two methods to calculate the red edge
of known reflectivity was recorded for normalization of the
inflection point (REIP). The linear interpolation method (Guyot
target measurements. The fieldwork was conducted between
and Baret, 1988) assumes that the spectral reflectance at the red
June 15 and July 15 in 2005. To minimize atmospheric
edge can be simplified to a straight line centered around a
perturbations and BRDF effects, spectral measurements were
midpoint between (i) the reflectance in the NIR shoulder at
made on clear sunny days between 11:30 a.m. and 2:00 p.m.
about 780 nm, and (ii) the reflectance minimum of the
chlorophyll absorption feature at about 670 nm. First, the
2.3 LAI measurements
reflectance value is estimated at the inflection point. Then, a
linear interpolation procedure for the measurements at 700 nm
In each subplot, LAI was non-destructively measured using a
and 740 nm is applied to estimate the wavelength corresponding
widely used optical instrument, the Plant Canopy Analyzer
to the estimated reflectance value at the inflection point:
LAI-2000 (LICOR Inc., Lincoln, NE, USA). A detailed
description of this instrument is given by LI-COR (1992) and
Welles and Norman (1991). In this study, measurements were
Rred edge = (R670 R780) / 2 (1)
taken either under clear skies with low solar elevation (i.e.,
within the two hours following sunrise or preceding sunset) or R red edge R 700 (2)
REIP linear = 700 + 40
under overcast conditions. The LAI measurements were taken
R 740 R 700
on the same day that the canopy spectral measurements were
made. To prevent direct sunlight on the sensor of LAI-2000,
samples of below- and above-canopy radiation were made in
where the constants 700 and 40 result from interpolation
the direction facing away from the sun (i.e., with the sun behind
between the 700 nm to 740 nm intervals, and R670, R700, R740
the operator), using a view restrictor of 45 . For each subplot,
and R780 are, respectively, the reflectance values at 670 nm, 700
reference samples of above-canopy radiation were determined
nm, 740 nm and 780 nm.
by measuring incoming radiation above the grass subplot (in an
open area). Next, five below-canopy samples were collected
The linear extrapolation method (LEM) (Cho and Skidmore,
and used to calculate the average LAI (Table 1).
2006) is based on the linear extrapolation of two straight lines
(Eqs. 3 and 4) through two points on the far-red (680 nm to 700
2.4 Chlorophyll measurements
nm) and two points on the NIR (725 nm to 760 nm) flanks of
the first derivative reflectance spectrum (D) of the red edge
A SPAD-502 Leaf Chlorophyll Meter (Minolta, Inc.) was used
region. The REIP is then defined by the wavelength value at the
to assess the leaf chlorophyll content (LCC) in each 1 m x 1 m
intersection of the straight lines (Eq. 5).
subplot. A total of 30 leaves representing the dominant species
472
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
content. The best performing indices and the band positions are
tabulated in Table 2.
Far-red line: D= m1. +c1 (3)
It can be observed from Table 2 that narrow band SAVI2 had
NIR line: D= m2. +c2 (4)
somewhat higher correlations than narrow band NDVI with the
studied variables. However, the coefficients of determination
between the grass characteristics and the indices were relatively
where m and c represent the slope and intercept of the straight
low. Studying regions where R2 0.6 for LAI and canopy
lines, respectively. At the intersection, the two lines have equal
wavelengths and D values. Therefore, the REIP, which is the chlorophyll content (CCC) revealed that LAI had a strong
wavelength at the intersection, is given by: influence on the selection of suitable bands for estimating
canopy chlorophyll content. The similarity in the observed
patterns is obviously due to the high correlation between the
( c1 c 2 ) two variables (not shown).
(5)
RIEP LEM =
(m1 m 2 )
2400
0.6
2200
2000
0.5
Partial least squares regression (PLSR) is a technique that 1800
reduces the large number of measured collinear spectral 0.4
Wavelength
1600
variables to a few non-correlated latent variables or factors 1400
while maximizing co-variability to the variable(s) of interest 0.3
1200
(Atzberger et al., 2003; Cho et al., 2007; Geladi and Kowalski, 0.2
1000
1986; Hansen and Schjoerring, 2003). The latent variables
800
represent the relevant information present in the measured 0.1
600
reflectance spectra and are used to predict the dependent
variables (here, biophysical and biochemical grass 600-***-**** 120*-****-**** 180*-****-**** 2400
Wavelength nm
characteristics). As with other linear calibration methods, the
Figure 1. 2-D correlation plots illustrating the coefficient of
aim is to build a linear model:
determination (R2) between narrow band SAVI2 and LAI.
Y=X + (6)
R2
Narrow
Variables [nm]
band VI
where Y is the mean-centred vector of the response variable
LAI NDVI 1105/1229 0.61
(grass characteristics), X is the mean-centred matrix of the
SAVI2 1998/1402 0.64
predictor (spectral reflectance), is the matrix of coefficients,
and is the matrix of residuals.
CCC NDVI 1141/1150 0.68
SAVI2 1211/1086 0.69
The optimum number of factors was estimated by leave-one-out
Table 2. Band positions and R values between the best narrow
cross-validation. A common way of using cross-validation for
band NDVI and SAVI2 (derived from 2-D correlation plots of
this estimation is to select the number of factors that minimizes
different data sets) and grass variables.
the RMSE (Geladi and Kowalski, 1986). To prevent collinearity
and to preserve model parsimony, the condition for adding an
For the best performing narrow band index, cross-validated R2
extra factor to the model was that it had to reduce the root mean
and relative RMSE (RRMSE = RMSE/mean) were computed
square error of cross-validation (RMSECV) by >2% (Cho et al.,
from linear regression models (Table 3). As can be observed
2007; Kooistra et al., 2004). In addition, coefficients of
from this table, compared with narrow band NDVI, narrow
determination (R2) between measured and predicted values in
band SAVI2 gave slightly higher R2 and lower RMSE values
the cross-validation were used to evaluate the relationships
for LAI and canopy chlorophyll content. The better
found. The PLSR analysis was performed using the TOMCAT
performance of SAVI2 compared with NDVI is probably due to
toolbox 1.01 within MATLAB (Daszykowski et al., 2007).
the fact that SAVI2 is less sensitive to external factors such as
soil background effects.
3. RESULTS
R2cv
Narrow
Variables RRMSEcv
3.1 Hyperspectral vegetation indices
band VI
NDVI and SAVI2 narrow band vegetation indices were LAI NDVI 0.60 0.34
calculated from the measured canopy reflectance spectra, using SAVI2 0.63 0.33
all possible two-band combinations. The coefficients of
determination (R2) between these narrow band vegetation CCC NDVI 0.67 0.36
indices and the grass canopy characteristics were computed. An SAVI2 0.68 0.35
illustration of these results is shown for LAI in the 2-D
correlation plot in Figure 1. The meeting point of each pair of Table 3. Performance of narrow band vegetation indices for
wavelengths in a 2-D plot corresponds to the R2 value of LAI predicting grass variables in Majella National Park, Italy.
and the vegetation index calculated from the reflectance values
in those two wavelengths. Based on the R2 values in the 2-D
correlation plots, band combinations that formed the best
indices were determined for LAI and canopy chlorophyll
473
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
R2cv
REIP method RRMSEcv
8
Linear LAI 0.49 0.39
7
interpolation
CCC 0.56 0.41
6
Estimated LAI (m m )
-2
5
2
Linear LAI 0.51 0.38
4
extrapolation
CCC 0.57 0.41
3
2
Table 4. Performance of red edge inflection point calculated
2
R =0.60
nRMSE=0.34
using different methods for predicting grass variables in Majella
1
National Park, Italy.
0
0 1 2 3 4 5 6 7 8
2 -2
Measured LAI (m m )
3.3 Partial least squares regression
Figure 2(a)
The relationships between grass variables and reflectance
3
spectra were modeled using PLSR. Cross-validated results
Estimated canopy chlorophyll content (g m )
-2
2.5
using the entire reflectance spectra as inputs are shown in
Figure 3. The optimal number of PLSR factors preventing over-
2
fitting was selected in two ways: (i) through visual inspection of
cross-validated RMSE versus the number of factors plots (not
1.5
shown), and (ii) by setting the condition that adding an extra
factor must reduce the RMSE (RMSECV) by >2%. The number
1
of factors in the final model were 4 for LAI and 5 for canopy
2
R =0.67
chlorophyll content models. Compared with other methods,
nRMSE=0.36
0.5
PLSR using entire reflectance spectra increased all R2 values
(R2 = 0.69, 0.74 for LAI and canopy chlorophyll content,
0
respectively) and decreased the Relative RMSE values
0 0.5 1 1.5 2 2.5 3
-2
Measured canopy chlorophyll content (g m )
(RRMSE = 0.32, 0.34 for LAI and canopy chlorophyll content,
Figure 2(b) respectively).
Figure 2(ab). Cross-validated prediction of grass variables in
Majella National Park, Italy, using narrow band NDVI. Left: 4. DISCUSSION
estimated LAI versus measured LAI; rights: canopy chlorophyll
content. The optimum wavebands are those reported in Table 2. The field experiment led to a large number of sample subplots
(191) with high variations in LAI. The canopy integrated
Figure 2 shows the relationships between the estimated and chlorophyll content (LAI x leaf chlorophyll content) strongly
measured LAI and canopy chlorophyll content using narrow reflects the variability of LAI and (to a lesser extent) leaf
band NDVI. From the figure, it seems that saturation starts to chlorophyll content, expressed by the high inter-correlation
occur for canopy chlorophyll content greater than 2 (g m-2) and between LAI and canopy chlorophyll content (not shown).
for LAI greater than 7(m2 m-2). Among the grass characteristics studied, canopy chlorophyll
content was most accurately estimated by nearly all of the
3.2 Red edge inflection point applied methods. The canopy chlorophyll content contains both
the structure and chlorophyll information of vegetation and can
The red edge inflection point (REIP) was calculated using two be accurately estimated by canopy spectral reflectance.
methods. As can be observed from the results reported in Table
4, the relationships between measured and estimated grass The relationship between measured and estimated LAI was
variables were not reliable using any of the methods. The R2 better explained by multivariate calibration methods (PLSR)
and relative RMSE of the grass variables obtained from the than by univariate methods such as narrow band vegetation
three methods were relatively similar. indices and REIP. This is because a two-wavelength index
utilizes only a limited amount of the total spectral information
Among the studied variables, estimation of canopy chlorophyll available in hyperspectral data (Lee et al., 2004).
content again yielded the highest R2 values and the lowest
relative RMSE. Compared with regression models developed The bands selected as the best combination of the vegetation
using the optimum narrow band indices, the REIP methods indices for LAI were found in the NIR to SWIR regions. This
produced somewhat lower accuracies.
index for LAI estimation. Moreover, the narrow band SAVI2
confirmed previous studies by researchers who suggested a
performed relatively well for canopy chlorophyll content. This
strong contribution by SWIR bands to the strength of
is due to the major influence of LAI in canopy chlorophyll
relationships between spectral reflectance and LAI (Cohen and
content and also to the fact that SAVI2 is relatively insensitive
Goward, 2004; Darvishzadeh et al., 2008; Lee et al., 2004;
to external factors such as soil background effects.
Nemani et al., 1993; Schlerf et al., 2005). Compared with the
narrow band NDVI, the narrow band SAVI2 gave somewhat
Although red edge has proved to respond more linearly to LAI
higher R2 and lower relative RMSE values for LAI. This result
and chlorophyll when compared with the classical NDVI, which
is in agreement with that of Broge and Leblanc (2001), who
often suffers from saturation problems (Danson and Plummer,
used simulated data and found SAVI2 to be the best vegetation
474
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
LAI was best estimated by partial least square regression
1995), in our study wavelengths within the red edge region
were almost absent. which utilize more than two wavelengths from the entire
spectral region (400 nm to 2500 nm) to estimate the
The PLSR model appears to be a powerful alternative to variable of interest.
univariate statistical methods (Darvishzadeh et al., 2008). SAVI2 is a potentially useful vegetation index for
Compared to the other investigated methods, it achieved extracting canopy variables such as LAI. However, the
relatively better results. It seems that important information will selection of appropriate wavelengths and bandwidths is
be lost by selecting only two wavelengths for narrow band important.
vegetation indices. Partial least squares regression provided the most useful
explorative tool for unraveling the relationship between canopy
LAI spectral reflectance and grass characteristics at canopy scale.
7
In summary, multivariate calibration methods, which until now
have only been used in a few cases concerning the remote
6
sensing of grasslands, can enhance estimates of different grass
Predicted LAI (m m )
-2
5
variables, and thus present new prospects for mapping and
2
monitoring heterogeneous grass canopies from air- and space-
4
borne platforms.
3
2
2
R =0.69
nRMSE=0 .3 2
ACKNOWLEDGEMENTS
1
1 2 3 4 5 6 7
We would like to acknowledge the assistance of the park
2 -2
Measured LAI (m m )
management of Majella National Park, Italy, and in particular
of Dr. Teodoro Andrisano. We extend our gratitude to Dr. Istiak
CCC
Sobhan for his assistance during the field campaign. Special
thanks go to Dr. Michal Daszykowski for his assistance in
Predicted canopy chloropyll content (g m)
-2
2.5
applying the TOMCAT toolbox and for his valuable comments.
2
1.5
REFERENCES
1
Atzberger, C., 1995. Accuracy of multitemporal LAI estimates
in winter wheat using analytical (PROSPECT+SAIL) and semi-
0.5
2
R =0 . 7 4
empirical reflectance models. In: Guyot, G. (Eds.), Proc.
n RMSE=0 . 3 4
0
Photosynthesis and Remote Sensing, EARSeL colloquium,
0 0.5 1 1.5 2 2.5
Montpellier, 28-30 August 1995, pp. 423-428.
-2
Measured canopy chlorophyll content (g m )
Atzberger, C., 1997. Estimates of Winter Wheat Production
Figure 3. Cross-validated prediction of grass variables in
through Remote Sensing and Crop Growth Modeling. PhD
Majella National Park, Italy, using the entire reflectance spectra
thesis, VWF Verlag, Berlin, Germany.
in partial least squares regression models. Left: estimated LAI
versus measured LAI; right: for canopy chlorophyll content.
Atzberger, C., 2004. Object-based retrieval of biophysical
canopy variables using artificial neural nets and radiative
transfer models. Remote Sensing of Environment 93 (1-2), 53-
67.
Estimation of biochemical and biophysical characteristics of
heterogonous grassland with mixtures of different grass species
Atzberger, C., Jarmer, T., Schlerf, M., K tz, B., Werner, W.,
is challenging in remote sensing (Roder et al., 2007), as the
2003. Spectroradiometric determination of wheat bio-physical
measured signal correspond to different grass species. In our
variables: comparison of different empirical-statistical
study, an indicator of this was the observed high variations in
approaches. In: Goossens, R. (Eds.), Remote Sensing in
the SPAD readings within a given subplot (not shown).
Transitions, Proc. 23rd EARSeL symposium, Belgium, 2-5 June
Nevertheless, by using hyperspectral remote sensing with a
2003, pp. 463-470.
large number of narrow spectral bands and powerful
multivariate regression techniques, the biophysical grass
Baret, F., Champion, I., Guyot, G., Podaire, A., 1987.
characteristics could be retrieved with acceptable accuracy.
Monitoring wheat canopies with a high spectral resolution
radiometer. Remote Sensing of Environment 22 (3), 367-378.
Broge, N.H., Leblanc, E., 2001. Comparing prediction power
5. CONCLUSION
and stability of broadband and hyperspectral vegetation indices
for estimation of green leaf area index and canopy chlorophyll
The most important conclusions that can be drawn from this
density. Remote Sensing of Environment 76 (2), 156-172.
study are as follows:
Broge, N.H., Mortensen, J.V., 2002. Deriving green crop area
index and canopy chlorophyll density of winter wheat from
Compared with LAI, canopy chlorophyll content was
spectral reflectance data. Remote Sensing of Environment 81 (1),
estimated with higher accuracy in all models.
45-57.
475
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
D., 2004. Leaf area index measurements in a tropical moist
Chen, J.M., Rich, P.M., Gower, S.T., Norman, J.M., Plummer,
forest: a case study from Costa Rica. Remote Sensing of
S., 1997. Leaf area index of boreal forests: theory, techniques,
Environment 91 (2), 134-152.
and measurements. Journal of Geophysical Research 102 (D24),
Kooistra, L., Salas, E.A.L., Clevers, J.G.P.W., Wehrens, R.,
Leuven, R.S.E.W., Nienhuis, P.H., Buydens, L.M.C., 2004.
Cho, M.A., Skidmore, A.K., 2006. A new technique for
Exploring field vegetation reflectance as an indicator of soil
extracting the red edge position from hyperspectral data: The
contamination in river floodplains. Environmental Pollution
linear extrapolation method. Remote Sensing of Environment
127 (2), 281-290.
101 (2), 181-193.
Lee, K.S., Cohen, W.B., Kennedy, R.E., Maiersperger, T.K.,
Cho, M.A., Skidmore, A., Corsi, F., van Wieren, S.E., Sobhan