Full text of DOI:10.1080/01431160050144947

This article was downloaded by: [University Of Maryland]
On: 17 October 2014, At: 16:29
Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number:
1072954 Registered office: Mortimer House, 37-41 Mortimer Street,
London W1T 3JH, UK
International Journal of
Remote Sensing
Publication details, including instructions for
authors and subscription information:
http://www.tandfonline.com/loi/tres20
Application of spectral
mixture analysis for terrain
evaluation studies
a
a
P. Casals-Carrasco , S. Kubo & B. Babu
a
Madhavan
a
GIS Laboratory , Keio University (SFC) ,
5322Endo, Fujisawa, 252-8520, Japan
Published online: 25 Nov 2010.
To cite this article: P. Casals-Carrasco , S. Kubo & B. Babu Madhavan (2000)
Application of spectral mixture analysis for terrain evaluation studies, International
Journal of Remote Sensing, 21:16, 3039-3055, DOI: 10.1080/01431160050144947
To link to this article: http://dx.doi.org/10.1080/01431160050144947
PLEASE SCROLL DOWN FOR ARTICLE
Taylor & Francis makes every effort to ensure the accuracy of all
the information (the “Content”) contained in the publications on our
platform. However, Taylor & Francis, our agents, and our licensors
make no representations or warranties whatsoever as to the accuracy,
completeness, or suitability for any purpose of the Content. Any opinions
and views expressed in this publication are the opinions and views of
the authors, and are not the views of or endorsed by Taylor & Francis.
The accuracy of the Content should not be relied upon and should be
independently verified with primary sources of information. Taylor and
Francis shall not be liable for any losses, actions, claims, proceedings,
demands, costs, expenses, damages, and other liabilities whatsoever
or howsoever caused arising directly or indirectly in connection with, in
relation to or arising out of the use of the Content.

This article may be used for research, teaching, and private study
purposes. Any substantial or systematic reproduction, redistribution,
reselling, loan, sub-licensing, systematic supply, or distribution in any
form to anyone is expressly forbidden. Terms & Conditions of access
and use can be found at http://www.tandfonline.com/page/terms-and-
conditions

int. j. remote sensing, 2000, vol. 21, no. 16, 3039–3055
Application of spectral mixture analysis for terrain evaluation studies
P. CASALS-CARRASCO*, S. KUBO and B. BABU MADHAVAN
Keio University (SFC), GIS Laboratory, 5322Endo Fujisawa, 252-8520 Japan
(Received 26 January 1998; in Ž nal form 6 May 1999)
Abstract. In this article, we describe an approach to calculate the spectral
mixture within pixels and classify multispectral images. The results are compared
with the classiŽ ed images by traditional supervised rules such as Maximum
Likelihood and appreciable results were accomplished. The method considers the
number of endmembers that form the scene spectra, followed by the determination
of their nature and Ž nally the decomposition of the spectra into their endmembers.
The only requirement for this method is a radiometrically corrected image because
the endmembers are directly selected from the image. To make the method
presented here more e cient, we propose to apply it only to the classes having
low accuracy after a traditional supervised classiŽ cation. Because the land cover
classes in this study are related to a geomorphological terrain unit, we propose
to mask the terrain units having problematic classes and decompose these units
into their endmembers.
A geomorphological analysis of the study area (Tonle Sap basin in Cambodia)
was made to establish the relationship between land cover, landforms and soils
through terrain mapping units. Then we performed a supervised classiŽ cation of
a Landsat Thematic Mapper (TM) image and of the same image merged with a
SPOT-panchromatic (PAN) image, based on the land covers corresponding to
the terrain mapping units. Then we masked a terrain unit having problematic
spectral classes and applied the spectral mixture analysis which allowed an e cient
separation of the land cover classes agglomerated in the preliminary classiŽ cation.
The result of this re-classiŽ cation was re-inserted into the Ž rst classiŽ cation and
was compared statistically with the results obtained in the preliminary classiŽ ca-
tion. We consider this procedure an e cient method to improve the results
obtained from a supervised classiŽ cation. The method can separate diŒerent land
covers that were agglomerated in the preliminary segmentation. In our case, the
classiŽ cation accuracy for the terrain unit used (the  uvial terrace) increases from
6
2% (using only the TM bands) and 69% (using TM1 SPOT) to 83%.
1.
Introduction
ClassiŽ cation techniques have been extensively applied to automatically extract
environmental information from remotely sensed images. Many approaches focus
on the elaborate use of spectral properties to improve the classiŽ cation accuracy
(Weismiller et al. 1977, Campbell 1987). Such classiŽ cation or so called ‘a per pixel
classiŽ er’ is fundamentally based on multispectral classiŽ cation techniques which
*Present address for correspondence: SSSA Unit, Space Applications Institute, Joint
Research Centre of the EC, 21020 Ispra (VA), Italy.
International Journal of Remote Sensing
ISSN 0143-1161 print/ISSN 1366-5901 online © 2000 Taylor & Francis Ltd
http://www.tandf.co.uk/journals

3040
P. Casals-Carrasco et al.
assign a pixel to a class through considering its similarities with the class or with
other classes (Jensen 1986, Gong et al. 1992).
Although some improvement techniques can help to get higher accuracy, it is
well accepted that the per pixel classiŽ ers such as the widely used Maximum
Likelihood ClassiŽ er (MLC), do not get high classiŽ cation accuracy from high spatial
resolution data (Toll 1984, Martin et al. 1988). Still visual analysis is often reported
to obtain higher accuracy (Martin 1989, Mas and Ramirez 1996).
It is well known that information concerning landforms and soils can be inferred
from satellite images through the analysis of texture, shadows, tone and location of
features. But the main information provided by the remotely sensed data in temperate
and tropical environments comes from the land cover types and their distribution.
However, diŒerences in natural or cultivated vegetation often indicate diŒerences in
terrain condition, because most developments tend to be located on the best sites
available for any given activity (Ryden 1997).
In this work, we applied a method to calculate the spectral mixture within the
pixels of the satellite data and classify the land cover patterns which can be related
to landforms and soil for a further environmental assessment and monitoring. An
improvement in the classiŽ cation accuracy was observed in the images resulting from
the spectral mixture analysis when compared with the more traditional classiŽ cation
approaches.
2.
Objectives
The primary aim of our work is to develop a method that permits the use of
operational earth observation satellite images for the detection and repeated mon-
itoring of vegetation and soil characteristics to expedite environmental assessment
application.
The assumption of this study is that patterns of topography, soils and vegetation
are so related that one can be used to predict the others. In this sense, the present
study investigates the relation between the spectral mixture classes and the terrain
characteristics.
3.
Study area
The study area is located in the northern part of Tonle Sap basin in Cambodia,
covering an areal extent of 1200 km2. It is bounded by the Phnom Kulen range in
the north and the lacustrine plain of Tonle Sap lake in the south. The central part
of the area is occupied by the historic resort of Angkor monument (Ž gure 1).
Phnom Kulen consists of thick deposits of sandstone which have built a cliŒ
overlooking the area. From Phnom Kulen towards the south there is a gradual
decrease in the elevation. The terrain is gently sloping towards the lake and composed
of sand, silt and clay of the Quaternary Period.
The study area shows varied land covers. Jungle occupies most of the mountain-
ous ranges and vicinity. The gently sloping terrain has upland crops (hill rice, manioc
and all kind of cash crops). However, in this area, due to the shifting cultivation,
there are a few patches of degraded forest with grasses and widely spaced trees. The
lower areas and northern part of the lake have paddy Ž elds (during the dry season
most of them are fallow) with scattered patches of orchards. The lacustrine plain is
occupied by a  ooded forest of rather poor  oristic status that can be considered as
shrubland during the dry season.
The study area has a tropical climate with a characteristic subdued equatorial

Spectral mixture analysis
3041
Figure 1. Index map of study area—Angkor, Cambodia.
regime. It has a dry season from December to February when the monthly rainfall
decreases down to 10 mm and a rainy season from March onwards which reaches a
peak in July and decreases in August, having a second peak in October.
4.
Data
Geocoded Landsat Thematic Mapper (TM) image of 31 January 1995 (Ž gure 2)
and a SPOT-panchromatic (PAN) stereopair of 22 and 25 February 1995 were used
for this study. The images were radiometrically corrected by the Remote Sensing
Technology Center (RESTEC), Japan. Other data used for this study are the
‘
Geological Map of Kampuchea’, 1988 and the ‘Carte mini e` re de Cambodge’ (Bureau
de Recherches G e´ ologiques et Mini e` res 1970), both at a scale of 1:1 000 000 and the
Land Cover Map of the integrated Resource Information Center (IRIC), 1995 at a
scale of 1:25 000.
5.
Methodology
A geomorphological analysis, relating landforms, soils and land cover, was carried
out based on the visual interpretation of the SPOT-PAN stereopair and the Landsat
TM false colour composite (473).
The data pre-processing such as georeferencing the SPOT-PAN data with respect

3042
P. Casals-Carrasco et al.
Figure 2. Landsat TM image, Siem Reap area, Cambodia, December 1995.
Figure 3. Geomorphological map derived from the interpretation of Landsat TM and
SPOT-P stereopair.

Spectral mixture analysis
3043
to Landsat TM and resolution merge were performed using an ERDAS Imagine
software run on a SUN Workstation.
Unsupervised and supervised classiŽ cation methods were applied to classify TM
data and TM merged with SPOT data. The band selection for the classiŽ cation of
the TM image was made analysing the correlation and covariance matrices, combin-
ing bands which have low correlation and high covariance. To integrate the geo-
morphological coverage with the classiŽ ed images and the land cover coverages,
ArcInfo GIS software was used.
The spectral mixture analysis was then tested versus the above classiŽ cation
approach, to see how it can help to improve the results.
The present approach attempts to model spectral re ectance signatures as a
mixture of few prototype spectra (endmembers) such that
n
n
R 5 SF RE 1 E and S F 5 1
ij
(1)
i
j
i
j
j=1
j=1
where R is the re ectance of the mixed spectrum in band i, RE the re ectance of
ij
the endmember spectrum j in band i, F denotes the fraction of endmember j and E
i
j
i
the residual error in band i. A unique solution is possible as long as the number of
spectral components does not exceed the number of bands plus one (Hill 1993).
Our work follows the trend introduced by Mathieu-Marni et al. (1996) who
attempted to overcome some restrictions of previous researchers who had already
applied a linear method to extract mixture percentages within pixels: usually the
endmembers of the mixture are selected from a bank of spectra or after a radiometric
Ž eld campaign. This means that the deŽ nition of the reference spectra is external
and does not only depend on the satellite image.
To overcome the limitation introduced by the utilization of a bank of spectra,
we have chosen our reference spectra that represents the endmembers directly from
the satellite image. The steps followed to perform the spectral mixture analysis are:
(
(
(
a) Ž nding the number of prototype spectra (endmembers) that form the scene,
b) determination of their nature and
c) decomposition of the scene into their endmembers.
To make the method presented here more e cient, we propose to apply the land
cover spectral unmixture only to the classes which have shown low accuracy in the
MLC-supervised classiŽ cation. Because the land cover classes are related to a geo-
morphologic terrain unit, we propose to mask the terrain units having problematic
classes and decompose these separated units in their endmembers.
6.
Results and discussion
A geomorphologic analysis for the study area was carried out as a Ž rst step to
understand the nature of the terrain. Geomorphology considers landforms, their
associated soils and vegetation and is used as a basis for planning and management
schemes (Cooke and Doornkamp 1990).
6.1. T errain analysis of the study area
From the visual interpretation of the study area with a SPOT-PAN stereopair,
a Landsat TM image and the help of the geological and land cover maps, a
geomorphological map of 11 landform categories (Ž gure 3) was drawn up. After

3044
P. Casals-Carrasco et al.
studying the correspondence between the landforms delineated in the geomorpholo-
gical map and the other parameters considered in this study (land cover, soil type
and lithology), six major terrain units were identiŽ ed (table 1). For the purpose of
comparison, the smaller landforms of the same genetic nature have been grouped as
a single major unit. Thus the structural landforms consisting of scarps, cuestas,
hogbacks, inselbergs, plateau surfaces and higher terraces were grouped together.
From table 1, we can see that all the structural landforms consist of sandstone
and have developed acid lithosols where the jungle is the land cover. No agriculture
is possible on lithosols (Mas and Ramirez 1996). The slope wash is made up of
sandstone of lower consolidation and has developed acid lithosols. Due to human
action (logging), the soil has degraded and the land cover has changed to open forest.
The footslopes made up of Pleistocene Quaternary sands have developed podzolic
soils. The characteristics of the soils along with the gently sloping morphology of
this unit made it suitable for upland crops (hill rice, manioc and all kind of cash
crops) and the soil can be considered as a humic podzol because, in general, it has
lost the very thin upper organic mineral layer keeping only the organic material.
However, in some places of this unit, shifting cultivation practices have developed.
In these places, a thin organic mineral layer is still preserved and the soil can be
considered more speciŽ cally as ferric humic podzol.
In the lower  uvial terrace, the Holocene Quaternary sediments have developed
agricultural hydromorphic soils with diŒerent moisture content. The water retention
properties of these soils are very suitable for the paddy Ž elds (where the moisture
content is higher) and orchards (where the moisture content is lower).
In the weathered sand deposits, the moisture content of the soils is low throughout
the year and hence no agriculture can be developed.
The lacustrine plain consists mainly of wetlands with high water concentration
during most of the year and the unit is not suitable for agriculture. A rather poor
Table 1. Correspondence among the main terrain parameters.
Landform
Lithology
sandstone
Soil type
acid lithosol
Land cover
jungle
Structural
landforms†
Slope wash
Footslopes
sandstone
degraded lithosol
open forest
Pleistocene
sediments‡
humic podzols
ferric/humic podzols
upland crops
shifting cultivation
Fluvial terrace
Holocene sediments§
agricultural hydromorphic
soils (higher
moisture content)
agricultural hydromorphic
soils (lower moisture
content)
paddy Ž elds
orchards
Weathered
sand deposits
Holocene sand
alluvial sandy soil
bare land
shrubland
Lacustrine plain lacustrine deposits
gleysoils
†
‡
Scarps, cuestas, hogbacks, inselbergs, plateau surfaces and higher terraces.
Sands, silts and clays, mainly of  uvial origin.
§
Silts and sands, organic material.

Spectral mixture analysis
3045
 ooded forest develops during the rainy season, and during the dry season, this unit
can be considered as shrubland.
6.2. Supervised classiŽ cation of L andsat T M and SPOT -PAN data
As explained previously, unsupervised and supervised classiŽ cation methods were
applied to classify TM data and TM merged with SPOT data. The band selection
for the classiŽ cation of the TM image was made analysing the correlation and
covariance matrices, combining bands which have low correlation and high
covariance.
An unsupervised classiŽ cation using ISODATA clustering was performed on
the selected band combinations. The band combination which obtained the best
separability among clusters was Ž nally selected (473).
A supervised classiŽ cation of the land cover was performed on the selected band
combination using the MLC (Ž gure 4). The MLC has generally been proven to be
the one that obtains the best results for classiŽ cation of remotely sensed natural
resource data (Mather 1987, Bolstad and Lillesand 1991, Shettigara 1991).
We wanted to investigate the spectral potential of multispectral image data
integration for automatic classiŽ cation. Landsat TM sensors give a spatial resolution
of 28.5 m. SPOT-panchromatic has one broad band with very good spatial resolution
(
10 m). Combining these two images to yield a new dataset with 10 m resolution has
been shown to provide the best characteristics of both sensors (Pouncey and
Schrader 1997).
The band combination used for the previous supervised classiŽ cation (473) was
merged with SPOT-PAN using the Brovey Transform. Because of the characteristics
of our images, this transform gives us better results than other methods used to
combine images from diŒerent sources such as the HSI transform. A supervised land
cover classiŽ cation was performed using the MLC (Ž gure 5).
The Integrated Resource Information Centre of Cambodia (IRIC) land cover
map and ground data collected during our Ž eld survey in the study area were used
for the accuracy estimation in the two classiŽ cations. The overall and average class
accuracy were estimated from the confusion matrix by calculating the total percentage
of pixels correctly classiŽ ed and the average of the accuracy obtained by class
(
table 2).
Although there is some improvement merging SPOT-PAN with the TM image,
some classes such as the open forest and the shrublands, have low accuracy. The
open forest was misclassiŽ ed with the shifting cultivation and shrubland class. It
seems that the statistical characterization of some classes used in the MLC was not
su cient to separate them accurately. Besides this, most of the units only achieved
moderate accuracy.
6.3. Spectral mixture analysis
The spectral mixture analysis applied in this study for the classiŽ cation of multis-
pectral images provided less biased maps of vegetation abundance and soil character-
istics. In our study, we have chosen the  uvial terrace unit as the geographical
domain in which to apply the method proposed (Ž gure 6). Although the  uvial
terrace is not the only terrain unit having low accuracy in the supervised classiŽ cation,
we have restricted our spectral mixture analysis to this terrain unit because is the
only one for which we have a detailed land cover information allowing us to perform
a proper assessment of our results.

3046
P. Casals-Carrasco et al.
Figure 4. Supervised classiŽ cation from Landsat TM (bands 4, 7, 3).
Figure 5. Supervised classiŽ cation from Landsat TM1 SPOT-P.

Spectral mixture analysis
3047
Table 2. Results of the supervised classiŽ cation.
Obtained accuracy (%)
Land cover class
Area (Ha)
% total area
TM
TM1 SPOT
Jungle
15 874
2303
34 431
1546
37 591
12 530
9131
7089
2779
12.88
1.87
27.93
1.25
30.49
10.16
7.41
5.75
2.26
87
20
79
41
62
61
63
40
60
57
93
20
91
62
61
62
61
40
91
64
Open forest
Upland crops
Shifting cultivation
Fallow Ž elds
Orchards
Bare land
Shrubland
Water bodies
Overall accuracy
First, the principal components of the analysis were selected to construct a basis
of the subspace into which the cloud of spectra is projected (Mathieu-Marni et al.
1996). principal components I and II were used because they account for the greatest
variability of the scene (Ž gure 7).
The cloud of spectra of the selected unit projected into the two-dimensional
subspace deŽ ned by the two Ž rst principal components shows that the  uvial terrace
has four main tendencies (Ž gure 7). These tendencies were masked and their corres-
ponding pixels were identiŽ ed in the image. The land cover of the identiŽ ed pixels
was veriŽ ed with the detailed land cover map and corresponded to (a) fallow Ž elds,
(
b) bare land, (c) orchards and (d) water.
Reference spectra for the re-classiŽ cation were chosen from the vertices of the
cloud of spectra because according to Mathieu-Marni et al. (1996) these vertices
correspond to pure class pixels. The reference spectra was localized on the detailed
land cover map and on the ground during our Ž eld survey. Thus our reference
spectra was taken from the satellite image itself and do not depend on other external
factors as it would be the case if the prototype spectra would have been taken from
a spectral library or from a radiometric Ž eld campaign.
The percentage of a reference spectrum inside the pixel was computed with the
spectral distance from this pixel to the reference spectrum. The spectral distance was
normalized with respect to the reference spectra value and the percentage of the land
cover was obtained rescaling the remaining fraction to 1.
F 5 1Õ (D /DN
r(i))
(2)
j
i
where F 5 fraction image of the endmember j, D 5 spectral distance from the pixel
j
i
to the reference spectra in band i, and DN 5 value in the channel i for the reference
endmember.
r(i)
This procedure was repeated for each reference spectra and Ž nally the resulting
fraction images summed again to 1:
n
F 5 1
(3)
ž
j
j=1
This has produced a set of land cover theme images (orchards, fallow Ž elds, bare
land and water theme images) showing the distribution of the land cover in the scene
(
Ž gures 8–11). The intensity of the pixels corresponds to the percentage of the land

3048
P. Casals-Carrasco et al.
Figure 6. Selected unit in the TM image for the determination of land use mixture.
Figure 7. Projection of the cloud of spectra of the selected unit onto the plan of the two Ž rst
principal components. The spectra shows four tendencies, Spectrum S1, fallow Ž elds;
S2, bare land; S3, orchards; and S4, water.
cover present in the pixel. Finally, every pixel is modelled as a linear combination
of these endmembers such that:
Pixel(i)5 a(land cover 1)i1 b(land cover 2)i 1 ... 1 z(land cover n)i
(4)
where pixel(i) is the value of the mixed spectrum in band (i) in every single pixel,
a1 b1 ... 1 z5 1 are the fractions of the endmembers present in the scene.
A unique solution is possible as long as the number of endmembers does not
exceed the number of bands plus one (Hill 1993).
Eventually, a new classiŽ cation image was produced for the area corresponding
to the  uvial terrace unit, showing these four land cover classes (Ž gure 13) and was
re-inserted into the TM classiŽ ed image. The mask classiŽ cation shows a considerable
improvement in the spatial distribution of the land cover classes. The coherence of
the mask classiŽ cation is shown by the correlation of the position of the re-classiŽ ed

Spectral mixture analysis
3049
Figure 8. Orchards theme image deduced from the percentage values.
Figure 9. Water theme image deduced from the percentage values.
pixels with their position in the preliminary classiŽ cation (Ž gure 14(a) and (b)). For
instance, the orchard pixels in the mask classiŽ cation are close to the orchard pixels
in the preliminary classiŽ cation. The same occurs for the remaining three classes.
7.
ClassiŽ cation accuracy estimation
To compare properly the diŒerent approaches applied in the present study,
emphasis has been given to estimate the accuracy of the four land cover classes
common in all the classiŽ cation approaches. The classiŽ ed and ground-based obser-
vation were compared and confusion matrix tables were generated for the Maximum
Likelihood ClassiŽ cation resultant images and the spectral mixture classiŽ ed image.
An overall accuracy of 83% in correctly delineating land cover has been achieved

3050
P. Casals-Carrasco et al.
Figure 10. Fallow Ž elds theme image deduced from the percentage values.
Figure 11. Bare land theme image deduced from the percentage values.
from the spectral mixture classiŽ cation approach. Moderate accuracy values were
observed while mapping TM bands alone (62%) and TM merged with SPOT (69%)
(
table 3).
For each error matrix, a Kappa (k) statistic was calculated as follows:
S Diagonal/NÕ S(Row totalÖColumn total)/N2
k5
(5)
1
Õ S(Row totalÖColumn total)/N2
where N is the total number of observations. The kappa statistic which is a non-
parametric measure of agreement between ‘ground truth’ and the classiŽ ed images
was used to rank the error matrices. A kappa value is calculated for each matrix

Spectral mixture analysis
3051
Figure 12. Four-theme classiŽ cation deduced from the percentage values (intensity of the
pixel corresponds to the percentage of the theme inside the pixel).
which is a measure of how well the classiŽ cation agrees with the reference data
(
table 4). ConŽ dence intervals were calculated for kappa using the approximate large
sample variance and following the procedure given by Congalton et al. (1983).
The rankings were considered to be signiŽ cantly diŒerent, or not, based on the
values obtained through the use of the following relationship:
|
k Õ k |
dk5
1
2
(6)
Ó(s2 1 s2)
1
Summaries of the interpretation test results are given in table 5.
2
We can see a signiŽ cant diŒerence between the use of the spectral mixture analysis
and the use of the MLC, indicating that the spectral mixture analysis gives a much
better interpretation.
8
.
Conclusions
We have presented a method for spectral mixture analysis of multispectral satellite
images and compared the results with other classiŽ ed images accomplished by MLC.

3052
P. Casals-Carrasco et al.
Figure 13. After the reclassiŽ cation of the  uvial terrace unit, the segmentation results are
reinserted in the Ž rst classiŽ cation.
Figure 14. Details of the classiŽ cation, (a) before the post-segmentation, (b) after the post-
segmentation.

Spectral mixture analysis
3053
Table 3. Percentages of accuracy obtained with the diŒerent classiŽ cation methods.
Spectral
mixture (%)
MLC/TM1 SPOT
(%)
MLC/TM
(%)
Land cover class
Orchards
Fallow Ž elds
Bare land
Water bodies
Overall accuracy
80
89
87
75
83
62
61
61
91
69
61
62
63
60
62
Table 4. Summary of the kappa statistics derived from paired error matrices for the four
land cover classes, common in the three classiŽ cations.
Rank
ClassiŽ cation method
Kappa statistic
I
II
III
Spectral mixture
MLC/TM1 SPOT
MLC/TM
0.7364
0.6671
0.5533
Table 5. Summary of the classiŽ cation method ranking for the four land cover classes
common in the three classiŽ cations.
Rank
ClassiŽ cation method
D Kappa
I
II
III
Spectral mixture and MLC/TM
Spectral mixture and MLC/TM1 SPOT
MLC/TM and MLC/TM1 SPOT
2.70
0.66
0.52
The results obtained with the land cover spectral mixture analysis are consider-
ably better than those obtained with the traditional supervised classiŽ cation methods.
The land cover spectral mixture analysis provides results at a level of precision and
spatial detail which is di cult to obtain by conventional mapping approaches.
The method applied for land cover spectral mixture analysis requires only
radiometric corrections of satellite images before the thematic analysis of the datasets.
For the application of this method, the suitable spectral endmembers are taken from
the image itself and does not require any input from a spectral library or spectral
Ž
eld measurements.
The potentially suitable spectral endmembers can be selected in the image with
the help of a large scale map or aerial photography. Thus the procedures used are
easy to standardize and they can be applied to a wide range of environments by
using the data from the most common types of sensors such as Landsat TM
and SPOT.
We have also developed an approach to establish the relationship among land
covers, landforms and soils through geomorphologic analysis, where the information
obtained from the land cover spectral mixture classes can be related with landforms
and soils, eventually providing the keys for environmental assessment and mon-
itoring. In this sense the mapping of the geomorphologic units, provide us with
geographical frames for which the land cover spectra is located in a reduced spectral
domain. This means that every geomorphologic unit has a limited number of corres-
ponding land covers and these cover entirely the unit which facilitates the application

3054
P. Casals-Carrasco et al.
of this method. Therefore although the technique applied here is not technically
complex, a good knowledge of the study area is required to understand the above-
mentioned relationship between land covers, landforms and soil and to make an
appropriate selection of the endmembers.
The technique suggested here can be applied for the improvement of the results
obtained by supervised classiŽ cation methods. Usually problematic classes are lim-
ited. We propose that it is appropriate to apply the method only for those problematic
classes and in this way, the spectral mixture analysis can be considered very e cient.
Acknowledgments
We would like to express our gratitude to Mr A. K. Sah for his useful comments
and discussions on the procedure and results of the work and to Mrs S. Haruyama
for sharing with us her knowledge of Cambodian geomorphology.
References
Adams, J. B., Smith, M. O., and Gillespie, A. R., 1989, Simple models for complex natural
surfaces: a strategy for the hyperspectral era of remote sensing. Proceedings of the
IGARSS’89 symposium (Vancouver: IEEE), pp. 16–21.
Bolstad, P. V., and Lillesand, T. M., 1991, Rapid maximum likelihood classiŽ cation.
Photogrammetric Engineering Remote Sensing, 57, 67–74.
Bureau de Recherches G eÁ ologiques et Mini e` res, 1970, Carte mini e` re de Cambodge, scale
1
:1000 000, DRME/AO mission, Cambodge.
Campbell, J. B., 1987, Introduction to Remote Sensing (Cambridge, Massachusetts: Guilford
Press).
Congalton, R. G., Oderwald, R. G., and Mead, R. A., 1983, Assessing LANDSAT classiŽ ca-
tion accuracy using discrete multivariate analysis statistical techniques.
Photogrammetric Engineering and Remote Sensing, 49, 1671–1678.
Cook, A., and Doornkamp, B., 1990, Geomorphology in Environmental Management (Oxford:
Oxford University Press).
Gong, P., Marceau, D. J., and Howarth P. J., 1992, A comparison of spatial feature extraction
algorithms for land use classiŽ cation with SPOT HRV data. Remote Sensing of
Environment, 40, 137–151.
Hill, J., 1993, Mediterranean ecosystem monitoring with Earth observation satellites. Paper
Presented at the European School of Climatology and Natural Hazards,
DesertiŽ cation in a European Context, Pueblo Acantilado, Spain, 6–13 October 1993.
Jensen, J. R., 1986, Introductory Digital Image Processing. A Remote Sensing Perspective
(
Englewood CliŒs, New Jersey: Prentice-Hall).
Martin, L. R. G., 1989, Accuracy assessment of LANDSAT based visual change detection
methods applied to the rural-urban fringe. Photogrammetric Engineering Remote
Sensing, 55, 209–215.
Martin L. R. G., Howarth P. J., and Holder, G. H., 1988, Multispectral classiŽ cation of
land use at the rural urban fringe using SPOT data. Canadian Journal of Remote
Sensing, 14, 72–79.
Mas, J. F., and Ramirez, I., 1996, Comparison of land use classiŽ cation obtained by visual
interpretation and digital processing. ITC Journal, 3–4, 278–283.
Mather, P. M., 1987, Computer Processing of Remotely Sensed Images: an Introduction
(
London: Wiley).
Mathieu-Marni, S., Leymarie, P., and Berthod, M., 1996, Removing ambiguities in a multi-
spectral image classiŽ cation. International Journal of Remote Sensing, 17, 1493–1504.
Pouncey, R., and Schrader, S., 1997, ERDAS Field Guide, 4th edn (ERDAS Inc.).
Ryden, A., 1997, Approaches to practical remote sensing for environmental studies: experiences
from a case study in Swaziland and southern Mozambique. IT C Journal, 2, 136–145.
Shettigara, V. K., 1991, Robustness of Gaussian maximum likelihood and linear discriminant
classiŽ er. In XIth IGARSS, Remote Sensing: Global Monitoring for Earth Management,
Helsinki, Finland, 3, 1830–1842.

Spectral mixture analysis
3055
Smith, M. O., Ustin, S. L., Adams, J. B., and Gillespie, A. R., 1990, Vegetation in deserts: I.
A regional measure of abundance from multispectral images. Remote Sensing of
Environment, 26, 1–16.
Toll, D. L., 1984, An evaluation of simulated TM data and LANDSAT MSS data for
determining suburban and regional land use and land cover. Photogrammetric
Engineering Remote Sensing, 50, 1713–1724.
Weismiller, R. A., Kristoff, S. J., Scholz, D. K., Anuta, P. E., and Momin, S. A., 1977,
Change detection in coastal zone environments. Photogrammetric Engineering and
Remote Sensing, 43, 1533–1539.