279
lability of multiband remotely sensed data, and can
lead to extraction of information not apparent on the
original imagery. One such technique is principal
component analysis (PCA), especially useful if the
bands are highly correlated. New images are produced
by calculating new, uncorrelated principal components
using a series of linear weights, called eigenvectors,
which are applied to the origional band values.
These principal components represent more efficient
ly the variance of the data.
A study of the eigenvalues for the El Guettar scene
indicates that the data are just about three dimensi
onal in structure with 99 per cent of the variance
being explained by the first three components (Table
4). A study of the eigenector weights attached to
six principal component images, derived for the study
area, using the six reflective TM bands. PCI is an
overall summary of the albedo while PC2 is the diff
erence between the visible/near IR and the mid-IR.
This explains the brightness of the image, the extra
ction of the areas of fan deposition, variation with
in the playa and structure in the mountains. PC3
(Fig. 8a) has a high band 4 weight which produces an
image similar to the 3/4 ratio displaying vegetation
while PC4 (Fig. 8b) is the difference between the two
mid-IR bands and picks out the structure in the solid
geology, fans and variation in the playa very clearly.
Too much noise in the remaining components prevent
any useful analysis.
Many remote sensing studies have used supervised
classification whereby the image is classified by the
user who defines training areas where the surface
properties are known. This allows the computer to
classify the image on the basis of the information
held in the training areas. This works very well in
vegetation studies (Townshend 1981) but in geomorph
ology it is very'- difficult to obtain unique spectral
responses for landforms. Therefore, supervised class
ification is not very suitable and the potential of
unsupervised classification has been tested. This
applies a clustering algorithm to the data and requi
res no previous knowledge of the area. In the algor
ithm used, the data was clustered according to a max
imum likelihood classification, and the only user in
volvement was to stipulate the starting number of
classes, the maximum and minimum percentage of the
image contained in any one class and the number of
iterations for clustering. The resulting image was
colour-coded, density sliced, and class statistics
were derived allowing the display of individual class
es and the proportion of the image contained in each
class.
Figure 9. Example of one cluster derived from an un
supervised classification of El Guettar using TM bands
3,4 and 5. This class relates to areas of solid
geology.
The test area was subjected to an unsupervised class
ification on a three band image (3, 4 and 5) which
produced 10 classes and 99 per cent of the image
classified. The results are very engouraging. Some
of the features which have been classified are areas
of solid geology, variation within the fans, playa,
alluvial plain, agriculture and the large ephemeral
channel but a number of meaningless classes were also
produced. The level of separation is visible in fig
ure 9 which shows how one class corresponds to areas
of solid lithology and oasis vegetation. Only rarely
did a unique class occur, usually a number of featu
res could be identified in the same cluster. This is
due to similarity of pixel values of different units
which leads to confusion within the feature space.
However, the technique shows great potential for geo-
morphological mapping especially if the user wishes
to extract the main geomorphic units in a new area of
interest.
Further information reqarding the imaae nrocessina
techniques discussed here can be found in Moik (1980)
Schowengerdt (1983) and Gillespie (1980).
6 CONCLUSIONS
It has been shown that although single band Thematic
Mapper imagery and standard false-colour composites
show excellent geomorphological detail, it is worth
noting that when dealing with bare, unvegetated sur
faces, typical of arid/semi-arid areas, the degree of
correlation between the bands is so high that most of
the bands are redundant. Much more information can be
extracted using digital image processing. Effective
techniques range from simple contrast stretching to
more complex multivariate techniques such as unsuper
vised classification and principal component analysis.
However, no single image processing technique is suit
able for all landforms. Best results are obtained
when there exists a specific relationship between
spectral response and the physical properties of the
phenomena under investigation. This relationship can
then be exploited by selecting the optimum image pro
cessing technique.
The resulting processed imagery are an invaluable
asset to anyone contemplating geomorphological mapp
ing especially in difficult terrain. Digitally enhan
ced satellite Thematic Mapper imagery is a powerful
tool for geomorphologists for studying processes and
for mapping. Not only does it afford a new perspect
ive from which to observe the Earth's surface but it
allows the development of new ideas regarding geomorph
ological processes, the origins and modifications of
landforms.
ACKNOWLEDGEMENTS
I wish to thank Dr. J.R.G. Townshend for his construe
tive comments during the preparation of this paper and
to the cartographic and photographic units of the
Geography Department, University of Reading. Arwyn
Jones is a NERC postgraduate research student GT4/
83/GS/87.
REFERENCES
Bailey, G.,J.Dwyer & Francica 1982. Evaluation of ima
ge processing of Landsat data for geologic interpre
tation of the Qaidam Basin, China. Second Thematic
Conf., Remote Sensing for Exploration Geology, Fort
Worth, Texas, p.555-577.
Brunsden, D.,J.Doornkamp & D.Jones 1979. The Bahrain
Surface Materials Resources Survey and its applicat
ion to planning. Geogr. Jour. 145:1-35.
Colwell, R. 1983. Manual of Remote Sensing, Vol I & II
American Soc. of Photogram., Falls Church, USA. 2nd
Edition.
Doehring, D. 1980. Geomorphology in arid regions. All
en & Unwin, London, p. 272.