32
Fig. 1 Location of test areas
- further, the algorithms should transform data in
dependently of variations in contrast. This faci
litates the construction of mosaics and enables
the interpretation of data across boundaries
between different Landsat records.
BASIC CONCEPT
In a first step the correlation matrix was calcula
ted to select three geologically significant bands,
which are needed for additive color coding (Tab. 1).
Two bands (5,6) are not useable fortnis concept.
The relativ low correlation of SWIR-band 5 (1.6 pm)
is caused by an overflow of the TM-detectors (on
LANDSAT 4 and 5) and is typical for high sun ele
vations (58° for this data of 04.17.85). Emission
signals of band 6 are mostly depending on relief
(sloping site) due to early morning overpass during
heating up phase.
Fig.2/3were calculated by use of TM band 1 (blue/
coded blue), TM band 4 (NIR/coded green) and TM
band 7 (SWIR/ coded red). Although there was a good
separability in brightness (albedo) and hues (domi
nant color frequencies), the interpretability of
the original composite was restricted by low satu
ration values. This deficiency is caused by well
balanced reflectance characteristics, related to
the lack of strong absorption bands associated with
most rock surfaces.
The saturation cannot be optimized by means of
additive (Red, Green, Blue) color processes e.g. in
the photolaboratory and instead, the following pro
cedure is used. Precondition is a digital transfor
mation into three new mutually independent compo
nents called Intensity (I), Hue (H) and Saturation
(S) (Haydn et al., 1982)
After a linear shift of saturation levels to the
high saturation part within the given 8-bit range,
the modified Saturation component and the original
Intensities and Hues are retransformed into RGB-
outputs.
In order to enhance subtle variations in structu
ral contrast, highpass-filtering was applied to the
data using a 3 x 3 matrix and added back to the
already I,H,S-calculated input components. The dis-
Tab. 1 Correlation matrix for As-Sirat area
ch 1
ch 2
ch 3
ch 4
ch 5 ch 6
ch 7
ch
i
1.0000
0.9685
0.9265
0.8965
0.7384 -0.0998
0.7762
ch
2
1.0000
0.9804
0.9556
0.7772 -0.1349
0.8437
ch
3
1.0000
0.9849
0.8239 -0.0971
0.8970
ch
4
1.0000
0.8650 -0.0910
0.9094
ch
5
1.0000 0.0721
0.9042
ch
6
1.0000
0.0447
ch 7 1.0000
advantageous effect of reduced spectral differences
can be compensated by the former Saturation increase
explained above (Bodechtel & Kaufmann 1985).
The last step is histogram computing and stretch
for display on screen or transmission via photowrite
system.
Fig. 2 shows the result, which offers optimum
conditions for lithological and structural mapping
within one single image product.
2.2 Classification
Whereas the results of the image optimization pro
cess mostly serve as a well suited image product,
for the following interpretation procedure the
classification methods lead to evaluation results
which are mainly produced by the computer. Most of
the classification procedures are working pixel-
oriented using only spectral information. The un
supervised methods combine the pixels with similar
features to spectral classes, for example with the
euclidean distance in the feature space as a cri
terion for the discrimination. The individual
pixels are then assigned to the class whose center
is nearest or whose distance is smaller than a
maximum distance. For the supervised methods accu
rate ground truth is needed to calculate the
features of the individual classes before classi
fication. A classification of all pixels to one
of the given classes will then be carried out using
a specific classification rule. The well known
maximum likelihood method, for example, is based
on the calculation of the likelihood with which
the individual pixels are members of these classes.
A classification is also assigned to the class with
the maximum likelihood (SWAN, DAVIS, 1978). To get
the likelihood, statistical descriptions of the
sample classes, such as the mean values in each
channel and the covariance matrix, are needed.
For the estimation of these statistics training
fields with known landuse are introduced. The main
problem of the supervised procedures is the se
lection of really representative sample areas for
the presented classes. Both kinds of classification
methods, the unsupervised and supervised pro-
ceuures, can also oe applied for the classi
fication with features, which aescribe the
texture and the shape of objects. A multispectral
classification with textural features for example,
car be carried out based on textural features as
additional channels, whereas the textural parame
ters have been calculated for each pixel in a
window fittet around each pixel. Well known textu
ral features are based on a statistical evaluation
of the grey dependencies between neighouring pixels
(HARALICK, SHANMUGAM, DINSTEIN, 1973). The calcula
tion of so called HARALICK-parameters for each of
the 7 Thematic Mapper bands leads to 7 additional
textural channels. For a subsequent supervised
maximum likelihood classification, with both the
spectral and the textural features, a highly so
phisticated channel selection or channel combina
tion procedure must be carried out. Furthermore,
a hierarchical classification procedure (WU, 1975)
should be applied to use the full information con
tents of all spectral and textural channels in a
cost effective way.
EXAMPLES
Fig. 4 shows an unsupervised classification of the
test area KARLSRUHE, with the euclidean distance
as discrimination criterion using channels 3/4.
Using a large euclidean distance (d=20 grey values)
only a few spectral classes will be discriminated,
whereas with smaller distances more spectral classes
occur. For the presented Thematic Mapper data
(color composite in fig. 3, date 7.7.1984) a small
euclidean distance (d = 7) seems to be a favourable
measure for the classification. Fig. 4 demonstrates
that this euclidean distance leads to more than 10
