33 
Fig. 5 a Feature space original data 
Fig. 5 b Feature space optimized data 
Fig. 5 c Feature space H/S components 
spectral classes. An advantage of the unsupervised 
classification is that information about the number 
and the spatial distribution of the spectral 
classes can be quickly achieved. Based on this 
classification the selection of training fields 
for the different spectral classes will now be 
simplified. Furthermore the unsupervised classifi 
cation refers to such classes which are spectrally 
homogeneous. In the remaining inhomogeneous areas, 
textural features should be calculated and con 
sidered as additional information in the super 
vised classification. 
The supervised classification of the KARLSRUHE 
area will be given in Fig. 7, oased on 8 main 
landuse classes, classified with a combination of 
channels 1/3/4. The classification accuracy de 
creases with another channel combination and also 
by using more than 3 channels simultaneously. A 
much better differentiation in the agricultural 
and wooded areas could be achieved through a com 
bination of well suited spectral channels. Also in 
the settlements a better discrimination could be 
expected using textural features. 
First results with different textural parameters 
confirm this speculation. More intensive investi 
gations will be done in the near future. 
The intention of this paper is the discussion of 
the classification results using different prepro 
cessed data, and not the detailed discussion of the 
optimal classification strategy. Therefore the 
comparison of the classification with spectral and/ 
or textural features will not be discussed, because 
the textural analysis of Thematic Mapper data is a 
separate subject. 
2.3 Comparative Discussion of the Classification Results 
Original data 
Tab. 2 shows that most of all 1972 pixels from 25 
training fields have been classified with a high 
accuracy (version a). A visual inspection of the 
classification result (Fig. 6a) indicates that the 
substantial geological units, such as the Schistose 
complex (right part of the imagery), the Basalt com 
plex (left above and in the center) and the Granite 
complex between Schistose and Basalt have been 
classified correctly. Otherwise a more detailed 
analysis of the result also shows miscalculations 
mainly near the strong relief in the left part of 
the imagery. To improve the classification in such 
areas one should take into account the topographic 
effects. 
Optimized data 
The classification of the enhanced data (IHS-trans- 
formation with saturation enhancement) results in 
a separation identical to the classification of the 
original data (Tab. 2). A comparison between the 
two feature spaces in Fig. 5, demonstrates that 
these classification results could be expected. The 
ellipses indicate that the IHS-transformation did 
not change their general arrangement and orientation. 
This transformation results in an optimization of 
the saturation component without any modification 
of the color frequencies and the total albedo of 
the different surfaces. 
Image optimization using IHS-transformation through 
enhancement of the saturation, can be characterized 
as a method which emphasizes the original color 
différencies without falsifying the information 
contents of the original data. 
IHS-components 
In this section we describe the classification using 
the IHS-components directly. These components are 
an intermediate product of the whole IHS-transfor 
mation (chapter 2.1). The classification imagery 
in Fig. 6b (and Tab. 2, Version c) show nearly the 
classification of the original data in Fig. 6 a.