12 
4 SELECTION OF TRAINING AREAS AND 
CLASSIFICATION METHOD 
The remote sensing data sets were 
geometrically but not radiometrically 
rectified. In order to cope with the 
well-known radiometric problems we refrained 
from the common interactive definition of 
training areas in the form of polygons. 
Instead, based on various selection criteria 
(e.g. forming part of a certain thematic 
map, slope gradient, slope aspect, 
shadowing, distance from nadir of image, 
etc.), pixels which in general are not 
neighbouring each other were defined as 
layered random samples of 20 %. In 
addition, this approach largely eliminates 
the influence of spatial autocorrelation of 
training pixels for the subsequent 
classification. For the classification 
process proper routines for discriminant 
analysis from the SPSS 9 software package 
were used. Figure 3 shows a schematic 
presentation of possible approaches for data 
evaluation. 5 
Figure 3• Schematic presentation of possible 
approaches to data evaluation. 
5 CLASSIFICATION RESULTS 
The percentage of correctly classified 
pixels referred to the unmodified "ground 
truth" of the respective thematic maps 
amounts to some 40 % in average, the extreme 
values being 100 and 6 %. Table 1 displays 
the percentual classification results of the 
categories of the lithological map, which 
with a classification accuracy of 31 % was 
the worst in reclassification. A detailed 
analysis of the misclassifications brought 
up a series of different possibilities for 
errors, some of which shall be presented 
here . 
A peculiarity of the reclassification of the 
lithological units was that they are 
sometimes sensor-exposed, i.e. they 
represent the earth's surface, and are 
sometimes covered by soil and vegetation. 
In the latter case, the rock units can only 
be recognized through the spectral 
information of the sensor-exposed vege 
tation. Therefore only those lithological 
units can be correctly reclassified, which 
are part of a allochthonous geo-ecosystem, 
where lateral processes, which would cause 
an atypical vegetation cover, play a minor 
role. In high mountain regions, however, 
the distribution of vegetation associations 
is, among others, also dependent on the 
duration of the snow cover and on the 
location within the macro- and microrelief, 
which both considerably control the 
migration of materials. Hence, auto 
chthonous geo-ecosystems are very rare. 
In the study area allochthonous geo-eco 
systems cover most of the area. So one 
finds the vegetation associations Arabidetum 
caeruleae and Seslerio-Caricetum 
sempervirentis, which indicate limestone 
rocks, also over prasinite, garnet muscovite 
schists and quartzite. This is due to the 
proved natural carbonitization of these 
habitats through carbonate-bearing surface 
run-off or through deposition of carbonate 
bearing eolian sediments. The above 
mentioned vegetation mosaics cause a 
relatively high inhomogenity of the stati 
stics of the training areas. Therefore the 
reclassification of lithological units 
contains apparent misclassifications, which 
can be explained by the variations in the 
vegetation cover. Nevertheless, it has to 
be stressed, that the type and the 
peripheral conditions of the random sample 
selection considerably influence the 
classification results. This understanding 
makes clear one of the direct limitations in 
the application of remote sensing for 
lithological mapping. This application 
field, however, presents itself as a test 
bed and stimulator for the theories about 
geo-ecosystems. 
If one only selects those areas where rocks 
form the earth's surface, the classification 
accuracy in the study area rises up to some 
50 %. Here it is mainly the similar 
spectral characteristics of the various 
lithological units which prevent better 
results. This is especially valid for 
quartzite, marble, dolomite and rauhwacke. 
If a fresh break of these rocks is studied 
in the laboratory, they show significant 
spectral differences; their weathered 
surfaces, however, can hardly be classified 
definitely. Therefore some lithological 
units with vegetation cover could be 
reclassified more precisely than others 
without, if they jointly formed part of an 
allochthonous geo-ecosystem. 
Several false classifications cannot be 
explained, e.g. the fact that 27,3 % of the 
class "recent alluvial material" and 20,9 % 
of the class "peat and bog" were classified 
as the class "quartzite, rhaeticite and 
graphitic schist". This is, to some extent, 
due to the unsystematic partial deformations 
in the lithological map, which only 
sometimes can be explained by inaccuracies 
within the mapping process. In general, the 
objects of geoscientific mapping will rather 
be distinguished by transitional than by 
well-defined boundaries. So it could be 
demonstrated by varying layering conditions 
and varying random samples, that in places 
the position and arrangement of the 
boundaries in the vegetation and soil map 
was better deducible from the formalized 
theory of catenas (cf. section 2) than from 
the complicated mosaic of distribution 
patterns. Therefore the drawing of 
boundaries will always be more or less 
problematic and, by no means, always be 
intersubjectively checkable. In the present 
study, the location of these boundaries in 
the map sheets without using an 
ortho-CIR-photo caused an additional 
geometric uncertainty. The aspects of 
cartographic generalization, which is 
already to some extent performed during 
Table 1. Cle 
pe 
No. 
Material 
1 
Gabbro-Amphih 
2 
Prasinite, Garni 
3 
Peridotite, Serp 
4 
Chlorite-Talcum 
5 
Quartzite, Rhae 
6 
Calcareous Mic 
7 
Garnet-Bearing 
8 
Dark Mica-Schis 
9 
Quartz-Rich Bre 
10 
Quartzite 
11 
Marble 
12 
Dolomite 
13 
Rauhwacke 
14 
Moraines: Ferna 
15 
Moraine of 1850 
16 
Peat and Bog 
17 
Ancient Rock Fa 
18 
Talus Fans 
19 
Recent Alluvial f> 
field mappi 
considerati 
example e.g 
to wrong g 
"quartzite, 
schists". 
The mutual 
"peat and 
material" c 
certain ge 
well as t 
extension 
time of fie 
remote se 
These chang 
the Gross, 
road). The 
in the lit: 
cartography 
result of 
meantime ai 
related to 
only been a 
effected a'. 
the same 
bro-amphibo] 
hornblende-: 
located so ; 
100 % accurs 
An additic 
different € 
classes in 
class "ta 
distinguishe 
common areal 
to some 10 % 
talus areas 
in classical 
term detrit 
used fro 
Nevertheless 
study area 
dif f erences 
vegetation m 
vegetation 
this class i 
16 % of th 
classified a 
It could bi 
congelifract 
mainly asigi