on 
ed 
ty 
ay 
st 
in 
us 
ce 
classification as a new class in order to be 
subsequently grouped with a class of 
approximate characteristics. 
These comparative computations and aggregation 
processes requiring considerable computation 
expenditure can formally be placed on a simple 
level of decision. The value of distance to be 
computed between 2 training area clusters is 
1 
t^ - (u-v)T * (A+B) ' * (u-v). 
Here, u and v represent the mean value vectors 
of the cluster samples to be compared, A and B 
their covariance matrices. 
The introduction of threshold values allows a 
very simple decision: if, e.g., 
t« 1, then the two testing areas are 
statistically equal and may be grouped 
before classification. If. e.g. 
t» 3, then the two testing areas are signifi- 
cantly different and stand for 
2 classes. If, e.g., 
1$t$3, then it is for a clear delimination of 
clusters concerned from others necessary 
to first introduce each testing area as 
independent into classification and to 
group it with the comparable ones 
afterwards. 
This approach prevents inadmissible grouping of 
training areas which are classified as identical 
by the interpreter, while actually being 
different. Thus, an unnecessary overexpansion of 
the clusters causing a decrease of the discri- 
mination capacity can be avoided. Beyond thet, 
controlled grouping of statistically equal, 
nearly normally distributed training areas freed 
from outliers can lead to larger and statisti- 
cally better substantiated units. It is to be 
expected that there is hardly any operator 
capable of such a performance. The clusters 
whose distance t is between ] and 3 can be 
grouped and assigned after classification when 
being interpreted. 
These are - all in all - the essential reasons 
for the lateral diagonals of the confusion 
matrices being occupied by zero and the main 
diagonal by values between 90% and 100%. The 
lateral diagonal elements occupied by zero mean 
that ambiguous classification results can be 
avoided. Deviations of 100 % in the case of main 
diagonal elements indicate an incomplete though 
free of conflicts training area classification. 
3. FURTHER PROCESSING OF THE CLASSIFICATION 
RESULTS 
Classification running automatically as depicted 
above provides very differentiated and safe but 
not yet interpreted results. Interpretation, 
which takes place together with the selection of 
training areas when proceeding interactively, is 
transferred to the end of the operational treat- 
ment in the case of automatic classification. 
Interpretation is performed by experts of the 
various disciplines concerned. 
913 
4. CONCLUSION AND OUTLOOK 
Automatic search for and grouping of training 
areas according to spectral and statistical 
aspects provides a substancial benefit as to: 
- objective, i.e. only spectrally proven 
training area selection 
- separation capability and thus secured variety 
of land uses as well as 
- free availability of the workstations which 
would otherwise be strongly occupied by these 
tasks. 
Excessive loading of the data with other back- 
ground knowledge which is not represented in the 
data does in this way not occur, hence reducing 
the risk of ambiguity normally caused by such 
overloading. Inclusion of multitemporal and 
multisensoral data in a simultaneous evaluation 
presupposes rectification accurate to the pixel 
in each image site, which is particularly 
difficult in moved terrain and has not been 
possible as yet. 
Inclusion of further data and additional know- 
ledge, e.g. via the interpretation process, will 
provide in the long run orientation towards 
artificial intelligence. This is the path we 
will follow. 
LITERATURE 
Schulz, B.-S., 1990. Analysis of the Statistical 
Prerequisites for Classification of Multi- 
spectral Data. Proceedings of the Symposium 
"Progress in Data Analysis", 
Comm. III of ISPRS, May 20-25, 1990, Wuhan, 
China