- verification of the tr aining sets - it is necessary to 
verify the labeling of training polygons and to test 
the separability of target classes 
- editing of the training samples using k - k x nearest 
neighbour rule 
- suggestion of appropriate spectral relations to be used 
for discrimination of target classes 
- knowledge base development (or updating) - the results 
of various tests are saved in the data base 
- generation of the classification algorithm 
- classification of the image data. 
Of course, classification method described has also some 
limitations: 
- the pixels have to be well illuminated - pixel with 
anomalous illumination (dense shadows) cannot be recog 
nized correctly 
- the pixels have to be relatively "pure" , the spectral 
properties are influenced by a single target class 
- the target classes can be recognized using the spectral 
properties only - the spatial and textural properties are 
not used. 
Verification of the training samples 
The collection of a suitable training samples and the decision 
in which classes may by classified the satellite image data 
create the serious problem. 
During the collection of training sets some of the training 
polygons are assigned to a certain class. It is necessary to 
verify, whether these polygons really belong to the same target 
class. Some methods solving the problem of unperfect labeling 
of training polygons (for normally distributed data) have been 
already investigated [ 2 ]. The decisions are made by 
comparison of mean values and covariance matrices. If we do not 
dispose with normal data distribution, then it is possible to 
use a method applying mutual information [ 3 ], [ 7 ]. 
Editing of training samples 
It has been shown, that the editing of training set improves 
the performance of the classifier. The k - k x nearest 
neighbour method is relatively simple. The k nearest neighbours 
from the whole training set are found for every sample. The 
tested sample is edited from the training set, when not being 
classified in accordance with its true class membership (when 
at least k x of its nearest neighbours do not belong to the same 
class as the tested sample). 
The suggestion of spectra1 relations 
The spectral relations characterize the shape of spectral 
reflectance curves in terms of certain inequalities to avoid 
the use of absolute values of individual features. The analyst 
can suggest an arbitrary spectral index using his empirical 
experience, studies of the literature or studies of spectral 
reflectance curves of target classes. The analyst suggests 
spectral indices, which seems to be typical for individual 
classes. Of course, the set of spectral indices from previous