327 
classification can be used. Ihis set has been saved in the 
spectral knowledge data base. In this case, the values of 
suitable thresholds T are updated. 
It is possible to use so called spectral transformations. 
They include linear combinations of spectral bands or various 
ratios of bands, e.g. brightness, greeness, yelowness [ 4 ] or 
PVI (perpendicular vegetation index - [ 5 ]). Many authors have 
found out that spectral transformations are correlated with 
vegetation growth parameters (which can correspond to various 
target classes). Dusek [ 6 j has shown that winter wheat 
vegetation indices (leaf area index, percent green ground 
cover) are highly correlated with Thematic Mapper band ratios. 
For example, the three band ratios, that produce the highest 
correlation with the leaf area index are: 
TM4 . TM5 / TM3 
TM4 . TM5 / TM1 
TM1 . TM4 . TM5 / TM2 . TM3 . 
The spectral knowledge base development 
The aim of this operation is to find the appropriate thresholds 
for spectral indices so that the resulting spectral relations 
will have a maximal power of discrimination. The spectral 
knowledge based system proposes a set of possible thresholds 
for every spectral index in an automatic way. Then, every 
target class (its training samples respectively) is examined, 
whether (and up to which degree) fulfil the spectral relation. 
The results are saved into the data base. This process may be 
divided into five steps: 
- calculation of the spectral index for all training 
samples 
- the histogram of computed values (for every class) is 
created 
- the threshold selection - all thresholds, which separate 
at least two target classes are selected 
- the test, which classes fulfil the spectral relation 
- the result is saved for later evaluation. 
Seneration of the c lass if ic at i onjl g or i t h m._ 
The classifier suggested is of binary tree character. A binary 
tree classifier assigns a class label to a sample by passing it 
through the tree from root to leaf. Leaves are labeled by class 
labels. The test of a spectral relation is performed in every 
node (except the leaves). Here, the features of pixel being 
classified are considered. If the test is fulfilled, then the 
test, corresponding to the right son, is performed. Otherwise, 
the testing continues in the left son. The label of target 
classes are assigned to a classified pixel at final level of 
the tree. 
The binary tree is constructed in an interactive way. The set 
of target classes is determined for every node (the root 
corresponds to the set of ail classes). Then, some spectral 
relations are chosen from the data base which are suitable for 
the separation of this set of classes. Two subsets S\ S2 of 
target classes are distinguishable on the basis of a spectral