tive spectral patterns. Fig. 1 gives an example of the supervised 
classification procedure with interactively selected samples - 
areas of 3x3 pixels for the different classes within 3 channels 
(1a to 1c) - and the classification result (1d) gained with a 
linear classifier. There are a lot of different classification 
procedures e.g. the maximum likelihood, regression function, 
nearest neighbourhood, minimum distance procedures of linear 
or higher ordered type. But none of them can be considered 
the best one in general. The quality of the classification re- 
sults depends more on the correctness of the selected samples 
which should be of sufficient number, represent the spectral 
characteristics of the specific classes, and have regard to 
the influences of the different sensing conditions. 
  
Fig. 1: Classification of multispectral data with inter- 
active selection of training samples