set of training areas must be available to describe the variation within a class. The 
quality of the results is mainly determined by a good selection of training areas and even 
minor errors may cause serious misclassifications. The effort to acquire the ground 
information is considerable. If these requirements are met, the actual classification is 
fast, about 1 - 2 hours for a complete LANDSAT scene, and further evaluation for example 
in a data base or in model calculations are easy. Fig. 2 shows as an example a supervised 
classification of the vicinity of Karlsruhe in Southern Germany. 
Unsupervised techniques offer a different set of potentials and depending on the specific 
algorithm used, the results may differ considerable. In Fig. 3 a clustering technique 
  
was used, which is based on a multidimensional histogram. Every class surrounds a 
Fig. 4 
maximum in the histogram, and the boundaries between classes follow the minima. It can 
be visualized as a mountain range where the peaks determine the cluster centers, the 
valleys the boundaries between classes. There are no restrictions placed on the number 
of classes and their shape, but a more general or more detailed separation can be 
reached setting a parameter. No specific ground information is needed for the 
clustering algorithm. But the different clusters must be labeled using their spatial vulconis 
distribution within the scene and other characteristics for example their spectral values, SUL Fed 
defining their position in the feature space. Again an interpreter can determine the selectec 
attributes of the clusters and label them accordingly, based on his experience with similar stretchit 
data. The algorithm may separate a land use class in different clusters.or within a cluster sequenc 
two or three land use classes, which are not well separable, may be combined. In this sequenc 
way spectrally different clusters may show up indicating a special subclass, e.g. an indicate 
additional type of land use, which was not expected in the scene. Grouping the these re 
different clusters according to their label can yield results which are very similar to a Black a 
supervised classification without requiring as input very detailed local ground information results c 
(fig. 3).This procedure is therefore very suitable if good ground information is not fo plan 
available, the number or type of classes is unknown or additional subclasses may become tion wa 
important. In addition statistical values for the clusters can be determined, e.g. the POTEN 
percentage of areq covered by a cluster within a scene. These h 
In the selection of areas for detailed geological investigations for example in exploration 1. Phot 
the problems are quite different. Within a research project related to plate tectonics simp 
we have to select promising areas for ground investigations of tectonic activities and not « 
related vulcanic rocks. Since the available geologic maps did not differentiate inte 
" P ERN 
REA = Wamu