To see the correlation between the vegetation classes and the elevation. We give the maximum likelihood classification and the best ICARE classification (imp=0.6) superimposed on the DEM.

The final classification appears to be quite well improved in comparison with the results of the usual supervised classification and is quite well correlated with the manually made map of the expert (average result for maximum likelihood classification being 58.60% and 67.31%, 71.88%, 65.86% for imp = 0.3 , 0.6 , and 0.9 respectively).

With imp = 0.6 the average increment of well classified samples is 13.28%. The classes for which the increments are the best are classes for which the samples correspond quite well with the corresponding knowledge (class 1 21%, class 3 49%, class 5 22%, class 7 38%) while those for which the increment is negative are samples which does not at all correspond to the given knowledge!

However one of the main disadvantages of the rule based approach is that each classification processing induces activation of all rules for all pixels in the image which produces about 10 minutes CPU time on a SUN SPARC 4 for the total application.