As can be seen, a lot of isolated pixels appear, which are obviously misclassified. Furthermore as this region is subject to "monsoon" and is hilly, the vegetation classes are correlated to orientation and to elevation. So some vegetation classes should not be found at low elevations or on east slope.

Classifications, in such a context, are never satisfactory if there are only based on image grey levels. It is obvious that other sources of information should be introduced, as the human expert does when he tries to obtain vegetation cartography from satellite images.

In fact, for this particular case the human expert placed the image "north upside" for example (this is obviously not necessary for a usual gray level based classifier of course!) and was able to evaluate the elevations directly looking on the image. The topology of vegetation homogeneous regions was also considered intuitively by the human expert (for example "elongated regions in large valleys" was characterizing some type of vegetation).

So it appears necessary to construct this type of information source and to combine it with the classical method only based on pixels' gray levels. It is also interesting to remark that for the human expert the main approach is based on his knowledge and on data which is not directly processed from the image (elevation, orientation, homogeneous regions and their spatial arrangements).