It seems necessary to introduce the expert knowledge about the localization of the searched classes in special contexts, and specially when the remote sensed image is taken over a heterogeneous region as the Palni hills.

These contexts must be based on accessible data sources such as elevation, slope, types of soils, human made structures (roads, urban areas) or natural structures (rivers). This knowledge may be assumed more or less complex.

For example:

• The knowledge "elevation always higher than 1000m" is very simple to manage. When you have a digital elevation model for each isolated pixel in the image you are able to compute immediately a degree of realization of the corresponding knowledge and in fact this degree has only two possible values ("realized", "not realized") as the context is a completely realized or absolutely not realized function of the elevation.
  This degree may be computed whatever are the neighbouring pixels! In our approach we call this type of knowledge "pointwise knowledge".

• The knowledge "often elongated regions located in large valleys" is more complex. In fact we have to manage with homogeneous regions in relation with another "object" (here a large valley).
  The problem here is that we cannot manage with regions as we have not yet classified the image and moreover the knowledge is not certain (the expert says "often" ) which means that other contexts are acceptable (even if less probable) and imprecise information may be present (here "large" valleys). In our approach we call this type of knowledge "structural knowledge". This type of knowledge is actually under evaluation.

In our example we shall take into account the first type of knowledge ("pointwise" knowledge)