The production rules of the knowledge base are applied to the initial fact data base which results from a supervised pre-classification and a GIS information. The result is a final fact data base .

The inference engine will proceed by "forward chaining": for each elementary fact, all production rules will be activated since each pixel is supposed to belong to each class with different certainty factors. It can be noted that a threshold on the certainty factors may be defined to limit the number of selectable rules.

So, for each pixel and each possible class, all the production rules are activated to define the new certainty factor. Note that uncertainty management appears at three levels :

• : uncertainty on the classification in class I,
• : uncertainty of a rule conclusion,
• : uncertainty on conclusion A of a rule...

For example :

I(x,y) class i,
selectable rule : IF class i THEN A ( )

For that pixel let us suppose is the certainty factor for realization of conclusion A.