Each effective input for the net, is the product of certainty factor of the premises, with their corresponding frequency degrees (0.8 for "often, +1.0 for always..):

• At the learning stage, the rules are learned with frequency degrees equal to 1, as if the rules were all defined by "always", and
• at the operating stage, these frequency concepts are introduced in order to represent the knowledge as it has been expressed by the expert.

The deactivation of an input is performed by assigning a zero value to .

After the presentation of the context of the pixel to all nets (corresponding to all rules), the system produces realization degrees of each rule. The result is contained in a vector S which will be combined with another vector R obtained by a neural network classification process using radiometric information.

These two types of data are considered as two different information sources on the same object, and their combination uses the Dempster 's orthogonal rule. The uncertainty factor, for the classification of one pixel in a given source is estimated by the separability of realization degrees among classes (Sep).

We consider the following notations:

•  is the result vector obtained by NN classification process , using symbolic knowledge.
• is the result vector obtained by NN classification process using image data (for our application, we used the radiometric satellite data).

R and S will be mapped to mass distribution vectors R' and S'. So,  and .

•  
• 
• 
  

with

The new mass distribution is computed with the Dempster orthogonal rule T=R'S':

We then define the support and plausibility for each labelling proposition:

• the support for a proposition A is defined as the sum of all mass assigned to the proposition and its subsets,
  
• while the plausibility of a proposition is defined as 1 minus the support assigned to any propositions in contradiction to proposition A. In algebraic terms we obtain :

The inference rule is expressed by the maximum support rule, in which the labelling proposition with the highest support is chosen.