In the second approach,

• the "initial" classification is produced via a neural networks technique and
  
  
• the expert knowledge is represented by a set of neural networks. Each neural network represents the context of a pixel, in relation with out-image information, then
  
  
• the Dempster orthogonal rule is used to combine the outputs of all nets concerned.

The aim of such an approach is to build reasoning and problem-solving capabilities within the frame of remotely sensed images and G.I.S (Geographical Information System).

In the GIS context, we want to express a problem by means of expert knowledge and to show the result by means of digital maps.

The system must be able to catch the expert knowledge about the best context for a particular problem solving. This knowledge is related to information that can be computed on the images or in the G.I.S.

The system must be able to deal with uncertain information and knowledge and produces a potentiality map for the solution of the problem.