Most pattern recognition applications using remote sensing imagery involve classifications from computed features :

• spectral information,
• spatial information such as texture and structure,
• color,
• multi-temporal information, ....

However image quality, pixel resolution, viewing conditions,... induce unsatisfactory results using the usual classification processes.

In fact, human experts manage with satellite images in order to produce thematic maps for example, using a large amount of "a priori" knowledge beyond that contained in the image concerning the studied area and their usual context. A human expert who wants to produce a vegetation map for example, will manage with three types of information sources :

• the remote-sensed images,
• available cartographic information (topography, soil quality...),
• knowledge of local vegetation types.

It therefore seems necessary to combine processed image information and "a priori" knowledge in order to improve image classification quality.