Multi-agent architecture for map features extraction in multi remotely sensed image fusion

The problem of map features extraction from satellite images requires the combination of several heterogeneous knowledge sources. These sources range from databased knowledge as well as the modeled expert's knowledge to the human expert interactions. The databased knowledge includes image filters and feature primitive extraction tools. The modeled expert's knowledge includes domain-based models (real, image and decision worlds), rules and strategies. The human expert interactions concern mainly the inputs of three-dimensional information and the validation processes. The global stream of information inferred by the previous knowledge sources should then be merged to lead to a single accurate decision, here the extraction and classification of map features. In order to keep as much independence and modularity as possible for each of the previous knowledge sources, a multi-agent architecture is developed. This paper presents this architecture in the context of a 3D linear planimetric map feature extraction and classification system (Palmyra).