Utilisation de la télédétection, des SIG et de l'intelligence artificielle pour déterminer le niveau de susceptibilité aux mouvements de terrain application dans les Andes de la Bolivie
The socio-economic impact of mass movements for our society is getting more and more serious.The loss of lives and economic losses are now ten times greater than they were at the beginning of the decade. In the hope of reducing these impacts, it is essential to adopt a preventive policy that will encourage mapping of mass movement susceptibility level (MMSL) in critical zones. However, this task is complex and only experts using present techniques can provide satisfactory results. To make possible the production of these maps by a larger number of individuals, we have developed an expert system called EXPERIM that uses remote sensing data and geographic information systems to facilitate the complex tasks without requiring the user to be highly competent in this field of study. This thesis presents the results obtained from a complete strategy developed for a region surrounding Cochabamba, Bolivia.The operational expert system prototype will soon be integrated within the watershed management program directed by the local executing organisation PROMIC.The knowledge acquisition and its expression in concrete terms constitute the principal axis of this research, while the results obtained are the heart of the EXPERIM expert system. These strategic steps aim to establish a knowledge base of data and rules that describe field conditions for each MMSL. We have been able to extract this information by using binary discriminant analysis of a MMSL map produced by an expert for a pilot zone called Cuenca Taquiäna, which is geoecologically representative of the 38 neighbouring watersheds. Using this technique, we were able to establish a sensitivity model that recreates the expert's map with a success rate of 89% and 78% when two or three MMS levels are used. Based on a detailed analysis of the susceptibility model it was evident that stability conditions are the result of the topographic, geologic and geomorphologic environments.The level of susceptibility was found to be independent of the vegetation condition. In order to apply the model to the surrounding watersheds, we integrated remotely sensed data within the spatial database to map the presence/absence of five essential geoecological units required by the susceptibility model. This was done using a hierarchical classification method. Three sensors were evaluated: Landsat, SPOT and RADARSAT. In the elaboration of this specific step, we evaluated the most efficient spectral band combinations within each image and between images for each of the five geoecological units. For each of the land cover types, the analysis shows that LANDSAT constitutes the most powerful sensor to map these units and that image fusion does not provide significantly better results when compared to the extra amount of work that this requires. Using remote sensing data instead of field data or airphotograph interpretation in watersheds where only topographic data are available decreases the level of accuracy by less than 10%.