The remotely sensed images are considered as an important source of information used in several domains such as environmental monitoring, disaster management and military intelligence. However, the appropriate information has to be extracted and represented to make efficient decision processes. In this thesis, we proposed a spatiotemporal, fuzzy and heavy ontology for modeling remote sensing images. The ontology is based on (1) the re-use of an existing ontology of domain which is considered as a kernel-ontology, and (2) the enrichment of this one in order to cover the domain of application and to take into account the specificities of remotely sensed images. The process of enrichment establish at various levels: conceptual, relational and axiomatic, basing on external resources like taxonomies, thesaurus, and knowledge extracted from experts and image archive. The proposed enrichment allows maximizing the capacity of the ontology on modeling and reasoning through the axioms. We propose to enrich the ontology with fuzzy logic in order to modeling the uncertainty in remotely sensed images. Furthermore, we propose a methodology for modeling and retrieving satellite images basing on their spatial knowledge. The main idea of our approach is that the use of spatial knowledge, reasoning and inference technique, can contribute to deduce the susceptibility of the scene at natural risks (erosion, flooding, fires, etc.). Our methodology takes in input a set of multi-sensor images representing a scene. It contains four modules: (1) Modeling of the scene, (2) fusion of image annotations, (3) similar case retrieval, and (4) reasoning and interpretation. The first module generates annotations which represents the semantic content of the satellite scene. The second module allows merging image annotations to have faithful information to the reality. The third module attempts to find similar cases to those of the annotated query to take advantage of cases, situations and past problems. The fourth module allows deducting the susceptibility of the image in a given natural phenomenon (Erosion, flooding, fires), basing on inferences on knowledge of domain (Ontology, knowledge of expert, knowledge of natural phenomena, etc.) and the result of the previous phases. Finally, we applied the methodology to deal with the problem of erosion to warn high-risk areas of the kef region, situated in northwest of Tunisia. Indeed, we developed a base of images representing the classes of the ontology (forest, lake, urban zone, etc.). Also, we developed a base of cases representing remote sensing images of the region of the Kef. In an "offline" process, we simulated the process of interpretation on requests to see if they present risks of erosion.
