Abstract
Satellite images have become a valuable source of information for Earth observation. They are used to address and analyze multiple environmental issues such as landscapes characterization, urban planning or biodiversity conservation to cite a few.Despite of the large number of existing knowledge extraction techniques, the complexity of satellite images, their large volume, and the specific needs of each community of practice, give rise to new challenges and require the development of highly efficient approaches.In this thesis, we investigate the potential of intelligent combination of knowledge representation systems with statistical learning. Our goal is to develop novel methods which allow automatic analysis of remote sensing images. We elaborate, in this context, two new approaches that consider the images as unlabeled quantitative data and examine the possible use of the available domain knowledge.Our first contribution is a hybrid approach, that successfully combines ontology-based reasoning and semi-supervised clustering for semantic classification. An inference engine first reasons over the available domain knowledge in order to obtain semantically labeled instances. These instances are then used to generate constraints that will guide and enhance the clustering. In this way, our method allows the improvement of the labeling of existing classes while discovering new ones.Our second contribution focuses on scaling ontology reasoning over large datasets. We propose a two step approach where topological clustering is first applied in order to summarize the data, in term of a set of prototypes, and reduces by this way the number of future instances to be treated by the reasoner. The representative prototypes are then labeled using the ontology and the labels automatically propagated to all the input data.We applied our methods to the real-word problem of satellite images classification and interpretation and the obtained results are very promising. They showed, on the one hand, that the quality of the classification can be improved by automatic knowledge integration and that the involvement of experts can be reduced. On the other hand, the upstream exploitation of topographic clustering avoids the calculation of the inferences on all the pixels of the image.