Thèse soutenue

Apprentissage automatique des classes d'occupation du sol et représentation en mots visuels des images satellitaires

FR  |  
EN
Auteur / Autrice : Marie Lauginie Lienou
Direction : Henri MaîtreMihai Datcu
Type : Thèse de doctorat
Discipline(s) : Signal et images
Date : Soutenance en 2009
Etablissement(s) : Paris, ENST

Mots clés

FR

Mots clés contrôlés

Résumé

FR  |  
EN

Land cover recognition from automatic classifications is one of the important methodological researches in remote sensing. Besides, getting results corresponding to the user expectations requires approaching the classification from a semantic point of view. Within this frame, this work aims at the elaboration of automatic methods capable of learning classes defined by cartography experts, and of automatically annotating unknown images based on this classification. Using corine land cover maps, we first show that classical approaches in the state-of-the-art are able to well-identify homogeneous classes such as fields, but have difficulty in finding high-level semantic classes, also called mixed classes because they consist of various land cover categories. To detect such classes, we represent images into visual words, in order to use text analysis tools which showed their efficiency in the field of text mining. By means of supervised and not supervised approaches on one hand, we exploit the notion of semantic compositionality: image structures which are considered as mixtures of land cover types, are detected by bringing out the importance of spatial relations between the visual words. On the other hand, we propose a semantic annotation method using a statistical text analysis model: latent dirichlet allocation. We rely on this mixture model, which requires a bags-of-words representation of images, to properly model high-level semantic classes. The proposed approach and the comparative studies with gaussian and gmm models, as well as svm classifier, are assessed using spot and quickbird images among others.