Extraction de réseaux linéiques à partir d'images satellitaires et aériennes par processus ponctuels marqués. (Line network extraction from satellite and aerial images by marked point processes)

This thesis addresses the problem of the unsupervised extraction of line networks (roads,rivers, etc.) from remotely sensed images. We use object processes, or marked point processes,as prior models. These models benefit from a stochastic framework (robustness w.r.t. noise,algorithms, etc.) while incorporating strong geometric constraints. Optimization is done viasimulated annealing using a Reversible Jump Markov Chain Monte Carlo (RJMCMC) algo-rithm, without any specific initialization.We first propose to model line networks by a process whose objects are interacting linesegments. The prior model is designed to exploit as fully as possible the topological propertiesof the network under consideration through potentials based on the quality of each interaction.The radiometric properties of the network are modeled using a data term based on statisticalmeasures.We then extend this model to more complex objects. The use of broken lines improves thedetection of network junctions and increases the accuracy of the extracted network.Finally, we propose a hierarchical model of hydrographic networks in which the tributariesof a given river are modeled by a process of broken lines in the neighborhood of this river.For each model, we accelerate convergence of the RJMCMC algorithm by using appropriateperturbations.We show experimental results on aerial and satellite images (optical and radar data) toverify the relevance of the object process models. Key-words: Stochastic geometry, marked point processes, simulated annealing, RJMCMC,line network extraction, aerial and satellite images.