Information extraction from high resolution SAR data for urban scene understanding

The high complexity and the increased importance of geometry in growing resolution Synthetic Aperture Radar (SAR) data in urban environments poses a limit on the usability of lattice-based scene models. Geometrical models based on marked point processes can be employed to provide better descriptions of the scene. A Gibbs potential based on a hierarchical Bayesian description of the direct model of the acquisition is defined on the process: an A-Priori measure of plausibility for the scene takes into account interactions between scene objects, while a Bayesian Likelihood term is based on the decomposition of scene objects into basic elements and on their mapping in the data space. Multiple reflections of the radar signals are considered and exploited. The resulting detectability measure is compared to a null hypothesis in a Likelihood ratio. The resulting posterior potential is optimized by Monte Carlo methods. The resulting algorithm is applied on a diverse set of single submeter resolution SAR intensity images on urban scenes, providing descriptions of the 3-d structure of the imaged urban areas in terms of separate objects.