Détection de changements et classification sous-pixelliques en imagerie satellitaire. Application au suivi temporel des surfaces continentales.

摘要

This thesis focuses on the land cover analysis and monitoring from remote sensing time series. The use of data with different resolution is critical for both a good discrimination and a good localization of the objects of interest. In this context, we propose two approaches for sub-pixelic classification and change detection, using very few a priori information. The first one is based on the definition of an energy function in a Bayesian framework. Given a number of classes, it enables an unsupervised estimation of the classification as a minimum of this energy function, through a simulated annealing algorithm. The second one is based on an a-contrario detection model with a stochastic algorithm that automatically selects the image subdomain representing the most likely changes. A theoretical and experimental analysis of the proposed approaches enabled to estimate their limitations and, in particular, to show their capability to deal with high resolution ratios. Actual applications are presented in the case of an agricultural scene of the Danubian plain (ADAM database).