Multitemporal image classification with automatic building of tree-structured MRF models

In this work we deal with the classification of remote-sensing images following a statistical approach. To take into account prior information on the class of images of interest we model the image as a tree-structured Markov random field (TS-MRF), so as to fit the intrinsic structure of the data. TSMRF models are defined recursively and, as such, lead to the formulation and solution of the segmentation task as a recursive problem, so that the original K-ary segmentation is decomposed into a sequence of reduced-dimensionality steps, and hence to a much simpler and more manageable segmentation. Here, we propose a method to build automatically the underlying tree structure of the model, based on a metric which compares class features in order to establish the hierarchical relationships among classes, and apply the technique to the segmentation of multitemporal remote-sensing images.