An evidential Markovian model for data fusion and unsupervised image classification

The authors deal with the fusion of information and the classification of images supplied by several sensors. By intrinsic characteristics of each sensor, the provided information is usually defined on a different set of hypotheses, called frames of discernment. An adapted formalism is needed to compute the fusion process. We resolve this problem of multi-sensor image fusion and classification in an evidential framework which is well adapted for the combination of knowledge defined on different frames of discernment. We present two models for merging available information, a non contextual and a vectorial model which is defined by using a Markov chain structure to represent a priori knowledge associated to labelling image. In the Markovian approach, the Markovian property is preserved after fusion, which enables us to apply standard classification algorithms. We adopt an unsupervised context in which parameter estimation is done by using a mixture distribution algorithm, the ICE algorithm. We apply these models to satellite images.