A mutual approach based on Markov random fields for multitemporal contextual classification of remote sensing images

Previous work has shown that the exploitation of temporal contextual information can improve significantly the performances of a classifier from the viewpoint of accuracy. In particular, Markov random fields (MRFs) have been found to provide a useful and theoretically well-established tool to integrate spatial and temporal contextual information in the classification process. As an alternative to the usual "cascade" approach adopted to deal with a temporal sequence of images, in this paper, a "mutual" approach is proposed that is based on a coupling of the MRF models defined for each single-time image. Under the form of appropriate energy functions, each single-time MRF model integrates three basic kinds of information: spectral, spatial and temporal contextual information. Spectral information is conveyed by the single-time posterior probabilities estimated by multilayer perceptron neural networks. The use of neural networks is motivated by their effective multisensor fusion capability. These estimates are also utilized to provide the initial classification maps required to start the iterated conditional modes algorithm. This algorithm is applied to find approximations of the MRF-MAP estimates, or in other words for energy minimization, because it represents a simple and computationally moderate solution for such a purpose. At each iteration, the previous MRF-MAP estimates of all the single-time images are exploited to update the MRF-MAP estimates for each single-time image. While spectral information for a given pixel do not change with iterations, both spatial and temporal contextual information are derived on the basis of the current estimate of the labels for the neighboring pixels. The determination of the MRF parameters, which play the roˆle of information source weights in the model, has always been a difficult issue. To deal with this problem, we propose a method based on "minimum perturbation" implemented by means of the minimum square error pseudo-inverse technique. Experimental results on a multitemporal data set consisting of two multisensor (Landsat TM and ERS-1 SAR) images are presented. The performances of the proposed "mutual" approach are compared with those obtained by a reference MRF-based classifier in terms of classification accuracy