A contextual classifier based on Markov random fields and robust M-estimates

Presents a maximum a posteriori classifier in conjunction with modified M-estimates. Since spatially adjacent pixels likely belong to the same class, the authors use the Markov random field model to characterize the discrete field containing the individual pixel classifications. The complexity of the model is restricted to Markov random field since each point is dependent only on its neighbors. Each pixel is classified into one of the classes in the possible class set such that a posteriori probability is maximized by the assigned class. Optimization is performed using iterated conditional modes to alleviate computational load. Iterations continue to update each individual pixel until no more updates occur for all pixels. In defining the conditional density functions of the observations, the authors use modified M-estimates because of their distributional robustness. Mean and covariance are, therefore, derived from a robust criterion. Consequently the modified M-estimates yield more accurate and robust initial classification than maximum likelihood estimates. The initial classification is of importance since the iterated conditional mode converges to a local minimum. Applying this approach to multispectral remote sensing data, the authors substantiate their method by comparing the overall classification performances with that of a conventional maximum likelihood classifier.