Estimation of context in random fields

When considering the problem of the classification of satellite data, several authors have shown the superiority of the contextual method over the blind (pixel by pixel) method. Determining the discriminating functions which define the best contextual strategy (in the Bayesian sense) requires knowledge of the distribution of the restriction of the random field which models the picture, to the given context. We propose an estimator of this distribution and demonstrate its good asymptotic behaviour in very weak hypotheses.