The first stage in the analysis of remotely sensed data is image segmentation and classification. The early approaches to this problem were based on the Bayesian decision rule for classifying pixels x on individual basis. Recent studies showed that the segmentation performance can be considerably enhanced by incorporating contextual information in decision making schemes. The contextual information is conveyed by neighbouring pixels x to that being classified. it can be of several types but herein we shall concentrate on the two main sources of context: spatial pixel category dependencies; two-dimensional correlations between pixels.

Commencing from the Bayes decision rule assign x to ω₁ if a general decision making scheme which incorporates both types of context has recently been derived [1]. This general scheme has been simplified under various assumptions so that it can be implemented in practice. One of the simplifying assumptions is that contextual information is conveyed only by adjacent neighbours to the pixel being classified. To extend this neighbourhood it is necessary to resort to probabilistic relaxation algorithms.

In this paper the relationship between contextual decision making schemes and relaxation algorithms is discussed. It is shown that the compatibility function of the conventional relaxation scheme [2] corresponds to the compound decision rule [3]. other compatibility functions suggested in the literature (e.g. Peleg [4]) are also related to the contextual classification method [1].