A Versatile Classifier Model for Multiobservational Analysis

The maximum likelihood decision rule, widely applied to the analysis of multispectral remote sensing data, can be generalized to handle a broad class of problems involving multiple observations. Such problems arise, for example, when it is desired to classify a location on the ground based on multiple passes over the site (temporal context); or to incorporate data from adjacent locations in the decision process (spatial context). The generalization is accomplished by redefining the classification objective and applying statistical decision-theoretic methods. As a simple example, if observations X1=X(t ) and X2=X(t2) are available from two satellite passes, then under appropriate assumptions it is possible to show that an optimal decision procedure is to classify the location into the class w2=w(t2) for which the posterior probability p(w2 1 x1,x2) is maximum. The trick, however, is to express this probability in terms of quantities which it is feasible to observe or estimate. Several applications of this classifier will be discussed together with experimental results of its use