ABSTRACT
Our methodology hinges on the ability of ancillary spatial data to identify and update the a priori probabilities of the ML algorithm. This involves the use of these data first to stratify images of urban areas according to spatial and contextual rules, and then to estimate the area of the urban classes within each tratum. These area estimates are then either directly inserted into the ML cia ifier as prior probabilities, or are u ed as part of an iterative process for creating and updating ML a posteriori probabilities. The use of ancillary spatial data to stratify feature space is a progres ion of work pioneered by Strahler (1980), who demonstrated improvement in classification accuracy of natural vegetation in Doggett Creek, California. Since then, the technique has been further elaborated by Haralick and Fu ( 1983), and upported by other work on the physical landscape. The mo t notable contributions include Skidmore and Turner ( 1988), who extracted class probabilities from grey-level frequency histograms, and Maselli et al. ( 1992), who formalized a parametric/non-parametric link between the ML discriminant function and prior probabilities. Apart from research by Barn ley et al. (1989), Mesev et al. (1996) and Mesev (1998), work in the urban phere eem to have been largely neglected, undoubtedly because of the spectral heterogeneity of
urban land use categories. It is exactly when classes are closely related, however, that prior probability estimates have the greate t effect (Mather 1985, Me ev 1998).
