Modified Maximum Likelihood Classifications of Urban Land Use: Spatial Segmentation of Prior Probabilities

Abstract

A technique for improving the classification of urban land use by modifying the standard maximum likelihood discriminant function within spatially segmented urban parameters is introduced. The methodology examines how population census data can be used to modify prior probabilities, not at the global or regional scales, but at the intra‐urban local level. This involves spatial segmentation of a Landsat TM image by areal census collection boundaries from which prior probabilities are generated with respect to varying proportions of three housing types. Consistent improvements over classifications produced by standard equal prior probabilities are evident on four settlements in southwest England. Total absolute error is lower under modified prior probabilities, but no one housing type has consistently lower area estimation error. Evidence suggests that high density housing is under‐estimated (less pixels classified), and conversely that low density housing is over‐estimated (more pixels classified).