The effects of viewing geometry on image classification

Abstract Thematic classification of remotely-sensed data assumes that class separability is aspatial. However, as classifiers discriminate between classes on the basis of their spectral responses this assumption is invalid since the latter vary spatially with the viewing geometry. Consequently classification accuracy is spatially variable. Classification accuracy statements can therefore over- and under-estimate inter-class discrimination at different locations within a scene and can therefore be misleading. This is illustrated by a classification of synthetic aperture radar data which had an aggregate accuracy of 75-7 per cent but also displayed a variation in accuracy between locations of up to 17-2 per cent.