Abstract
Unsupervised contextual image classification of land-cover categories is discussed by taking mixel information into account. From the knowledge that most of mixels locate in boundaries of land-cover categories, we first detect edge pixels and remove them from the image to reduce influence of mixels. Then, we make classes by clustering spectrum observed at the remaining pixels which are considered as pure pixels. We here introduce a new measure of spatial adjacency of the classes. The classes are aggregated into categories by the use of the adjacency measure. Further, class-labeling of the pure pixels are updated by a novel technique based on the Markov-Random-Field model of the image. Finally, the mixels are allocated to the categories, and the category-labeling of the mixels are updated similarly. Thus, all the pixels are assigned to one of the land-cover categories. We apply the proposed method to LANDSAT TM images. The method shows an excellent performance compared with the noncontextual maximum likelihood method.
