抄録
Most of the pixel-by-pixel supervised classification methods for remotely sensed multispectral images are quite inefficient, because the discrimination procedure is applied to every pixel even in a uniform area consisting of the data belonging to only one class. As they make no use of the spatial properties of objects, they do not achieve high accuracy, either. As for the “per field” classification method, there is no established technique for segmentation of an image into “fields”.
In this paper is proposed a classification method where the spatial resolution for classification is adaptively varied according to the local uniformity of the image. The adaptive division of the area to be classified is realized on the assumption that the image has hierarchical (pyramid) structure. It is achieved by a two-step procedure, each step of which makes use of a statistical test of local uniformity of the image.
The classification accuracy and the processing time by this method are compared with those by the pixel-by-pixel maximum likelihood method which is most often used for classification of remotely sensed images. The comparison shows that this method improves the performance of classification both in efficiency and in accuracy.
