A modified probabilistic relaxation approach to land-cover classification

Abstract

The probabilistic relaxation (PR) method uses neighborhood information in an iterative manner to reduce ambiguities caused by single-step operators. This method has rarely been applied in remote-sensing multispectral classification. One reason is that tasks such as land-cover classification involve a large number of classes spread over a large spatial area. To use the PR method in such situations, one needs a considerable amount of random-access memory (RAM) and computation time to process the large volume of image data. A modified PR method, in which the RAM and computation requirements are reduced, is presented. SPOT HRV (high resolution visible) data were used in the test, and twelve land cover classes were obtained. A preliminary comparison shows that both the original and the modified PR methods produce higher classification accuracies than a maximum-likelihood classification.