KNOWLEDGE-BASED SEGMENTATION OF LANDSAT IMAGES

In this paper a knowledge-based approach for Landsat image segmentation is proposed. Landsat images are complex; therefore semantic knowledge of the image and processes generating the image should be exploited to improve the reliability of segmentation procedures. We attack the image segmentation problem by extracting kernel information from the input image to provide an initial interpretation of the image, and utilizing a knowledge-based hierarchical classifier to discriminate between major land-cover types in the study area. The knowledge necessary to drive the hierarchical classifier is derived partially from training data. A more detailed interpretation of the image is then produced using a spatial clustering technique, the previously extracted kernel image information, and spectral and spatial rules which make up the knowledge base of the hierarchical classifier. Experimental results from a limited set of test images yield promising results. Topics for future research include the development of more spatial rules and use of prior map information in the image-segmentation procedure.