Use of textural features in the analysis of Landsat images

The use of remotely sensed data for land-cover categorization has been well demonstrated in recent years. l In addition to the use of conventional methods, such as windshield surveys and lowand high-altitude aerial photographs, satellite data are being employed by several planning agencies for creating land-use and land-cover data bases to assist in decisionmaking. Human interpretation of aircraft and satellite imagery is still used extensively and is essential for certain applications (e.g., lineament studies in geology). However, automatic interpretation of the image data using digital computers significantly complements human analysis and is advantageous in providing consistency, labor savings, and timely availability of interpreted data for postprocessing and inclusion into data bases. The major differences between human and machine interpretation of images are as follows. A human interpreter uses color, spatial variation of color, and relative location of objects (i.e., spectral, textural, and contextual features) in an image simultaneously. Starting with a synoptic view of the illlage, the human interpreter partitions that image into homogeneous segments and associates labels with each of them. It is difficult to design algorithms for machine interpretation which simulate the human photointerpreter. However, machine analysis has the advantage of high resolution not achievable with human interpretation. Although it is possible to incorporate spectral, textural, and contextual features into machine analysis, the most commonly used classification algorithms employ only the spectral data and assign categories (or class labels) to single resolution elements (pixels) independently of other pixels. The accuracy of classification maps is a principal concern for users. Studies of classification algorithms have been conducted which compare ground-truth maps with the classifica.tion maps generated by machine interpretation of Landsat multispectral data using only the spectral features at individual pixel locations. For example, in Reference 2, confu'sion is shown to exist between certain urban and agricultural cover types. This is due partially to the similarity in the spectral characteristics and partially to the differences in human and computer characterization of classes (e.g., when a region with large residential lots is encountered, human interpretation for the entire region is "urban,"