Date of Award
Master of Arts (MA)
Geography and Geology
Digital image classification is a technique to extract land cover information from imagery using certain classification schemes. Additional information, either in a map, digital or verbal format, from multi-sources other than satellite imagery can be integrated into a classification scheme to improve its performance. The research in this field is essential because data interpretation has long been a weak link between the functions of two powerful systems: data acquisition by the remote sensing system and data storage, renewal and retrieval by the geographic information system - that has been more and more involved in the geographical research with the development of new concepts and technology. Statistical models are flexible in incorporating information from multi-sources into a classification procedure to extract land cover information from satellite data. When a statistical model is applied to the urban scene, however, it is difficult to incorporate any spatial and ancillary information into the classification procedure because of their unpredictable nature in an urban scene and the difficulty in defining them into a meaningful format for a machine. In this thesis, the probabilistic relaxation model is used with Landsat TM data of the Omaha metropolitan area. The purpose of this approach is to improve the overall classification accuracy from a Gaussian Maximum Likelihood (GML) classifier by incorporating the contextual information contained in the neighborhood of a central pixel into the classification procedure. Prior to classification, ten land cover types determined visually from the image were combined into six classes by comparing their relative positions in a 6-dimensional classification space. After the the completion of GML classification on the original image data, the relaxation model is applied to adjust the class membership probabilities; after each adjustment, the classified map is compared with the reference map to assess the improvement of overall classification accuracy. This process is continued until the maximum accuracy is obtained. The error pattern during the iteration is analyzed and it is found that most error pixels occur at the boundaries between classes or they are located inside the dominant classes as individual parcels. Part of these errors have been corrected through the continuous adjustment of probabilities by iteration. The rationale of all statistical models used in this study and their mathematical meanings are reviewed in this thesis. It is concluded from the actual operation of the relaxation procedure on the Landsat TM imagery of the Omaha metropolitan area that the spectral separability checking has minimized the spectral ambiguity and the relaxation model has reduced much of the identity ambiguity. The result from this combined effort is the increase of overall classification accuracy from 77.01% of the original GML classification to 88.49% after the 13th relaxation iteration. The successful operation of the procedure has shown that it has a potential to be used in a computer-assisted land use monitoring system for the purpose of resource management.
Lu, Yafei, "The improvement of digital image classification over an urban area: A probabilistic relaxation approach" (1989). Student Work. 587.
Files over 3MB may be slow to open. For best results, right-click and select "save as..."