COMPUTER INFORMATION EXTRACTION FROM RADAR IMAGES (REMOTE SENSING SYSTEMS)
As remote sensing systems begin to make radar imagery more available, it becomes increasingly important to realize and exploit the full information content of the media. The goals of this investigation are threefold: (1) to assess the applicability of commonly used image manipulation tools to radar imagery, (2) to investigate and the quantify the added value brought to the classification problem by radar imagery, and (3) to investigate the value of local spatial features, derived from a radar image, in the classification undertaking.^ The first portion of the study investigates the statistical properties of radar and visible image data. Three different digital radar data sets are examined and their autocorrelation functions computed. The autocorrelation functions are found to be substantially lower than those of visible images. The effects of resampling on imagery with this lower correlation coefficient are found to be more severe than for visible imagery. The implication of these findings on information extraction from radar images is that spatial properties, which contribute a significant amount of information to classification success, are diminished.^ In the second and most significant portion of the investigation, the relative abilities of visible texture measures and radar roughness measures to separate bare fields of different roughnesses are examined. The visible photographs and multi-frequency multi-angle multi-polarization microwave data set were acquired by the University of Kansas in 1975. Some 104 textural features, including co-occurrence, Fourier transform, differences-of-averages, gray level run lengths, autocorrelation and Mitchell's max-min measures, were used in the classification of digital segments of the visible images. A proposed set of 390 microwave roughness features including models from inverse scattering theory as well as heuristic measures, were extracted from the microwave data set. Feature success was determined through the combined use of Fisher and Bhattacharyya distances. The single-feature results indicate that the microwave features are superior in differentiating the field roughnesses. However, for higher dimensionalities the benefit of the combined use of microwave and visible features is indicated.^ The third portion of the investigation compares the relative abilities of spectral and textural features applied to visible and radar images of the same areas on the ground, acquired from aircraft and satellite altitudes, to identify specific ground materials. In the low altitude aircraft comparison, the radar textural features were the best performers in a classification experiment while the visible spectral values offered the least separability. For the higher altitude data, the visible texture features provided the superior performance.^ General conclusions are as follows: (1) The simultaneous use of visible and microwave sensors offers an alternative to researchers in the artificial intelligence and remote sensing communities who are attempting to duplicate human perception and recognition (Scene understanding). (2) The information value of a number of simple features has been demonstrated. (3) The information value of spatial features applied to radar imagery has been shown to be diminished through the resampling process. (4) A systematic method of feature evaluation has been demonstrated for comparing the value of large numbers of features of different wavelengths. ^
Engineering, System Science
LAWRENCE DANIEL ALEXANDER,
"COMPUTER INFORMATION EXTRACTION FROM RADAR IMAGES (REMOTE SENSING SYSTEMS)"
(January 1, 1981).
Dissertations available from ProQuest.