Knowledge-based classification of SAR images

Classification of SAR images remains a challenge. One of the best efforts to date [Van Zyl and Burnette, 1992] only classifies three categories: ocean, dense city, and dense vegetation while training and testing the classifier on the same image. Using JPL AirSAR images a technique has been developed to classify each pixel as either of 4 classes: city, flat (water, pavement), short vegetation, or tall vegetation. The technique uses the backscattered powers in the co- and cross-polarized channels at L- and C-bands as well as the phase difference between the co-polarized channels. Several different measures of these basic parameters were formed to aid in the discrimination process. For instance, the texture [Ulaby, et al., 1986] of each pixel at each frequency and polarization was estimated using a 5×5 pixel mask; also, various ratios of parameters were formed. The decision strategy was developed from measures on a single image, then tested on another image of another area, with the accuracy of the classifier being checked against a land-use map. The method presented is intuitive and easily extended to finer classification categories given appropriate ground truth and image data