Research On Extended Terrain Classification In High-resolution, Single-look And Single Polarimetric SAR Image

This theme is research on extended terrain classification that is from Automatic Terrain Classification(ATC)technology. The purpose of this theme is to classify SAR images into four regions:man-made targets(vehicles and buildings),natural targets(trees and shrubbery),background(field and grass)and shadow(the target shadow and mirror scatter). On the one hand,the processed results can be directly used as image product,which offer necessary parameters for expert's interpretation to build up an independent auxiliary interpreting system. On the other hand,the results can be introduced into SAR Image interpreting system as the replacing target detection and discrimination models with layer ROIs(region of interest)in order to provide potential target chips. Therefore,the work of this thesis is a very practical and pivotal part in the SAR image interpretation technology,which is still a very challenging task.Because of the full target texture and the acute speckle of the microwave-imaging machine,especially for the single frequency,single look and single polarization SAR images,It is very difficult to extract image characteristics in the image pixel level. Therefore,it is a puzzle to study the classification model,increase the class number and improve the detection ratio according to the classic classification algorithms.In theory,Based on the characteristic detection probability model of SAR image,the ability of distinguishing two connected resolution unit and detecting point target is discussed firstly. Then we do conclude that the high precision classification algorithms can build up based on pixel level,which combine three characteristics:the main characteristic(the RCS statistical distribution),two assistant characteristics(the shadow and structure model). Secondly,the practical problems of the three pattern classification technologies(statistical pattern recognition,neural network,fuzzy neural network)for the SAR image are analyzed. Finally,some evaluating measures for image quality and classification are given.In algorithms,classification algorithms are divided into two cases:one for known statistical distribution model and the other for unknown statistical distribution model. Four classification algorithms,the Bata-Prime statistic model fusing Quadratic Gamma classifier,based on SAR image RCS reconstruction and space position mode,on the Mixed Double Hint Layers RBFN(MDHRBFN)model and on the Self-adapt Fuzzy RBFN(AFRBFN)Model,are derived. The problems,including how to further improving the class ratio of the Bayes decision,decreasing the dependence on the statistical model and directly providing the adapted algorithm with samples,are solved. Different algorithms can preferably solve theclassification problem in the different application background respectivelyIn the end,the system software frame is presented. Based on the real MSTAR data,the four algorithms are evaluated.