Classification Of Polarimetric SAR Image Based On SVM And RBF Neural Network

Polarimetric synthetic aperture radar (PolSAR) is a new system radar that canmeasure targets in full-polarization, which realize the improvements of capability forclutter retraining, target detection and terrain classification by obtaining multi-channelPolSAR images. The classification of PolSAR images is basic and forehand issue of theinterpretation for PolSAR images in current, the aiming of this classification is to utilizemeasured data of multiple polarization to accurately labeling the corresponding terrainclass of each pixel. Herein, feature extraction and classifier designing are two keyproblems influencing the classification accuracy of PolSAR images. Devoting tosolving them, the dissertion has done the following works:First, we analyze the the scientific signification and application value of thePolSAR system. Furthermore, the historical development, researching status andfundamental theories of PolSAR images classification techniques are alsosystematically summarized.Second, some researches are conducted on the polarimetric features extraction.Eighteen polarimetric characteristics including SPAN,degree of polarization and so onare extracted based on the features acquired by the simple transforms of the measureddata. Meanwhile, aiming at the features based on target decomposition, we employseveral methods, like the Pauli decomposition， Cloude decomposition, Freemandecomposition, etc., to extract thirteen various polarimetric features. Additionally, themaximum likelihood (ML) distance of the polarimetric G0distribution (i.e.,) is alsointroduced to construct the feature set, which help form a comprehensive feature set forthe next classification task of PolSAR images.Third, an algorithm of PolSAR image classification is proposed based on targetdecomposition and support vector machine (SVM). Several kinds of scattering featuresof PolSAR image are firstly extracted to form the feature vector of each pixel in theimage via two decomposition methods, namely, Cloude decomposition and Freemanone. Then, the SVM is trained with the feature vector of the pixels in the sample areas.Next, with the feature vector of pixels ready for classification as the input of the trainedSVM, the classification of PolSAR image trends to be completed. The experimentalresults on the measured PolSAR images demonstrate that the proposed method is able toutilize effectively the complementary information of various features and shows a fineclassification precise.Finally, A method of classification for PolSAR images based on the radial basisfunction (RBF) neural network is proposed in this paper. On the basis of constructingthe classification feature set（contain the ML distance of the polarimetric G0distribution and some normal characteristics） of PolSAR image, the designing of classifier iscompleted by using the samples to train RBF neural network. The experimental resultsof classification on some measured PolSAR images show that the proposed method haswell performance of keeping the image details.