Unsupervised Classification Algorithm Based On Polarimetric Sar Images

Because of the high 2-Dimensional resolution of the synthetic aperture radar (SAR), it has been consider as one of the main methods for radar classification and recognition. Polarimetric SAR combines the high space resolution of the SAR and the polarimetric information of polarimetric radar, can reveal the scattering feature further, and improve the accuracy of the SAR image classification, which is one of the hotpoint in the field of SAR application. This dissertation mainly researches the method of reducing speckle for polarimetric SAR images and unsupervised classifying the polarimetric SAR images. The main contributions of the dissertation are as follows:1. Introduce the fundamental theory of electromagnetic wave polarimetry including the description of the EM wave's polarization states, representation of the polarization states on maps, the Sinclair scattering matrix and Mueller matrix, and summary the popular algorithms of polarimatric SAR classification.2. Analyze the speckle in SAR/polarimetric SAR image, including the reason and effect of the speckle on application of SAR/polarimetric SAR image, and the statistic feature of the speckle. Summary the typical methods to reduce the speckle and introduce the standard of speckle reducing. Finally, based on the polarimetric Lee filter technique, adaptive window speckle reducing method for polarimetric SAR image is presented in this paper.3. Analyze the theory on polarimetric decomposition and the advantage and disadvantage of entropy/alpha plane polarimetric SAR classification method. Discuss the unsupervised classification method based on the Wishart distribution. Based on this, a new dynamical cluster unsupervised classification method for the polarimetric SAR image is presented to overcome the shortcoming of the traditional methods, which can not adjust the class number automatically. Finally, the feasibility of this new method is verified by polarimetric SAR images.4. Analyze the technique of the support vector machines(SVM), and employ the SVM method to the classification of polarimetric SAR images. Based on this, a new for the polarimetric SAR image is proposed by combining the SVM and polarimetric decomposition techniques. Finally, the feasibility of this new method is verified by polarimetric SAR images.5. Analyze the effect of the number of on the classification error. Based on the principle of minimum support vector number, we present a new unsupervised classification method by combining the SVM and Quad-tree techniques, and the effect of different training sets select is discussed by polarimetric SAR images.