Max-Margin Classifiers For PolSAR Image

Polarimetric Synthetic Aperture Radar(Pol SAR)is a multi-channel,multi-parameter and high resolution imaging equipment.It records scattering echoes of different polarization modes by vectors and achieves rich terrain information.With a lot of investment of Pol SAR system in more and more countries,a large number of radars with full polarization measurement were developed,greatly expanding the applications of radars.There are many tasks in understanding and interpretation of Pol SAR images,including denoising,classification,detection,identification and so on.Among them,Pol SAR image classification palys a significant role.Results of Pol SAR image classification can be used as a map or an intermediate step,and have been widely used in agriculture,military,disaster relief,and other fields.Besides,the Pol SAR data are large scale,the training samples are limited,and the data are affected by the speckle.Therefore,Pol SAR image classification is an essential,challenging and hot problem.In this thesis,to achieve effective classifiers,we study the Pol SAR image and the max margin criterion.Different classifiers of max margin characteristic are designed based on polarimetric and spatial information.Meanwhile,the max margin criterion preserves the robust performance of classifier.The designed classifiers are summarized as follows:1.Least Squares Support Vector Machine(LS-SVM)achieves high accuracy in small size sample classification,but it cannot deal with large scale problem.Inspired by the compressive sensing(CS)theory,a new classifier based on compressed dictionary and LS-SVM is therefore proposed to deal with large scale problems.The coefficients of support vectors can be recovered from a few measurements if LS-SVM is approximated to sparse structure.Using the known Cholesky decomposition,we approximate the given kernel matrix to represent the coefficients of support vectors sparsely by a low-rank matrix that we have used as a dictionary.The proposed measurement matrix being coupled with the dictionary forms a compressed dictionary that proves to satisfy the restricted isometry property(RIP).Our classifier has the quality of low storage and computational complexity,high degree of sparsity and information preservation.2.Taking the special properties of Pol SAR data into account,we propose a classifier based on discriminative sensing matrix and spatial-Wishart kernel.Firstly,according to the distribution of Pol SAR data and spatial neighbourhood,we propose a spatial-Wishart kernel to describe the Pol SAR data,reduce the noise of speckle and improve the spatial consistency.Then,we design a discriminative sensing matrix,inspired by the idea of linear discriminative analysis,which aims to minimizing the within class scatter and maximizing the between class scatter,simultaneously.The discriminative sensing matrix largely reduces the scale of the optimation problem and keeps the information related to the classification as much as possible.Finally,the discriminative compressed dictionary is constructed based on the sensing matrix and the spatial-Wishart kernel.The sparse support vector coefficients are achieved by the novel pruning strategy with only one step and thus the sparse classifier is obtained with a large margin.3.Weighted Wishart distance learning,shorted for W2-based distance learning is proposed as a max margin classifier for Pol SAR image classification.Unlike learning in LS-SVM,the new classifier requires no modification or extension for problems in multiway classification.It can be seen as the logical counterpart to LS-SVM in which k-NN classification replaces linear classification.The input of the classifier is a pair of training samples.Compared with the traditional supervised classifier,it no longer depends on the sample label totally but the similar/unsimilar information of the pair of samples.Furthermore,the classifier aims to adjust the Wishart distance by enhancing discrimination as well as exploiting spatial information.The proposed distance learning keeps samples within the same category close and separates samples from the different classes far apart.It is effectively implemented by solving a linear programming.Input of W2-based distance learning is called weighted Wishart feature,which is designed specifically for Pol SAR data to describe the Wishart distribution,achieve regional consistency and reduce speckle noise.Weight is estimated according to an adaptive window,where homogeneous samples are derived based on a connected region and extracted edge information.With this feature,W2-based distance learning is a whole scheme to adjust the Wishart distance.4.Considering the high cost of manually labeling,we propose a self-supervised classifier for Pol SAR image.The self-supervision means automatically extracting training samples based on polarimetric information,without human interaction and then inputing them to the supervised classifier.The classifier has the advantage of both supervised and unsupervised methods,liberates the labor of labeling increased data and makes supervised classification in a fully automated framework.Samples generated for each class are taken as the candidate and the training samples are extracted from candidate according to the proposed credibility.Furthermore,a rectifying method is designed to achieve robust classification if the label according to candidate and the supervised method is conflicted.