Polarimetric SAR Image Classification Based On Sparse Representation And Deep Bayesian Learning

In recent years,Polarimetric Synthetic Aperture Radar(Pol SAR)is one of the significant research directions in Remote Sensing field.Pol SAR has great advantage over the traditional unipolar single-channel Synthetic Aperture Radar,such as the characteristics of all-day and all-weather,larger image scope,higher resolution,which is diffusely used in agricultural production,urban planning,natural disaster prediction etc.Under the background of it,Pol SAR image classification,the key to Pol SAR image interpretation,should be given sufficient study.Machine learning and deep learning theory are important tools for studying Pol SAR image classification.In our thesis,we study the classification of Pol SAR image on accout of the theory of sparse representation,Bayesian regularization and DBN,including the following three aspects:The first method is based on sparse representation and super-pixel generated by fixed Simple Linear Iterative Clustering(Simple Linear Iterative Clustering,SLIC).The method include the following steps: first use the K-SVD algorithm to get the learning dictionary,and then use the algorithm of SLIC to obtain the super-pixel,meanwhile,express it with joint sparse representation,utilize Orthogonal Matching Pursuit algorithm to get sparse coefficient,and finally accounding to the minimum reconstruction error to determine the super-pixel class.Sparse representation of the classifier classification process is pixel-wise,while our method introduces the super-pixel,which is processed unit,combined with the joint sparse representation classifier,improving the classification accuracy and efficiency.The second method is based on Bayesian regularized Restricted Boltzmann Machine(RBM).In our method,Bayesian regularization is introduced into the objective function of the Boltzmann machine,which is the mean of the weight's square sum.The Deep Belief Network(DBN)is constructed with the improved RBM.Firstly,extract features from the coherent matrix obtained 9-dimensional feature,the Cloude decomposition obtained7-dimensional feature,and the rest 3-dimensional featur obtained from Pauli decomposition,and then construct the 19-dimensional eigenvector,using the training sample to obtain the parameter of DBN.It has been proved that the proposed method has achieved high classification accuracy.The third method is based on multi-feature fusion and improved DBN.The classification method is based on the fourth chapter,introducing the texture features,the polarized scattering matrix algebraic operation of the characteristics.All the extracted features are merged as the constructed eigenvector,and the improved DBN be trainned by selected samples.It has been proved that this method not only improves in terms of accuracy,but also keeps the details of the margin information more complete.