Sparse Representation Classification Of PolSAR Image Based On Orthogonal Gaussian Random Matrix

Polarimetric Synthetic Aperture Radar(Pol SAR)image has the characteristics of all day,all-weather and wide observation area.The multi-channel characteristic of Pol SAR image can receive more scattering features and polarization information,which is very beneficial to improve the classification accuracy of Pol SAR image.Sparse Representation Classification(SRC)algorithm is one of the mainstream image classification algorithms because of its advantages of simple algorithm,easy design and implementation.However,here are still some problems in the design of SRC algorithm,such as the measurement matrix does not satisfy the RIP condition and the non-negative condition of the high-dimensional sparse coefficient vector is not considered,which leads to its low classification accuracy.In order to solve the above problems,therefore,the SRC algorithm of Pol SAR image based on Orthogonal Gaussian Random Matrix(OGRM?SRC)algorithm is proposed.The details are as follows.(1)In order to make the SRC algorithm satisfy the RIP condition,the OGRM?SRC algorithm is proposed to achieve accurate classification of Pol SAR image.Firstly,in order to reduce the speckle noises,the Pol SAR image after reciprocity is processed by multilooks processing and Refined Lee filtering.The Pol SAR image is decomposed by four polarimetric decomposition methods to obtain its polarimetric features.As a result,polarimetric feature image can be obtained.Secondly,in order to make the SRC algorithm satisfy the RIP condition,OGRM is generated by multiplying the Gaussian Random Matrix and orthogonal matrix.Based on the polarimetric features and OGRM,the observation dictionary is constructed with the typical ground targets samples of the polarimetric feature image.Non-negative high-dimensional sparse representation coefficient vector for each pixel in the polarimetric feature image is calculated by Orthogonal Matching Pursuit(OMP)algorithm with the observation dictionary.Finally,the residual of the pixel to be processed relative to each atom in the observation dictionary is calculated,and the minimum reconstruction residual is used as the classification standard.The OGRM?SRC algorithm is used for classification.The optimal classification results are obtained and the accuracy is evaluated.(2)In order to further improve the classification accuracy of OGRM?SRC algorithm,the polarimetric features are analyzed and selected to determine the optimal polarimetric feature combination.Based on four polarimetric decomposition methods,the correlation between the polarimetric features and the scattering mechanism of typical ground targets of study area is analyzed in detail by using box plots,Cloude-Pottier plane scatter plots and power mean scatter plots.The optimal polarimetric feature combination are selected according to the criteria of feature selection factor,feature judgment factor,H/? plane,A/? plane,H/A plane,mean and standard variance.Taking the above optimal polarimetric feature combination as input,the OGRM?SRC algorithm is used to achieve the accurate classification of ground targets,and the accuracy of the classification results is evaluated.