Polarimetric SAR Image Classification Based On Conditional Random Field Model

Polarimetric synthetic aperture radar(SAR)image classification can interpret the environment information timely and accurately,and thus promote the application of polarimetric SAR system in the fields of remote sensing,such as mapping,ocean research and battlefield detection.Therefore,polarimetric SAR image classification is a key research topic in the field of microwave remote sensing.Conditional random field(CRF)is a powerful tool for polarimetric SAR image classification.How to mine the polarization features with high discrimination and generalization abilities,and how to build an accurate and universal statistical model of polarimetric SAR data are two key techniques of CRF model for polarimetric SAR image classification,which could improve the classification accuracy.In this thesis,the feature mining and statistical modeling of polarimetric SAR data are studied.Then,in the CRF framework,the effective polarization feature information,scattering statistical information and spatial correlation information of polarimetric SAR images are incorporated based on Bayesian information fusion to achieve accurate classification of polarimetric SAR images.The main work of this thesis can be summarized as follows.1.A feature mining method of polarimetric SAR image based on generalized robust principal component analysis(GRPCA)is proposed to improve the discrimination ability of polarimetric SAR image features.The main work is as follows.1)Firstly,GRPCA uses simple linear iterative clustering algorithm to obtain super-pixels,which is the preprocessing of feature mining.Then,according to the consistency of super-pixels,the low-rank sparse decomposition of polarimetric SAR data is carried out by robust principal component analysis to decouple the polarization feature information from the speckle and thus improve the robustness of polarization feature against speckle.2)In view of the non-Gaussian statistical characteristics of polarization feature,GRPCA uses generalized principal component analysis algorithm to reduce feature redundancy and amount of data,and obtain the low dimensional features with discrimination ability in the meaning time.Experimental results demonstrate that GRPCA algorithm is effective in feature mining of polarimetric SAR images.2.Statistical modeling of polarimetric SAR data based on Wishart-generalized gamma(WGΓ)distribution is studied.To achieve robust and accurate estimation of texture component parameters in the WGΓ distribution,we propose a maximum likelihood and logarithmic cumulants(ML-LC)method to estimate the generalized gamma distribution(GΓD)parameters.The ML-LC method constructs a novel scale-independent shape parameter estimator in the log-domain based on the Mellin transform and maximum likelihood estimation,and estimates the GΓD parameters based on the multistart local search,gradient descent,and bisection methods,rather than solving the system of highly nonlinear equations in the traditional estimations.In this way,the ML-LC method realizes accurate estimation of the texture component parameters,thus improving the accuracy of the WGΓ distribution.The goodness-of-fit tests demonstrate that the ML-LC method is universal and accurate in estimating GΓD parameters.The classification experiments verify the effectiveness of WGΓ distribution for the statistical modeling of polarimetric SAR image.3.We propose a novel CRF model that can incorporate polarization features,scattering statistics,and spatial correlation based on Bayesian information fusion,named as SCRF-GRPCA,for polarimetric SAR image classification.The main work is as follows.In the framework of random fields model,GRPCA algorithm proposed in Chapter 3 is used to extract discriminative features and construct the unary potential of SCRF-GRPCA model,thus measuring the class probability of each pixel.Then,the WGΓ distribution model is used to construct the likelihood term in the SCRF-GRPCA model,in which the distribution parameters are estimated by the ML-LC method proposed in Chapter 4.Then,the SCRF-GRPCA model achieves a Bayesian fusion of polarization features obtained by GRPCA,scattering statistics modeled by WGΓ distribution,and spatial correlation,thus capturing Pol SAR image information in a more complete manner and improving the classification performance further.The experimental results verify the effectiveness of SCRF-GRPCA model for polarimetric SAR image classification.