PolSAR Image Classification Based On Statistical Distribution And Random Field Model

Polarimetric synthetic aperture radar(PolSAR)transmits and receives radar signals alternately in horizontal polarization and vertical polarization.The measured PolSAR data is not the backscattering coefficient of targets,but the complex scattering matrix or the complex covariance matrix.Thus,PolSAR data can obtain more comprehensive and abundant information,and then establishes the significant basis for further studing the scattering information of targets.Along with the successful development and launch of spaceborne and airborne PolSAR systems,considerable PolSAR images become available,and then provide necessary supports for images analysis and interpretation.As a practical and critical step of automatic interpretation,classification of PolSAR image aims at assigning pixels in image into different classes according to the characteristics of the classifying elements.It can reveal the structrue and the nature of images,and is the basis of automatic target recognition and detection for PolSAR system.Such a process is attracting a growing interest in civil and military applications.In this dissertation,we commit ourselves to the study of PolSAR image classification techniques.For PolSAR images,several classification technologies on the basis of the statistical distribution and the random field model are proposed for improving the flexibility of statistical models,providing stronger noise immunity and edge preservation,implementing fast and unsupervised image classification,and realizing effective feature extraction and fusion schemes.The main contents of this dissertation are summarized as the following five parts.In the first part,focusing on the greater flexibility of the generalized Gamma distribution(G?D)in the statistical modeling,we propose the Wishart-generalized Gamma(WG?)distribution for PolSAR data,and then apply it to real PolSAR image classification.The WG?distribution releases the limitations of Wishart,K and Kummer U distributions in modeling PolSAR data,and provides better fitness for different land cover types of homogeneous,heterogeneous,and extremely heterogeneous terrains.In the proposed model,we respectively utilize the complex Wishart distribution and the G?D to model the speckle component and texture component,and then derive its closed-form expression based on the polarimetric product model.Then,we propose a parameter estimation of the WG?distribution based on the method of matrix log-cumulants(Mo MLC).Finally,we introduce it into the Markov random field framework to classify PolSAR images.The experiments demonstrate the effectiveness of the WG?distribution.In the second part,based on the studies of the correlations among statistics and the different contributions in the model inference,we propose the mixture WG?-MRF(MWG?-MRF)model for PolSAR image classification.The proposed algorithm well maintains the correlations among statistics,and fuses the spatial-contextual and edge information of images.Thus,it can simultaneously provide strong discriminating ability,good noise immunity and edge locations.In the mixture model framework,we first propose the WG? mixture model based on the WG? distribution proposed in the first part.In each law of the MWG?-MRF model,we construct the interaction term by the ratio of exponentially weighted average(ROEWA)and the multilevel-logistic(MLL)model,and the likelihood term by WG? model,so that each law of the MWG?-MRF model can achieve an energy function.Then,we get the mixture energy function by fusing the energy functions of every law,and further obtain the MWG?-MRF model.Finally,we design a parameter estimation based on the iterative conditional estimation(ICE),and realize classification.Experiments on simulated data and real PolSAR images demonstrate the effectiveness of the MWG?-MRF model.In the third part,to effectively fuse the fuzziness of polarized scattering mechanisms in classification,we propose the fuzzy triplet discriminative random fields(FTDF)model.The proposed model can well maintain the dominant scattering mechanism(DSM)information in classification,thus obtaining more promising results.First,we analyze and define the fuzziness of polarized scattering mechanisms,and utilize an auxiliary random field to describe it.Then,under the guidance of the auxiliary random field,we construct the FTDF model.For the pixels with specific DSMs,the FTDF model combines the multiple features of PolSAR data by kernel k-means(KKM)to enhance the classification.And for the pixels with fuzzy DSMs,the FTDF model introduces the fuzzy c-means(FCM)algorithm to promise the classification.Finally,we realize the FTDF-based PolSAR image classification by the iterative conditional models(ICM)algorithm.Experimental results demonstrate that the proposed FTDF can provide higher accuracy,and is suitable for the classification of PolSAR images with complex scenes.In the fourth part,to enhance the computational efficiency of the traditional pixel-based discriminative random field model,we propose a novel classification algorithm for PolSAR images with the superpixel-based hybrid discriminative random field(sp-HDRF)model.The proposed algorithm can capture more direction and spatial information with superpixel segmentation,and can also reduce the number of classifying elements and the complexity of model.Thus,the proposed algorithm can provide stronger noise immunity,and enhance the computational efficiency simultaneously.In the sp-HDRF model,the simple linear iterative clustering(SLIC)algorithm,which is modified by introducing the ROEWA operator and the symmetric revised Wishart distance,is utilized to obtain a superpixel graph.Then,on the superpixel graph,the sp-HDRF model is defined.The unary potential is constructed in the same way as the FTDF model in the third part,the pairwise potential considers an adaptive neighbour system,and the data term is modeled by the complex Wishart distribution.Finally,based on the ICM algorithm,the sp-HDRF model is utilized to implement classification.Experiments demonstrate the effectiveness of the sp-HDRF model,providing an idea for fast PolSAR image classification.Focusing on the fact that the it is generally to obtain the number of clusters in real PolSAR image,the fifth part proposes the Dirichlet process mixture model and Markov random fields with similarity measure(DPMM-SMMRF)model for unsupervised PolSAR image classification and segmentation.The proposed method effectively combines the advantages of the Dirichlet process mixture model(DPMM)and the MRF,thus it can simultaneously provide good noise immunity and recognize the number of clusters automatically.First,based on the similarity measure between the neighboring polarimetric covariance matrices,the MRF and DPMM models are utilized to construct the prior distribution.And the likelihood term is constructed by multi-dimensional Gaussian distribution(MGD)based on the multi-dimensional polarimetric features.Then,we construct the conditional distribution of the DPMM-SMMRF model,and derive the conditional posterior probabilities of the class labels and model parameters under the Bayes framework.With the conditional posterior probabilities,we propose a detailed sampling procedure based on the Gibbs sampling.Fanilly,the classification is iteratively executed,until the label field is converged.Experiments on real PolSAR images demonstrate the effectiveness of the DPMM-SMMRF model,thus being a good prospect for application in for unsupervised PolSAR image classification and segmentation.