PolSAR Image Classification Via Joint Clustering And Sparse Representation

Polarimetric Synthetic Aperture Radar(Pol SAR) can analyze scattering waves of ground objects in detail, and provide efficient features and scattering information of targets. Therefore, it has been widely used in both military and civil fields. Pol SAR land cover classification is the key issue of the understanding and interpretation of Pol SAR data. A large number of supervised and unsupervised classification methods have been proposed. However, those methods mainly focus on the statistical and scattering characteristics of Pol SAR data, and regard little the features of Pol SAR image itself. Therefore it cannot make full use of information involved in the Pol SAR data. In this dissertation, we have made thorough research to Pol SAR image features, and combined polarimetric characteristics with texture characteristics for better classification results. In the spatial-polarimetric domain, aiming at the existing problems in clustering analysis and sparse representation of Pol SAR image, we make further study on feature extraction, supervised and unsupervised classification methods.The main work and innovations of this dissertation are summarized as follows:(1) Existing spectral clustering methods only consider the spatial or polarimetric information of Pol SAR data, therefore cannot efficiently describe the physical characteristics of Pol SAR image. Furthermore, they are sensitive to scaling parameters. To solve these problems mentioned above, a novel spectral clustering ensemble method with combined similarity measures is proposed for Pol SAR land cover classification. The proposed method combines the complementary information from the spatial and polarimetric domain of Pol SAR data, and applies the combined similarity measures to construct the similarity matrix of spectral clustering. It makes full use of the information involved in the spatial-polarimetric joint domain. The proposed method also takes the advantages of spectral clustering ensemble strategy, not only avoiding the sensibility of spectral clustering to scaling parameters, but also improving the robustness of classification method. The experimental results demonstrate that the proposed method improves the classification performance and region harmony, achieves stable results.(2) Aiming to solve the problems of high computation complexity and classification instability of spectral clustering algorithm, a dictionary learning and spectral clustering ensemble method is proposed for Pol SAR land cover classification. The proposed method improves the performance of traditional spectral clustering in the aspects of polarimetric feature extraction and sample selection of Nystr?m approximation algorithm. Texture features and polarimetric features are combined to improve the performance of spectral clustering algorithm. Laplacian eigenmap(LE) is applied to reduce the dimensionality of feature matrix, which maps the high-dimensional features to low-dimensional feature space to reduce the computation complexity of spectral clustering. Dictionary learning based sample selection method of Nystr?m approximation is applied to select typical sample points. Thus it can increase the information involved in sample points and avoid the instability classification results caused by random sampling. Compared with existing spectral clustering methods, the experimental results of Pol SAR image demonstrate that, the proposed method has advantages in robustness, region homogeneity and edge preservation.(3) In the traditional spectral clustering, k-means algorithm is used to find the cluster centers in a new low-dimensional feature space mapped by spectral clustering. However, k-means algorithm is ensitive to initialization and easily falls into local optimum. To address this issue, a novel immune clonal spectral clustering method(ICSC) in the joint-domain is proposed for Pol SAR image classification. The proposed method combines the complementary advantages of spectral clustering and immune clonal algorithm: 1) the dimensionality reduction of spectral clustering reduces the dimension of feature matrix, therefore improves the efficiency of immune clonal algorithm to find global optimization solution obviously; 2) immune clonal algorithm can obtain global optimum solution with a large probability and improve the classification results. The experiments results show the feasibility and efficiency of the proposed method.(4) Traditional affinity propagation(AP) clustering algorithm employs Euclidean distance as similarity measures, therefore cannot effectively describe the complex distribution of Pol SAR data. To address this issue, a novel AP clustering method based on combined manifold distance(CMD-AP) is proposed for Pol SAR image classification. The proposed method divides the Pol SAR image into superpixels. It not only takes the spatial relationship between pairwise pixels into account, but also reduces the computation complexity of AP algorithm. Meanwhile, Wishart-derived manifold in the spatial domain and Euclidean-derived manifold in the polarimetric domain are combined to form the manifold distance measures and construct the similarity matrix. Thus, it depicts the manifold structure of Pol SAR image in the spatial-polarimetric joint domain more detailedly. The experimental results show the feasibility of the proposed method.(5) Existing dictionary learning algorithms only consider the global sparsity of data, yet ignore the spatial structure of data. Moreover, their high computation complexities lead to the difficulty of dealing with large-scale image data. Taking the information of Pol SAR image in the spatial-polarimetric domain into account, a novel DL and sparse representation classification method is proposed for Pol SAR image classification. Firstly, the spatial-polarimetric manifold based fast AP clustering algorithm is employed to learn an over-complete dictionary. It reduces the time complexity of dictionary learning meanwhile keeping the manifold structure of Pol SAR data. Then locality-constrained linear coding method is adopted to extract the sparse coding features of Pol SAR image in the spatial and polarimetric domain respectively. Those features mentioned above are combined to form the spatial-polarimetric sparse coding features for better classification result. Experimental results demonstrate that the proposed method can improve the classification accuracy, and efficiently reduce the computation complexity of classification method.