Hyperspectral Image Classification With Spectral-Spatial Information Mining And Sparse Representation Learning

Remote sensing technology has become a significant means of human access to information, of which hyperspectral remote sensing plays an essentially important role due to its characteristics and broad applications. Classification is an important task of hyperspectral image understanding and interpretation, and also one of the foundations of practical applications in many fields, such as environmental monitoring, precision agriculture and military reconnaissance. With the development of remote sensing acquisition technology, spatial resolution and spectral resolution of hyperspectral image improve a lot, helping gain more information than ever, but also bringing new challenges: how to mine more discrimination information, especially spectral and spatial information; how to solve the changes of intraclass and interclass difference; how to fuse the information of multiple features, etc. To solve these problems, some new methods based on characteristic of hyperspectral image and sparse representation learning are proposed in this dissertation, and the main research work can be summarized as follows:1. In order to improve the classification performance for hyperspectral iamge, a novel sparse representation method based on spectral information divergence has been presented. The proposed method exploits the spatial correlation across neighboring pixels, data spectral characteristic and the sparsity of the sparse representation simultaneously. Thus, it describes spectral variability, similarity, and discrimination of HSI more effectively Experimental results have demonstrated that the proposed method yields more accurate classification maps. Therefore, it provides a new way to improve the classification methods based on sparse representation.2. Hyperspectral images usually have complex content and chaotic background, while the combination of multiple kinds of features would be helpful for the classification task. Thus, we presented a joint sparse representation classification method with multifeature combination for hyperspectral imagery. Once getting several complementary features(spectral, shape and texture), the proposed model simultaneously acquires a representation vector for each kind of feature and imposes the joint sparsity row,0l-norm regularization on the representation coefficients. The regularization can enforce the coefficients to share a common sparsity pattern, whichpreserves the cross-feature information. To further improve the classification performance, we incorporate contextual neighborhood information of the image into each kind of feature. Compared with state-of-the-art algorithms, it has been proved that the proposed algorithm with much less memory requirements performs faster than those on real hyperspectral images, while provides the same(or even better) accuracy.3. Although multifeature joint sparse representation model povided a new way to fuse multiple feature informance, different features have unequal contributions to the final decision. In order to get better classification results, a new weighted multifeature sparse representation methods has been presented to take the contributions of different feature representation tasks to the final decision into consideration. Furthermore, kernel joint sparse representation model is presented to handle nonlinearity in the data. Kernel model projects the data into a high-dimensional space to improve the separability, achieving a better performance than the linear version. At the same time, we incorporate contextual neighborhood knowledge into the learned models. Experiments on several real hyperspectral images indicate that the proposed algorithms with much less memory requirements perform significantly faster than state-of-the-art algorithms, while exhibit highly competitive classification accuracy.4. In order to make good use of complementary information from different features, a novel joint sparse representation classification method with class-level sparse constraint has been presented for hyperspectral image classificaion. The proposed model enforces pixels in a small region of each type features to share the same sparsity pattern, at the same time, the pixels described by different features have freedom to adaptively choose their own appropriate atoms, but still belong to the same class. Thus, the proposed model not only preserves the spatial information by joint sparse constraint but also utilizes additional complementary information from different features by class-level sparse constraint. Furthermore, we also kernelize the model to handle nonlinearity in the data. And a new version of simultaneous orthogonal matching pursuit is proposed to solve the aforementioned problems. Experiments on several real hyperspectral images indicate that the proposed algorithms provide a competitive performance when compared with several state-of-the-art algorithms.5. Ensemble learning shows significant potential benefits to the classification ofhyperspectral image. However, the ensemble strategy remarkably influences the classification results, which include determining the minimum number of classifiers and assigning advisable weights associated with each classifier. In order to solve this problem, a novel sparse ensemble learning method with spectral-spatial knowledge has been presented for hyperspectral image classification. It considers the ensemble strategy under sparse recovery framework, where the solved non-zero coefficients reveal the importance of the selected classifier, from which a compact and effective ensemble learning system can be derived. Moreover, the spatial information is incorporated into the classification to develop a spectral-spatial joint sparse representation based ensemble learning algorithm for more accurate classification of hyperspectral images. Experimental results on several real hyperspectral images show that the proposed sparse ensemble system can achieve better performance than traditional ensemble learning methods using all classifiers, and it largely improves the efficiency in testing phase.