| In the last three decades, hyperspectral remote sensing has already played an important role in the various military and civilian fields. However, due to its high complexity, how to effectively extract material information from large amount of data and achieve satisfactory classification results is still a key problem for hyperspectral remote sensing. In recent years, sparse representation has attracted much attention in the fields of image processing and pattern recognition due to its efficiency and simplicity. In particular, the theory of sparse representation has been applied to remote sensing images, especially for hyperspectral imagery classification and good results have been achieved. In this paper, based on the Sparse Representation-based Classification(SRC) framework, spatial contextual information has been integrated. Furthermore, a Fisher criterion-based Gabor Cube selection algorithm has been proposed. Through a multi-task sparse representation strategy, the classification performance has been greatly improved. The main contributions of this thesis lie in:Firstly, a new spectral-spatial-combined SRC method has been presented to jointly consider the spectral and spatial neighborhood information of each pixel to explore the spectral and spatial coherence by the SRC method. As the test samples are processed one by one when calculating the sparse representation coefficients,this kind of operation has omitted the spatial information among samples. Hence, in our method, training samples were used to encode each test sample to get the spectral-based sparse representation, while the neighborhood pixels of the test sample were used to get the spatial-based sparse representation. Then both the spectral and spatial information were combined together to identify the sample. After all the test samples have been computed by the above step, the classification map can be obtained. Compared with the map obtained by the standard SRC method, the spatial information has been incorporated. Obviously, in order to acquire a more stable and consistent map, the spectral-spatial procedure can be repeated several times and the classification accuracy can be gradually improved. Furthermore, a fast interference-cancellation operation has been applied on the proposed method to improve the time efficiency. Compared with the other algorithms, such as the K-nearest neighbor(KNN), support vector machine(SVM), SRC, the experimental results have shown that the proposed method can achieve better classification performance for small training sample sets.Secondly, the Fisher discrimination criterion has been used to select Gabor cube features. The method is to compute the ratio of the between-class and within-class distance, which is then used to choose the representative Gabor cubes. The experimental results have demonstrated that the introduced method can select those representative features, as well as remove redundance and interference, and has a positive impact on the classification.Finally, we proposed a method which combines the Fisher feature selection and Multi-task Joint Sparse Representation Classification(MTJSRC) algorithm. After several representative Gabor feature cubes using Fisher discrimination criterion for each class have been picked out, the multi-task framework has been used to jointly represent each test sample. Compared our method with KNN, SVM, MTJSRC, experimental results have shown the feasibility and efficiency of the method on high-spatial-resolution hyperspectral image data sets. Specifically, the proposed approach can not only achieve higher classification accuracies, but also reduce the computation complexityIn the end, we have summarized the advantages and disadvantages of the proposed algorithms, and pointed out the direction of the further work。... |
PDF Full Text Request