A Study Of Hyperspectral Image Classification And Unmixing Based On Sparse Representation

Hyperspectral image classification technology and spectral unmixing technology play extremely important roles in remote sensing area’s development, and have been a key research direction of scholars both in domestic and foreign. In recent years, with the development of sparse regression technique, some scholars have begun to apply it to the hyperspectral remote sensing image processing, especially for hyperspectral image classification and spectral unmixing. Among the previous studies, some typical papers which combined spatial information and sparse regression open up a few new directions in the hyperspectral image classification and spectral unmixing. These algorithms achieved better classification and unmixing results than traditional algorithms, but they still exist disadvantages, for example, these methods failed to make full integration between the spectral information of source image and the spatial information. In the parallel region of hyperspectral image, there exists several types of correlations between the neighboring pixels which may contain same materials, and the fractional vectors also have similarity. Based on the above characteristics, neighboring pixels of the center pixel could be used to add the spatial constraint to the classification and unmixing models. Because of the combination of the spatial information and spectral information, the new models can improve accuracy of classification and unmixing technology. Main contributions of this thesis are given as follows:1. In the parallel region of hyperspectral image. There exists several types of correlations between the neighboring pixels which may contain same materials, thus the neighboring pixels may be labeled to the same class. Therefore, this characteristic can be used to improve the performance of the classification techniques. In this thesis, a new sparse classification algorithm based on First-Order Neighborhood System Weighted constraint is proposed. In the proposed algorithm, the pixels in the first-order neighborhood system are used to constrain the center pixel and the smaller the difference between the center pixel and its neighbor in the first-order neighborhood system, the larger the proportion in the constraint. To test the classification performance of the new algorithm, common hyperspectral images obtained by AVIRIS and ROSIS sensors are used in the experiment. The experimental results show that the proposed algorithm has a smoother classification map and higher classification accuracy.2. In this thesis, we describe the classic sparse regression formulation firstly and then propose a new unmixing algorithm which involving in more complete spatial information on sparse unminxing formulation for hyperspectral image. The new algorithm integrates the spectral and spatial information in a Bayesian framework which is introduced to exploit the spatial-contextual information effectively. Moreover, the Markov Random Fields can model the spatial correlation efficiently, the new method employ a Markov Random Fields based on the Bayesian framework. Compared with other sparse regression based linear unmixing methods, our experimental results showed that the method proposed in this paper not only improves the characterization of mixed pixels but also obtains better accuracy in hyperspectral image unmixing.