Research On Classification Of Hyper-spectral Image

Hyperspectral image contains abundant spectral information and spatial information. By analyzing these informations, people could identify object categories on the hyperspectral image. Hyperspectral image classification is a very pivotal part in the practical applications. Nowadays, hyperspectral image classification is widely used in agriculture, military, marine management, geological exploration and so on. Hyperspectral image classification technology has become an important part of modern science and technology. In this thesis, the hyperspectral image classification information is extracted from two aspects of spectral information and spatial information. This thesis proposes two algorithms: a sparse representation based on normalized Euclidean distance and an adaptive neighborhood Markov random field based on edge detection. The main contents of this thesis indicated as follows.1. This thesis proposed an improved method called sparse representation based on normalized Euclidean distance for hyperspectral image classification. The normalized Euclidean distance is introduced into the traditional sparse representation algorithm to overcome the problem of traditional sparse representation does not take into account the category information. The reconstruction error is recalculated by the normalized Euclidean distance in the expectation-maximization algorithm model. Moreover, the diagonal variance matrix in the normalized Euclidean distance is adaptively updated by the expectation-maximization algorithm after each classification. Simulation results show that the performance of the proposed method is better than others.2. An adaptive neighborhood Markov random field(MRF) based on edge detection in hyperspectral image classification is proposed in this thesis. This algorithm used an adaptive Markov random field to improve the hyperspectral image classification accuracy. In this algorithm, the principal component analysis(PCA) is used to reduce the dimensionality of the hyperspectral image. Canny operator is adopted to detect the edge of the grayscale image which is produced by normalizing the first dimension data of PCA. A strategy of four directions of freedom is used to determine the size of the neighborhood with the test pixel as the center. Finally, the MRF is used to determine the label of the central pixel, and the final classification result is obtained from this algorithm. Experimental results show that the proposed method can effectively improve the classification accuracy of hyperspectral images.