Study On Methods Of Hyperspectral Imagery Processing Based On Sparse Representation

A hyperspectral sensor data deluge is beginning to swamp today’s sensing systems, especially with a rapidly growing spectral and spatial resolution for hyperspectral sensing technology. As a result the data processing ability is far behind the data acquisition ability, which leads two aspects of challenges. One is the difficulties to store and transmit massive data. Another one is how to reveal the useful information from the massive data. Recent years, sparse representation(SR) theory is one of the hot topics in the field of signal processing. SR has been widely applied in signal and image processing, computer visual and pattern recognition areas. There’s an inspiring trend to apply the SR theory to analyze high-dimensional hyperspectral image in remote sensing area.This thesis studies the hyperspectral data acquisition and information processing problems, which includes spectral imaging, hyperspectral image classification and target detection based on sparse representation.Firstly, we studies spectral imaging based on SR considering hyperspectral data acquisition. Compressive coded aperture scheme is a typical spectral imaging solution based on SR. However this scheme will encounter non-negative constraints of the compressed measurement matrix in the physical implementation. Therefore a bipolar compression solution of hyperspectral imaging is presented in this thesis for the disadvantage of traditional scheme. The proposed scheme employs two-channel observation structure to achieve bipolar observation of compression observation matrix by calculating a difference of non-negative observation between two channels. Compared with traditional scheme, the proposed scheme solves the inconsistency between the compression imaging theory and actual physical constraints. Furthermore the scheme matches the theory characteristics of compressed observation matrix, which implies that the proposed scheme can maintain the structure and information of the original signal better and get better reconstruction result. Simulated experiments then demonstrate the effectiveness of the scheme presented in this thesis.Then for the problem of hyperspectral data information extraction, this thesis studies the hyperspectral image classification problem based on sparse representation. This thesis firstly studies the sparse representation and classification decision for hyperspectral classification problem. Then the ability of constructed dictionary in SR model is furture studied. In conventional models, the dictionary consists of training samples from each class, which cannot capture the relative difference of similar classes. This thesis proposes a fisher discriminative dictionary learning(FDDL) based hyperspectral classification method. The proposed method enlarges the interval between similar classes and simultaneously constructs a new subdictionary from original subdictionary in dictionary learning process. Moreover the sparseness property of the dictionary learning process is considered which leads the learned dictionary performing better discriminative and reconstructive ability. Extensive real data experiments show that the proposed method preforms better.Finally, we further study the hyperspectral imagery target detection problem based on sparse representation. Considering that target prior information cannot be obtained in real target detection application, especially in military field. This thesis proposes both in spectral and spatial domain the local sparseness difference based hyperspectral target detection, which is based on the research of spatial-spectral joint sparsity. Sliding dual window used in proposed method effectively captures the nonnegative sparse characteristic of target on the background spectral dictionary. Then the nonnegative sparse characteristics of spatially local neighborhoods are cumulated to detect candidate targets. Extensive experiments on both simulated and real data demonstrate the efficiency of the proposed algorithm.