Sparse representation based hyperspectral image compression and classification

Abstract

Abstract This thesis presents a research work on applying sparse representation to lossy hyperspectral image compression and hyperspectral image classification. The proposed lossy hyperspectral image compression framework introduces two types of dictionaries distinguished by the terms sparse representation spectral dictionary (SRSD) and multi-scale spectral dictionary (MSSD), respectively. The former is learnt in the spectral domain to exploit the spectral correlations, and the latter in wavelet multi-scale spectral domain to exploit both spatial and spectral correlations in hyperspectral images. To alleviate the computational demand of dictionary learning, either a base dictionary trained offline or an update of the base dictionary is employed in the compression framework. The proposed compression method is evaluated in terms of different objective metrics, and compared to selected state-of-the-art hyperspectral image compression schemes, including JPEG 2000. The numerical results demonstrate the effectiveness and competitiveness of both SRSD and MSSD approaches. For the proposed hyperspectral image classification method, we utilize the sparse coefficients for training support vector machine (SVM) and k-nearest neighbour (kNN) classifiers. In particular, the discriminative character of the sparse coefficients is enhanced by incorporating contextual information using local mean filters. The classification performance is evaluated and compared to a number of similar or representative methods. The results show that our approach could outperform other approaches based on SVM or sparse representation. This thesis makes the following contributions. It provides a relatively thorough investigation of applying sparse representation to lossy hyperspectral image compression. Specifically, it reveals the effectiveness of sparse representation for the exploitation of spectral correlations in hyperspectral images. In addition, we have shown that the discriminative character of sparse coefficients can lead to superior performance in hyperspectral image classification.