SPARSE REPRESENTATION BASED LOSSY HYPERSPECTRAL DATA COMPRESSION

Sparse representation is capable of modeling signals as linear combination of a few atoms from a pre-trained dictionary. It allows learning an adaptive dictionary that leads to highly sparse nature in the representation of signals. In this paper, sparse representation is deployed in a lossy hyperspectral data compression framework. Dictionaries that exploit spectral correlation, as well as both spectral and spatial correlations are trained using online dictionary learning. A hyperspectral data is then represented using the learned dictionary via sparse coding. The resulting sparse coefficients are encoded to formulate the final bit stream. Experimental results on a number of hyperspectral datasets show that the proposed approach is indeed competitive to wavelet based methods, such as 3D-SPIHT, in terms of rate-distortion performance.