Sparse Representation-Based Hyperspectral Data Processing: Lossy Compression

This paper presents a method for lossy hyperspecral data compression based on sparse representation. The idea is to learn a dictionary that induces sparsity in the coefficient vectors that represent new input signals. The energy compaction feature of such sparse coefficient vectors is then evaluated in a lossy hyperspectral data compression framework. Experimental results on a number of hyperspectral data show that this approach is effective in hyperspectral data compression, and comparable to some of the state-of-the-arts data compression schemes, such as JPEG2000 with multiple component transformations and three-dimensional-set partitioning in hierarchical trees. Specifically, using the proposed framework, dictionaries that exploit spectral correlation, or spectral and spatial correlations, are trained using online dictionary learning. A hyperspectral data is represented using the learned dictionary via sparse coding. The resulting sparse coefficients are then encoded to formulate the final bit stream. The proposed framework allows using a base dictionary trained offline, or incorporating an update to the base dictionary, to achieve more adaptivity.