A TRAINING BAND SELECTION ALGORITHM FOR SPARSE REPRESENTATION BASED HYPERSPECTRAL DATA COMPRESSION

摘要

Due to imaging with the same area, hyperspectral data has strong spectral correlation, and thus it’s reasonable to use a fixed dictionary to represent all bands sparsely. We have proposed a compression scheme for hyperspectral data using sparse representation in our previous work. For the dictionary learning stage in the scheme, different training band brings different compression performance. In this paper, we focus on training band selection, and an algorithm based on the variance of spectral correlation coefficient is proposed to select the optimal training band, and to make the learned dictionary be more effective for sparse representation. Experimental results reveal that the compression performance can be improved by our proposal and the compression scheme will be modified with our proposal.