Spectral super-resolution based on matrix factorization and spectral dictionary

Abstract

Spectral information in hyperspectral imagery (HSI) directly acquired by sensors, commonly with surplus bands and redundant information, takes high memory and transmission costs, resulting in reduced spatial resolution and aggravated spectral mixture. Therefore, the desired high spectral resolution HSI can be obtained via spectral super-resolution after acquiring original HSI with lower spectral resolution but relatively higher spatial resolution. In this paper, we proposed a spectral super-resolution method based on spectral matrix factorization and dictionary learning. High and low spectral resolution HSIs are assumed to have the same spatial resolution and share the same spectral signatures. So abundances of low spectral resolution imagery can provide high spatial information, while its endmembers can supply accurate spectral characteristics. Then several high spectral resolution HSIs in 2-D forms are utilized to train a spectral dictionary which contains both high spatial resolution information and high spectral resolution information. Finally, the desired spectral enhancement results are achieved through the use of spatial fidelity constraint. Experiments on Sandigo dataset indicated the superiority of our proposed method.