Hyperspectral sparse unmixing based on a novel adaptive total variation regularization

Abstract

Hyperspectral unmixing realizes the unmixing by determining the pure substances (endmembers) and their proportion (abundances) in the mixing pixels. Most of the previous sparse unmixing methods are effective in the case of high signal-to-noise ratio. However, these methods do not make full use of the spatial and spectral information of hyperspectral data, and their unmixing effect is not good enough in the case of high noise level. To solve this problem, in this paper, two new models are proposed. One is a hyperspectral sparse unmixing model using an adaptive total variation regularization (SU-ATV). Considering that different regions have different spatial structures, the adaptive total variation (ATV) regularization adaptively adjusts the horizontal difference and vertical difference of TV, which can better optimize the efficiency of TV to enhance the anti-noise performance. The other is an extended model (called WSU-ATV), where a spectral weight term is added to enhance the sparsity of abundance. To solve these two models, an alternating direction method of multipliers (ADMM) is presented. Experimental results on both synthetic data and real data show that the proposed models have better accuracy and anti-noise performance, compared with other related sparse unmixing methods. Besides, the WSU-ATV performs better than the SU-ATV owing to the spectral weight.