Blind Hyperspectral Unmixing Based on Graph Total Variation Regularization

Remote sensing data from hyperspectral cameras suffer limited spatial resolution, in which a single pixel of image may contain information several materials the field view. Blind unmixing is process identifying pure spectra individual (i.e., endmembers) and their proportions abundances) at each pixel. In this article, we propose novel blind model based on graph total variation (gTV) regularization, can be solved efficiently by alternating direction method multipliers (ADMM). To further alleviate computational cost, apply Nyström to approximate fully connected small subset sampled points. Furthermore, adopt Merriman-Bence-Osher (MBO) scheme solve gTV-involved subproblem ADMM decomposing gray-scale into bitwise form. A variety numerical experiments synthetic real images are conducted, showcasing potential proposed terms identification accuracy efficiency.