First-Order Graph Trend Filtering for Sparse Hyperspectral Unmixing

In semisupervised unmixing, mixed pixels in a hyperspectral image (HIS) can be decomposed into corresponding abundances based on the known endmember library. The HSI contains important spatial information about the land cover, which can help enhance the performance of hyperspectral unmixing (HU). In this letter, we proposed first-order graph trend filtering (GTF) for sparse unmixing (SU) to explore and utilize spatial information more effectively and accurately. The proposed method adaptively constructs the first-order graph difference operator (GDO) from the original data and then uses double reweighted $\ell _{1}$ -norm regularization to promote the sparsity of the abundances. The results of experiments on simulated and real datasets show that the proposed algorithm can more accurately utilize the spatial structure and outperform competing methods.