A Multi-Attention Autoencoder for Hyperspectral Unmixing Based on the Extended Linear Mixing Model

Hyperspectral unmixing, which decomposes mixed pixels into the endmembers and corresponding abundances, is an important image process for further application of hyperspectral images (HSIs). Lately, unmixing problem has been solved using deep learning techniques, particularly autoencoders (AEs). However, majority them are based on simple linear mixing model (LMM), disregards spectral variability in different pixels. In this article, we present a multi-attention AE network (MAAENet) extended LMM to address issue real scenes. Moreover, networks ignore global spatial information HSIs operate pixel- or patch-wise. We employ attention mechanisms design spatial–spectral (SSA) module that can deal with band redundancy extract features through correlation. noticing always intersection materials, novel sparse constraint homogeneity designed constrain abundance abstract local features. Ablation experiments conducted verify effectiveness proposed structure, SSA module, constraint. The method compared several state-of-the-art methods exhibits competitiveness both synthetic datasets.