Spatial-Spectral Attention Sparse Unmixing Network Based on Algorithm Unrolling

Abstract

Hyperspectral unmixing technology is increasingly developing in the direction of artificial intelligence. Researchers are working on reducing the complexity of the unmixing network while improving the unmixing performance. In this paper, combined with the iterative formulation of the SUnSAL algorithm, an efficient unmixing network with interpretability is proposed. This method maps the SUnSAL algorithm to the convolutional neural network, and uses an attention mechanism for hyperspectral data, allowing the network to pay more attention to the calculation of important data. Specifically, we connect three variable update layers in series to realize iterative calculation, and the network scale is small and has strong interpretability. The experimental results show that the unmixing ability of the network is significantly better than the traditional sparse unmixing algorithm in three different of signal-to-noise ratios.