Band Mask Network with Spatial-Spectral Fusion for Hyperspectral Unmixing

Hyperspectral technology combines imaging and spectral detection, which can realize the fine monitoring of ground samples. However, the mixed pixel phenomenon restricts the further development of hyperspectral technology. Hyperspectral unmixing (HU) is currently the most effective method to deal with mixed pixels and has become a research hotspot in the hyperspectral field. In recent years, deep learning (DL) has been widely used in the field of HU due to its strong feature learning ability. However, the existing DL networks usually input all bands directly for simple identity mapping, resulting in inadequate use of spectral information. In addition, the use of spatial information is limited to the original scale, and the ability to obtain multi-scale features is limited. In this regard, we propose a band mask network (BM-Net) with spatial-spectral fusion for HU. For spectral information, a novel random band mask (RBM) network is constructed to characterize the key features of the original signal through nonidentity mapping learning. For spatial information, a multi-receptive field connection module is designed to coordinate multi-view information from different scales. Comparison experiments on one simulated and two real data sets show that the proposed BM-Net has optimal accuracy in both endmember extraction and abundance inversion.