A Spectral-Spatial Attention Autoencoder Network for Hyperspectral Unmixing

Convolutional neural-network-based autoencoders, which can integrate the spatial correlation between pixels well, have been broadly used for hyperspectral unmixing. Nevertheless, these methods are hindered in their performance by the fact that they treat all spectral bands and spatial information equally in the unmixing process. In this article, we propose a spectral-spatial attention autoencoder network to solve the mixing pixel problem of the hyperspectral image. First, we design a spectral-spatial attention module, which refines spectral-spatial features by sequentially superimposing the spectral attention module and spatial attention module. The spectral attention module is built to select useful spectral bands, and the spatial attention module is designed to filter spatial information. Second, we exploit the geometric properties of endmembers while considering abundance sparsity. We introduce minimum volume and sparsity regularization terms into the loss function. Extensive experiments indicated that our method achieved competitive unmixing results compared with other advanced unmixing approaches.