A Dual-Attention Hyperspectral Unmixing Network Based on CNN and Swin Transformer

Hyperspectral unmixing aims to decompose mixed pixels into pure substances (i.e., endmembers) and their corresponding proportions (i.e., abundances). To achieve high-precision interpretation of mixed pixels, this paper combines the respective advantages of convolutional neural networks (CNN) and Transformer, and proposes a hyperspectral dual-attention unmixing network based on CNN and Swin Transformer. The network adopts a dual-branch structure of CNN and Swin Transformer, used to extract local detail features and global contextual information, respectively. On this basis, parallel spatial attention and channel attention modules are introduced to further explore discriminative features. Finally, a decoder is used to linearly reconstruct the extracted features, achieving precise unmixing of mixed pixels. Experimental results on two real hyperspectral datasets demonstrate the effectiveness and superiority of the proposed method.