Hyperspectral Unmixing Network Based on Superpixel Token Double Attention Mechanism

Hyperspectral unmixing techniques based on deep learning have gained increasing attention and achieved significant progress. However, the unmixing methods relying solely on convolutional neural networks or transformer networks are insufficient to effectively capture global information, thereby affecting the accuracy of unmixing. To better leverage the global information contained in hyperspectral images, we propose a novel hyperspectral unmixing network based on a Superpixel Token Double Attention (STDA) mechanism in this article. The proposed method integrates the strengths of superpixel segmentation and attention mechanisms to capture both local and global information more effectively. By segmenting hyperspectral images into superpixels, which group similar pixels spectrally, the local coherence of the hyperspectral images is preserved and the boundaries are delineated more precisely. To further enhance the unmixing performance, two complementary attention modules are introduced: one is a cross-superpixel attention mechanism that propagates information across superpixel boundaries, enabling long-range dependency capture. The other is an intra-superpixel attention mechanism that focuses on local information exchange within each superpixel. This combination ensures that the proposed network model can leverage both global context and local details, addressing the limitations of previous unmixing methods that struggle with either long-range dependencies or local feature interactions. Experimental results demonstrate significant improvements in abundance estimation and endmember extraction compared to several state-of-the-art methods and validate the effectiveness of the proposed STDA method in hyperspectral unmixing tasks.