GGCT-Net: A Dual-Branch Gated Graph Convolution and Grouped Cross-Attention Transformer Network for Hyperspectral Unmixing

In recent years, deep learning-based hyperspectral unmixing techniques have garnered increasing attention and achieved significant progress. However, existing deep learning methods exhibit notable limitations: CNNs struggle to correlate spectrally similar yet spatially dispersed pixels, while Transformers lack the ability to distinguish between spectral and spatial features when processing high-dimensional data. To address these, this article proposes a dual-branch gated graph convolution and grouped cross-attention Transformer network (GGCT-Net). Specifically, this network designs a Spatial-Spectral Cross-Attention Block (SCAB) within the Grouped Cross-Attention Transformer encoder (GCA-Transformer) to process spectral and spatial features separately. Within the gated graph convolution encoder (G-GCN), a graph structure based on spectral similarity is constructed to correlate spectrally similar yet spatially dispersed pixels, and deep feature interactions are achieved through a gated cross-attention fusion module (GCAF). Experimental results on both simulate and real hyperspectral datasets demonstrate the proposed method’s effectiveness and superiority compared to various hyperspectral unmixing approaches.