Dual-stream feature fusion network with transformer for hyperspectral unmixing

Recent advancements in hyperspectral unmixing based on deep learning have been significant. However, methods relying solely on convolutional neural networks or Transformers often fail to adequately capture both local and global information from the data, adversely affecting unmixing accuracy. In this paper, we propose a novel hyperspectral unmixing network, termed DFFT-HU, aimed at enhancing the accuracy of hyperspectral unmixing. The network architecture features an encoder-decoder structure, comprising one branch that employs a densely connected network and another that utilizes a multi-scale convolutional network. The features from both branches are fused to enable the network to effectively capture global contextual information via the Transformer's multi-head self-attention mechanism. Experimental results from three datasets—Samson, Apex, and Jasper Ridge—demonstrate that our method significantly outperforms existing techniques in terms of end-member abundance estimation and spectral reconstruction, as evidenced by improvements in average RMSE and SAD metrics.