DCANet: dual-stream coupled attention network for hyperspectral unmixing

Hyperspectral unmixing (HU) aims to decompose mixed pixels into a set of endmember spectra and their corresponding abundance fractions, is a fundamental problem for the refined analysis of hyperspectral imagery. However, the complex spectral variability and nonlinear mixing effects prevalent in real-world scenarios impose limitations on the accuracy and interpretability of traditional linear models. While deep learning-based methods have achieved significant progress in recent years, they still face challenges such as high computational overhead, insufficient utilization of pure-pixel priors, and weak physical interpretability. To address these issues, we propose a Dual-stream Coupled Attention Network (DCANet). This network facilitates feature interaction through two complementary branches: a mixed-pixel stream and a pure-pixel stream. It also incorporates a novel cross-attention mechanism to dynamically inject pure-pixel priors into the abundance learning process. Furthermore, to overcome the computational bottleneck associated with large-scale hyperspectral images, we design a lightweight Spatial Reduction (SR) strategy that reduces the attention complexity from O(N2) to O(NZ). On the Samson dataset, SR decreases the overall model FLOPs from 1.44G to 1.11G (a 23.2% reduction), while the attention module itself is theoretically reduced by 99.78%. Experimental results demonstrate that DCANet surpasses existing methods in abundance estimation accuracy, endmember discrimination, and computational efficiency.