A multi-domain dual-stream network for hyperspectral unmixing

Hyperspectral unmixing is of vital importance within the realm of hyperspectral analysis, which is aimed to decide the fractional proportion (abundances) of fundamental spectral signatures (endmembers) at a subpixel level. Unsupervised unmixing techniques that employ autoencoder (AE) network have gained significant attention for its exceptional feature extraction capabilities. However, traditional AE-based methods lean towards focusing excessively on the information of spectral dimension in the data, resulting in limited ability to extract endmembers with meaningful physical interpretations, and achieve uncompetitive performance. In this paper, we propose a novel multi-domain dual-stream network, called MdsNet, which enhances performance by incorporating high-rank spatial information to guide the unmixing process. This approach allows us to uncover pure endmember data that is hidden within the original hyperspectral image (HSI). We first apply superpixel segmentation and smoothing operations as preprocessing steps to transform the HSI into a coarse domain. Then, MdsNet efficiently handles multi-domain data and employs attention generated from the approximate domain to learn meaningful information about the endmembers’ physical characteristic. Experimental results and ablation studies conducted on Synthetic and real datasets (Samson, Japser, Urban) outperform state-of-the-art techniques by more than 10% in terms of root mean squared error and spectral angle distance, illustrating the effectiveness and superiority of our proposed method. The source code is available at https://github.com/qiwenjjin/JAG-MdsNet.