A model-driven multi-feature fusion network for hyperspectral unmixing with spectral variability

Spectral variability and spatial-spectral characteristics are two important aspects that have been extensively investigated in hyperspectral unmixing. Although some recent methods attempt to consider both aspects simultaneously, spatial-spectral feature fusion and spectral variability modeling are often handled as two separate modules, which limits the physical consistency of variability characterization. To address this issue, this work conducts multi-feature collaborative modeling by jointly leveraging spectral features, spatial features, and variability-aware features, enabling unified characterization of spectral variability at both global and local scales. Building upon this formulation, we propose a novel Multi-Feature Linear Mixing Model (MFLMM) and develop a model-driven deep network, termed MFLMM-Net, to achieve interpretable estimation of the model parameters. Moreover, recognizing that spectral variability tends to exhibit local consistency within spatial-spectral neighborhoods, we impose smoothness constraints during feature extraction not only across adjacent spatial locations but also along neighboring spectral bands. Extensive experiments on multiple synthetic and real hyperspectral datasets demonstrate that the proposed MFLMM-Net achieves more favorable unmixing performance than several state-of-the-art methods.