MDFPF-Net: Multidomain Feature Perceptual Fusion Network for Hyperspectral Unmixing

Abstract

Hyperspectral unmixing (HU) is crucial for retrieving subpixel information. However, existing deep learning (DL) methods typically rely on original spatial–spectral information and struggle to fuse the complementary advantages of multidomain features, resulting in endmember confusion and inaccurate abundance estimation. To address these challenges, a multidomain feature perceptual fusion network (MDFPF-Net) is proposed for HU. The network constructs a parallel feature extraction framework across the spatial, spectral, and frequency domains. Specifically, the spatial–spectral feature extraction (SSFE) module adopts a dual-branch design to capture distinct spatial and spectral attributes, while a frequency-domain multibranch attention (FMBA) module is introduced to selectively suppress redundant responses and enhance frequency components with stronger separability under the guidance of fine-grained texture information, yielding a high-fidelity representation of material distributions. To enhance cross-domain complementarity, a cross-domain dynamic perception (CDDP) module is designed. It uses adaptive context-aware convolutional kernels to achieve semantic alignment and collaborative optimization between both spatial–spectral and spatial–frequency features, strengthening endmember spectral discriminability and accurately characterizing abundance distribution discrepancies. Then, learnable parameters are used to adaptively integrate the dual-path outputs, producing deeply fused multidomain features. Finally, a semantic-enhanced abundance generator (SEAG) is embedded within a Transformer encoder–decoder architecture. Endmember-wise collaborative attention enhancement together with lightweight constraints is applied to further improve unmixing accuracy. Extensive experiments on synthetic and real-world datasets demonstrate that MDFPF-Net achieves superior performance, yielding lower root-mean-square error (RMSE) and spectral angle distance (SAD) than state-of-the-art methods in most cases. The source code is available at https://github.com/Nuist-HSI-Group/MDFPF-Net