WF-Mamba: Weighted Fourier State-Space Models With Adaptive Spatial-Frequency Synergy for Hyperspectral Unmixing

Abstract

Hyperspectral unmixing (HU) is a crucial technique that aims to estimate subpixel-level ground object information in remote sensing images. Existing deep learning-based unmixing methods, especially autoencoders (AEs), typically operate only within the original spatial–spectral feature domain. However, due to the impact of noise and spectral variability, relying solely on the original domain makes it difficult to fully extract effective features and leaves the models highly susceptible to interference. To address these issues, we present an innovative dual-stream spatial-frequency state-space model (SSM), termed WF-Mamba, for enhanced HU. A physics-driven snake-like scanning (SS) mechanism is introduced in WF-Mamba to remodel the data serialization process, which strictly preserves the intrinsic neighborhood contiguity and 2-D topological priors. To extend the feature domain and overcome the frequency-domain blindness of pure spatial models, a macroscopic weighted Fourier neural operator (WFNO) branch is constructed. By dynamically recalibrating global low-frequency modes in the complex domain, WFNO functions as an advanced physical filter to actively suppress high-frequency noise. Furthermore, an adaptive synergistic fusion (ASF) mechanism is proposed for WF-Mamba to break the bottleneck of rigid feature concatenation. More robust features can be obtained by ASF by autonomously balancing the sharp local spatial boundaries extracted by Mamba with the globally purified smoothness from WFNO via dynamic residual gating. The fused robust features are then mapped to estimate accurate abundances and reconstruct the image. Extensive experiments on one synthetic and three real hyperspectral datasets demonstrate that the proposed WF-Mamba achieves competitive performance compared to advanced methods.