FS-FlowNet: Frequency–Spatial Dual-Domain Residual Flow Network for Remote Sensing Dehazing

Atmospheric haze and cloud cover severely degrade remote sensing image quality, adversely affecting critical tasks like object recognition and change detection. Existing dehazing methods face significant challenges with such images: traditional physical models often cause over-dehazing and color distortion, while single-stage deep learning models generalize poorly due to fixed regression paths. To address these limitations, this paper proposes FS-FlowNet, a dual-domain frequency-spatial residual flow network for remote sensing image dehazing. We innovatively combine frequency-spatial domain processing with multi-stage ordinary differential equation optimization within an efficient end-to-end architecture. First, during coarse dehazing, an enhanced U-Net encoder–decoder conducts multi-layer feature extraction with skip connections to capture haze-related features while preserving spatial structures. Next, a dual-domain module extracts amplitude and phase spectra via FFT, using a dual-branch (frequency-spatial) strategy to improve high-frequency detail. We employ a flow matching algorithm to cast dehazing as a multi-stage flow-field optimization problem; Here, a fourth-order Runge–Kutta (RK4) solver is used to activate/solve the underlying ordinary differential equations, yielding smooth and physically plausible flow transformations. Extensive experiments on the real world benchmark and multiple synthetic datasets demonstrate superior performance in both real and synthetic scenarios.