FSDNet: Joint Frequency-Spatial Modulation Network for Remote Sensing Image Dehazing
Xiaohui Dong, Yuanming Lu, Haoming Su, Chengchao Wang · IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing · 2026
Remote sensing image dehazing is a critical preprocessing step for numerous downstream Earth observation tasks, including object detection, postdisaster damage assessment, and urban expansion monitoring. However, this task remains inherently challenging due to haze-induced atmospheric scattering, which severely degrades key visual attributes—particularly global contrast, edge definition, and high-frequency structural details. Such degradation directly impairs the reliability and accuracy of subsequent analytical pipelines. While conventional spatial-domain convolutional neural networks (CNNs) have achieved notable performance, they often suffer from excessive smoothing of fine-scale edges and textures. More recent Transformer- and diffusion-based approaches, though promising, incur prohibitive computational overhead and memory demands—limiting their practical deployment in resource-constrained remote sensing workflows. To bridge this gap, we propose FSDNet, a lightweight joint frequency-spatial modulation network that explicitly decouples and jointly optimizes spatial and frequency representations. Central to our design is the frequency-spatial dual-branch block (FSDBlock): the spatial branch leverages a multiscale pyramidal perception module—integrating depthwise and strip convolutions in parallel—to efficiently capture both global context and local structural cues across heterogeneous receptive fields; the frequency branch (FGB) operates in the Fourier domain, employing an adaptive amplitude mask guided by FFT to selectively attenuate haze-corrupted low-frequency components. Furthermore, we introduce a dedicated frequency reconstruction loss ( $L_{\text{FFT}}$ ) that enforces fidelity between the predicted and GT amplitude spectra, thereby preserving essential frequency-domain structures. Comprehensive evaluations on multiple benchmark remote sensing dehazing datasets consistently demonstrate that FSDNet achieves strong restoration performance while maintaining low computational cost, memory usage, and inference latency.