Frequency-Decoupled and Collaborative Dehazing Network for Remote Sensing Images

Abstract

Remote sensing image (RSI) dehazing is a crucial preprocessing step for Earth observation, which aims to restore the covered textures from images degraded by nonhomogeneous distributed haze. Existing Transformer-based methods have powerful global context modeling and static feed-forward networks (FFNs). However, they struggle to explicitly preserve high-frequency textures and lack the adaptability to varied haze distributions and scene contents. To address these challenges, we propose a frequency-decoupled and collaborative dehazing network (FDCDNet), optimized by a dual-domain contrastive loss. Built upon a U-shaped framework, FDCDNet employs stacked frequency-decoupled and collaborative blocks (FDCBs) to progressively distill and refine haze-free representations. The FDCB comprises a frequency-decoupled and enhancement module (FDEM) and a frequency collaborative FFN (FCFFN). The FDEM explicitly decouples features into high- and low-frequency components, then conducts detail enhancement through a local convolution mixer (LCM), and performs holistic haze removal via a cross-domain attention (CDA) mechanism. The FCFFN incorporates a heterogeneous group of experts and enables spatial, channel, and frequency fusion via a dynamic gating mechanism. Furthermore, a dual-domain contrastive loss is leveraged to reinforce feature discriminability by leveraging both spatial- and frequency-domain consistencies, guiding the network to generate more natural and structurally coherent results. Experiments on challenging remote sensing datasets demonstrate that FDCDNet achieves state-of-the-art performance in both quantitative metrics and visual quality. In particular, FDCDNet achieves a 1.75-dB peak signal-to-noise ratio (PSNR) improvement on the StateHaze1k-thick set compared to MB-TaylorFormerV2.