Gradient-Aware Physics-Guided Network for Remote Sensing Image Dehazing

Remote sensing image dehazing is critical for restoring spectral fidelity in Earth observation tasks. While vision Transformers (ViTs) excel at global context modeling, preserving high-frequency structural details essential for sub-scale object interpretation remains a challenge. To address this, we propose a gradient-aware physics-guided network (GPNet) that integrates global context understanding with fine-grained structure recovery. Specifically, we design a gradient-aware Transformer Block (GTB). Unlike standard designs, we introduce a gradient-aware feed-forward network (GFN) to substitute the conventional feed-forward layer. By incorporating difference convolutions, this module explicitly models anisotropic gradient priors, effectively mitigating structural smoothing.To adaptively handle spatially variant degradation, a haze query module (HQM) is embedded at the bottleneck, leveraging learnable prototypes to identify and decouple latent haze patterns from deep content features. Furthermore, a physics-guided enhancement module (PEM) is hierarchically embedded to enforce physical consistency, which modulates the restoration process based on the atmospheric scattering model to prevent color distortion. Extensive experiments on the StateHaze1K dataset demonstrate that GPNet achieves state-of-the-art performance, surpassing existing methods with a 23.77 dB PSNR and improved visual structural fidelity.