DSCH-Net: Diffusion-State-Contextual Hybrid Network for Physics-Inspired and Direction-Aware Dehazing of Remote Sensing Imagery

Atmospheric haze reduces contrast and suppresses fine structures in remote sensing imagery, degrading the reliability of downstream mapping and detection tasks. Existing CNN and attention-based dehazing methods often struggle to capture long-range dependencies efficiently and may introduce color shifts or halo artifacts under non-uniform haze. DSCH-Net is proposed as a physics-inspired and direction aware state-space network for single-image dehazing. The framework combines a four-directional state-space operator for linear-time global context modeling with a PDE-driven diffusion module that injects explicit gradient–divergence priors to refine structural details. A residual multi-dilation unit recovers fine textures, while a selective fusion gate stabilizes encoder–decoder interactions to preserve spatial consistency. Extensive experiments on synthetic and real remote sensing benchmarks show that DSCH-Net achieves consistent state-of-the-art performance, reaching up to 32.20 dB PSNR and 0.980 SSIM on RSID and delivering strong results across RICE1/2, Haze1K, and DHID. Visual evaluations further demonstrate clearer sky gradients, sharper boundaries, and improved preservation of thin structures compared to recent CNN and Transformer models. DSCH-Net offers a compact, end-to-end, and resolution-scalable solution suitable for practical remote sensing workflows under adverse weather conditions. Code and data are available at https://github.com/CPEKMUTT/DSCH-Net.