Cross-layer texture injection for remote sensing image dehazing oriented to downstream object detection
Zihao Meng, Wenwu Zheng, Yifei Sun · International Journal of Remote Sensing · 2026
Haze degrades the radiometric fidelity and interpretability of optical remote sensing images, especially high-resolution aerial scenes with small, dense targets. We propose CLTI-Net, a cross-layer texture injection network for dehazing and downstream object detection. Its Cross-Layer Texture Injection Module extracts high-frequency residual structures from encoder features and injects them into decoder features to mitigate texture loss during downsampling. A lightweight coordinate-attention mechanism models directional spatial dependencies under non-uniform haze. We construct DOTA-Haze using a physically motivated scattering model while retaining object annotations, and evaluate downstream performance with a frozen YOLOv8 detector. Experiments on real and synthetic remote sensing datasets show competitive restoration quality and improved detection under thin and thick haze. These results indicate that explicit preservation of high-frequency structures benefits both dehazing quality and task-relevant remote sensing interpretation.