An RGB difference prior for aerial remote sensing image dehazing with a DCP enhancement learning network

Hazy conditions significantly impair the utility of aerial remote sensing (ARS) images. Although existing dehazing methods show promising results, their effectiveness is often limited by unclear assumptions about the degradation principles. In this paper, we propose a novel dehazing prior, termed the RGB Difference Prior (RGB-DP), which is based on image channel differences and complements the dark channel prior (DCP). The RGB-DP demonstrates broader applicability across various hazy scenarios, including clear sky conditions. In addition, we design a DCP Enhancement Learning Network (DEL-Net) with two distinct branches: an enhancement (ENC) branch and a prior branch. The ENC branch is designed to mathematically reinforce the prior branch, which aids in preserving details. The two branches are integrated to recover haze-free images through a supervised training process. To further improve performance, we propose a new RGB Difference Loss (RGB-DL), which reduces errors arising from inaccurate DCP assumptions. Experimental results on the proposed ARS Road Extraction Hazy Dataset (REHD) and the SateHaze1k dataset demonstrate that our method achieves the highest average PSNR of 27.31 dB and SSIM of 0.888. Additionally, real-world image dehazing tests highlight the superior generalizability of our method.