A Wavelength-Dependent Haze Distribution Prior Network for Remote Sensing RGB Image Dehazing

Haze in remote sensing imagery often exhibits irregular spatial distributions and highly uneven concentrations due to complex atmospheric conditions and large observation scales. Moreover, haze exhibits different scattering and attenuation behaviors across RGB wavelengths. However, most existing dehazing methods are developed under the assumption of spatially uniform haze, which is frequently violated in real remote sensing scenarios. In this article, a wavelength-dependent haze distribution prior dehazing method is proposed, which embeds a haze distribution prior to the dehazing learning process. First, a haze distribution modeling network (HDMNet) is introduced to estimate the wavelength-dependent haze distribution map by jointly modeling atmospheric light and the wavelength-dependent transmission map, where the transmission map serves as an implicit representation of the haze concentration. Furthermore, a multibranch interaction network (MBINet) is designed to integrate the wavelength-dependent haze distribution map into the RGB image dehazing process, enabling effective improvement for haze removal under complex and nonuniform conditions. Finally, a collaborative dehazing loss is designed to guide HDMNet to focus on regions where haze severely degrades image quality, utilizing the residual differences between dehazed RGB images and haze-free RGB images. The proposed method is evaluated on multiple synthetic hazy datasets and a real-world hazy dataset, demonstrating great improvement over state-of-the-art methods in both quantitative metrics and visual quality.