Remote Sensing Image Dehazing: A Systematic Review of Progress, Challenges, and Prospects

A comprehensive new survey from researchers at multiple institutions provides the first unified analysis of AI techniques for clearing haze, fog, and thin clouds from satellite and aerial imagery. The team, led by Heng Zhou, systematically categorized the field's evolution across three stages: from traditional handcrafted physical priors, to data-driven deep learning restoration, and finally to today's hybrid physical-intelligent generation models. They evaluated more than 30 representative methods spanning CNNs, GANs, Transformers, and diffusion models across five public datasets using 12 different metrics.

Cross-domain experimental results reveal that recent Transformer- and diffusion-based architectures deliver significant performance gains, improving structural similarity (SSIM) by 12-18% and reducing perceptual errors by 20-35% on average compared to earlier approaches. Crucially, the review highlights that models incorporating explicit physical constraints—like atmospheric transmission or airlight modeling—achieve superior radiometric consistency, reducing color bias by up to 27%. This is vital for downstream scientific applications where accurate surface reflectance data is required.

The authors identify five key open challenges for future development: dynamic atmospheric modeling, multimodal data fusion, lightweight deployment for edge devices, scarcity of high-quality paired training data, and handling joint degradations. They outline a roadmap toward building trustworthy, controllable, and efficient (TCE) dehazing systems. The team has made all reviewed resources—including source code, benchmark datasets, evaluation metrics, and reproduction configurations—publicly available to accelerate research and practical deployment in fields like agriculture, urban planning, and climate monitoring.