Deep learning for fog detection and prediction: A systematic review with an adapted quality assessment framework
Isti Ma’atun Nasichah, Mohammad Isa Irawan, Budi Setiyono · Array · 2026
Fog severely disrupts transportation systems, underscoring the need for reliable detection and short-term prediction. This study presents a systematic review of deep learning approaches for fog detection and prediction using satellite and CCTV imagery. It maps datasets, architectures, preprocessing, and evaluation strategies. It also introduces a domain-specific adaptation of the PROBAST+AI framework to assess methodological quality and operational applicability in remote-sensing contexts. Following the PRISMA 2020 protocol, 203 records published between 2016 and 2025 were screened in Scopus, yielding 25 eligible studies. Most studies used geostationary satellite imagery, including Himawari, FY-4, and MODIS. There is also increasing multimodal integration of meteorological variables. Model architectures evolved from CNN and U-Net to transformer-based and probabilistic approaches. PROBAST+AI ratings revealed recurrent high or unclear risks in the analysis domain. These stemmed from small sample sizes, potential data leakage, and limited external validation across satellites, regions, and night–dawn periods. Calibration and uncertainty reporting were minimal. Datasets remained geographically concentrated in East Asia. The adapted framework shows that model credibility is shaped more by data provenance and evaluation design than by model complexity. These findings highlight persistent challenges. These include imbalanced datasets, temporal leakage risks, and scarce cross-sensor validation. At the same time, the review provides a reproducible foundation for improving transparency, methodological rigor, and operational reliability in future fog-detection and prediction models.