Automated night-time fog detection and masking using machine learning from near real-time satellite observations

Fog significantly reduces visibility, impacting transportation and safety, especially in United Arab Emirates (UAE) during winter months. This study develops a machine learning (ML) approach for automated fog detection and masking from near real-time SEVIRI Satellite observations. We evaluate six ML models across four training strategies: (1) supervised training using SEVIRI nighttime microphysics Red-Green-Blue (RGB) pixels with Meteorological Aerodrome Reports (METAR) station labels; (2) as (1) but adding the three infrared channels; (3) k-means labels derived from Night Microphysics RGB; and (4) a fusion of station-labeled and k-means-labeled data. Among the models, the eXtreme Gradient Boosting (XGBoost) performs best overall. Using the same fog events analyzed by Weston and Temimi (2020), the fusion approach (Approach 4) with XGBoost more sharply delineates fog boundaries, accurately captures “fog holes”, and reduces false alarms and missed detections—including during marginal/light-mist episodes—relative to the thresholding method, with notable improvements over inland deserts and along the coast. At Abu Dhabi, station-trained models achieve a Probability of Detection of ∼0.73 with a False Alarm Ratio of ∼0.11; the fusion approach maintains strong detection skill with competitive false-alarm rates while improving spatial coherence. Regional case studies over Qatar and Saudi Arabia demonstrate that the trained model generalizes across the Arabian Peninsula. The workflow executes in seconds and relies only on three infrared channels, avoiding auxiliary reanalysis inputs and supporting near-real-time operations. These results show that combining complementary labels from stations and clustering substantially enhances satellite-based fog masking, providing a practical pathway for operational monitoring and a foundation for short-term nowcasting in arid environments.