Algorithm for continual monitoring of fog based on geostationary satellite imagery

This study presents an algorithm for the detection of fog and low stratus (FLS) over Europe based on the infrared bands of the SEVIRI (Spinning Enhanced Visible and InfraRed Imager) instrument on board the Meteosat Second Generation geostationary satellites. As the method operates based on the SEVIRI infrared observations only, it is expected to be stationary in time and thus can provide a coherent and detailed view of FLS development over large areas over the 24 h day cycle. The algorithm is based on a gradient boosted tree machine learning model that is trained with ground truth observations from METeorological Aerodrome Report (METAR) stations and the SEVIRI observations at bands centered at 8.7, 10.8, 12.0, and 13.4 µm wavelengths. The METAR data used here comprise a total number of 2 544 400 data points spread over the winters (i.e., 1 September to 31 May) of the years 2016–2022 and 356 locations across Europe. Among them, the data points corresponding to 276 stations and the winters of 2016-2018 and 2019-2021 (∼ 45 % of all data points) were used to train the algorithm. The remaining data points comprise four independent datasets which were used to validate the algorithm's performance and applicability to time spans and locations within the study area (i.e., Europe) that extend beyond those covered by the data points used for the algorithm training, as well as to compare the algorithm's accuracy at the locations of METAR stations with that of the existing state-of-the-art daytime FLS detection algorithm Satellite-based Operational Fog Observation Scheme (SOFOS). Validation of the algorithm against the METAR data showed that the algorithm is well suited for the detection of FLS. Specifically, the algorithm is found to detect FLS with probability of detection (POD) values ranging from 0.70 to 0.82 (for different inter-comparison approaches) and false alarm ratios (FARs) between 0.21 and 0.31. These numbers are very close to those achieved by SOFOS for differentiating FLS from other sky conditions at the tested locations and time spans. These results also showed that the technique's applicability in the study region extends beyond the particular locations and time spans covered by the data points used for training the algorithm.