Characterization and forecast of a unique fog and low-level clouds event: microphysics measurements, mesoscale modeling and machine learning

We present a study of the microphysics, mesoscale and synoptic conditions of a rare radiation-fog and low-level clouds event (hereafter, FC event), and build numerical tools to forecast it. The FC event developed in the south-eastern Mediterranean region during January 3-6, 2021. The FC formed during nighttime from south to coastal areas and dissipated at morning hours leaving low-clouds only. The synoptic conditions were dominated by Red Sea Troughs at the surface without cyclonic upper air circulation, suitable for radiation fog development. The following methods were combined to analyze the event and to evaluate the feasibility of accurately numerical forecasting it: 1. in-situ measurements consisting of Forward Scattering Spectrometer Probe FSSP-100, surface meteorological stations and radiosoundings, 2. satellite-retrieved IR and visible imagery [by the Spinning Enhanced Visible and InfraRed Imager (SEVIRI) instrument (https://www.eumetsat.int/seviri) on-board of Meteosat Second Generation (MSG) satellites], 3. high resolution (1-km grid size) Weather Research and Forecast model (WRF) with Real-Time Four-Dimensional Data Assimilation (RTFDDA) model forecasts, 4. post-processing of the model forecasts with simple and machine learning (ML) algorithms. The micro-physical analysis involved measurements of droplet size distribution and visibility range, allowing the calculation of liquid-water content and effective diameter of droplets. The measured visibility range was 90 m. The droplet diameter main mode was 1-2 micrometers, followed by another one around 6 micrometers. Typical liquid-water content values were 0.01-0.025 g/m³. These measurements were in agreement with the classification of the satellite imagery as “small drops fog/low-clouds”. FC forecasting by numerical-weather-prediction models is still challenging, as microphysics parameterizations are too crude. Therefore, we developed two post-processing algorithms based on basic model-forecast variables: wind speed, dew-point temperature and relative humidity at 1000 and 975 hPa vertical levels. The first post-processing algorithm identified FC based on a combination of thresholds of the aforementioned model variables (“Thresholds Algorithm”, TA). It was verified against satellite imagery and independent in-situ observations. The second is a Gradient Boosted Tree (GBT) ML post-processing algorithm in which the aforementioned model variables served as features and satellite imagery as label in the training process. Verification of the GBT algorithm was performed by cross-validation against satellite imagery and against independent in-situ observations, too. Both, the TA and GBT algorithms proved useful to identify FC areas. The GBT algorithm over-performed the TA algorithm during early morning hours, though it overestimates FC areas during the late morning. The combination of the GBT algorithm and TA is able to remove this inaccuracy providing an optimal strategy to post-process model forecasts. While the satellite imagery cannot distinguish between surface fog and low-level clouds, the post-processed model, does show differences between the two analyzed vertical levels, providing the possibility of determining the vertical extent and level of the phenomenon whether fog or low-level clouds.