Understanding and Modeling Marine Fog in Areas Offshore from Atlantic Canada

Abstract

In recent years, machine learning (ML) has gained popularity in the field of weather forecasting, particularly in areas where numerical weather prediction (NWP) models face challenges. One such area is fog prediction. A key factor in improving fog forecasts is the accurate representation of microphysics processes, which play a crucial role in fog formation and dissipation. In this study, we propose a post-processing approach that combines the Weather Research and Forecast (WRF) model with a machine learning classifier to distinguish between fog and no-fog conditions. This research builds on previous work conducted during the Fog and Turbulence Interactions in the Marine Atmosphere (FATIMA) project, where a 3-day marine fog forecast over the Yellow Sea was provided using WRF. The project highlighted the importance of microphysics parameterization schemes in accurately forecasting marine fog. Our study spans eleven years, from 2012 to 2023, focusing on the months from April to August each year. We utilize the WRF Preprocessing System (WPS) with initial and boundary conditions provided by ERA5 reanalysis data. The features of our ML model include 2-meter temperature, U and V wind components, 2-meter relative humidity, surface pressure, and the month, day, and hour. The target variable is hourly reported visibility data from Navigation Canada on the Environment and Climate Change Canada (ECCC) website for the locations of St. John’s, Newfoundland and Labrador, and Yarmouth, Nova Scotia in Canada. When tested with data in 2024, our ML model demonstrates better performance compared to predictions based solely on the liquid water content (LWC) from the WRF model. We used various metrics to evaluate the classification, with the F1 score being the most important metric. By using the ExtraTreesClassifier model, we were able to obtain a 11% increase in the F1-score (0.69 vs. 0.62)and a 2% increase in accuracy (0.90 vs. 0.88) compared to the WRF model at St John’s. A similar performance was noted in Yarmouth. The accuracy there increased by 4% (0.86 vs. 0.82) and F1-score by 11% (0.59 vs. 0.53). This approach shows promise in enhancing the accuracy of fog prediction, offering valuable insights for aviation, marine operations, and transportation safety.