Marine And Coastal Fog: Forecasts and Evaluations

Abstract

Accurate marine fog forecast is of importance for human activities in the coastal regions and over the ocean. However, it is still a challenge due to lack of observation and unsatisfactory model performance. This dissertation includes three studies on forecasting marine fog. The first study has summarized the performances of three different models during the FATIMA (Fog And Turbulence Interactions in the Marine Atmosphere) field campaign on Sable Island, Nova Scotia, including two WRF (Weather Research and Forecasting) models and a COAMPS (Coupled Ocean/Atmosphere Mesoscale Prediction System) model. It is found that the models can perform differently on fog prediction with similar errors in temperature, dew point, relative humidity, wind speed and wind direction. Additional tests show that the WRF model can be improved by adjusting the horizontal and vertical domains, and a spin-up time is necessary for forecasting fog.

The second study has compared the liquid water content and droplet concentration observations from the field campaign to WRF variables in two periods. It is found that the WRF model produces liquid water contents up to 0.6 g m-3 while the observation has up to 0.3 g m-3, partly due to the updraft speed in the microphysics scheme being too high and causing a high activation rate towards droplets. The surface temperature of the island, not included in the GFS (Global Forecast System) data but assigned by WPS (WRF Preprocessing System) is also too high, causing the fog to incorrectly dissipate in daytime.

The last study has used XGBoost, a machine learning model to post-process WRF output. XGBoost is trained with the ERA5 (ECMWF Reanalysis v5) data of St. John’s, Newfoundland and Labrador, and Yarmouth, Nova Scotia. XGBoost can predict fog or clear with given 2 m temperature and relative humidity, 10 m U and V winds, land surface pressure, their corresponding values one hour ago, hour, day and month. Tests using forecasts of the features from WRF in 2024 found that XGBoost improved the recall by up to 0.07 without decreasing the precision, compared to using WRF only. This shows the potential to combine the strength of both models.