Open-Loop Generative AI Nowcasting of Dense Marine Fog Visibility from the FATIMA Grand-Banks Campaign Measurements Using Multiple Lookback Windows

Reduced visibility due to fog poses significant safety and operational risks in marine environments, emphasizing the need for accurate short-term nowcasting. Typical machine learning forecasting models fail to capture temporal meteorological dependencies that influence fog dynamics, due to the reliance on data assimilated measurements for closed-loop time series prediction. This work introduces a novel multiple lookback window (MLW) architecture that utilizes the recurrent neural network model of Gated Recurrent Units (GRU) to nowcast dense fog visibility (Vis < 400 m) time series. This architecture facilitates learning the short- and long-term dependencies in meteorological data for non-data assimilation (open-loop) time series prediction. Prediction intervals were obtained using Bayesian approximation at 95% confidence interval. Vis time series nowcasts were obtained using autoregressive generation at increasing lead times, without relying on the assimilation of future observations and were evaluated across various fog Vis conditions characterized by its coefficient of variation. Marine fog Vis conditions and measurements were collected using instrumentation mounted on the Research Vessel Atlantic Condor from the FATIMA (Fog and turbulence interactions in the marine atmosphere) campaign in July 2022 in the Grand Banks and Sable Island areas of the North Atlantic region of Canada. The MLW-architecture with the GRU model outperformed the naïve persistence model nowcast for open-loop nowcasting dense fog Vis, with a mean RMSE of 11.829±1.260 (2.96% error at 400 m) and Skill Score of 0.073±0.061. The best nowcasting occurred at the 10 and 20 min lead times, where the mean Skill Score across both lead times was 0.132±0.082 and the RMSE was 7.013±1.105m (1.75% error at 400 m) and in high variability conditions with a mean SS of 0.387±0.037 and RMSE of 6.990±1.990 (1.75% error at 400 m). Thus, the proposed model is robust and practical for dense fog Vis nowcasting.