A Deep Learning-Based Method for Sea Fog Nowcasting over the Southern Coast Area of the Shandong Peninsula

The methodologies for researching and predicting sea fog predominantly rely on numerical simulations. Nonetheless, the presence of initial condition errors and model flaws frequently constrains the precision of sea fog nowcasting using numerical simulations. Focusing on the southern shore of the Shandong Peninsula, a region with frequent sea fog occurrences, this study proposes a nowcasting approach for sea fog area based on deep learning (DL) technology and develops a corresponding DL forecasting model. The model is trained utilizing five forecasting variables: temperature, wind, specific humidity, land-sea mask, and sea fog mask. The forecasting findings for 13 instances of sea fog along the southern coast of the Shandong Peninsula indicate that the average Equitable Threat Score (ETS) and Probability of Detection (POD) for 3-hour predictions are roughly 0.50 and 0.75, respectively. During a typical sea fog event, a number of sensitivity experiments are performed to evaluate the forecasting model′s responsiveness to different forecasting variables. The results demonstrate that increasing wind speed caused the forecasted fog area to dissipate, while alterations in humidity similarly affect fog formation, indicating that the DL forecasting model effectively accounts for the impact of various physical factors on sea fog development. However, it was noted that the model′s response to certain physical factors, such as sea surface temperature, was less pronounced. Additionally, errors in satellite-retrieved sea fog mask impacted the model's forecast performance, highlighting the need for further model refinement. Further study is essential to investigate the influence of integrating other forecasting variables on the model′s efficacy, and comparisons with numerical forecasting outcomes are necessary to assess the model′s practical application.