Research on Sea Fog Prediction in the Northwest Pacific Using Machine Learning

The Northwest Pacific (NWP) is a region with the highest frequency of sea fog occurrence and most widely distributed fog area in the world oceans. The sea fog prone area is located in the mid-latitudes where ocean navigations are getting more and more active. Sea fog is heavy obstacles for navigation due to low visibility in fog. The techniques of sea fog forecasting in the NWP still remains unsatisfactory compared with marginal seas so far. Using observations from ICOADS and ERA5 data from 2013 to 2021, this study tried to make sea fog prediction model in the NWP based on machine learnings. The distribution characteristics of sea fog in the NWP were analyzed and compared with China offshore fog areas. Based on analysis of sea fog occurrence, 12 key factors were identified by mutual information (MI) method, including sea surface temperature (SST), surface air temperature (SAT), SST-SAT, humidity, wind speed and direction, etc. In addition, geographical coordinates (latitude and longitude information) were also taken into consideration as factors. Four machine learning models were constructed for sea fog prediction, employing resampling techniques to address the extreme imbalance between foggy and non-foggy samples. The results demonstrated a significant enhancement in model performance by resampling, especially with oversampling, and a decline in performance was noted upon the removal of the geographical coordinates. Among the four evaluated models, the eXtreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), and Convolutional Neural Network (CNN) exhibited similar performance metrics, with threat score (TS) scores about 0.3 and the Decision Tree (DT) showed more stable results. Regarding individual case performance, the XGBoost model outperformed the others, showing the highest agreement with the fog area range observed in satellite images. This study reveals the complexities of sea fog formation in the NWP and provides a scientific basis for sea fog prediction in vast expanded ocean areas.