Data-to-data translation-based nowcasting of specific sea fog using geostationary weather satellite observation

Dense sea fog events are responsible for many traffic accidents and fatalities. Research has demonstrated the difficulty in distinguishing sea fog from clouds. This study presents a sea fog nowcasting method that uses brightness temperature (BT) at a mid-wave infrared (MWIR) 3.7 mu m bands and the BT difference between MWIR and long-wave infrared 10.8 mu m bands of communication, ocean, and meteorological satellite data through a data-to-data (D2D) translation based on a conditional generative adversarial networks technique. Sixty fog events that occurred in the Yellow Sea from 2017 to 2020 were used to develop the model. The D2D prediction model for predicting sea fog from 15-min to 2-h ahead with 15 min intervals was trained by stacking individual prediction time datasets in an array of 512 x 512 x 4 pixels for training and 512 x 512 x 8 pixels for testing. Otsu's thresholding method was used to determine sea fog pixels. The sea fog prediction model showed excellent statistical scores, such as the probability of detection = 0.923, false alarm ratio = 0.032, critical success index = 0.895, Heidke skill score = 0.909, and post agreement = 0.968 for 1 h-ahead of a sea fog forecast. Consequently, our study can make a scientific and operational contribution to fog forecasting as useful fog-related data, in addition to other numerical prediction data and ground observation data.