Improved ocean fog detection and prediction over the Yellow Sea using geostationary satellite and numerical model data based on machine learning

Abstract

This dissertation aims to improve the detection and prediction of ocean fog in order to enhance maritime and coastal safety and reduce socioeconomic losses. This dissertation comprises four chapters. Chapter 1 provides an overview of the background and objective of the dissertation. Chapter 2 proposes an improved ocean fog detection model by utilizing a machine learning technique on the infrared channels obtained from the Himawari-8 satellite. The developed ocean fog detection model not only demonstrated high accuracy with an F1 score of 97.93% and a proportion correct of 98.59% throughout the day but also showed spatiotemporal continuity in comparison to the operational product. Furthermore, it employs Shapley additive explanation (SHAP) analysis to explain the contributions of different input variables. As a result, the function changes of each input variable according to night/day and the usefulness of the 12μm channel were revealed. Chapter 3 introduces an improved prediction model for accurately forecasting ocean fog in the short term. This model utilizes meteorological data from the Local Data Assimilation and Prediction System (LDAPS) model, sea surface temperature data from the Hybrid Coordinate Ocean Model (HYCOM), and accumulated shortwave radiation (SWR) data from Himawari-8. The AutoGluon model utilizes these inputs to predict ocean fog up to six hours in advance with exceptional accuracy, surpassing the performance of the operational LDAPS model with an F1 score of 75% and a proportion correct of 79.8%. This model emphasizes the crucial role of SWR in predicting ocean fog and demonstrates the potential of integrating satellite observations with numerical simulations to enhance reliability. Chapter 4 presents final conclusions regarding the dissertation's contribution to future research and maritime safety. These conclusions are based on the high-performance results and the interpretation of the role of input variables. In addition to the conclusions, the study also proposes possible future research ideas regarding object-based ocean fog detection and self-qualifying ocean fog prediction.