Short-Term and Long-Term Prediction of South China Sea SST Based on Multiple Meteorological Factors and Machine Learning

Sea surface temperature (SST) is a vital component of the climate system, and its spatiotemporal variations significantly influence global climate and ecological equilibrium. In this paper, based on the ERA5 reanalysis data, three machine learning algorithms, namely, Random Forest (RF), XGBoost and LightGBM, are used to construct short-term and long-term SST forecast models for the South China Sea. The input feature variables include seven meteorological and hydrological variables such as SST, 10m u-component of wind (U10), 10m v-component of wind (V10), 2m dewpoint temperature (d2m), 2m temperature (t2m), mean sea level pressure (SLP), and total cloud cover (TCC). Correlation analysis revealed that these meteorological factors are significantly correlated with SST, with the strongest correlations observed for 2-meter dew point temperature and 2-meter air temperature. Model performance is assessed using metrics such as Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and the coefficient of determination (R²). The results indicate that the RF model exhibits the highest accuracy for both short-term and long-term forecasting models. Furthermore, this study explores high-resolution SST forecasting models for the South China Sea, revealing that total cloud cover (TCC) contributes more to SST predictions than sea surface salinity (SSS), and the model performs well across most areas of the South China Sea (excluding coastal regions), achieving forecasts with a lead time of at least 20 months. These findings demonstrate the feasibility of machine learning algorithms for SST prediction, providing an efficient approach to understanding future SST changes and their potential impacts, while emphasizing the necessity of integrating multiple meteorological factors to enhance forecasting accuracy.