Forecasting summer marine heatwaves in the South China Sea using explainable machine learning models

Abstract

Marine Heatwaves (MHWs), classified as extreme oceanographic meteorological phenomena, have profound effects on ecological systems and socio-economic activities within the South China Sea (SCS). Multiple machine learning (ML) algorithms, specifically Random Forest (RF) and Ridge models, are employed to predict summer Sea Surface Temperature Anomalies (SSTA) and occurrence of MHWs in the SCS. The ML forecast results are also compared with climatology forecasts, persistence forecasts, and the European Centre for Medium-Range Weather Forecasts Sub-seasonal to Seasonal (ECMWF S2S) hindcasts. The results reveal that for regression forecasting with a one-week lead time, the Ridge model performs the best among all models. However, for longer lead times, all models tend towards climatology forecasts. When it comes to the classification forecasting, the RF model demonstrates superior performance to predict extreme MHWs compared to typical MHWs. Moreover, the feature importance is measured via the mean decrease in impurity (MDI) incorporated in the RF model to identify the predictors that contribute to predicting the MHWs. The SHapley Additive exPlanation (SHAP) algorithm is additionally employed to evaluate both the positive and negative impacts of these predictors on the model. SSTA and the Indian Ocean Basin-Wide (IOBW) index are the most two critical forecasting factors influencing the occurrence of MHWs, exhibiting a pronounced positive contribution to the model. The insights derived from this study are expected to provide strong support for the MHWs early warning system in the SCS and provide essential information for the conservation of marine ecosystems and the management of fisheries resources.