Improving 24-h surface-drifter trajectory prediction around the Korean Peninsula using a multi-source machine learning framework

Predicting surface-drifter trajectories is vital for material transport studies, marine hazard response, and analyses of upper-ocean circulation. However, this task remains challenging because winds, tides, mesoscale currents, and waves interact nonlinearly in ways that are difficult to represent in conventional drift-forecasting frameworks. This study developed a multi-source machine learning framework for drifter trajectory prediction around the Korean Peninsula. The framework uses oceanic model outputs from HYCOM, TPXO, and SCHISM, together with atmospheric and wave forcings from ERA5, as environmental input variables. An Extreme Gradient Boosting (XGBoost)-based surrogate model was trained to estimate the zonal and meridional components of drifter velocity, and the resulting velocity estimates were time-integrated to reconstruct 24-h trajectories. Model performance was evaluated against independent drifter observations and compared with a SCHISM-based OpenDrift simulation, while SHapley Additive exPlanations (SHAP) analysis was used to examine regional variations in input-variable attributions. Applied to the marginal seas around the Korean Peninsula, the proposed approach improved 24-h trajectory prediction, reducing distance RMSE from 9.73 to 7.49 km (23%) and increasing the normalized cumulative Lagrangian separation skill score (NCLS-SS) from 0.47 to 0.57 (21%) over the independent test set. Performance gains were observed across the tide-dominated western, mixed southern, and eddy-influenced eastern subregions. The SHAP results further revealed systematic regional contrasts in the attribution patterns of the environmental inputs across these dynamical regimes. These results indicate that the proposed framework offers a practical and interpretable approach for short-term surface-drift forecasting in heterogeneous coastal and marginal-sea environments.