Phase-resolving wave surface prediction under typhoon conditions using hierarchical spatiotemporal deep learning model

Accurate phase-resolved prediction of nearshore wave fields is needed for coastal hazard assessment. This study develops a deep learning framework with boundary water levels as input which directly establishes the mapping between boundary water-level sequences, bathymetry, and the spatiotemporal surface-elevation field, without feeding previous predictions back into the model. Meanwhile, this study proposes a hierarchical spatiotemporal deep learning model integrated with FourierGRU to address the excessively long training time. A coupled SWAN-SWASH system generated 472 samples from 22 typhoon events. U-RNN and hierarchical model were evaluated with surface elevation and SWH. All models reproduced the dominant wave-train patterns. U-RNN achieves the highest prediction accuracy, with an SWH RMSE of 0.187 m and PCC of 0.992. The hierarchical model reduced training time from more than 15 days to 0.9 days and retained an SWH MAE of 0.175 m, demonstrating a favorable balance between accuracy and efficiency. The study demonstrates that the deep learning model with boundary water levels as input enables fast, long-horizon prediction of phase-resolved nearshore wave fields. The proposed hierarchical spatiotemporal model further provides a practical, low-cost technical solution for operational forecasting, engineering design optimization, and rapid decision-making in coastal and offshore engineering.