Enhancing Coastal Wave Forecasts Around the Korean Peninsula with Time-Series Decomposition and Spatio-Temporal Deep Learning

The accurate prediction of significant wave height (SWH) is essential for maritime safety, coastal management, and disaster preparedness. Using ERA5 reanalysis data, a deep-learning-based modeling framework was proposed to predict the SWH in the seas surrounding the Korean Peninsula. As a baseline model, a Simple Ocean Prediction Network (SMOP-Net) was constructed based on a Convolutional Long Short-Term Memory architecture, reflecting the structure of conventional spatiotemporal prediction models for wave data. To improve the predictive performance, an enhanced model, the time-series Decomposition-based Ocean Prediction Network (TDOP-Net), was developed by incorporating a time-series decomposition process into the SMOP-Net input layer. In this process, raw wave data were decomposed into long-term trends, seasonal patterns, and residual components, which were then provided as parallel input channels for prediction. Both models were trained using data from 2016 to 2020 and evaluated using independent data from 2021. The results showed that the TDOP-Net consistently outperformed the SMOP-Net across all sub-areas, with improvements in prediction accuracy reaching up to 30% in the Yellow Sea, where coastal complexity and limited spatial resolution often challenge numerical models. The performance gain from decomposition was particularly pronounced in dynamically variable regions and under shorter input windows. Sensitivity experiments further indicated that a 12-h input window provided the best trade-off between predictive accuracy and computational efficiency. Overall, the proposed TDOP-Net enhanced both accuracy and robustness, making it suitable for real-time operational coastal hazard early warning systems, especially in marginal seas subjected to rapid environmental variability.