Validation-Guided Development of a Virtual Buoy for Coastal Wave Reconstruction: Feature Ablation and Historical High-Wave-Event Augmentation

Abstract

Reliable nearshore wave information is essential for harbor engineering, coastal-hazard mitigation, and marine operations, yet field observations are often limited by sparse monitoring networks, data gaps, and few high-wave samples. Using hourly observations from three marine stations and one tide gauge, this study develops a multi-station virtual-buoy framework to reconstruct significant wave height, peak period, and wave direction. Random Forest is used as the primary model, and an independent validation set guides comparisons of feature modules, coordinate representations, lagged inputs, static spatial features, and alternative algorithms; the test set is retained for final hold-out evaluation. Validation results show that wave and wind inputs form a consistently competitive configuration: wind provides the most consistent auxiliary benefit, although the absolute improvements are generally modest, whereas current and tide provide limited and less consistent additional benefit. Hs is comparatively less sensitive to lag configuration, whereas Tp and Dir benefit more consistently from short-term lagged inputs; static features add little. Historical high-wave-event-window augmentation substantially improves Hs reconstruction during unseen high-wave events, although the most severe peak remains underestimated. Overall, validation-guided feature selection and targeted historical-event augmentation improve reconstruction under both ordinary and high-wave conditions while retaining a relatively simple model architecture.