Hindcasting Extreme Significant Wave Heights Under Fetch-Limited Conditions with Tree-Based Models

Abstract

Accurately hindcasting waves in semi-enclosed, fetch-limited basins remains challenging for reanalysis models, which tend to underestimate storm peaks near the coast. We developed interpretable ML models for Rijeka Bay (northern Adriatic) using only wind observations from two land-based wind stations to predict buoy Hm0 measurements spanning 2009–2011 (testing) and 2019–2021 (training and validation). The tested tree-based models included Random Forest, XGBoost, and Explainable Boosting Machine. This study introduces a novel approach in the literature by employing weighted schemes and feature engineering to enhance the predictive performance of interpretable, low-complexity machine learning models in hindcasting waves. Representing wind direction as sine–cosine components generally reduced RMSE and BIAS relative to traditional speed–direction inputs, while an exponential sample weight scheme that emphasized storm waves halved extreme Hm0 underprediction without inflating overall RMSE. The best-performing model, a Random Forest model, achieved an RMSE of 0.096 m and a correlation of 0.855 on the unseen test set—30% lower overall RMSE and 50% lower extreme wave RMSE than the MEDSEA and COEXMED hindcasts. Additionally, the underprediction was reduced by 90% compared to these reanalysis models. The method offers a computationally lightweight, transferable supplement to numerical wave guidance for coastal engineering and harbor operations.