An optimized boosting-based machine learning model for predicting maximum scour depth around spur dikes

Spur dikes, installed perpendicular or at an angle to the riverbank, guide the current, curb bank erosion, and keep channels safely navigable. Unfortunately, the flow disturbances they create can also excavate deep local scour holes that threaten a structure's stability. To anticipate this risk, this study developed and compared four boosting-based machine-learning models: Gradient Boosting Machine (GBM), Adaptive Boosting (AdaBoost), Categorical Boosting (CatBoost), and eXtreme Gradient Boosting (XGBoost), for predicting the maximum scour depth (ds). GBM proved the most reliable, yielding a Pearson correlation coefficient of 0.989, a Nash–Sutcliffe efficiency of 0.975, an index of agreement of 0.963, and very low errors (RMSE = 0.219; MAPE = 9.64%). Its superiority over published models and established empirical formulas highlights both the hydraulic soundness and the practical value of this data-driven approach, positioning GBM as a robust tool for the design and maintenance of spur dike structures.