Prediction of maximum scour depth downstream of bed sills using integrated machine learning algorithms

Abstract

Accurate prediction of maximum relative scour depth (y_s/H_s) is critical for hydraulic infrastructure resilience. This study advances scour depth prediction downstream of bed sills by establishing three ensemble models—Random Forest (RF), Gradient Boosting Decision Tree (GBDT), and XGBoost—trained on a comprehensive dataset combining 328 standardized flume experiments (clear-water and live-bed conditions) with 73 field measurements. Validation using 29 field datasets from Maso River reveals MAE reductions of 32.5 %, 28.7 %, and 30.2 % for RF, GBDT, and XGBoost, respectively, compared to laboratory-trained models, translating to at least 31.6 % higher accuracy than traditional empirical approaches. Comprehensive sensitivity analysis identifies four dimensionless parameters as critical predictors, ranked by their relative importance to scour development: morphological transition coefficient (a₁/H_s) > sediment sorting coefficient (a₁/ΔD₉₅) > weir spacing ratio (L/H_s) > channel slope (S₀). By integrating lab and field data, this approach enhances scour prediction accuracy for fluvial risk management.