An explainable machine learning framework for bedload transport prediction in vegetated and bare channels

Accurate prediction of bedload transport in vegetated channels is crucial for ecological restoration and river engineering but remains challenging due to strong nonlinearity. While machine learning (ML) has succeeded in bare-bed conditions, a systematic investigation for complex vegetated flows remains lacking. This study compiled 396 flume experiments from multiple sources and built a unified data-driven workflow to predict unit-width bedload transport rates under both vegetated and unvegetated conditions. Two independent ML algorithms (ANN and SVR) and two ensemble ML methods (XGBoost and RF) were evaluated. Bayesian optimization was employed for hyperparameter tuning, and a nested 5-fold cross-validation framework with stratified sampling and sample weighting was used to specifically address the prediction of engineering-critical extreme transport events. By comparing different input combinations, the study identified the full set of directly measured hydrodynamic, sediment, and geometric variables as the optimal input configuration. Benchmarking against representative empirical formulas revealed a substantial performance advantage of the data-driven models over traditional approaches. Among them, XGBoost achieved the optimal comprehensive performance. Sensitivity analysis, SHapley Additive exPlanations (SHAP), and Sobol indices revealed that hydrodynamic variables, particularly flow velocity and energy slope, are the dominant predictors. Furthermore, the analysis indicates that the vegetation density parameter makes a smaller direct contribution to model output, as its influence on sediment transport is primarily exerted indirectly by modifying the flow structure. This offers a novel engineering insight: ML models trained on bare-bed data possess the potential to be applied to vegetated channels if representative near-bed hydraulic parameters are available.