Unified scour depth prediction for culverts and sluices using random forest and model-based sensitivity analysis

Abstract Accurate prediction of scour depth downstream of water conveyance structures such as culverts and sluices is critical for ensuring infrastructure safety and for mitigating environmentally damaging bed degradation, sediment mobilization, and habitat disturbance. Although these structures share similar hydraulic and geometric characteristics, previous studies have typically analyzed them separately, often leading to inconsistent conclusions regarding the most effective predictive models and dominant parameters. This study presents a unified framework that evaluates multiple machine learning (ML) models—Random Forest (RF), Artificial Neural Network (ANN), Group Method of Data Handling (GMDH), and Support Vector Regression (SVR)—for predicting maximum scour depth downstream of both sluices and culverts. A Model-based Sensitivity Analysis (MBSA) approach is introduced to quantify the relative importance of input parameters without model retraining. Results show that RF significantly outperforms other ML models and empirical equations in terms of predictive accuracy, confirmed by statistical testing across 100 independent trials. Furthermore, the MBSA consistently identifies the densimetric Froude number ( F d ) as the most influential parameter governing scour development. These findings demonstrate RF’s robustness for scour depth prediction and highlight MBSA as a reliable, physically interpretable tool for understanding parameter influence in hydraulic systems with direct implications for environmentally resilient river management and design.