Evaluation of Traditional and Ensemble ML Algorithms for Scour Depth Prediction: Performance, Error Distribution, and Gaussian Noise Robustness

Reliable prediction of scour depth is essential for hydraulic design, yet its nonlinear dependence on flow and sediment parameters often limits the accuracy of empirical formulations. This study develops a comprehensive machine-learning framework to model scour depth using a dataset of 450 samples. Two families of ML models were employed: (i) traditional techniques—Decision Tree and Support Vector Regression (SVR), and (ii) ensemble-based techniques—Random Forest, Bagging Regressor, AdaBoost, and Gradient Boosting. Model performance was evaluated using multiple statistical and graphical diagnostics, including scatter plots, residual distributions, cumulative relative-error curves, frequency histograms, Taylor diagrams, and train–test comparisons.

To assess robustness, controlled Gaussian noise perturbations (2.5–15%) were synthetically induced in the input variables, and ten Monte-Carlo trials were performed for each noise level. For every model, the lower bound, upper bound, and mean R² values were computed, enabling a stability-based comparison. Ensemble models demonstrated substantially higher accuracy and noise-tolerance than traditional approaches. Gradient Boosting and Random Forest consistently exhibited the highest coefficient of determination, narrowest error bands, and least sensitivity to perturbations, whereas SVR and Decision Tree showed wider deviation ranges.

Overall, the findings confirm that ensemble learning—particularly boosting-based methods—provides a more accurate, robust, and generalizable tool for scour prediction compared to standalone ML models. The proposed framework establishes a reproducible methodology that integrates predictive accuracy with noise-resilience, making it suitable for practical hydraulic engineering applications.