Predicting scour depth in a meandering channel with spur dike: A comparative analysis of machine learning techniques

In this research, an assessment of scour depth prediction in meandering channels with spur dikes is made employing machine learning approaches. Efficient determination of the scour depth is therefore vital in the prediction of morphologic aspects and structural stability. The input parameters include sinuosity (S), spur dike locations (Ld), and porosity (P) with experimental data from sinusoidal flumes. Four machine learning models; Extreme Gradient Boosting (XGBoost) with Particle Swarm Optimization (PSO) XGBoost-PSO, Random Forest (RF), k-Nearest Neighbors (k-NN), and Decision Tree-Neural Network (DT-NN) were used and compared. The findings demonstrate an R-value of 0.995 in the case of RF model while XGBoost-PSO gave second-best accuracy with R = 0.988. The results of the SHAP analysis illustrated that porosity and sinuosity are significant factors affecting scour depth (Ds/Yn, Ds: scour depth, Yn: water depth) and had moderate importance assigned to spur dike location. Kernel density plots further supported the RF model regarding error distribution consistency. Even though, both XGBoost-PSO yielded better results because of hyperparameter tuning, k-NN and DT-NN had less precise outcomes specifically predicted for progressive hydraulic procedures. Taylor's diagram even revealed greater accuracy of prediction by RF. Hence, a proper selection of appropriate machine learning models remains the first step in estimating scour depth sufficiently for flood and erosion control.