Comparative Study of Machine Learning Techniques for Prediction of Scour Depth around Spur Dikes

Spur dikes are hydraulic structures constructed within rivers to mitigate bank erosion, enhancing the stability of riverbanks. It has received considerable attention and is being adopted throughout the world. However, scouring around spur dikes affects their hydraulic performance and stability. Therefore, a precise prediction of scour depth fluctuations is essential for developing reliable and cost-effective designs for spur dikes. This study applies two types of regression-based techniques (linear and non-linear regression) and explicit expression (multivariate adaptive regression splines, M5P tree, and group method of data handling) based machine learning techniques to predict scour depth around spur dikes. For this purpose, a database of 154 experimental observations has been collected from existing literature, of which 80% and 20% of observations have been used for training and testing subsets, respectively. The root mean square error (RMSE), coefficient of determination (R-square), coefficient of correlation (CC), and mean absolute error (MAE) have been used to compare the performance and accuracy of these techniques. The multivariate adaptive regression splines (MARS) technique has shown the highest accuracy (RMSE = 0.0883, R-square = 0.9869, CC = 0.9934, and MAE = 0.0485). It is the best machine learning technique for predicting scour depth around spur dikes as compared to the M5P tree (M5P), group method of data handling (GMDH), non-linear regression (NLR), and linear regression (LR). In addition, two regression-based equations and three explicit expression-based equations are derived for predicting scour depth in any instance.