Prediction of Maximum Scour Depth near Spur Dikes in Uniform Bed Sediment Using Stacked Generalization Ensemble Tree-Based Frameworks

Abstract

The scouring process near spur dikes could jeopardize the stability of riverbanks. Thus, accurate estimation of the maximum scour depth near spur dikes is crucial in river engineering. However, due to the complexity of the scour phenomenon around these structures, it has been a challenge to accurately estimate the maximum scour depth. Few efforts have been made to develop a machine learning (ML) approach for such a purpose. In this study, two novel multilayer stacked generalization frameworks are developed to model the scour depth near a spur dike in a uniform sediment condition. Stacked tree-based frameworks consist of three standalone ML approaches, including multivariate adaptive regression spline, multigene genetic programming, and kernel extreme learning machine as the first layer and the boosting regression tree (BRT) and bagging regression tree (BGT) as meta learners. A total of 186 data points were collected from previous experimental studies, and 32 flume experiments were further conducted under clear water conditions at the Indian Institute of Technology (IIT) Roorkee Laboratory in this study. The performances of the models were assessed using various statistical metrics [e.g., correlation coefficient (R), root-mean-square error (RMSE), and mean absolute percentage error (MAPE)], graphical criteria, and some existing empirical equations. The modeling results demonstrated that the stacked BRT and BGT frameworks were superior to all the standalone ML approaches (R=0.9786, RMSE=0.1654, and MAPE=11.68 for BRT; and R=0.9742, RMSE=0.1831, and MAPE =11.29 for BGT). This finding was also confirmed by a comparison between the empirical correlations and the artificial intelligence models. In addition, the sensitivity analysis proved that the mean sediment size ratio (L/d50) was the most influential variable in estimating the scour depth near spur dikes.