Estimation of Time-Dependent Pier Scour Depth Using Ensemble and Boosting-Based Data-Driven Approaches

The scour phenomenon around the vertical piles in rivers and oceans can have a significant impact on the stability of the structures. As a result, accurate prediction of the scour depth forms an important challenge in the design of piles. Various empirical approaches proposed in the literature are often confined to specific environmental and bed conditions. So, when such empirical approaches are applied to a new environment, they either underestimate or overestimate the scour depth, which may lead to improper design of the piles. This study aims to develop two data-driven approaches: extra trees regressor (ETR) and extreme gradient boosting regressor (XGBR), which are ensemble and boosting-based machine learning-based approaches, respectively, to estimate the temporal variation of pier scour depth with non-uniform sediments under clear water conditions. The motivation behind using a boosting and an ensemble-based approach is that they provide superior results as compared to standard machine learning-based approaches. The dataset is compiled using various sources from existing literature. For each of the data-driven approaches, nine different combinations of features (shallowness of the flow, sediment coarseness, densimetric Froude number, sediment particle size distribution, pier Froude number, and three different dimensionless time scales) are tried in order to determine the best combination that can be used for prediction of scour depth. Both extra trees regressor and XGBR excel at prediction of the scour depths, but extra trees regressor performs better in most of the models as compared to XGBR. The highest r² and NSE across nine models for extra trees regressor are 0.956 and 0.9544, respectively, while in the case of XGBR, the highest r² and NSE across nine models are reported as 0.9474 and 0.9461, respectively.