Data-driven approaches in predicting scour depths at a vertical seawall on a permeable shingle foreshore

Abstract

Despite recent research efforts in advancing machine learning (ML) tools to predict nearshore characteristics at sea defences, less attention has been paid to ML algorithms in predicting scouring characteristics at vertical seawalls. In this study, four ML approaches were investigated, including gradient boosting decision trees (GBDT), random forest (RF), support vector regression (SVR), and ridge regression (RR). These approaches were utilised to predict scour depths at the toe of an impermeable vertical seawall in front of a permeable shingle slope. The developed ML algorithms were trained and tested (70% for training and 30% for testing) using the scouring datasets collected from laboratory tests performed on seawalls in a 2D wave flume at the University of Warwick. A novel hyperparameter tuning analysis was performed for each ML model to tailor the underlying dataset features while mitigating associated data overfitting risks. Additionally, the model training process demonstrated permutation feature importance analysis to reduce overfitting and data redundancy. The model predictions were compared with the observed values using the coefficient of determination (R 2) score, root mean square error (RMSE), and Pearson correlation R-value. Consequently, the RF and GBDT methods accurately predicted scour depths at the toe of vertical seawalls with shingle foreshores. This study produced data, information, and a model that could directly or indirectly benefit coastal managers, engineers, and local policymakers. These benefits included forecasting scour depths and assessing the impact on the structural integrity of the sea defences in response to the threat imposed by extreme events, which are essential for the sustainable management of coastal protections and properties behind such structures in coastal areas.