The Role of Machine Learning, Incorporating Wind Data, to Predict Coastal Wave Overtopping

Anthropogenic climate change is increasing wave overtopping frequencies. Reliably predicting overtopping is important for minimising casualties and economic losses. Empirical formulas, such as EurOtop, can provide accurate
overtopping predictions based on physical model experiments. EurOtop requires nested regional process-based models to provide the input of nearshore coastal conditions. These models are computationally expensive, restricting their application in operational forecasting. EurOtop excludes wind speed and direction; significant variables influencing overtopping. This study explores the role of training random
forests, using annual overtopping observations, to predict overtopping at two study sites: Dawlish and Penzance (UK). We compare the performance of random forests with a EurOtop based model, while investigating the performance of including hourly measured wind speed and direction. The random forests estimated over topping and non-overtopping with approximately 80% and 95% accuracy, respectively. Including wind speed and direction enhanced the random forests’ predictive performance (R2 = > 0.70, F1 = > 0.80). Random forest outperformed the predictive precision and accuracy made by EurOtop at Dawlish (F1 = 0.81, F1 =
0.65, respectively) and Penzance (F1 = 0.86, F1 = 0.07, respectively). Random forests generated predictions rapidly (<4 s) and generalised well, demonstrating great potential for implementation in different overtopping hotspots worldwide, provided overtopping observation availability.