Comparison of machine learning algorithms in predicting wave overtopping discharges at vertical breakwaters

<p>Sea defences such as vertical breakwaters are critical marine infrastructures that safeguard communities and properties behind the structure from coastal flooding arising from wave-induced overtopping. In the context of future climate change, the frequency and magnitude of extreme wave events that threaten the functional performance of these defense lines and cause flooding is expected to increase. The capacity to reliably predict mean overtopping rates and individual overtopping volumes at these structures is therefore critical in deriving the tolerable limits of overtopping hazards. Common approaches for predicting overtopping rates at sea defences (such as vertical seawalls) have typically relied on physical, empirical and numerical methods. Notwithstanding the accuracy of these approaches, they are often complex and determining reliable predictions requires considerable expertise and time. Of late, the use of soft computing techniques such as Machine Learning (ML) algorithms have been employed to predict overtopping rates with comparable accuracy to the more common approaches. A significant advantage of ML methods are associated with their straightforward construction that can efficiently use existing databases, such as EurOtop 2018, in their training and testing to produce satisfactory results.</p><p>Research to date has, for the most part, focused on the application of  ML algorithms (such as decision tree and artificial neural network) to predict overtopping rates at sea defenses.  However,  the trade-offs of these methods (e.g., altered performance from missing values in the database) have not yet been investigated. Here, we investigate the application of two advanced ML methods, a Gradient-Boosting based Decision Tree (GBDT), and a feed forward based Artificial Neural Network (ANN) framework. Both algorithms were trained and tested using the CLASH database to predict mean overtopping rates at seawalls.  The CLASH database for this study comprises more than 1500 overtopping entries and a train-test split of 70% and 30%, respectively, was applied. Hyperparameter tuning was performed on the GBDT algorithm to refine the outputs. A provision was included in the ANN algorithm for it to detect, check and impute missing values as ANN does not implement when there is missing values and also imputing for a large number of missing values may negatively impact the performance of GBDT models.  </p><p>Results of this study revealed that the GBDT algorithm, overall, performed marginally better the ANN algorithm. The root-mean-squared errors (RMSE) for the GBDT and ANN models were 0.50 and 0.52, respectively. The Pearson R values for the GBDT and ANN algorithms were 0.92 and 0.90, respectively, confirming a strong correlation between the predicted and measured overtopping discharges for methods. Additionally, by permutation importance analysis, the GBDT algorithm was shown to be capable of identifying influential overtopping parameters, with significant wave height and crest-freeboard being shown to be significant in this study.  </p><p><strong>Keywords:</strong> Machine Learning, Wave Overtopping, Climate Resilience. Climate Change</p><p> </p>