Reliable prediction of wave overtopping volumes using Bayesian neural networks

Artificial neural networks have been used successfully for generating predictions of wave overtopping volumes, which are required to appropriately design coastal structures as well as support associated flood risk analyses. It is particularly important to assess the uncertainty associated with such predictions, given the complexity of the modelling problem and the difficulty in obtaining accurate measurements of the large number of variables needed to estimate wave overtopping. The bootstrapping method previously used to estimate the uncertainty associated with ANN overtopping predictions does not, however, fully capture the total prediction variance. In this paper, Bayesian ANN methods are used to improve the reliability and robustness of wave overtopping predictions and to provide accurate estimates of the associated prediction uncertainty.