Development of a Bayesian networks-based early warning system for wave-induced flooding

Abstract

Coastal flooding prediction systems can be an efficient risk-reduction instrument. The goal of this study was to design, build, test, and implement a wave-induced flooding early warning system in urban areas fronted by sandy beaches. The system utilizes a novel approach that combines Bayesian Networks and numerical models (SWAN + XBeach) and was developed in two phases. In the development phase, firstly, the learning information was generated including the creation of oceanic conditions, modeling overtopping discharges, the haracterization of the associated im pacts (no, low, moderate and high) in pedestrians, urban components and buildings, and vehicles, and secondly, the Bayesian Networks were designed that surrogated the previously generated information. After their training, the conditional probability tables were created representing the foundation to make predictions in the operational phase. This methodology was validated for several historical events which hit the study area (Praia de Faro, Portugal), and the system correctly predicted the impact level of around 80% of the cases. Also, the predictive skills varied depending on the level, with the no and high impact levels overcoming the intermediate levels. In terms of efficiency, one simulation (deterministic) of coastal flooding for 72 h by running SWAN + XBeach operationally would take more than two days on a one-logical processor workstation, while the current approach can provide quasi-instantaneously predictions for that period, including probability distributions. Moreover, the two-working phase approach is very flexible enabling the inclusion of additional features such as social components representing a powerful tool for risk reduction in coastal communities.