A Bayesian network approach to modelling rip-current drownings and shore-break wave injuries

Abstract. A Bayesian network (BN) approach is used to model and predict shore-break-related injuries rip-current drowning incidents based on detailed environmental conditions (wave, tide, weather, beach morphology) the high-energy Gironde coast, southwest France. Six years (2011–2017) of boreal summer (15 June–15 September) surf zone (SZIs) were analysed, comprising 442 (fatal non-fatal) drownings caused by rip currents 715 shore-break waves. Environmental at time SZIs train two separate networks (BNs), one for other wave injuries. Each BN included so-called “hidden” exposure hazard variables, which are not observed yet interact with several (environmental) in turn limit number edges. Both BNs tested varying complexity using K-fold cross-validation multiple performance metrics. Results show a poor fair predictive ability models according different Shore-break-related appear more predictable than selected predictors within BN, as systematically performed better BN. Sensitivity scenario analyses address influence data variables their interactions exposure, resulting life risk. Most our findings line earlier SZI physical hazard-based work; that is, warm sunny days light winds; long-period waves, specifically high tide steep profiles; near low near-shore-normal incidence strongly alongshore non-uniform morphology. The also provided fresh insight, showing risk approximately equally distributed between (variance reduction Vr=14.4 %) (Vr=17.4 %), while water user waves much important (Vr=23.5 (Vr=10.9 %). Large found decrease beachgoer hazard, this currents. Rapid change elevation during large tidal range was result incidents. We advocate such BNs, providing understanding risk, can be developed improve public safety awareness campaigns, parallel development skilful anticipate high-life-risk days.