Estimating extreme river discharges in Europe through a Bayesian Network
Abstract. Large-scale hydrological modelling of flood hazard requires adequate extreme discharge data. Models based on physics are applied alongside those utilizing only statistical analysis. The former requires enormous computation power, while the latter are most limited in accuracy and spatial coverage. In this paper we introduce an alternate, statistical approach based on Bayesian Networks (BN), a graphical model for dependent random variables. We use a non-parametric BN to describe the joint distribution of extreme discharges in European rivers and variables describing the geographical characteristics of their catchments. Data on annual maxima of daily discharges from more than 1800 river gauge stations were collected, together with information on terrain, land use and climate of catchments that drain to those locations. The (conditional) correlations between the variables are modelled through copulas, with the dependency structure defined in the network. The results show that using this method, mean annual maxima and return periods of discharges could be estimated with an accuracy similar to existing studies using physical models for Europe, and better than a comparable global statistical method. Performance of the model varies slightly between regions of Europe, but is consistent between different time periods, and is not affected by a split-sample validation. The BN was applied to a large domain covering all sizes of rivers in the continent, both for present and future climate, showing large variation in influence of climate change on river discharges, as well as large differences between emission scenarios. The method could be used to provide quick estimates of extreme discharges at any location for the purpose of obtaining input information for hydraulic modelling.