Flood models typically produce flood maps with constant return periods in space, without considering the spatial structure of flood events. At a large scale, this can lead to a misestimation of flood risk and losses caused by extreme events. A stochastic approach to global flood modelling allows the simulation of sets of flood events with realistic spatial structure that can overcome this problem, but until recently this has been limited by the availability of gauge data. Previous research shows that simulated discharge data from global hydrological models can be used to develop a stochastic flood model of the United States (Wing et al., 2020) and suggests that the same approach can potentially be used to build large scale stochastic flood models elsewhere but this has not so far been tested. This research therefore focuses on using discharge hindcasts from global hydrological models to drive stochastic flood models in different areas of the world. By comparing the outputs of these simulations to a gauge-based approach, we analyse how a model-based approach can simulate spatial dependency in large scale flood modelling outside of well-gauged territories such as the US. Based on data availability we selected different areas in Australia, South Africa, South America and Europe for the analysis. The results of this research show that the performance of a model-based approach in the different continents is promising and the errors are comparable to the results obtained in the United States by Wing et al. (2020). In the United States, with this magnitude of errors, the loss distribution obtained using the model-based approach is near identical to the one produced by the gauge-based method. This suggests that this method could be used in other regions to characterize losses. Using a network of synthetic gauges with data from global hydrological models would allow the development of a stochastic flood model with detailed spatial dependency, generating realistic event sets in data-scarce regions and loss exceedance curves where exposure and vulnerability data are available.

A series of recent flood events in Canada affecting areas around lakes and reservoirs have highlighted the need to explicitly represent such features in large scale flood models. Water level fluctuations in lakes are traditionally modelled using detailed hydrological models designed – as far as possible – to represent the actual physical processes that take place. This approach, while appropriate for local-scale studies in data-rich areas, is not applicable for large-scale flood modelling where data availability for model calibration and validation is often severely limited. This paper explores two methodologies, one statistical and one physically based, designed to approximately predict the increase in the water level of lakes in Quebec (Canada) using only limited morphological information about the lakes and the estimated discharge entering the water body during a flood event. Of the two methods, the statistical approach proved to be the most applicable to a large-scale modelling framework as it exhibited lower errors whilst being considerably easier to implement in a semi-automated modelling chain.