Abstract

Large-scale socioeconomic studies of the impacts of floods are difficult and costly for countries such as Canada and the United States due to the large number of rivers and size of watersheds. Such studies are however very important to analyze spatial patterns and temporal trends to inform large-scale flood risk management decisions and policies. In this paper, we present different flood occurrence and impact models based upon statistical and machine learning methods over 31,000 watersheds spread across Canada and the US. The models can be quickly calibrated and thereby easily run predictions over thousands of scenarios in a matter of minutes. We find it is possible to calibrate Generalized Additive Models that yield out-of-sample predictive capabilities that are comparable to tree-based methods, but are interpretable and allow for extrapolation, which is very important for projections over future climates. The modelling structure also allows for spatial dependence across basins due to climate and hydrological dynamics between a river and its tributaries. To illustrate the capabilities of the model, we present the geographical distribution of the modelled average annual number of people displaced due to flooding in Canada and the US, as well as various scenario analyses. The model can therefore be used by a broad range of end-users ranging from climate scientists to economists who seek to translate climate and socioeconomic scenarios into flood probabilities and impacts measured in terms of population displaced.