Duy Anh Alexandre, Chiranjib Chaudhuri, and Jasmin Gill-Fortin
Research and Development Group, Geosapiens Inc., Quebec City, Canada
Context
The estimation of extreme discharge levels corresponding to very low probabilities of occurrence in river catchments is crucial for flood risk management. These values are used in hydraulic models to simulate water flow and calculate their equivalent flood depths. River discharge often reach high levels simultaneously at multiple nearby locations, because of the connectivity of the river networks and the large-scale underlying climatological drivers like precipitation. At the same time, a flood event never affects all locations with the same flood magnitude: the return periods of discharges seen during a same flood event are essentially different. Therefore, considering the spatial dependence of fluvial flooding is very important to avoid overestimating the associated losses when assessing flood risk. Adopting an event-based approach where spatial dependence is explicitly modelled allows to calculate financial losses which are deemed more reliable.
Data and methodology
At Geosapiens, we have developed a statistical model which captures the spatial pattern of observed flood events, based on the theory of multivariate statistics. We relied on station daily discharge data to conduct our analysis. Data were collected from multiple sources (GRDC, USGS Survey, Environment Canada and Quebec Atlas Hydroclimatique), then quality-checked before analysis.
The model contains a machine-learning component that harnesses information from 130 different catchment-specific attributes (topography, land cover, soil types, climate) to predict spatial correlation between extreme discharge at ungauged catchments (where station measurements are lacking). In this way, spatially coherent flood events with a complete spatial coverage can be synthesized. The spatial extent of our model covers all North America. We have divided the continent into 14 hydrologically similar regions with each then further subdivided into river networks and catchments with an average area of 10 km2. This delineation into subregions and the locations of available gauge stations are shown in figure 1.
Figure 1: The 14 hydrological regions and locations of the gauge stations
Highlighted results
Our model underwent extensive testing, and all components were validated by comparing the observed and simulated flood events. Spatial dependence between nearby catchments is well-captured, as the model is able to simulate clusters of extreme discharge with magnitude and spatial extent similar to the observed events. . Figure 2 depicts the different flood patterns that our model can reproduce. The prediction of flood dependence at ungauged catchments gives remarkable results, as cross-validation of the machine-learning model yields a R2 score of 0.97.
Figure 2: Four simulated events in British Columbia with catchments experiencing extreme discharge shown in red
Benefits
The Geosapiens Fluvial Flood Event Catalogue offers significant benefits to insurance and reinsurance companies by providing a comprehensive tool for flood risk assessment and management. By incorporating spatial correlations in extreme discharge events, the catalogue enables a more accurate representation of flood geographic variability, aiding in the estimation of potential impacts across regions. The synthetic event catalogue, spanning thousands of years, allows for robust Probable Maximum Loss (PML) estimation, crucial for setting premiums and reserves. Additionally, the event-based approach and simulation capabilities enhance loss estimate reliability, facilitating informed risk management strategies and financial planning, including the calculation of insurance and reinsurance layers, attachment points, limits, and excess-of-loss.
To learn more, please consult this document: egusphere-2024-442.pdf (copernicus.org)