When high-resolution data is not enough to validate numerical models
In the field of flood modelling, there is a persistent belief that the mass collection of field data following a flood is the ultimate key to validating numerical models. A study conducted by researchers at the HECE laboratory at the University of Liège demonstrates that this view is incomplete. Drawing on the July 2021 floods in the Vesdre valley, they reveal that the quality, collection method and processing of data matter just as much as their quantity.
In July 2021, the Vesdre Valley was hit by floods of historic proportions. Towns such as Verviers and Pepinster suffered considerable damage. In the weeks that followed, the Wallonia Public Service (SPW) mobilised nearly 200 staff to record over 16,000 high-water marks on building facades, creating one of the most detailed post-flood datasets ever compiled worldwide. The team at the HECE at the University of Liège put to the test the idea that more field data would be sufficient to ensure reliable validation. Using their 2D hydrodynamic model WOLF, which they developed, the researchers simulated the event over nearly fifty kilometres with a resolution of up to two metres. "The model achieved a critical success index of 0.86 and an average error of 0.56 metres for maximum depths," explains Pratik Chakraborty, a PhD student at HECE. "These are remarkable results, but they mask a more nuanced reality."
The problem lies in the fact that in a numerical model, buildings are impermeable blocks, whereas maps derived from observations interpolate water levels continuously, including through buildings. Directly comparing these two types of map is like comparing apples and pears. In order to “reconcile” these data, the researchers propose four methods to make these maps comparable: adding buffer zones around buildings, filling gaps in the simulated map, removing built-up areas (the footprint of a building, which is a vertical projection that includes overhangs and projections) from the observed map, or propagating the free surface through built-up areas whilst respecting the terrain. “The results are quite striking,” continues the PhD student. “Depending on the method, validation scores can vary by more than 20% in certain areas. No method is free from bias.”
Resolution and topography: choices with far-reaching consequences
Nine simulations combining three grid resolutions (2, 5 and 10 metres) and three topographic resampling operators show that the transition from 2 m to 10 m significantly degrades the results, particularly in urban areas where narrow streets are no longer accurately represented. The researchers also recommend improving the geolocation of high-water marks, which are too often placed within building footprints, where the model does not accurately calculate water depth.
"We thought we had found the final piece of the puzzle," explains Benjamin Dewals, a hydraulic engineer at ULiège, "but when we put it in place, we discovered that there are still others missing. Having more data is necessary, but the way in which it is collected, documented and used is just as crucial." While the WOLF model developed by ULiège researchers performs excellently, the study reveals that the choice of comparison method— t observed and simulated data—strongly influences the results, on which decisions are based in practice. This work demonstrates that the availability of high-resolution data is not a magic bullet for validating flood models.
Scientific reference
- Chakraborty, P., Dessers, C., Archambeau, P., Pirotton, M., Erpicum, S. & Dewals, B. (2026). The mirage of the silver bullet: Exploring the limitations of high-resolution data in flood model validation. Journal of Hydrology, 665, 134578. doi.org/10.1016/j.jhydrol.2025.134578
Contacts
This research was conducted in part as part of the Interreg ResiRiver project (https://resiriver.nweurope.eu/), co-funded by the European Regional Development Fund (ERDF) and the Walloon Region.
