Hydrological modelling of the exceptional floods of July 2021 in eastern Belgium
In July 2021, exceptional floods struck Western Europe, causing more than 220 deaths and €32 billion worth of damage. In Belgium, the Vesdre and Amblève catchment areas were the worst affected. A study conducted by researchers at the University of Liège presents the first detailed hydrological analysis of these catchment areas, comparing several models to reconstruct the event. The results show that no single approach is sufficient to predict such an extreme flood, highlighting the need to combine multiple models and rainfall datasets.
T
riggered by the low-pressure system Bernd, the floods of July 2021 – the advance of which was slowed by a high-pressure system over Eastern Europe – caused almost stationary rainfall for several days. In Belgium and Germany, hundreds, if not thousands, of buildings were swept away, destroyed or partially damaged. In the Vesdre catchment area, cumulative rainfall averaged 175 mm, with a local peak of 293 mm over five days. The Amblève catchment area received lower but still significant amounts, averaging 120 mm. “At some rain gauges, the amounts recorded over two days were nearly double the volume associated with a 200-year return period,” explains Christophe Dessers, a physics engineer at the University of Liège. “The estimated peak flow at Eupen exceeded four times the 100-year return flow, confirming the absolutely exceptional nature of the event.”
Faced with such vastly different data between the two catchment areas, the researchers decided to adopt separate strategies. “In the Vesdre catchment, most of the measuring stations were damaged or washed away, making it impossible to observe the flood wave directly,” continues Pierre Archambeau, a hydraulic engineer at ULiège (HECE / School of Engineering). “However, a grid-based runoff model, developed in-house, enabled us to reconstruct the missing hydrographs.” This model was manually calibrated using water levels in the reservoirs of the Vesdre and Gileppe dams, supplemented by indirect validation via a 2D hydraulic model. The extent of the flooding we simulated corresponds well with field observations, with nearly 90 per cent of the flooded areas correctly predicted by the model.”
Conversely, in the Amblève catchment, almost all stations remained operational, providing an ideal setting for comparing several models: an improved version of the grid-based model (taking groundwater flow into account) and two 0D hydrological models (GR4H and VHM), used for rainfall-runoff modelling. These three models, calibrated against the 2021 event, successfully reproduced the flood with a high degree of accuracy (Nash-Sutcliffe coefficients (NSE) ranging from 0.89 to 0.99). “However, simply transferring the runoff model calibrated for the Vesdre to the Amblève proved unsuitable, due to the significantly lower rainfall volumes received by the latter catchment.”
The determining factor in data quality
The study highlights the major influence of rainfall data quality on model results. Four radar products were compared. The real-time product (RadQPE), provided by the IRM, considerably underestimates rainfall volumes, with a local peak of 39 per cent lower in the Vesdre catchment compared with the post-event products (Radflood21, Radclim). From a hydrological perspective, the corrected post-event products provide very similar estimates and improve the modelling results. To give an order of magnitude, this new rainfall version generates 50 per cent more discharge at Chaudfontaine than the rainfall estimated during the event. This finding is particularly relevant for operational flood management, where only real-time data is available.
A key finding of the study concerns the difficulty of transferring models between events of different intensities. When the models are calibrated using six historical floods (1991–2011), with peak flows ranging from 242 to 374 m³/s at Martinrive, they significantly underestimate the peaks of the 2021 flood (661 m³/s at the same point), with discrepancies of up to 45 per cent. Conversely, models calibrated against the extreme event of 2021 show varying trends for historical floods, but all tend to overestimate the summer flood of 1998. These results illustrate the complexity of the hydrological processes at play during extreme events and the limitations of calibration based on a single type of event.
This study constitutes the first in-depth hydrological analysis of the two Belgian catchment areas most affected by the historic floods of July 2021. It demonstrates that no single model, whether grid-based or 0D, consistently proves superior in all scenarios. The WOLFHydro platform, developed at the University of Liège, offers a modular framework that allows different models to be combined within a single simulation environment, which is essential for carrying out this type of analysis. “This study reinforces the importance of combining different rainfall products and hydrological models to construct envelope curves and cope with events that are statistically so rare,” concludes Christophe Dessers. At a time when such extreme events are set to recur in the context of climate change, this research provides an essential foundation for improving flood forecasting and the design of risk reduction measures.
Scientific reference
- Dessers, C., Archambeau, P., Erpicum, S., Dewals, B. & Pirotton, M. (2026). Hydrological modelling of the 2021 mega-flood in the east of Belgium. Journal of Hydrology, 667, 134901. https://doi.org/10.1016/j.jhydrol.2025.134901
