Ferrari et al. (2026) Real-time flood forecasting in lowland rivers exposed to levee breaches using GPU hydrodynamics

Identification

• Journal: Scientific Reports
• Year: 2026
• Date: 2026-09-23
• Authors: Alessia Ferrari, Giulia Passadore, Elena Crestani, Mattia Pivato, Renato Vacondio, Paolo Mignosa, Daniele Pietro Viero, Luca Carniello
• DOI: 10.1038/s41598-026-71401-9

Research Groups

• Department of Engineering and Architecture, University of Parma, Italy
• Department of Civil Environmental and Architectural Engineering, University of Padova, Italy

Short Summary

This study develops and validates an integrated, GPU-accelerated flood forecasting system capable of predicting real-time floods, including those caused by levee breaches, with lead times of up to two days in fast-reactive lowland river basins.

Objective

• To develop and validate an integrated, real-time flood forecasting system that can predict inundations in lowland areas, including those triggered by levee breaches, with lead times of up to two days.

Study Configuration

• Spatial Scale: Pilot site in Northern Italy (Lamone River basin). Hydrological domain: 510 square kilometers upstream catchment. Hydrodynamic domain: 60 kilometer reach of the Lamone River and 1 kilometer of the Marzeno River, including 700 square kilometers of potentially floodable areas. Topography described by a 4 meter Digital Terrain Model.
• Temporal Scale: Forecast lead times up to 48 hours for civil protection actions. Meteorological forecasts from ECMWF with a 10-day horizon (90 hours at hourly resolution). Hydrological model warm-up period of 15 to 60 days. Hydrodynamic model warm-up period of 2 days. Simulation of a real flood event on March 14–15, 2025. Hydrological model calibrated using data from 2008–2024.

Methodology and Data

• Models used:
  • Operational Platform: Delft-FEWS
  • Rainfall-runoff model: RHYME (River HYdrological ModEl)
  • Two-dimensional hydrodynamic model: PARFLOOD (solves 2D shallow water equations, GPU-accelerated, implemented in CUDA C++)
  • Levee breach model: Geometric approach prescribing time-dependent trapezoidal geometry (final width 100 meters within 3 hours).
• Data sources:
  • Meteorological forecasts: European Centre for Medium-Range Weather Forecasts (ECMWF).
  • Observed meteorological data: Hourly rainfall from 38 meteorological stations, daily average temperatures and daily cumulative potential evapotranspiration from the ERG5-Eraclito dataset.
  • Topography: 4 meter Digital Terrain Model (DTM) from a 2023 LiDAR survey.
  • Hydrological observations: Streamflow at the Reda gauging station for calibration.
  • Hydrodynamic validation data: Stage hydrographs from gauging stations (Reda, Pieve Cesato, Mezzano) and satellite-derived extent of the September 2024 flood event.

Main Results

• The integrated system accurately reproduced the March 14–15, 2025 flood event in the Lamone River basin.
• Hydrological model performance improved significantly as observed meteorological data replaced forecasts, with peak discharge error for the Lamone River decreasing from -91.2% to 0.0% and peak timing error from 10 hours to 0 hours.
• Hydrodynamic model simulations converged towards observed water levels, with Nash-Sutcliffe Efficiency (NSE) at Reda increasing from 13.3% to 83.1% as forecast inputs were replaced by observations.
• The complete forecasting chain provided 3-day flood scenarios within less than 3 hours of computational time on an NVIDIA A100 GPU, achieving speed-ups of approximately 96 for base simulations and 87 for breach scenarios.
• Hypothetical levee breach scenarios identified critical stretches and provided high-resolution inundation maps and hazard assessments (total depth indicator D). One breach scenario (B) resulted in high and very high hazard levels (D ≥ 1 meter) over approximately 20% of the flooded areas.
• Flood propagation was rapid, with an average of 30% of the total flooded area inundated within the first 3 hours and over 80% within 24 hours of a breach opening.
• The system provided substantial lead times for breach-induced flood scenarios: approximately 1 hour for 1-hour inundation, 12 hours for 12-hour inundation, and 23–24 hours for 24-hour inundation. The complete flood evolution could be simulated with lead times ranging from 55 to 63 hours.

Contributions

• Development of an integrated, real-time flood forecasting system that explicitly incorporates levee breach scenarios, representing a significant advancement over conventional forecasting chains.
• Demonstration of the capability to provide high-resolution (4 meter grid) flood inundation maps and hazard assessments (water depths, flow velocities, and arrival times) with lead times of up to two days, even in fast-reactive river basins.
• Leveraging GPU-accelerated hydrodynamic modeling (PARFLOOD) to achieve computational efficiency, making real-time application feasible without compromising spatial resolution.
• Enabling proactive flood risk mitigation by identifying at-risk areas and evaluating both potential and engineered levee failures in advance, supporting critical decision-making for civil protection.
• Implementation of the entire modeling workflow within a single operational platform (Delft-FEWS), which reduces procedural complexity during emergencies and enhances operational usability.

Funding

• L’Oréal Italia and UNESCO for the “For Women in Science Young Talents Italy” initiative (for Alessia Ferrari).
• CINECA award under the ISCRA initiative for high-performance computing resources and support.

Citation

@article{Ferrari2026Realtime,
  author = {Ferrari, Alessia and Passadore, Giulia and Crestani, Elena and Pivato, Mattia and Vacondio, Renato and Mignosa, Paolo and Viero, Daniele Pietro and Carniello, Luca},
  title = {Real-time flood forecasting in lowland rivers exposed to levee breaches using GPU hydrodynamics},
  journal = {Scientific Reports},
  year = {2026},
  doi = {10.1038/s41598-026-71401-9},
  url = {https://doi.org/10.1038/s41598-026-71401-9}
}