ECMWF's forecast model now routes rivers and floods globally in operations

Why it matters

Many weather and climate models handle rivers and floodplains poorly, even though floods and droughts are among the most damaging hazards. Rivers inside the same Earth system model give consistent land water data for forecasts, monitoring and machine-learning training. River flow also acts as a check on the whole land water cycle, as it exposed faults in the model.

What they did

The authors coupled the CaMa-Flood river model to ECMWF’s land surface scheme, ecLand, inside the Integrated Forecasting System. They refactored the code, added a basin-based hybrid MPI/OpenMP parallelisation and single-precision computing, and set up river initial states. The system went through ECMWF’s research-to-operations testing. They show uses in global water monitoring, flood case studies in Queensland, the Zambezi and Pakistan, wetland methane, and land water diagnostics.

Key findings

• As of January 2026, river analyses and 15-day ensemble river forecasts run operationally every 12 hours at 3 arcmin resolution with hourly time steps.
• Adding the river model did not change the atmosphere component: bit-identical behaviour was shown for the 1-way coupling configurations.
• Single precision for variables that don’t need strict water balance nearly doubled computational speed. The basin-based split limits the number of MPI nodes to 16 with bifurcations, or 30 if some are ignored.
• In the 2025 Queensland floods, ensemble forecasts issued 7, 5 and 3 days ahead showed a flood peak at Sellheim around 3 to 4 February. They captured it closely by three days ahead, with a slight underestimate of peak size and an earlier arrival than observed.
• In the Zambezi case, finer river resolution gave smaller flooded areas (2,525 km2 at 6 arcmin, 1,533 km2 at 3 arcmin, 912 km2 at 1 arcmin) and lower exposed population estimates.
• Spurious river flood peaks in West Africa revealed a dry bias interacting with soil moisture data assimilation, and a fix was included in Cycle 49r1.

Limitations

• The 2-way coupling, where floods feed back to land and atmosphere, is still experimental. The operational system routes runoff to rivers one way only.
• Ignoring river bifurcations to use more MPI nodes introduces approximations in discharge and water levels, and the largest connected basins cap how far the parallelisation can scale.
• The case studies are illustrative examples, and some diagnostics such as runoff at different resolutions and ocean freshwater coupling are still ongoing investigations. Forecast archives are for internal use, with public release planned for the next upgrade.

Glossary

• CaMa-Flood: A global hydrodynamic model that routes runoff through river networks and gives discharge, water level and flood extent.
• ecLand: ECMWF’s land surface model, the land component of the Integrated Forecasting System.
• 1-way coupling: A set-up where information flows in one direction only, here from land to rivers, with no feedback.
• Return period: The average number of years between floods of a given size, for example a 50 year flood.