GloFAS v5 calibration required historical meteorological forcings: precipitation, 2-metre average temperature, and reference evapotranspiration.

Model calibration had the purpose to achieve the highest accuracy in reproducing time series of in-situ historical observations.

Similarly to the previous versions, GloFAS v5 made use of in-situ historical observations of discharge. The use of in-situ historical observations of reservoirs inflow, outflow, and volume was a novel element introduced in GloFAS v5 calibration.

This page described the datasets used for GloFASv5 calibration.

Meteorological forcings

The meteorological forcings used for calibration were provided by C3S ERA5. OS LISFLOOD requires as input total precipitation, 2-metre average temperature, and reference evapotranspiration (water, soil, vegetation).

Meteorological forcings used in calibration covered the time interval 01/01/1975 to 31/12/2023, with daily temporal resolution. The very first 5 years (01/01/1975-31/12/1979) were used for model spin-up. The end date (31/12/2023) was defined for consistency with the discharge observation dataset available for calibration (details below).

C3S ERA5 data were upsampled from their native resolution (0.25 degrees) to the resolution of GloFASv5 (0.05 degrees or 3arcmin) using bilinear delaunay interpolation scheme, as implemented in the OS LISFLODD pre-processor pyg2p (it is here noted that GloFASv4 used bilinear interpolation).

In particular, GloFASv5 calibration and reanalysis dataset benefited of an amended version of ERA5 precipitation data. Spurious high-intensity rainfall values at a single grid point can negatively affect the quality of C3S ERA5 precipitation dataset (Hersbach et al., 2020¹). These were removed following the methodology explained in Mastrantonas et al. Met. App [submitted]², as implemented in the OS LISFLOOD utility rainbomb.

The outcomes of previous studies on the quality of C3S ERA5 dataset were taken into account: a notable example is the analysis of Zsoter et al. (2020)³ on the quality of the precipitation dataset for selected basins (more details are provided below).

Finally, similarly to the previous GloFAS versions, reference values of evapotranspiration were computed following the Penmann-Monteith method, which is implemented in the OS LISFLOOD pre-processor LISVAP.

Discharge observations

GloFASv5 calibration made use of daily discharge data from in-situ gauge stations. Data collection and preparation was largely performed in 2024. The interval of observed data was 01/01/1980-31/12/2023, thus allowing the tuning and verification of model parameters for a potential time span of 43 years.

In 2024, the CEMS Hydrological Data Collection Centre database counted over 15,500 stations with discharge data: detailed information about hydrological data collection, validation, and post-processing is provided in the CEMS Hydrological Data Collection Centre – Annual Report 2024 (Garcia Padilla, M., et al, 2026⁴).

River gauge stations were accurately geolocated on the 0.05 degrees resolution drainage network. Model river drainage network is derived from a digital elevation model, which is unavoidably affected by approximations. Consistency between real and model drained area for each gauge station is essential to allow reliable comparisons between observed and modelled discharge values. A detailed explanation of the relevance of this step is provided here. Observed discharge data time series were subject to automatic and manual quality checks: some time series were manually corrected to remove outliers or flat liners. 

GloFASv5 calibration stations were selected according to the following criteria:

1. Drainage area larger than 500 km². This threshold was defined for consistency with the resolution of the model set-up and of the meteorological forcings.
2. Discrepancy between real drainage area (as indicated by the data provider) and drainage area computed according to the GloFAS local drainage network lower than 10%.
3. A minimum number of daily discharge observations of 365 x 4 (i.e. 4 years equivalent) within the period 01/01/1980-31/12/2023.
4. Good data quality (where possible, outliers and flat-liners were removed).
5. Stations located close (less 500 km² inter-catchment drainage area) to another station and having the same data quality, but shorter time series were excluded.
6. Where stations were available on all the tributaries upstream of a confluence and then on the main river, stations located on tributaries were preferred to enable parameter calibration close to the runoff generation area.
7. Stations located immediately downstream of reservoirs and highly affected by reservoir operations were generally excluded: reservoir parameters were calibrated independently, according to the methodology explained in this page. Where possible, use of stations upstream of reservoirs providing information on reservoir inflow and/or stations further downstream showing a more natural behavior was preferred.
8. Stations affected by physical processes not included in the model set-up (e.g. river diversions) were excluded to avoid spurious values of calibrated parameters and degraded performances in forecast.
9. Stations covering areas and years with well-known issues in the meteorological forcings were excluded to avoid spurious values of calibrated parameters. A notable example is provided by stations in the Congo and Chad basins: stations with available discharge data older than 01/01/2000 were excluded to account for the local inaccuracy of the precipitation dataset used in calibration (following the analysis of Zsoter et al. 2020³).

Some of the stations that did not fulfill criteria 1, 2, or 3 were still used if they allowed to improve the spatial coverage in data scarce areas.  The use of these stations was validated via careful assessment of the calibration results. Specifically, the following exceptions were allowed:

• 488 stations for which information on the real drainage area was not provided or deemed not reliable.
• 31 stations had less than 500 km2 drainage area, of which only 1 smaller than 400km², 10 smaller than 450km². 
• 42 stations had less than 365x4 daily data, of which 36 had more than 365x3 daily data. 

The selection process led to the identification of 5379 calibration points, with an increase of over 160% compared to GloFASv4 (1995 stations):

• 646 calibration points were in Africa (+226 stations compared to GloFASv4);
• 1314 in Asia (+1064);
• 1314 in Central-North America (+697); 
• 1298 in South America (+862);
• 743 in Europe (+532);
• 64 in Oceania (+3).

The drained area of the 5379 stations entailed 51.6% of the quasi-global (-180,180,90,-60) domain. GloFASv4 calibration stations entailed 47.5% of the global domain: the large increase in stations number of GloFASv5 allowed model calibration for an increased number of head-catchments and inter-catchments, thus enabling parameter calibration closer to the areas where runoff generation occurs.

Figure 1 shows the spatial distribution of the calibration points and of the total extent of GloFASv5 area with available gauged data.

Figure 1 – Calibration stations: the yellow points are the calibration stations used in GloFAS v5, the black points are the calibration stations used for GloFAsv4. The area drained by the calibration stations is colored in orange or blue: in orange the area included for the first time in GloFAS v5, in blue the area included in both GloFAS v4 and GloFAS v5.  In grey the areas not covered by calibration points. The insets in North-West and South-East Africa highlight examples of extension of calibrated area in GloFASv5. 

The temporal extent of the observed time series varies across the global domain. Figure 2 shows the length () of the observation time series for each calibration point: number of daily measurements in equivalent number of years. Figure 3  shows the histogram distribution of the length of the time series: 57.6% of the calibration points had at least 20 years (equivalent number) of observations.

Figure 2 – GloFAS v5 calibration stations: length of the observation time series in years, spatial distribution. The points in pink were included to increase the spatial coverage of the calibration.

Figure 3 –  GloFAS v5 calibration stations: histogram plot of the length of the observation time series.

Reservoir observations

The use of reservoir information for model calibration is a novelty of GloFASv5 compared to the previous versions.

Specifically, observed time series of discharge inflow, discharge outflow, and reservoir volume were required to calibrate the parameters of the reservoir modelling routine. Further to the observed time series, a set of attributes were required to regionalize the parameters of the reservoir module.

Observed time series for the 3 variables could be collected for 360 reservoirs; all the reservoirs are listed in the GRanD (now Global Dam Watch) dataset. As shown in Figure 5, 123 reservoirs were in the USA, 59 in Mexico, 48 in Brazil, and 130 in Spain. The time interval was the same as the in-situ discharge observations, from 01/01/1980 to 31/12/2023. 

Figure 5 – Reservoir dataset: the size of the dots is proportional to the reservoir total storage.

Reservoir attributes were required for the 360 reservoirs listed above but also for all the 1486 reservoirs included in the semi-global modelling domain. Specifically, the attributes were classified in the following categories:

1. Climatic indices: average temperature, potential evapotranspiration, precipitation, snowfall; aridity and seasonality indices; high and low precipitation frequency and duration.
2. Reservoir/dam characteristics: reservoir storage capacity, dam height and length, reservoir depth, degree of regulation, degree of disruptivity.
3. Water use: hydropower, flood protection, fishing, irrigation, navigation management, recreation, and/or water supply.

Attributes (1) were taken from GloFASv5 meteorological forcings; (2) and (3) from GRanD.

More details on data collection, and reservoir data sets can be found in Casado-Rodríguez et al. (2026)⁵ ,  Casado-Rodríguez et al. (2025a, 2025b)⁵ ,  Steyaert et al. (2021).

References

¹ Hersbach H, Bell B, Berrisford P, et al. The ERA5 global reanalysis. Q J R Meteorol Soc. 2020;146:1999–2049. https://doi.org/10.1002/qj.3803

² Mastrantonas et al., Postprocessing suspiciously high localized precipitation in ERA5 for improved hydrological simulations, Met. App. [submitted] 

³ Zsoter et al., 2020, Trends in the GloFAS-ERA5 river discharge reanalysis. ECMWF Tech. Memo. 871

⁴ Garcia Padilla, M., Garcia Sanchez, R., Jiménez Molina, A., Márquez Arroyo, M., Serratosa Márquez, A. et al., CEMS Hydrological Data Collection Centre – Annual Report 2024, Publications Office of the European Union, Luxembourg, 2026, https://data.europa.eu/doi/10.2760/7145824 , JRC145381.

⁵ Casado-Rodríguez, J., Disperati, J., Grimaldi, S., and Salamon, P.: Benchmarking reservoir operation schemes for large-scale hydrological models, EGUsphere [preprint], https://doi.org/10.5194/egusphere-2026-904, 2026.

⁶ Casado-Rodríguez, J., Disperati, J., & Salamon, P. (2025a). ResOpsUS+CARS: Reservoir Operations US and CAtchment and Reservoir Static attributes (1.0) [Data set]. European Commission - Joint Research Centre. https://doi.org/10.5281/zenodo.15978041

⁷ Casado-Rodríguez, J., Disperati, J., & Salamon, P. (2025b). ResOpsBR+CARS: Reservoir Operations Brazil and CAtchment and Reservoir Static attributes (1.0) [Data set]. European Commission - Joint Research Centre. https://doi.org/10.5281/zenodo.16096623

⁸ Steyaert, J.,  Condon, L., Turner S., & Voisin., N. (2021). ResOpsUS (Version 2) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.6612040