Selection of stations

GloFAS v5 verification activities are based on the river gauging stations used in the GloFAS v5 calibration, which was carried out by the Joint Research Centre (JRC) of the European Commission (see GloFAS v5 calibration methodology and parameters). The verification is based on 5,378 calibration stations, located in 119 countries. Each station is located on the GloFAS v5 river network and has daily river discharge observations available to the GloFAS team. The same set of stations is used for the GloFAS v5 hydrological model performance assessment and for the hydrological model performance layer in the map viewer:

• Length of observation record: No minimum record length is applied in the verification. The observation records for all stations start between January 1979 and January 2020 and end between January 1983 and January 2026. 1,044 stations have observations in 2025. From the first to the last observation, the records cover 3.1 to 47 years (median 35.6 years). Taking gaps into account, the amount of data available per station ranges from 1.7 to 47 equivalent years (median 30.9 years). An equivalent year is 365.25 days with data, so a 20-year record with 50 % gaps counts as about 10 equivalent years.
• Short records: 21 stations have fewer than 4 equivalent years of observations. Even though the scores will not be robust for such short records, which users should be aware of, they are kept so that users can still get a first impression of the model performance at these stations.
• Lake, reservoir and human influence: Unlike the GloFAS v4 general hydrological model performance assessment, which used station networks without large lake or reservoir influence, no stations are excluded on this basis in GloFAS v5. All stations are verified, including those with lakes or reservoirs upstream. Scores at stations with a strong lake, reservoir or human influence should be interpreted with care, as these catchments are more difficult to model. A specific group of 228 stations with reservoirs built during the observation record is described in the last section of this page.

Verification period

The GloFAS v5 hydrological model performance assessment and the evaluation layer in the map viewer are based on the GloFAS v5 historical river discharge reanalysis, available from January 1979 to December 2025. The data can be accessed via EWDS https://ewds.climate.copernicus.eu/datasets/cems-glofas-historical?tab=download.

The verification uses the whole reanalysis period with all available river discharge observations, from 1 January 1979 at the earliest to 31 December 2025 at the latest. Both the general performance analysis and the GloFAS hydrological model performance layer in the map viewer (placeholder - GloFAS hydrological model performance web product) used this period.  The observations are the most recent ones available in the GloFAS hydrological observation database at the time of the verification (autumn 2026). For each station, the scores are computed over all days on which both the reanalysis and the observations are available. Gaps in the observations are not filled. There is no separate verification period. The model spin-up years covered an earlier period, from 1975 until the end of 1978, and are not part of the verification.

The reanalysis is the simulation from which the GloFAS flood thresholds (return periods) are derived. The verification therefore shows how well the simulation that underpins the forecast flood thresholds reproduces the observed river discharge.

Performance scores

GloFAS hydrological performance verification is done against the river discharge observations available to the GloFAS team. The hydrological model performance analysis is based on the modified Kling–Gupta efficiency (KGE′; Gupta et al., 2009; Kling et al., 2012; ideal value is 1):

\( KGE' = 1 - \sqrt{(r - 1)^2 + (\beta - 1)^2 + (\gamma - 1)^2}, \qquad \beta = \frac{\mu_s}{\mu_o}, \qquad \gamma = \frac{\sigma_s / \mu_s}{\sigma_o / \mu_o} \)

where s and o are the simulated and observed river discharge, μ is the mean and σ the standard deviation.

The three components of KGE′ are also shown:

• Pearson correlation (r) highlights timing errors through the strength of the linear relationship between the simulated and observed time series. It ranges from −1 to 1, with 1 as the ideal value.
• Bias ratio (β) represents the volume errors, as the ratio of the mean simulated to the mean observed discharge. It ranges from 0 to +∞, with 1 as the ideal value. Values above 1 mean the simulated volume is too large, values below 1 that it is too small.
• Variability ratio (γ) shows the errors in the variability of the flow, as the ratio of the coefficients of variation of the simulated and observed discharge. Using the coefficient of variation makes it independent of the bias. It ranges from 0 to +∞, with 1 as the ideal value.

In addition, three further scores are used in GloFAS v5:

• Nash–Sutcliffe efficiency (NSE; Nash and Sutcliffe, 1970) measures how much better the simulation is than simply using the mean observed discharge. It ranges from −∞ to 1, with 1 as the ideal value; 0 means the simulation is no better than the observed mean. It is very sensitive to errors in high flows.

\( NSE = 1 - \frac{\sum_{i=1}^{n}(s_i-o_i)^2}{\sum_{i=1}^{n}(o_i-\mu_o)^2} \)

• Jensen–Shannon divergence (JSD; Lin, 1991) measures how different the distributions of the simulated and observed discharge are, regardless of their timing. Following Ficchì et al. (2026), it compares the histograms of the logarithm of the discharge, so it gives weight to low flows as well as to medium and high flows. It ranges from 0 to 1, with 0 as the ideal value (identical distributions). Values below 0.1 generally indicate an adequate representation of low flows (Ficchì et al., 2026).
• Joint Divergence Kling–Gupta efficiency (JDKGE; Ficchì et al., 2026) extends KGE′ with the JSD as a fourth, equally weighted component, adding a penalty for differences in the flow distribution. It ranges from −∞ to 1, with 1 as the ideal value.  

\( JDKGE = 1 - \sqrt{(r - 1)^2 + (\beta - 1)^2 + (\gamma - 1)^2 + JSD^2} \)

Stations with reservoirs off in the calibration and on in the reanalysis

At 228 stations, the reservoirs upstream of the station were built after the period covered by the observations used for the calibration of the model. These stations are spread all over the globe: 61 in Asia, 47 in Africa, 39 in North America, 38 in South America, 24 in Europe, 11 in Central America and 8 in Oceania.

At these stations, the calibration and the reanalysis are set up differently:

• Calibration – reservoirs off: The observations available for the calibration describe the natural flow regime before the reservoirs existed. The model parameters are calibrated against these observations, so the reservoirs were switched off in the calibration to match them.
• Reanalysis – reservoirs on: The reanalysis is used to derive the flood thresholds (return periods) for the GloFAS forecasts. It must represent present-day, regulated conditions, so all reservoirs are always switched on.

As a result, the reanalysis at these stations is partly compared with observations of a flow regime that no longer exists, and its scores can be lower than those of the calibration run. This is expected. It does not mean that the reanalysis, or the flood thresholds derived from it, are less reliable for present-day conditions.

To show the effect, both simulations are compared at each of the 228 stations against the same observations over the same period: the observations and the period used in the calibration. The difference in skill is measured as ΔKGE = KGE′(reanalysis) − KGE′(calibration):

At most of these stations (178 of 228), both simulations perform about the same, which means the reservoirs have little effect at the station. At 45 stations, the reanalysis scores lower, at 31 of them by more than 0.10. The median KGE′ is 0.70 for the calibration run and 0.64 for the reanalysis. Where the reservoirs have a large effect, the reanalysis typically shows damped flood peaks and higher low flows than the observations from before the reservoirs were built. This lowers the variability ratio γ and, as a result, KGE′.

An interactive viewer shows each of these stations with its scores for both simulations and its hydrograph over the calibration period: https://dev.globalfloods.eu/reservoir-stations-viewer/.

The scores of these stations in the evaluation layer cover the full verification period, so they mix observations from before and after the reservoirs were built. They should be interpreted with the same care.

 References

• Ficchì, A., Bavera, D., Grimaldi, S., Moschini, F., Pistocchi, A., Russo, C., Salamon, P. and Toreti, A. (2026). Improving low and high flow simulations at once: An enhanced metric for hydrological model calibration. *EGUsphere* [preprint]. https://doi.org/10.5194/egusphere-2026-43
• Gupta, H. V., Kling, H., Yilmaz, K. K. and Martinez, G. F. (2009). Decomposition of the mean squared error and NSE performance criteria: Implications for improving hydrological modelling. *Journal of Hydrology*, 377(1–2), 80–91. https://doi.org/10.1016/j.jhydrol.2009.08.003
• Kling, H., Fuchs, M. and Paulin, M. (2012). Runoff conditions in the upper Danube basin under an ensemble of climate change scenarios. *Journal of Hydrology*, 424–425, 264–277. https://doi.org/10.1016/j.jhydrol.2012.01.011
• Lin, J. (1991). Divergence measures based on the Shannon entropy. *IEEE Transactions on Information Theory*, 37(1), 145–151. https://doi.org/10.1109/18.61115
• Nash, J. E. and Sutcliffe, J. V. (1970). River flow forecasting through conceptual models part I — A discussion of principles. *Journal of Hydrology*, 10(3), 282–290. https://doi.org/10.1016/0022-1694(70)90255-6