-
Created by
Michela Giusti on Jul 11, 2025
15 minute read
Official reference number service contract: C3S2_410_Lot2
History of Modifications
Acronyms
1. Scope of the document
This document presents the setup and performance of an ensemble of hydrological models used for the hydrological seasonal forecasts. The performance for both the historical simulation and the forecast skills are presented for each of the contributing hydrological models.
2. Hydrological models
2.1. The E-HYPE model
The E-HYPE model (Donnelly et al., 2016) is based on the continuous process-based hydrological model HYPE (HYdrological Predictions for the Environment; Lindström et al. 2010), which simulates components of the catchment water cycle and water quality. The model is semi-distributed, in which a river basin may be subdivided into multiple sub-catchments, which are further subdivided into hydrological response units (HRUs) based on soil type and land use classes. A conceptual image of E-HYPE is presented in Figure 1. It has conceptual routines for the major hydrological surface and subsurface processes (e.g., snow/ice accumulation and melting, evapotranspiration, surface and macropore flow, soil moisture, runoff generation, groundwater fluctuation, routing through rivers and lakes), land management , (irrigation, abstractions), and nutrient turnover (diffuse and point source releases, solid-matter and dissolved sub-surface pools, plant uptake, riverine transport) that are controlled by a number of parameters that are often linked to physiography to account for spatial variability and estimated through calibration.
The model parameters can generally be categorized into HRU and general parameters. The HRU parameters are either soil or land use type dependent. Unique parameter values are estimated for each soil or land use type and are applied throughout the model domain. The general parameters, such as the routing parameters, are sub-catchment scale parameters and may be assigned constant values throughout the model domain or estimated separately for different parameter regions. They can also be estimated using a regionalization approach as functions of catchment.
The model setup for the pan-European domain is referred to as E-HYPE (Donnelly et al. 2016), and is contributing with five different setups: four perturbed hydrological parameter calibrations at a catchment scale, and a 5-km gridded version. All E-HYPE models are driven by daily mean temperature and precipitation.
Figure 1: A conceptual diagram of the E-HYPE model.
2.2. E-HYPEcatch
E-HYPEcatch is the original E-HYPE model with irregular polygons that delineate the sub-catchment. The version used here covers an area of 8.8 million km2 and is subdivided into 35,408 sub-catchments with an average size of 248 km2. A range of physiographic data sources were used in setting up the model (Donnelly et al. 2016). Note that some catchment with poor spatial cover from the reference data are removed, leaving 29,986 catchments in the final data set. The "id"-values in each NetCDF file provides the "subid"-identifier in the shapefiles. The shape files for the E-HYPEcatch models can be downloaded here: http://doi.org/10.5281/zenodo.581451.
In order to account for the uncertainty in the estimated model parameters, a multi-model ensemble was generated next to the standard setup of E-HYPEcatch (Hundecha et al. 2016). This was performed by first identifying the most sensitive parameters for the model, calibrated using the procedure described in Hundecha et al. (2020). The dominant soil and land use parameters corresponding to the eight catchment groups, along with the general parameters were sampled using the SAFE sensitivity and uncertainty analysis toolbox (Pianosi et al. 2015) to generate 10,000 randomly distributed samples with the Latin hypercube strategy. The model performance was assessed in terms of NSE.
Parameter samples whose model performance was above a subjectively defined threshold (top 1% of the 10,000 samples) were accepted based on a GLUE-type parameter identification approach (Beven and Binley, 1992). In the next step, the top 1% samples were combined across the eight groups to provide narrower identifiable parameter ranges and 15,000 sets were sampled within these ranges with the same strategy as before. These new global sets were then evaluated against all calibration stations, and parameter sets were chosen as in the previous step, providing a number of acceptable global parameter sets. Finally, out of the ten best performing sets, three representative setups were selected as the multi-model ensemble, herein referred to as E-HYPEcatch01, E-HYPEcatch06 and E-HYPEcatch09.
2.3. E-HYPEgrid
E-HYPEgrid has been setup using the 5 km drainage network of EFAS. In order to set up E-HYPEgrid, the following data was obtained from EFAS: flow direction, mean elevation, fraction of water bodies, lake-mask, and stream channel length within each grid cell as well as the total upstream area draining to each grid cell. Each grid cell is treated as a sub-catchment, as in the standard E-HYPEcatch model. Hydrological response units are defined within each grid cell as a combination of land use and soil type, using the same classification as E-HYPEcatch.
In order to exploit HYPE’s capability of modeling lake dynamics with detailed natural and regulation lake processes, large lakes are allowed to extend over several grid cells. To accommodate for this, lake areas were added as an additional computation grid point during the E-HYPEgrid setup and calibration phase using the European and global lake and reservoir databases. Large lakes in E-HYPEgrid cover one or several grid cells completely, and surrounding grid cells partially, while smaller lakes and elongated river dams cover several grid cells partially. Since model results are calculated only for the outlet point of the lake, this point is representing the whole lake and any partial land computation cell will be replaced by the lake cell, following the reasoning that the water flows through the lake unit. To deliver E-HYPEgrid results on the original EFAS grid, the lake values have replaced the land values when the lake is only partial, following the reasoning that the water flows through the lake unit.
2.4. The VIC-WUR model
The Variable Infiltration Capacity (VIC) model (Liang et al., 1994) is a semi-distributed macroscale hydrological model and forms the fundamentals of the VIC-WUR model. The VIC-WUR (version 2.0.0) model is an extension of the VIC-5 model and includes additional modules to account for anthropogenic influences (Hamman et al., 2018, Droppers et al., 2020). A conceptual overview of the model is presented in Figure 2. The model has been setup in natural flow conditions and water balance mode with three thermal nodes. Lakes and frozen soils are not simulated. A single elevation band is used, because heterogeneity in topography is preserved by the high spatial resolution of the model. Sub-daily timesteps of 6 hours are used to solve the energy and water balance for each cell individually, i.e. there is no water or energy exchange between cells. The snow module also uses a 6 hourly timestep and calculates the snow variables in the surface and vegetation layers. Soil parameters and vegetation characteristics (Lawrence and Chase, 2007) are used from respectively 0.5- and 0.05-degrees gridded data and remapped to the model domain using ECMF nearest neighbor algorithm. The parameter file includes 3 vertical soil layers and 16 vegetation classes. There is surface runoff whenever precipitation occurs under saturated topsoil conditions, while precipitation under unsaturated conditions infiltrates into the soil layers. The upper two soil layers interact with the canopy layer and provide water for transpiration. The third and deepest soil layer is assumed to be outside the root zone and accounts for subsurface runoff (baseflow). VIC-WUR uses an internal routing module (Lohmann et al, 1996, 1998) to simulate discharge based on the unit hydrographs of both stream inflow and outflow. The unit hydrographs are computed from the EFAS flow direction, channel length, mean elevation (slope) and the SCS dimensionless unit hydrograph using a flow velocity of 1.5 m/s and diffusion value of 2000 m/s2. The calibrated model parameters have been taken from the global calibration performed by Nijssen et al. (2001). The infiltration parameter, thickness of the second soil layer and soil hydraulic parameters has been calibrated by each climate zone to match the total annual flow volume and shape of the mean monthly hydrograph. The parameters are then transferred to the global basins based on the representing climate zone.
Inputs to the VIC-WUR model are 6 hourly temporal resolution of temperature, precipitation, solar radiation, thermal radiation, wind speed, surface pressure and humidity. Because data at this high temporal resolution are not available from the observational data set EFAS-Meteo, the daily resolution data were disaggregated using different methods depending on the use case.
Figure 2: Schematic presentation of the water and energy fluxes in the VIC-WUR model. Source: https://vicwur.readthedocs.io/
3. Evaluation of river discharge
The E-HYPE models (catchment and grid versions separately) and the VIC-WUR model were evaluated for their performance of simulating river discharge when forced with the EFAS-Meteo data set (Ntegeka, 2013), using a set of stations that are independent of the stations used for calibration of the models.
3.1. Method for performance evaluation
To assess the model performance at each of the discharge gauging stations used for model evaluation, we use three model efficiency measures: the Nash-Sutcliffe Efficiency (NSE; Nash and Sutcliffe, 1970), Kling-Gupta Efficiency (KGE; Gupta et al., 2009), and percentage model bias (Pbias). In addition, the model performance in terms of reproducing the extremes of daily flow is assessed by evaluating how well the model captures the 5th and 95th percentiles of the long-term daily flow. All models are driven by the EFAS-Meteo forcing data over the period 1991-2016, and are evaluated at 323 selected validation stations across Europe, which are independent for all calibrated models.
3.2. Evaluation results
Figure 3 presents an overview of the KGE, NSE and Pbias results for each of the hydrological models. All models perform generally well for continental scale models, with median scores above or near 0.5 in KGE, and NSE generally above 0.5. The exception is VIC-WUR with slightly lower scores, and higher percentage bias.
Figure 3: Boxplots of the scores across all stations of KGE (top), NSE (middle) and percentage bias (bottom) for each of the contributing hydrological models.
Figures 4, 5, and 6 present maps of the performance for each validation station and each hydrological model for KGE, NSE and Pbias, respectively. The E-HYPE models have strong similarities in the spatial pattern of both KGE ans NSE values, and differences between the models are visible in rather few locations across the continent. The E-HYPEcatch00 and E-HYPEgrid models resemble each other because of the similar calibration, and the other E-HYPEcatch(01, 06, and 09) models are alike due to the separate calibration strategy. The main visible difference between these two sets of E-HYPE models are in the percentage bias over central Europe, where the E-HYPEcatch(01, 06, and 09) have lower bias (Figure 6). VIC-WUR has similar spatial pattern of KGE and NSE as the E-HYPE models, but with generally lower values, and especially a higher frequency of the very low (red) values (see e.g. NSE in Figure 5), and higher percentage bias in central Europe (Figure 6).
Figure 4: Evaluation scores of KGE for all validation stations and each hydrological model.
Figure 5: Evaluation scores of NSE for all validation stations and each hydrological model.
Figure 6: Evaluation scores of Pbias (%) for all validation stations and each hydrological model.
4. Evaluation of the forecast skill
4.1. Method for skill evaluation
A perfect forecast approach is used for evaluating the hydrological seasonal reforecasts, i.e., they are evaluated with respect to their performance against a reference historical simulation (Pechlivanidis et al., 2020). This approach is not dependent on observed records, and allows analysis of the forecasting skill over all modelled locations, although it is an evaluation of the theoretical accuracy, and not the actual accuracy.
The evaluation of the reforecasts is performed on monthly means for the 1993–2016 period, and for each start month and lead months of the forecasts. The first 25 members of each system ECMWF-s51 and CMCC-s35 are used.
Four different performance measures are calculated for each sub-basin or grid point, initialization month and lead time: The Fair Continuous Rank Probability Skill Score (fairCRPSS), the tercile fair Ranked Probability Skill Score (tercile fairRPSS), fair Brier Skill Score for lower and upper terciles (fairBSSbn and fairBSSan respectively). All skill scores range between -infinity and 1, with 1 being the optimum.
All metrics are calculated using two R-packages: easyVerification (https://cran.r-project.org/web/packages/easyVerification/index.html), and SpecsVerification (https://cran.r-project.org/web/packages/SpecsVerification/index.html), using the functions FairRpss, FairCrpss, and FairBrierSs(ens, ens.ref, obs, tau = NA).
4.2. Skill evaluation results
A few representative results are presented to highlight the general performance across different skill scores, variables and hydrological models. For reasonas of readability, the main conclusions are written in bold text within the caption of the figures. Line plots present the median skill score across the complete domain, and the maps present the skill score for each separate calendar month as they are written to the NetCDF files.
4.2.1. Skill scores
Figure 7: The panels show the different skill scores for each initialization month, each with the six lead months over the year, for the different skill scores for river discharge from E-HYPEcatch00 for the two meteorological forecast systems ECMWF-s51 and CMCC-s35. The general time-dependence is similar for all scores, and across meteorological systems, with some variations in magnitude.
4.2.2. Variables
Figure 8: The panels show the skills for each initialization month, each with the six lead months over the year, for the CMCC-s35 driven E-HYPEcatch00 forecasts with the CRPSS score. River discharge and runoff naturally follow a similar annual cycle as they are strongly related, whereas soil moisture has a different annual cycle of the skills, with lower values in the drier summer months. The snow water equivalent has zero skill in the summer months when few catchments have any snow.
4.2.3. Models
Figure 9: The panels show the skills for each initialization month each with the six lead months over the year for the river discharge from the different hydrological models for the ECMWF-s51 forecast system. The hydrological models have generally similar skills with only minor differences in the magnitudes. The E-HYPE based models have a generally similar performance across all variables, but the VIC-WUR model shows clearly better performance for the first few lead months compared to the E-HYPE models. VIC-WUR shows better skill for the winter months, whereas the E-HYPE models are better for the spring months.
Figure 10: The panels show the skills for each start month each with the six lead months over the year for the runoff from the different hydrological models for the ECMWF-s51 forecast system. The hydrological models have generally similar skills with only minor differences in the magnitudes. The E-HYPE based models have a generally similar performance across all variables, but the VIC-WUR model shows clearly better performance for the first few lead months compared to the E-HYPE models. VIC-WUR shows better skill for the winter months, whereas the E-HYPE models are better for the spring months.
Figure 11: The panels show the skills for each start month each with the six lead months over the year for the soil moisture from the different hydrological models for the ECMWF-s51 forecast system. The hydrological models have generally similar skills with only minor differences in the magnitudes. The E-HYPE based models have a generally similar performance across all variables, but the VIC-WUR model shows clearly better performance for the first few lead months compared to the E-HYPE models.
Figure 12: The panels show the skills for each start month each with the six lead months over the year for the snow water equivalent from the different hydrological models for the ECMWF-s51 forecast system. The hydrological models have generally similar skills with only minor differences in the magnitudes. The E-HYPE based models have a generally similar performance across all variables, but the VIC-WUR model shows obvious problems with the forecasts of snow water equivalent with overall low skill scores. The differences in VIC-WUR are likely related to differences in the temperature forcing of the model where VIC-WUR uses disaggregated 6-hourly values based on daily minimum and maximum temperature for the historical simulation, while the sub-daily data are available from the meteorological forecasts. Secondary effects can also be related to similar differences in the radiative forcing.
Figure 13: Maps of the CRPS skill score for river discharge from E-HYPEgrid and VIC-WUR for the first three months of the forecast from ECMWF-s51 system with the March start month. The VIC-WUR model shows higher skills in the very north of Europe, and overall similar values as E-HYPEgrid (all E-HYPE models show similar skill scores) in continental Europe.
Figure 14: Maps of the CRPS skill score for snow water equivalent from E-HYPEgrid and VIC-WUR for the first three months of the forecast from CMCC-s35 system with the March start month. E-HYPEgrid (all E-HYPE models show similar skill scores) shows higher skill scores, also covering a larger part of the domain, than VIC-WUR.
5. References
Beven K, Binley A (1992) The Future of Distributed Models: Model Calibration and Uncertainty Prediction. Hydrol. Process. 6 (3): 279–98.
Donnelly, C., Andersson, J. C. M., & Arheimer, B. (2016) Using flow signatures and catchment similarities to evaluate the E-HYPE multi-basin model across Europe. Hydrol. Sci. J. 61(2), 255-273. doi:10.1080/02626667.2015.1027710
Droppers, B., W. H. P. Franssen, M. T. H. van Vliet, B. Nijssen, and F. Ludwig. 2020. Simulating human impacts on global water resources using VIC-5. Geosci. Model Dev. 13:5029-5052. https://doi.org/10.5194/gmd-13-5029-2020
Gupta, H.V., et al., 2009. Decomposition of the mean squared error and NSE performance criteria: Implications for improving hydrological modelling. Journal of Hydrology, 377, 80–91. doi:10.1016/j.jhydrol.2009.08.003
Hamman, J. J., Nijssen, B., Bohn, T. J., Gergel, D. R., and Mao, Y. (2018) The Variable Infiltration Capacity model version 5 (VIC-5): infrastructure improvements for new applications and reproducibility, Geosci. Model Dev., 11, 3481-3496, https://doi.org/10.5194/gmd-11-3481-2018.
Hundecha, Y., Arheimer, B., Donnelly, C., and Pechlivanidis, I. (2016) A regional parameter estimation scheme for a pan-European multi-basin model. J. Hydrol.-Regional Studies. 6, 90-111, doi: 0.1016/j.ejrh.2016.04.002
Hundecha, Y., Arheimer, B., Berg, P., Capell, R., Musuuza, J., Pechlivanidis, I., & Photiadou, C. (2020). Effect of model calibration strategy on climate projections of hydrological indicators at a continental scale. Climatic Change, 163(3), 1287-1306.
Lawrence, P. J., & Chase, T. N. (2007). Representing a new MODIS consistent land surface in the Community Land Model (CLM 3.0). Journal of Geophysical Research: Biogeosciences, 112(G1).
Liang, X., D. P. Lettenmaier, E. F. Wood, and S. J. Burges (1994), A simple hydrologically based model of land surface water and energy fluxes for general circulation models, J. Geophys. Res., 99(D7), 14415–14428, doi:10.1029/94JD00483.
Lindstöm G, Pers C, Rosberg J, Strömqvist J, Arheimer B (2010) Development and testing of the HYPE (Hydrological Predictions for the Environment) water quality model for different spatial scales. Hydrol Res 41(3-4):295-319. doi:10.2166/nh.2010.007
Lohmann, D., R. Nolte-Holube and E. Raschke (1996). "A large-scale horizontal routing model to be coupled to land surface parametrization schemes." Tellus A 48(5): 708-721.
Lohmann, D., E. Raschke, B. Nijssen and D. P. Lettenmaier (1998). "Regional scale hydrology: I. Formulation of the VIC-2L model coupled to a routing model." Hydrological Sciences Journal 43(1): 131-141.
Nash, J.E. and Sutcliffe, J.V., 1970. River flow forecasting through conceptual models. Part I. A discussion of principles. Journal of Hydrology, 10, 282–290. doi:10.1016/0022-1694(70)90255-6
Nijssen, Bart, et al. (2001), 'Predicting the Discharge of Global Rivers', Journal of Climate, 14 (15), 3307-23.
Ntegeka V, Salamon P, Gomes G et al (2013) EFAS-Meteo: A European daily high-resolution gridded meteorological data set for 1990–2011. Report EUR, 26408.
Pechlivanidis, I. G., Crochemore, L., Rosberg, J., & Bosshard, T. (2020). What are the key drivers controlling the quality of seasonal streamflow forecasts?. Water Resources Research, 56(6), e2019WR026987.
Pianosi F, Fanny S, Wagener T (2015) A Matlab Toolbox for Global Sensitivity Analysis. Environ Modell Softw 70: 80–85.
Related articles
-
TCWV HOAPS: Algorithm Theoretical Basis Document (ATBD) (Copernicus Knowledge Base)
-
Temperature statistics for Europe derived from climate projections documentation (Copernicus Knowledge Base)
-
