Hydrological Projections

Global Hydrological Models

Four well-established global hydrological models (GHMs) participated in CO-MICC:

  • The Community Water Model (CWatM) version 1.04, developed at the International Institute for Applied Systems Analysis (IIASA), in Austria
  • The Lund-Potsdam-Jena managed Land (LPJmL) model version 5.0, developed at the Potsdam Institute for Climate Impact Research (PIK), in Germany
  • The ORganising Carbon and Hydrology In Dynamic EcosystEms (ORCHIDEE) land surface model version ORCHIDEE_gmd-2018-57, developed at the Dynamic Meteorology Laboratory (LMD) of the Pierre Simon Laplace Institute (IPSL), in France
  • The Water Global Assessment and Prognosis model version 2.2d (WaterGAP2.2d), developed at the Institute of Physical Geography (IPG) of the Goethe University Frankfurt (GUF), in Germany

These models simulate the terrestrial water cycle with its water flows, water storage compartments and human water use sectors (except ORCHIDEE, which neglects human impacts on water flows and storages), on the global scale. They can assess past and future impacts of climate change on freshwater systems. Despite their common applications, these GHMs were developed within different modelling frameworks and have distinct foci, leading to fundamental differences in model structure and model parameterization.

LPJmL and ORCHIDEE are used by both the global hydrological and the global vegetation communities, as not only can they be classified as GHMs, but they also fall into the category of Dynamic Global Vegetation Model (DGVM). Unlike the other GHMs, LPJmL and ORCHIDEE have an advanced representation of vegetation dynamics by means of a dynamic vegetation module, which simulates the evolution of vegetation with climate. In addition to vegetation composition and distribution, they also simulate the terrestrial carbon and water cycles. LPJmL considers both natural and agricultural ecosystems, whereas ORCHIDEE is restricted to natural ecosystems.

CWatM and WaterGAP2.2d are global water resources and water use models. They are comparable in terms of structure, parameterization and the main motivations behind their development. They were designed with the purpose of quantitatively assessing water availability, water demand and environmental needs. A major focus in these models is the representation of human water use in multiple sectors (irrigation, domestic, industrial and livestock) as water abstractions from surface water and groundwater, and reservoir operations. They consider multiple land cover classes, but have no dynamic vegetation module.

All four models were part of the model evaluation phase of the project. However, only three out of four, namely CWatM, LPJmL and WaterGAP2.2d, were used to generate project-specific multi-model ensembles (MMEs) of future impacts of climate change on freshwater systems. These GHMs were run over a reference period (1979-2010) and over 30-year time periods centered from 2020 to 2085 in 5-year time steps. The inclusion of ORCHIDEE as part of the CO-MICC impact models is expected to happen at a later stage.

CWatM, LPJmL and WaterGAP2.2d run at daily temporal resolution over a global grid divided into 0.5° by 0.5° cells (55 km by 55 km at the equator and ~3000 km2 grid cell). The grid covers the global continental area except for Antarctica. Outputs are generated as daily, monthly or annual spatially explicit time series. Each GHM was run according to two different modelling schemes regarding the role of increasing atmospheric CO2 concentrations through the adaptive response of vegetation: one scheme including the vegetation response to this changed condition and the other one assuming an absence of this response. These two modelling schemes made it possible to include the impacts of a non-climatic driver, namely the rising CO2 concentrations, on freshwater systems in the MMEs.

As a DGVM, LPJmL includes a dynamic vegetation module and is thus able to explicitly simulate the active response of vegetation, and its implications on land hydrology. CWatM and WaterGAP2.2d, on the other hand, do not include a dynamic vegetation module and are thus unable to explicitly simulate the active response of vegetation. Nevertheless, as a workaround, these “offline” impact models can mimic the hydrological implications of vegetation response to rising CO2 concentrations (as represented by the Representative Concentration Pathways) by adapting the equation used to calculate potential evapotranspiration (PET).

Moreover, the models were run under two different modes:

  • the anthropogenic mode (standard mode), which includes both climate- and human-induced impacts, and
  • the naturalized mode, which neglects human water abstractions as well as the construction and operation of artificial reservoirs.

Most of the hazard indicators in the data portal were generated from the outputs of the anthropogenic model runs. On the other hand, naturalized runs were required to generate the indicators related to blue water production and naturalized streamflow (see Hazard Indicators).

Model evaluation

The performance of the GHMs was evaluated prior to their usage to simulate future freshwater availability conditions. As is common practice in the hydrological modelling community, the performance of the models was assessed by testing their capacity to reproduce past conditions. The assumption behind is that, if the models are capable to reproduce past conditions with an acceptable degree of accuracy, then they can be trusted to simulate future conditions with a similar degree of accuracy. Concretely, model estimates of streamflow were compared to in situ observations at gauging stations in the CO-MICC focus regions (see Fig. 1 and Table 1) over historical time periods.

Figure 1: CO-MICC focus basins and gauging stations used for model evaluation. The delineated basins correspond to the Ebro basin (Spain), the Moulouya basin (Morocco and Algeria), the Chelif basin (Algeria) and the Wadi Majardah basin (Algeria and Tunisia). The stations are shown as red triangles. The drainage network of the hydrological models in shown in blue.
Figure 1: CO-MICC focus basins and gauging stations used for model evaluation. The delineated basins correspond to the Ebro basin (Spain), the Moulouya basin (Morocco and Algeria), the Chelif basin (Algeria) and the Wadi Majardah basin (Algeria and Tunisia). The stations are shown as red triangles. The drainage network of the hydrological models in shown in blue.

The observations were obtained from the data portal of the Global Runoff Data Centre (GRDC). The models were forced with the daily-resolution observed global climate data set from the Global Soil Wetness Project Phase 3 (GSWP3), based on the reanalysis data set 20CR and using the bias targets GPCC, GPCP, CPC-Unified, CRU and SRB. The evaluation period considered was 1914-1980 because it is the earliest time period with sufficient coverage of observations for all stations. The earliest possible period was chosen to reduce the water-related anthropogenic activities that are not considered in the GHMs, assuming that the hydrology of these basins is less impacted by humans during that period.

Table 1: Attributes of gauging stations used for model evaluation
Station nameGRDC station codeLongitudeLatitudeRiver nameDrainage area (km2)Elevation (m.a.s.l.)
Fraga62266500.35 41.52Cinca9612100.0
Seros62266000.42 41.45Segre 1278285.0
Castejon6226300-1.69 42.18Ebro 25194265.0
Tortosa62268000.52 40.81Ebro
8423025.0
Zaragoza6226400-0.88 41.66Ebro 40434189.0
Dar El Caid1308600-3.32 34.24Moulouya 24422325.0
Sidi Belatar11041500.27 36.02Chelif
437502.0
Sloughia12015009.52 36.58Majardah2089567.0

The first stage of model evaluation consisted in evaluating how well the models can reproduce the mean seasonal variability and the annual variability of the observed streamflow.

In general, the models showed a poor performance in terms of mean seasonal streamflow variability. The poor fit between observations and model estimates could be due to several factors. For instance, it could stem from the climate input data used by the models, from how the models simulate runoff generation and/or from how the models simulate water use and dam operation. Another factor could be the presence of water transfers, which are not considered by the models. GHMs are not well enough informed to simulate the local water-related anthropogenic activities. Nevertheless, the focus of the CO-MICC project is not on simulating the local anthropogenic activities but on translating global climate signals into hydrological responses.

Unlike for mean seasonal streamflow variability, the models showed a good performance in terms of annual variability. Based on the Kling-Gupta efficiency (KGE) metric, the models perform better regarding the correlation and variability than the bias (Table 2). Furthermore, the comparison between the annual streamflow simulated in anthropogenic mode to the one obtained in naturalized mode showed smaller differences between the two modes than for the mean seasonal streamflow. From this, it was inferred that the annual streamflow variability is likely less influenced by human impacts (dams, water use) and more influenced by climate variability than the mean seasonal variability. Given that the focus is on translating global climate signals into hydrological responses, this implies that the annual streamflow variability is a more suitable indicator for evaluating the reliability of the outputs generated by the GHMs.

Table 2: Goodness of fit between observed and modelled annual streamflow in eight gauging stations based on the Kling-Gupta Efficiency (KGE). The KGE consists of three individual components, namely the correlation (r), bias (beta) and variability (gamma). Values of the KGE, r, beta and gamma closer to 1 suggest a better fit.
ModelParameterFragaSerosCastejonTortosaZaragozaDar El CaidSidi BelatarSloughia
WaterGAP2KGE0,760,8480,6190,50,238-0,844-5,52-4,158
r0,9240,8920,9370,7120,8410,9020,9650,788
Beta0,9790,9091,3640,8651,7281,7594,565,289
Gamma0,7740,9451,0950,6151,1612,6786,4623,857
LPJmLKGE0,6110,730,6680,450,392-5,827-10,476-9,973
r0,9160,8820,8820,6140,8120,9390,9660,96
Beta0,7321,2381,3021,3911,5686,4148,649,647
Gamma0,731,0471,0731,0261,1135,1599,5637,756
CWatMKGE0,2730,5220,7830,5730,414-2,518-6,347-6,214
r0,8580,7430,9320,6590,8430,9290,9780,955
Beta0,5590,8921,1971,21,5613,6756,2966,86
Gamma0,4390,6111,060,8381,0693,2836,0935,206
ORCHIDEEKGE0,6710,5580,6870,2850,705-1,665-2,757-1,889
r0,8490,7310,8860,5020,7760,9160,9990,97
Beta1,2181,3320,7730,8080,9672,9923,9112,929
Gamma0,8040,8850,8180,5240,812,7693,3753,15

With that in mind, the second stage of model evaluation consisted in looking into two additional indicators, namely the interannual variability of annual streamflow (Fig. 2) and the sensitivity of annual streamflow to precipitation variability (Fig. 3).

In general, the GHMs agree reasonably well with the interannual variability of the observed annual streamflow. Moreover, the GHM-based annual streamflow response to precipitation variability is close to the observed annual streamflow response. The conclusion drawn from these results was that the GHMs are capable of translating the climate signal into a hydrological response, with some uncertainty.

Figure 2: Comparison between observed and modelled streamflow interannual variability for eight gauging stations. Observations were collected from the online data portal of the Global Runoff Data Centre (GRDC), 56068 Koblenz, Germany. Model estimates were computed by CWatM, LPJmL, ORCHIDEE and WaterGAP2.2d. Qm is the mean annual streamflow and Qi is the annual streamflow at time i.
Figure 2: Comparison between observed and modelled streamflow interannual variability for eight gauging stations. Observations were collected from the online data portal of the Global Runoff Data Centre (GRDC), 56068 Koblenz, Germany. Model estimates were computed by CWatM, LPJmL, ORCHIDEE and WaterGAP2.2d. Qm is the mean annual streamflow and Qi is the annual streamflow at time i.
Figure 3: Comparison between observed and modelled sensitivity of annual streamflow to precipitation variability for eight gauging stations. Observations were collected from the online data portal of the Global Runoff Data Centre (GRDC), 56068 Koblenz, Germany. Model estimates were computed by CWatM, LPJmL, ORCHIDEE and WaterGAP2.2d. Qm and Pm are the mean annual streamflow and mean annual precipitation, respectively. Qi and Pi are the annual streamflow and annual precipitation at time i, respectively.
Figure 3: Comparison between observed and modelled sensitivity of annual streamflow to precipitation variability for eight gauging stations. Observations were collected from the online data portal of the Global Runoff Data Centre (GRDC), 56068 Koblenz, Germany. Model estimates were computed by CWatM, LPJmL, ORCHIDEE and WaterGAP2.2d. Qm and Pm are the mean annual streamflow and mean annual precipitation, respectively. Qi and Pi are the annual streamflow and annual precipitation at time i, respectively.