The representation of hydrological processes in land surface schemes (LSSs) has recently been improved. In this study, the usability of GCM runoff for river discharge modeling is evaluated by validating the mean, timing, and amplitude of the modeled annual discharge cycles against observations. River discharge was calculated for six large rivers using runoff, precipitation, and actual evaporation from the GCMs ECHAM5 and Hadley Centre Global Environmental Model version 2 (HadGEM2). Four methods were applied: 1) accumulation of GCM runoff, 2) routing of GCM runoff, 3) routing of GCM runoff combined with temporal storage of subsurface runoff, and 4) offline hydrological modeling with the global distributed hydrological model PCRaster Global Water Balance (PCR-GLOBWB) using meteorological data from the GCMs as forcing. The quality of discharge generated by all four methods is highly influenced by the quality of the GCM data. In small catchments, the methods that include runoff routing perform equally well, although offline modeling with PRC-GLOBWB outperforms the other methods for ECHAM5 data. For larger catchments, routing introduces realistic travel times, decreased day-to-day variability, and it reduces extremes. Complexity of the LSS of both GCMs is comparable to the complexity of the hydrological model. However, in HadGEM2 the absence of subgrid variability for saturated hydraulic conductivity results in a large subsurface runoff flux and a low seasonal variability in the annual discharge cycle. The analysis of these two GCMs shows that when LSSs are tuned to reproduce realistic water partitioning at the grid scale and a routing scheme is also included, discharge variability and change derived from GCM runoff could be as useful as changes derived from runoff obtained from offline simulations using large-scale hydrological models.
The transport of water through rivers to oceans was previously often neglected in general circulation models (GCMs; Miller et al. 1994). In the last decade it has been recognized that surface hydrology and river flow play an important role in the global climate system. For instance, the freshwater influx to oceans changes their salinity and consequently may affect ocean circulation and convection (Arora 2001). Furthermore, the hydrological cycle influences feedback mechanisms between land surface and atmosphere (Kite 1998). Full inclusion of these feedback mechanisms requires the presence of land surface schemes (LSSs). These LSSs regulate the partitioning of precipitation into evaporation, storage, and runoff and hereby determine the division of radiant energy in latent and sensible heat fluxes (Van den Hurk et al. 2005; Pappenberger et al. 2009).
In addition to near-surface temperature, runoff is one of the few outputs of an LSS that can be validated with observed data. Therefore, the performance of LSSs, and of GCMs containing an LSS, is frequently evaluated by translating the gridcell runoff into river discharge and comparing the obtained discharges with observations (Graham and Jacob 2000; Arora 2001; Pappenberger et al. 2009). Resolutions of recent LSSs now approach the resolutions of macroscale hydrological models and the inclusion of hydrological processes can be as complete as in hydrological models (Clark and Gedney 2008; Hagemann and Gates 2003).
Therefore, in addition to using it as means of validation of LSSs, discharge obtained by routing GCM runoff may as well have a hydrological value. Moreover, in large-scale hydrological impact studies, GCM runoff may even be useful as a replacement for discharge obtained from hydrological modeling, as is currently common practice (Sperna Weiland et al. 2011; Alcamo et al. 2007; Fowler and Kilsby 2007). When GCM runoff is used, a closed hydrological cycle is guaranteed and inconsistencies such as the double inclusion of hydrological processes in the modeling chain (e.g., storage of water in soil, snow, and canopy and calculation of actual evaporation fluxes) can be avoided. At the same time, computation times are reduced, which enables the analysis of larger ensembles of GCMs. Although the main purpose of a GCM is not the reproduction of realistic runoff fluxes, GCM runoff fields have been used directly in hydrological change studies before (e.g., Manabe et al. 2004). Other studies routed GCM runoff fields to obtain river discharge and to calculate future discharge changes (Arora and Boer 2001; Milly et al. 2005; Nohara et al. 2006).
In this study, we investigate the hydrological value of GCM runoff. This hydrological value is defined as the suitability of GCM-generated runoff fields for river discharge modeling in hydrological impact studies. To this end, observed annual mean discharge, discharge extremes, and timing and amplitude of the annual discharge cycle should be correctly reproduced using GCM runoff. In this study, a first step of this assessment is made by comparing discharges derived from the runoff fields of two GCMs [ECHAM5 and the Hadley Centre Global Environmental Model version 2 (HadGEM2)] with observations. In this comparison we distinguish between correct implementation of local runoff-generation mechanisms and the effect of the inclusion of a river routing scheme. To this end, discharge is generated using the following four methods: 1) daily accumulation of GCM runoff, 2) routing of GCM runoff, 3) routing of runoff in combination with attenuation of subsurface runoff in a groundwater reservoir, and 4) offline hydrological modeling with meteorological data from the two GCMs. In the analysis of results, step-by-step comparisons of the different methods are made. The comparison of daily accumulated runoff with routed runoff illustrates the effect of runoff routing. The effect of additional storage in a groundwater reservoir is illustrated by comparison of direct routing of the complete runoff flux with the inclusion of a delay by temporal storage of the subsurface runoff before routing. The offline hydrological model is used to illustrate the influence of different land surface schemes.
2. Methods and data
1) Hydrological data
This study focuses on six large river basins: the Amazon, Brahmaputra, Lena, Mississippi, Rhine, and Zambezi. The selected rivers cover multiple continents, a wide range of climate conditions, and they differ in total catchment area and degree of regulation. Figure 1 shows the geographical location of the rivers together with some basin characteristics. Below follows a short description of the basins. The Amazon originates in the Andes Mountains and flows through Brazil to its mouth at the Atlantic Ocean. Along its course, of about 6400 km, around 1000 tributaries join the river. The Amazon is at its narrowest (~1.5 km) at the Obidos discharge station, where the river depth is 60 m. Here the annual cycle has a typical high-flow period from November to June. The river course is almost completely natural with limited water regulation and use.
The Lena is located in Russia and originates in the mountains near Lake Baikal. It has a length of approximately 4500 km and ends with a large delta in the Arctic Sea. Winter temperatures can be as low as −70°C and precipitation amounts are small (200–400 mm). The river is free from ice for five months a year and has a typical cycle with limited-to-zero flow (caused by freezing) during the winter and high flow during the summer. As for the Amazon, the river course is almost completely natural with limited water regulation and use.
The Rhine originates in the Swiss Alps. The river flows through Germany and drains through the Netherlands into the North Sea. It has a length of 1230 km. Many (Alpine) tributaries join the river in Switzerland and Germany. The annual cycle is characterized by a low winter discharge, a high discharge in spring originating from snowmelt, and again a low end-of-summer discharge. The river is highly regulated and embanked.
The Zambezi rises in northwestern Zambia and flows through Angola, Botswana, and Mozambique to the Indian Ocean. The basin is partly located in the tropics. In the lower part, located in the valley, temperatures (~40°C) and, consequently, evapotranspiration can become very high. Annual average basin precipitation is around 600 mm. Discharge is at its maximum in March–April and diminishes in October–November to less than 10% of the maximum flow. River (in)flow is regulated by several dams and, in the populated areas, water use influences river flow in the dry season.
The Mississippi rises at Lake Itasca, Minnesota. From there it flows south where it is joined by its major tributaries, the Missouri and Ohio Rivers. The river ends at the Gulf of Mexico after a course of 3780 km. Temperature and precipitation amounts vary widely over the basin. In subarctic northern Minnesota, water is stored as snow in winter (temperatures of −12°C). Annual rainfall amounts vary spatially from 750 to 1800 mm. The highest monthly discharge is observed in March and mainly originates from the Ohio River. The river is highly regulated with many levees, embankments, and reservoirs for flood control and temporal water storage.
The Brahmaputra originates in the glaciated areas of the Himalaya in Tibet (China). The river has a length of 2900 km and traverses China, India, Bhutan, and Bangladesh. In Bangladesh it merges with the Ganges and the two rivers form a delta. The climate of the basin is monsoon driven and has a wet season from June to September that accounts for 60%–70% of the annual rainfall. The river course is still very natural and changes continuously over time, especially in the lower Brahmaputra delta (http://www.britannica.com).
2) GCM data
Daily global time series from the GCMs ECHAM5 (Roeckner 2006) and HadGEM2 (Johns 2009) are obtained from the World Data Centre for Climate (WDCC 2010) for the current climate experiment 20C3M (period 1961–90). The study is restricted to the GCMs ECHAM5 and HadGEM2 because of limited online availability of GCM runoff fields. Table 1 shows the GCM parameters used in this study (all parameters are cell-specific values).
3) Reference meteorological data
For historical comparison, two reference precipitation datasets have been included. Two datasets were selected because the quality of precipitation data is always hampered by measurement errors and no true dataset exists (Fiedler and Döll 2007).
The Global Precipitation Climatology Project (GPCP) precipitation dataset (Beck et al. 2005, 181–190) is a global monthly gridded precipitation dataset for the period 1951–2000 with a resolution of 1.0°. The gridded precipitation fields are derived from the Global Precipitation Climatology Centre precipitation database, which contains global validated station data. This precipitation dataset has only been used for the validation of the GCM precipitation since other variables, required to perform a hydrological model run, are not provided.
(ii) CRU monthly time series downscaled to daily time scale using ERA-40 reanalysis
In addition to precipitation, the second meteorological dataset contains all required variables (e.g., precipitation, temperature, and reference crop potential evaporation) to perform a reference hydrological model run. This dataset is created by downscaling the Climatic Research Unit (CRU) TS2.1 time series (New et al. 2000) to daily values with the 40-yr European Centre for Medium-Range Weather Forecasts (ECMWF) Re-Analysis (ERA-40) dataset (Uppala et al. 2005; see also Van Beek 2008 and Sperna Weiland et al. 2010). For the calculation of potential evaporation, the required radiation and wind speed were not available and monthly time series and data from the CRU Community Land Model, version 1 (CLM1) dataset have been used (New et al. 1999).
b. Land surface schemes of selected GCMs
In this study, ECHAM5 version 2.02a with a resolution of T63L31 (approximately 1.5°) is used. ECHAM5’s land surface scheme consists of four water reservoirs: snow intercepted at the canopy, snow stored at the surface, rainwater intercepted by the canopy, and soil water (Roeckner et al. 2003; Hagemann et al. 2006). Evapotranspiration is limited by water availability and stomatal resistance. Storage of rain and snow in the canopy is limited by the size of the interception reservoir. Snow leaves the canopy through slipping (wind dependent) and melting (temperature dependent). All water passing through these storage reservoirs becomes available to the soil water reservoir. The soil water reservoir is a single bucket with subgrid variability for the calculation of the maximum storage capacity. This storage capacity decreases according to a probability density function that defines, based on the water content, which fraction of the cell is saturated. All rainfall falling on this saturated fraction is transformed into runoff (Dumenil and Todini 1992). Drainage (e.g., subsurface runoff) from the soil water reservoir depends on the percentage of the maximum storage capacity occupied. Water is released through a slow drainage process when storage is between 5% and 90% of the maximum storage capacity and through a fast drainage process when storage is between 90% and 100% of the maximum storage capacity.
The land surface resolution of HadGEM2 is approximately 1.25° latitude by 1.875° longitude. The land surface scheme is similar to the Met Office Surface Exchange Scheme (MOSES1) (Cox et al. 1999; Clark and Gedney 2008). It contains six water storage components: snow mass, canopy water, and four soil layers in which thicknesses increase with depth. Evaporation is calculated using an extended Penman–Monteith equation that includes conductive heat transport through the soil. The evaporation flux consists of evaporation from the canopy store, transpiration from vegetation, bare soil evaporation, and sublimation of snow. Snowmelt and throughfall infiltrate into the soil and infiltration excess runoff occurs when the infiltration rate exceeds the saturated hydraulic conductivity. According to Clark and Gedney (2008), this surface runoff flux is rare on the model’s grid scale. Although the model does include a procedure to calculate an exponential subgrid distribution of rainfall, the saturated hydraulic conductivity is uniform over a grid cell and most of the time not exceeded by the infiltration rate.
Vertical moisture fluxes between the soil layers are modeled using an approximation of the Richard’s equation, where the soil hydraulic characteristics follow Clapp and Hornberger (1978). Drainage from the bottom soil layer forms the subsurface runoff flux. Drainage rates equal the constant conductivity of this bottom soil layer and the boundary condition is given by free drainage. The complexity of the hydrological schematization is comparable to that of distributed hydrological models. In addition to the hydrological part, the land surface scheme contains a vegetation module and a module that calculates soil thermodynamics. Both modules interact with the hydrological module through evapotranspiration rate, soil hydraulic conductivity, and the freezing of soil water.
c. River discharge generation
We consider the following methods for generation of river discharge from GCM output.
1) Runoff accumulation (ACCU)
In this method, the cell-specific total GCM runoff flux (MRRO) is accumulated at a daily time step. All water that becomes available in a catchment is, within the same time step, transported along a drainage network to the catchment outlet. This results in daily discharge time series for all individual grid cells. The cell-specific daily discharge amount is the sum of the runoff that becomes available in the cells’ upstream catchment area in the specific time step. Here, the upstream area consists of cells upstream of the cell in question, which are connected to the same river branch of the drainage network. The drainage network is based on the 30′ global drainage direction map (DDM30) dataset and has a vertical resolution of 0.5° (Döll and Lehner 2002).
2) Routing of total runoff (MRRO)
Within this method, cell-specific total runoff (MRRO) is accumulated for each individual time step and, as an extension of the above method, routed along a drainage network using the kinematic wave equation. The drainage network is based on DDM30 and has a vertical resolution of 0.5° (Döll and Lehner 2002). It is extended with lakes, wetlands, and reservoirs for which storage volumes have been obtained from the Global Lakes and Wetlands Database (GLWD; Lehner and Döll 2004). The discharge–stage (Q–h) relationships are used to calculate the discharge from reservoirs and lakes. No reservoir scheme has been implemented in this version of the model. Flow velocities are calculated using the kinematic wave equation (Chow et al. 1988) implemented in the PCRaster framework (Wesseling et al. 1996). A disadvantage of the kinematic wave equation is its simplicity. Complex dynamic processes, such as backwater effects, are not included. Furthermore, the method assumes that the slope of the water surface is parallel to the slope of the bottom of the water course. This criterion is not always fulfilled in reality. Yet, important advantages of the application of the kinematic wave equation in this study are its calculation speed, which enables running the model on a global scale, and its limited input data requirements, which is very useful for data-sparse regions of the world.
Sensitivity of the routing parameters has not been fully tested. Based on literature, the value of the manning coefficient for the floodplains is set to 0.1 and the manning coefficient for channels is set to 0.04 (Chow et al. 1988). The length of a river segment is equal to either the length of the diagonal or the length of the grid cell, depending on the flow direction. To correct for the difference between modeled and real-world river length, a global average tortuosity multiplication coefficient of 1.3 has been implemented.
Initial conditions were obtained by running the model multiple times with a climatology for the period 1961–90. This climatology was derived from the CRU dataset, which had been downscaled to daily values with the EA-40 reanalysis dataset. Steady-state conditions were reached at the moment the global annual average catchment outflow differed less than 0.5% between two consecutive annual climatological runs.
3) Routing and attenuation of subsurface runoff (MRROSS)
In this method, the GCM subsurface runoff (MRROSS) is temporarily stored in a linear groundwater reservoir. Groundwater release from the reservoir is regulated with cell-specific reservoir constants that have been derived from a global lithology map (Dürr et al. 2005) following Van Beek and Bierkens (2009). After release, the groundwater flux is added to the GCM surface runoff flux (MRROS) and the combined flux is routed along the drainage network as described in section 2c(2). Initial groundwater reservoir storage, which is required for this method, has been obtained by back-to-back simulations with 10 yr of GCM subsurface and surface runoff. These runs were repeated until dynamic steady-state conditions were achieved for the soil and groundwater stores (e.g., less than 1% difference in catchment and reservoir outflow for consecutive 10-yr runs).
4) Hydrological modeling with PCR-GLOBWB (EVSPSBL)
In the final method, the global distributed hydrological model PCRaster Global Water Balance (PCR-GLOBWB) (Bierkens and Van Beek 2009; Van Beek and Bierkens 2009) has been used to derive river discharge from the actual evaporation, precipitation, and temperature calculated by the GCMs. The model has a resolution of 0.5°. Each cell consists of a canopy layer, a snow layer, two vertical soil layers, and one underlying groundwater reservoir (i.e., the same reservoir as described in the MRROSS method). Subgrid parameterization is used to represent surface water, short and tall vegetation, and to calculate saturated areas for surface runoff as well as interflow. The amount of water entering the cell is prescribed by the GCM precipitation flux and can be stored as canopy interception or snow. Precipitation is stored as snow when temperature is below 0°C. Snow melts at temperatures above 0°C. Meltwater and throughfall are passed to the surface, where they either infiltrate into the soil or become surface runoff (Dumenil and Todini 1992). Exchange of soil water is possible between the soil layers and groundwater reservoir in both up- and downward directions depending on soil moisture status and groundwater storage. The GCM actual evaporation flux is subtracted from open water, bare soil, and vegetation compartments depending on vegetated fractions of cells and crop factor values. For the cell fractions without open water, we apply the following rules: if the prescribed GCM actual evaporation flux exceeds the amount of water available on the canopy and in the bare soil, the remainder is subtracted from the first and second soil layer through transpiration. If the GCM actual evapotranspiration amount exceeds the actual water availability in the cell, the actual evapotranspiration is reduced to the total water availability.
Runoff consists of noninfiltrating meltwater, saturation excess surface runoff, interflow, and baseflow. The model is run at a daily time step. For each time step, the water balance is calculated for the individual cells and runoff is accumulated and routed along the drainage network as described in section 2c(2) for the MRRO method. An extensive description of the hydrological model can be found in Van Beek et al. (2011).
5) Reference hydrological model run (CRU)
As an additional reference for validation of the GCM-derived annual discharge cycles, PCR-GLOBWB has been run with the combined CRU ERA-40 dataset [section 2a(3); Sperna Weiland et al. 2010]. In this reference run, the model is forced using temperature, precipitation, and potential evaporation calculated from the CRU data with the Penman–Monteith equation (Van Beek 2008). Actual evapotranspiration is derived from potential evaporation and moisture conditions within the hydrological model.
a. Effect of biases in different components of the hydrological modeling chain
1) Bias in precipitation
Figure 2 shows bar charts with the percentage deviation of ECHAM5, HadGEM2, and CRU mean annual basin-average precipitation (PR) from observed annual mean GPCP PR. Figure 3 displays Taylor diagrams (Taylor 2001) for comparison of the mean annual cycles of ECHAM5, HadGEM2, and CRU PR with the observed mean annual cycle of GPCP PR. The Taylor plots show, for all rivers except the Zambezi, that CRU PR best resembles GPCP PR. This indicates that both measurement-based historical datasets are comparable. And, since the GPCP dataset only consists of precipitation, the CRU dataset can best be used as reference forcing for the hydrological model.
HadGEM2 PR overestimates observed GPCP PR for all selected catchments and ECHAM5 overestimates GPCP PR for all catchments except the Amazon and Zambezi. For the Amazon, both GCMs overestimate the amplitude of the annual cycle, but their correlation with observed PR is above 0.9. For the Brahmaputra, ECHAM5 shows a high correlation (>0.95) with observed PR, yet this GCM overestimates the amplitude. HadGEM2 has a lower correlation (<0.8) and a comparable root-mean-square difference (RMSD). The Taylor plot of the Lena is comparable to the Taylor plot of the Brahmaputra. For the Mississippi, the correlation between HadGEM2 and observed PR is low (≈0.4) and the RMSD is high. ECHAM5 performs somewhat better with lower RMSD values and higher correlation (>0.8). For the Rhine, correlation with observed PR is low for both GCMs (<0.3). In addition, both GCMs overestimate the amplitude of the annual cycle. The Zambezi correlation with observed PR is high (>0.9) for both GCMs. Moreover, both GCMs have a smaller RMSD from GPCP PR than the CRU PR.
2) The effect of the precipitation bias on simulated discharge
Figure 4 shows bar charts with the annual average percentage bias of ECHAM5-, HadGEM2-, and CRU-derived discharge (Q) from observed GRDC Q. Figure 5 displays Taylor plots for the annual average Q cycle obtained from GCM data with the different discharge-generation methods. Observed GRDC Q has been used as a reference. In this section, the effect of the PR biases on discharge modeled with PCR-GLOBWB is estimated for the two GCMs [i.e., annual average PCR-GLOBWB modeled basin Q from HadGEM2 (Hp) and ECHAM5 (Ep) data]. Under- and overestimations of ECHAM5 PR are mirrored in modeled Q (plus and minus signs of PR and Q differences are the same). However, underestimations are exaggerated in Q. For the Zambezi, the underestimation of observed Q is a result of the high actual evaporation flux [e.g., ECHAM5 gives a runoff coefficient of only 0.11, while a runoff coefficient of more than 0.2 has been derived from observed data (Fig. 6)]. This observation-based runoff coefficient is calculated by dividing observed basin discharge (e.g., discharge measured at the most-downstream gauge in the catchment) by basin-total GPCP precipitation (Fig. 6). Furthermore, for the Zambezi, the conversion of PR to Q heavily diminishes the correlation with the observed annual cycle. Here, river discharge may be biased by water use, the model’s overestimation of actual evapotranspiration, and the presence of dams, which are not schematized in detail in the routing model.
The correlation between the annual cycles of the two GCMs with the observed annual cycle is lower for Q than for PR (0.6 versus 0.9) for the Amazon as well. This is most likely due to the complexity of the recirculation of precipitation that is being converted to actual evapotranspiration and becomes available for precipitation again within the Amazon basin (McGuffie and Henderson-Sellers 2005). For the Brahamaputra, the Taylor plots are different for ECHAM5 and HadGEM2 PR. Yet, the Taylor plots for modeled Q are comparable for the two GCMs, although different from the plot for CRU Q. The comparison of Taylor diagrams for PR and Q for the Mississippi indicates a similar reduction of differences between the two GCMs as obtained for the Brahmaputra. Correlations with observed discharge are unexpectedly high for the Mississippi, because the river is highly regulated and this regulation is only included in the hydrological model in a limited form. Yet, the annual discharge cycle is influenced by the temporal storage of precipitation as snow in the mountainous regions. This temperature-dependent temporal storage is likely to increase the correlation while transitioning from precipitation to discharge. For both GCMs for the Rhine, and for HadGEM2 for the Lena, the correlation with the observed Q cycle is larger than the correlation with the observed PR cycle. This is also caused by temporal storage of precipitation as snow within the hydrological model. The improvement is largest for HadGEM2. Although it has a significant snow component as well, the temperature impact on the Brahmaputra discharge is small because of the strong influence of the Monsoon, which coincides with the melt season. The large overestimation of annual average precipitation in combination with a high runoff coefficient (Fig. 6) results in an overestimation of the annual average discharge in the Brahmaputra basin.
3) The effect of kinematic wave routing
In this section, we evaluate the importance of river routing in the reproduction of observed timing and the annual discharge cycle. To this end, the ACCU and MRRO methods are compared [e.g., accumulated basin runoff from HadGEM2 (Ha) and ECHAM5 (Ea) and kinematic-wave-routed basin runoff from HadGEM2 (Hr) and ECHAM5 (Er) in Fig. 5]. For ECHAM5, the correlation between the observed and modeled annual discharge cycle is highest for all rivers when kinematic wave routing of surface runoff is applied. The difference between routed and accumulated runoff is smallest for the Rhine, where the through-routing-introduced discharge delay is smallest because of the small catchment size. The difference is also small for the Zambezi, where for all methods applied to ECHAM5 data, deviations from observed discharge are large. The largest differences between accumulated and kinematic-wave-routed runoff are obtained for the Amazon. Here, simple accumulation of runoff results in a negative correlation with the observed annual discharge cycle, whereas routing introduces the required travel time, which in turn results in a realistic annual cycle.
For the Lena for HadGEM2, kinematic wave routing does not result in an increase in the correlation with the observed annual cycle. The snowmelt-driven discharge rise in spring, which characterizes this basin, occurs too late in the year for the ACCU method and within the MRRO method this delay is actually increased. For HadGEM2, a large increase in the correlation (0–0.6) is only found for the Amazon. Yet, routing reduces RMSD values for both HadGEM2 and ECHAM5 runoff.
4) The effect of subsurface flow delay
To evaluate the effect of a delay of subsurface runoff by temporal storage in a linear reservoir, we compare the MRRO method (kinematic wave routing) and the MRROS method (kinematic routing and groundwater reservoir storage). In the Taylor plots (Fig. 5), these model realizations are displayed as routing HadGEM2 (Hr) and ECHAM5 (Er) and routing and storage HadGEM2 (Hs) and ECHAM5 (Es). Overall, differences between the MRRO and MRROS methods applied to the ECHAM5 runoff fluxes are small because of the relatively small percentage of subsurface runoff within the total runoff flux (Fig. 7a). For HadGEM2, the MRROSS method results in an underestimation of the amplitude of the annual hydrological cycle for all six basins. The subsurface flux is relatively large within HadGEM2 (Fig. 7a). As a consequence, the temporal storage of subsurface runoff in a groundwater reservoir is too large and it introduces an undesired delay in the annual Q cycle, which in turn results in a large baseflow that reduces seasonal discharge variation. Surface runoff is rare for HadGEM2 because saturated hydraulic conductivity is assumed to be uniform over a grid cell (Clark and Gedney 2008) and no subgrid variable saturated areas concept is included. In most humid regions (Northern Hemisphere, Amazon, and Southeast Asia) the percentage subsurface runoff in HadGEM2 is more than 70% of total runoff. Therefore, for most basins, the MRRO method applied to HadGEM2 data results in higher correlation with observed Q than the MRROS method. Recent attempts have been made to include subgrid variability based on TOPMODEL principles, where soil moisture and runoff are related to topography (Clark and Gedney 2008; Beven and Kirkby 1979).
5) The effect of using different land surface hydrology schemes
In this section, the influence of using different LSSs is tested. To this end, the MRRO method (Er and Hr) is compared with PCR-GLOBWB runs forced with either ECHAM5 or HadGEM2 meteorological data (Ep and Hp). The land surface modeling within PCR-GLOBWB clearly outperforms ECHAM5 for the Lena and Rhine. Lower RMSD values are obtained and, particularly for the Lena, the correlation with the observed annual cycle is higher. Differences are smaller for the Mississippi and Amazon, but still the PCR-GLOBWB runs result in lower RMSD values and standard deviations closer to observed. Differences are small for the Zambezi and Brahamaputra.
PCR-GLOBWB outperforms the MRRO method applied to HadGEM2 for the Amazon, Rhine, and Zambezi. Correlation coefficients with the observed annual cycle are comparable for the two methods; however, direct routing of GCM runoff results in an overestimation of the amplitude and larger absolute deviations from observed values. For the Lena and Mississippi, differences are small and the LSS of HadGEM2 outperforms PCR-GLOBWB for the Brahmaputra.
Notable is the small difference between global patterns of the PCR-GLOBWB modeled percentage subsurface runoff obtained from either ECHAM5 or HadGEM2 meteorological data (Fig. 7b). This indicates that the formation of surface and subsurface runoff is influenced mainly by the definition and parameterization of the LSS, or water balance model, and that the meteorological forcing has only a minor impact.
b. Autocorrelation for time lags of several days
Figure 8 displays correlograms for the Amazon and Rhine at the most-downstream discharge stations in the catchments (i.e., Obidos and Rees). The autocorrelation of the discharge of the Rhine decreases more quickly with an increasing time lag than that of the Amazon. The smaller Rhine catchment has a higher sensitivity to variations in surface runoff, subsurface runoff, and precipitation. Because of this sensitivity, the difference between methods is larger for the Rhine. The results of the PCR-GLOBWB runs, forced with meteorological data from either GCM, are comparable. This exemplifies again the importance of the parameterization of the partitioning of runoff in subsurface and surface runoff and the delaying influence of the groundwater reservoir.
For the Amazon, the correlation coefficient remains above 0.9 for time lags of multiple days for most methods. The ACCU method for ECHAM5 is the only exception. For this method, the correlation coefficient drops below 0.65 after three days. The autocorrelation is relatively small because the ECHAM5 runoff flux is accumulated over the entire catchment on a daily basis and because the cell-specific runoff itself has a larger day-to-day variability than the HadGEM2 flux because of a smaller subsurface component. This shows that the delays introduced by groundwater storage and channel travel time are the most important elements in explaining discharge autocorrelation.
The hydrographs of the Rhine and Amazon (Fig. 9) illustrate that the importance of routing on day-to-day variability depends on catchment size. Note that in Fig. 9 hydrographs are shown for a random year and cannot be compared one by one, since, because of their chaotic nature, a given year within a GCM is only a possible representation of a random year in that period. Therefore, the hydrographs only give an indication of performance for timing and the reproduction of extremes. For the Amazon, the hydrograph generated with the MRRO method from ECHAM5 runoff shows a realistic reduction in day-to-day variability, while with the ACCU method, day-to-day variability is too high. Furthermore, when kinematic wave routing is applied, the peak discharge occurs later and a relatively constant baseflow is reached at the end of the summer. Even though ECHAM5 has a small baseflow component, the delay introduced by routing is that large that realistic hydrographs are obtained with the MRRO method. Although routing reduces the day-to-day variability for the Rhine as well, the variability remains larger than for the Amazon. For the Rhine, peak flows of more than 20 000 m3 s−1 are calculated with both the MRRO and ACCU methods, while the maximum observed discharge at Lobith is 12.600 m3 s−1 (Te Linde et al. 2010). In addition, discharge values below 1 m3 s−1 are obtained for low-flow periods. This is clearly the result of an underestimation of the baseflow component, which in a smaller river is not compensated for by residence time in the drainage network. The hydrographs obtained using the MRRO and MRROSS methods are comparable, mainly because of the small subsurface runoff flux. Notable is the small difference between the hydrographs of the MRROSS method and PCR-GLOBWB for the Rhine.
The Rhine hydrographs obtained from HadGEM2 data with the MRRO and ACCU method show a large resemblance to one another. Compared to ECHAM5, there is less day-to-day variability and the peak discharges are lower. This difference can mainly be explained by the large subsurface runoff flux in HadGEM2 for the Rhine catchment (93% of total runoff versus 36% of total runoff in ECHAM5) and by the small effect of routing within this relatively small basin. Because of the large baseflow component, the MRROSS method results in the most constant river flow throughout the year. This constant river flow best resembles the observed flow. We note, however, that the small variation in observed seasonal cycle is also a result of river regulation, which is only included in the routing scheme in a limited form. For the Amazon, the MRROSS method results in the smallest seasonal variation with the highest summer baseflow, as was observed for the Rhine as well. The hydrographs obtained using the MRRO and ACCU method show the lowest baseflow and for these methods baseflow is closest to observed summer discharge. However, this is mainly due to the overestimation of observed discharge by HadGEM2 as a result of too much rainfall. The ACCU method applied to HadGEM2 data results in a lower day-to-day variability than the ACCU method applied to ECHAM5 data. Still, for HadGEM2 the ACCU method is the method with the highest day-to-day variability.
4. Conclusions and discussion
In this study, we evaluated the suitability of GCM runoff (from the GCMs ECHAM5 and HadGEM2) for hydrological impact studies. To this end we used four different discharge-generation methods, ranging from simple GCM runoff accumulation, to runoff routing and advanced hydrological modeling. For all four methods, generated discharges show deviations from observed discharge due to biases in GCM runoff, precipitation, and evaporation fluxes and the absence of water use in the models.
Routing and temporal storage in a groundwater reservoir, as included in the MRROSS and PCR-GLOBWB methods, introduce a delay for both GCMs (Oki and Sud 1998). This delay reduces discharge peaks to realistic values and it introduces constant baseflows.
Too-high peak flows and day-to-day variability are obtained with the ACCU method, which lacks discharge routing. Differences between this method and the other methods are larger for the Amazon than for smaller catchments, such as the Rhine. However, we only conducted our analysis for relatively large catchments using GCM data and the value of GCM or RCM runoff should be assessed on smaller scales as well. For such an analysis, additional runoff downscaling methods based on soil properties and topography (Beven and Kirkby 1979) might be useful. Furthermore, the analysis is restricted to two GCMs. It would be interesting to extend the analysis when runoff fields become available for more GCMs.
The schematizations of the LSSs of ECHAM5 and HadGEM2 are comparable to the hydrological model PCR-GLOBWB, at least in the detail and complexity of the hydrological processes included. However, the incoming precipitation is divided differently over the water balance components (e.g., surface runoff, subsurface runoff, and actual evapotranspiration) because of a different parameterization and formulation of the subsurface hydrological processes. In addition, the resolution of PCR-GLOBWB is higher. In HadGEM2, the low resolution in combination with the lack of subgrid parameterization results in a large subsurface runoff flux. This affects the annual discharge cycle and may influence the surface energy balance of the GCM as well. In ECHAM5, the low resolution is partly compensated for by subgrid parameterization of surface runoff, yet here the fraction of groundwater runoff is probably too small.
The Taylor plots show that PCRGLOB-WB outperforms the other runoff-generation methods. For ECHAM5, this is the case for all six rivers—for HadGEM2, for the Mississippi, Lena, and Rhine. Nevertheless, this study also illustrates that differences between PCR-GLOBWB and the MRRO and MRROSS methods can be small. These two methods are promising because of their limited computational demand.
So far, land surface schemes are typically tuned to reproduce the surface energy balance correctly. However, the analysis of these two GCMs and comparison with results from a large-scale hydrological model show that, when GCM runoff generation is additionally tuned with discharge observations and a routing scheme is added, discharge derived from GCM runoff can be as suitable as discharge derived from runoff calculated by an offline hydrological model for large-scale studies.
We acknowledge the Global Runoff Data Centre for providing the global discharge time series. Furthermore, we are grateful to the three anonymous reviewers; their comments greatly helped us to improve the quality of the manuscript.