1 FEBRUARY 2010 HURKMANS ET AL.

679

Changes in Streamflow Dynamics in the Rhine Basin under Three

High-Resolution Regional Climate Scenarios

RUUD HURKMANS, WILCO TERINK, REMKO UIJLENHOET, AND PAUL TORFS

Hydrology and Quantitative Water Management Group, Wageningen University, Wageningen, Netherlands

DANIELA JACOB
Max Planck Institut für Meteorologie, Hamburg, Germany

PETER A. TROCH
Department of Hydrology and Water Resources, The University of Arizona, Tucson, Arizona

(Manuscript received 2 February 2009, in final form 21 August 2009)

ABSTRACT

Because of global warming, the hydrologic behavior of the Rhine basin is expected to shift from a combined
snowmelt- and rainfall-driven regime to a more rainfall-dominated regime. Previous impact assessments have
indicated that this leads, on average, to increasing streamflow by ;30% in winter and spring and decreasing
streamflow by a similar value in summer. In this study, high-resolution (0.0888) regional climate scenarios
conducted with the regional climate model REMO (REgional MOdel) for the Rhine basin are used to force
a macroscale hydrological model. These climate scenarios are based on model output from the ECHAM5–
Max Planck Institute Ocean Model (MPI-OM) global climate model, which is in turn forced by three Special
Report on Emissions Scenarios (SRES) emission scenarios: A2, A1B, and B1. The Variable Infiltration
Capacity model (VIC; version 4.0.5) is used to examine changes in streamflow at various locations throughout
the Rhine basin. Average streamflow, peak flows, low flows, and several water balance terms are evaluated for
both the first and second half of the twenty-first century. The results reveal a distinct contrast between those
periods. The first half is dominated by increased precipitation, causing increased streamflow throughout the
year. During the second half of the century, a streamflow increase in winter/spring and a decrease in summer is
found, similar to previous studies. This is caused by 1) temperature and evapotranspiration, which are con-
siderably higher during the second half of the century; 2) decreased precipitation in summer; and 3) an earlier
start of the snowmelt season. Magnitudes of peak flows increase during both periods, and the magnitudes of
streamflow droughts increase only during the second half of the century.

1. Introduction peratures will cause more precipitation to fall as rain

instead of snow and will cause the snowmelt season to
According to the latest Intergovernmental Panel on
start earlier in the year (Barnett et al. 2005). As far as the
Climate Change (IPCC) assessment report (Adger et al.
Rhine basin is concerned, this amounts to an increase in
2007), all currently available global climate models agree
temperature of 1.08–2.48C by 2050, causing the hydro-
on an increase in global mean temperature of 1.18–6.48C
logical regime of the Rhine to shift from a combined
over the twenty-first century. Associated with this global
rainfall- and snowmelt-driven system to a more rainfall-
warming, the hydrological cycle is expected to intensify,
dominated system (Barnett et al. 2005; Adger et al. 2007).
causing precipitation to fall in more extreme events
The Rhine basin is densely populated and heavily
(Trenberth et al. 2003). In addition, higher air tem-
industrialized, and the river Rhine has the highest traf-
fic density in Europe as an inland waterway. Its water
is used for many purposes, for example, agriculture,
Corresponding author address: Ruud Hurkmans, Hydrology and
Quantitative Water Management Group, Wageningen University, industry, domestic water supply, and (hydro-) power
P.O. Box 47, 6700 AA, Wageningen, Netherlands. generation (Kwadijk and Rotmans 1995; Middelkoop
E-mail: [email protected] et al. 2001). Because of all the economic value that is

DOI: 10.1175/2009JCLI3066.1

Ó 2010 American Meteorological Society

680 JOURNAL OF CLIMATE VOLUME 23

concentrated in the basin, it is very vulnerable to hy- Institute Ocean Model (MPI-OM) and downscaled in
drologic extremes, both floods and droughts (Kleinn two steps by the regional climate model REMO (Jacob
et al. 2005). It is therefore not surprising that many cli- 2001). To capture extreme precipitation events, and thus
mate change impact assessments have been carried out also extreme peak flow events, a high spatial resolution is
for the Rhine basin in recent years (e.g., Graham et al. important because extreme precipitation is usually con-
2007; Kwadijk and Middelkoop 1994; Kwadijk and vective in nature and its spatial extent relatively small.
Rotmans 1995; Lenderink et al. 2007; Middelkoop et al. Moreover, as part of the catchment is very mountainous,
2001; Pfister et al. 2004; Shabalova et al. 2003; Zierl and orography is relatively complex and a high model reso-
Bugmann 2005). Most of these studies are based on lution allows more of this complexity to be resolved
climate model output, which are based on scenarios explicitly. An additional advantage of the climate sce-
describing greenhouse gas emissions. Emission scenar- narios used in this study is the fact that they are transient,
ios range from a simple doubling of CO2 concentrations spanning the entire twenty-first century. The studies that
(e.g., Middelkoop et al. 2001) or a gradual increase of were mentioned above typically used a time slice of
CO2 (e.g., 1% yr21 in Shabalova et al. 2003) to the more 30 yr, often the period 2071–2100.
complex scenarios developed by IPCC in the Special In this study, we employ three high-resolution climate
Report on Emissions Scenarios (SRES; Nakićenović change scenarios, according to three SRES scenarios:
and Swart 2000). SRES scenarios are based on several B1, A1B, and A2. We compare the scenarios with results
alternative developments of energy technology and con- of a reference model run spanning the period 1950–2000.
tain not only CO2 but also other greenhouse gases and All atmospheric time series are corrected for their
aerosols. Because the spatial resolution of global climate model bias based on observations of precipitation and
models (GCM) is generally too low for hydrological temperature (section 3a). As a macroscale hydrological
applications (.100 km), downscaling occurs either by model, we use the Variable Infiltration Capacity model
statistical methods [as was done, e.g., by Kwadijk and (VIC; Liang et al. 1994). This model was applied to the
Middelkoop (1994); Kwadijk and Rotmans (1995)] or by Rhine basin at a spatial resolution of 0.058 (;5.5 km), as
nesting a regional climate model (RCM) in the GCM described by Hurkmans et al. (2008). Compared to other
over the domain of interest (e.g., Graham et al. 2007; models that are typically used for climate change impact
Shabalova et al. 2003; Middelkoop et al. 2001; Kleinn studies and that are generally more simple, VIC has the
et al. 2005). The downscaled climate data are then fed to advantages that 1) its parameterization of evapotrans-
some hydrological model to obtain river discharges. piration is physically based; 2) it has a detailed de-
Because RCMs often give different results (e.g., van den scription of snow accumulation and melt processes; and
Hurk et al. 2005), often multiple RCMs are used. For 3) subgrid variability of land use, elevation, and in-
example, Graham et al. (2007) used results from all filtration is taken into account. For these reasons, VIC
RCMs that were compared in the Prediction of Regional was found to be less sensitive to parameter settings than
Scenarios and Uncertainties for Defining European the more simple water balance models mentioned above
Climate Change Risks and Effects (PRUDENCE) pro- (Hurkmans et al. 2008). This is an important advantage
ject (Christensen et al. 2007) to investigate the effects because it is questionable whether conceptual model
of climate change on Rhine River discharge. All the as- parameters determined under current climate condi-
sessments mentioned above qualitatively agree in their tions will be valid throughout the twenty-first century
results. More discharge is projected in winter and spring under a changing climate. From the model results, we
because of increased precipitation and snowmelt (130% investigate the effect of climate change on average and
according to Lenderink et al. 2007 and Shabalova et al. extreme streamflow, both in terms of peak flows and
2003), and less discharge is projected in summer because streamflow droughts. By analyzing streamflow at several
of increased evapotranspiration, less snow buffering, and locations throughout the river basin and the distributed
decreased precipitation (230% to 240% according to model results, the spatial pattern of the effects of climate
the same studies). changes is investigated. Finally, changes in various com-
In the studies described above, most of the employed ponents of the water balance such as evapotranspiration,
regional climate scenarios have spatial resolutions of 25– snowmelt, and baseflow are analyzed.
50 km. By means of a second downscaling step, higher
spatial resolutions can be obtained. From the Max 2. Data and model
Planck Institut für Meteorologie (MPI-M) in Hamburg,
a. Study area
Germany, climate scenarios are available for the Rhine
basin at a spatial resolution of 0.0888 (;10 km). These The Rhine basin is a major river in western Europe.
scenarios are based on the GCM ECHAM5–Max Planck It originates in the Swiss Alps and drains portions of
1 FEBRUARY 2010 HURKMANS ET AL. 681

FIG. 1. (left) Location of the Rhine basin and (right) elevations in the basin. The outline of the basin and the subbasins represent the
discretization of the basin in VIC (at a resolution of 0.058). Source: Hurkmans et al. (2009).

Switzerland, Germany, France, and the Netherlands a linear and a nonlinear segment. Routing of baseflow
before draining in the North Sea. The total area covered and surface runoff occurs through a simple algorithm
is about 185 000 km2. After crossing the German–Dutch developed by Lohmann et al. (1996), which was applied
border at Lobith, the river splits into three branches. in combination with VIC before (e.g., Lohmann et al.
Therefore, we only take into account the area upstream 1998a,b). The VIC model has been used in previous
of Lobith (Fig. 1). studies as a land surface parameterization in climate
models (e.g., Lohmann et al. 1998c; Pitman et al. 1999;
b. Hydrological model
Liang et al. 1998) but also for hydrological purposes. For
The VIC model (Liang et al. 1994) is a distributed, example, Hamlet et al. (2007) used the model for climate
macroscale hydrological model. It is developed to pro- variability impact assessments and Matheussen et al.
vide land surface boundary conditions for climate (2000) and Hurkmans et al. (2009) investigated land use
models and therefore solves evapotranspiration and the changes using VIC.
associated turbulent fluxes in a physically based way. The VIC model was applied to the Rhine basin in
The coupled water and energy balances are solved, in a previous study (Hurkmans et al. 2008) at a spatial
combination with the Penman–Monteith equation resolution of 0.058 and a temporal resolution of 3 h. Two
(Penman 1948; Monteith 1965), to calculate evapo- modifications of the VIC model that were introduced in
transpiration. Subgrid variability in land use and eleva- a recent study dealing with land use changes (Hurkmans
tion is accounted for by dividing each grid cell into tiles et al. 2009) are used in this study as well. First, in sparsely
based on fractional areal coverage. Subgrid variability of vegetated areas bare soil evaporation is allowed to oc-
topography is accounted for by means of an exponential cur at the same time as transpiration and canopy evap-
relation between the fraction of the grid cell that is oration. Second, a parameter is introduced in urban
saturated and the infiltration capacity. The soil column is areas to increase the amount of surface runoff relative
divided in three soil layers. Drainage from the surface to to baseflow. For more details about these modifications,
the lowest layer is gravity-driven and downward only. we refer to Hurkmans et al. (2009). Atmospheric forc-
Baseflow depends on the soil moisture content of the ing for the model is described in section 2c. Land use
lowest soil layer through a function that consists of data were obtained from the Pan-European Land Cover
682 JOURNAL OF CLIMATE VOLUME 23

TABLE 1. Overview of atmospheric datasets used in this study.

Dataset Driving model Spatial resolution Time resolution Period

Validation ECMWF 0.088 Hourly 1979–2003
Reference ECHAM5–MPI-OM 0.088 Hourly 1950–2000
Scenarios ECHAM5–MPI-OM 0.088 Hourly 2001–2100
Observations — Subbasin Daily 1961–95

Monitoring (PELCOM) project (Mücher et al. 2000) at 15-yr European Centre for Medium-Range Weather
a spatial resolution of 1 km, and soil data are taken from Forecasts (ECMWF) Re-Analysis (ERA-15) reanalysis
the global Food and Agriculture Organization (FAO) dataset (http://www.ecmf.int; 1979–93), extended up to
dataset described by Reynolds et al. (2000). Based on 2003 using operational ECMWF analysis data. Because
percentages of sand and clay from this dataset, soil the ERA data are (partly) based on observations, these
textures are classified and the default parameters from data can be used to calibrate the hydrological model,
the VIC Web site (http://www.hydro.washington.edu/ as was done in Hurkmans et al. (2008). The reference
Lettenmaier/Models/VIC/) are assigned to each class. dataset is based on a REMO model run driven by the
general circulation model ECHAM5–MPI-OM, as are
c. Atmospheric data
the three scenario datasets. Because the reference da-
All atmospheric datasets that are used in this study are taset is not constrained by observations, it only resem-
model output from the regional climate model REMO bles the current climate in a statistical sense. It cannot,
(Jacob 2001) as provided by MPI-M in Hamburg, Germany. therefore, be used for hydrological model calibration.
A validation dataset, a reference dataset, and three sce- Because it is based on the same models as the scenarios,
narios datasets were available. The main characteristics it can be compared with the scenarios to obtain a climate
of all datasets are shown in Table 1, and the way they are change signal. This is not the case for the validation
used in this study is shown schematically in Fig. 2. The dataset because it is based on a different driving model
validation dataset is an extended version of the atmo- (i.e., ERA data).
spheric dataset that was used in Hurkmans et al. (2008) The spatial resolutions of both ERA and ECHAM5–
and Hurkmans et al. (2009). It is based on the global MPI-OM data are too low for hydrological applications

FIG. 2. Schematic representation of all datasets used in this study and how they are used.
1 FEBRUARY 2010 HURKMANS ET AL. 683

(.100 km). They are, therefore, dynamically downscaled seasonally varying correction factor, and Hay et al.
using the REMO model in two steps. First, intermediate (2002) used the gamma distribution to match observed
REMO output with a resolution of 0.448 (about 50 km) is and simulated precipitation values. In this study, we
obtained, which is downscaled using a second step to use a relatively simple method that was applied to the
0.0888 [about 10 km; Jacob et al. (2008) and S. Kotlarski Meuse basin by Leander and Buishand (2007). In addi-
2008, personal communication; MPI-M]. tion, Shabalova et al. (2003) applied it to the Rhine basin
Both driving ‘‘models,’’ ERA data and ECHAM5– to obtain a plausible climate scenario. An advantage of
MPI-OM, have structural modeling errors compared this method is that not only the mean but also the tem-
with observations. Prior to using the data, this model poral variability is corrected. The method corrects pre-
bias needs to be corrected for using observations. In- cipitation and temperature separately and leaves other
terpolated observations were available over the Rhine input variables for VIC (i.e., radiation, humidity, wind
basin (Sprokkereef 2001) for this purpose. Because the speed, and air pressure) unaffected. Because no obser-
spatial and temporal resolutions of these observations vations are available for these variables at this scale and
are too low, they cannot be used to force the VIC model. resolution, it is not possible to correct for them. More-
Therefore, we use the validation and reference datasets over, the effects on the resulting streamflow induced by
described above to represent the current climate con- errors occurring in these variables are relatively small
ditions. The bias corrections are described in detail in compared to biases in precipitation and temperature.
section 3a. Because the three scenario datasets are based Observed precipitation and temperature for the period
on the same modeling chain as the reference dataset, we 1961–95 are available on a daily basis for each of 134
assume that the model bias is similar for both the ref- subcatchments in the Rhine basin (Sprokkereef 2001).
erence dataset and the scenarios and apply the same Therefore, and because the model bias was found to vary
correction to both. spatially (in accordance with Leander and Buishand
The three scenarios are based on three different CO2 2007 and Lenderink et al. 2007), the bias correction is
emission scenarios as defined by the IPCC in the SRES applied to daily average temperature and daily accu-
(Nakićenović and Swart 2000): the A2, A1B, and B1 mulated precipitation for each subbasin individually.
scenarios. Each of those scenarios is based on a storyline Because the reference dataset and validation datasets
of energy and technology development. The A2 sce- (Table 1) are based on different models, a bias correc-
nario refers to a world with continuously growing pop- tion is applied to both datasets separately. For both
ulation and very regionally oriented economic growth. corrections, the maximum period of overlap between
A1B refers to a globalized, very rapidly growing econ- data and observations is used to estimate the correction
omy with fast introduction of new technologies that are parameters. For the validation dataset this period is
balanced between fossil fuel intensive and sustainable 1979–95, whereas for the reference dataset it is 1961–95.
and clean. Population in A1B grows rapidly until the For both precipitation and temperature, a 5-day clima-
middle of the century and declines thereafter. B1, fi- tology is derived from observations and model output by
nally, refers to a globalized, rapidly growing population calculating the average and standard deviation over all
but with changes in economic structure with an envi- available years, that is, 17 yr for the validation and 35 yr
ronmental emphasis and fast introduction of clean and for the reference dataset. A window of 30 days before
efficient technologies. For more details about these story- and 30 days after the considered 5-day period is used
lines, we refer to Nakićenović and Swart (2000). to calculate the ‘‘smoothed’’ average and standard de-
viation. The average and standard deviation are thus
calculated over 65 3 17 (validation dataset) or 65 3 35
3. Methodology (reference dataset) data points. The correction is ap-
plied separately for each of the 73 five-day periods in the
a. Bias correction
resulting climatology.
A problem with the use of RCM output is the fact that Precipitation is corrected according to a nonlinear
simulated precipitation and temperature differ system- function (Leander and Buishand 2007):
atically from observed values (Christensen et al. 2008).
These biases can vary among RCMs and are also geo- P* 5 aP b , (1)
graphically dependent (Jacob et al. 2007), and, to obtain
realistic streamflow simulations in the present climate, where P is the uncorrected precipitation, P* is the cor-
they should be corrected (Lenderink et al. 2007). Dif- rected precipitation, and a and b are two parameters.
ferent methods exist to correct for these biases. For First, b is fitted for the coefficient of variation (CV)
example, Lenderink et al. (2007) used a spatially and of the modeled precipitation to match the CV of the
684 JOURNAL OF CLIMATE VOLUME 23

FIG. 3. Results of the bias corrections for the (a),(b) validation and (c),(d) reference datasets. Monthly averaged precipitation [(a) and
(c)] and temperature [(b) and (d)], both weighted by subbasin area, are shown for observed, uncorrected, and corrected precipitation and
temperature.

observed precipitation. Second, a is fitted such that the uncorrected precipitation and temperatures for both the
average modeled precipitation fits the average of the validation and the reference dataset. It appears that the
observed precipitation (using the value of b obtained in correction works very well in terms of monthly averages.
the first step). Especially for temperature, corrected values almost ex-
The correction of temperature is done in a linear way, actly match the observed ones. The mean absolute error
based on the mean and standard deviation. For each (MAE) decreased from 1.158 to 0.158C for the validation
subbasin, the corrected daily temperature T mod* is cal- dataset and from 0.768 to 0.168C for the reference data-
culated as set. For precipitation, the results are not as good as for
temperature, but still the monthly means were improved
s(T obs ) significantly: for the validation dataset the MAE decreased
* 5 T obs 1
T mod (T T obs ) 1 (T obs T basin ),
s(T basin ) mod from 8.5 to 3.6 mm month21, and for the reference data-
(2) set the MAE decreased from 19.5 to 4.6 mm month21.
Figure 4 shows spatial patterns of the difference between
where Tmod is the uncorrected daily modeled tempera- modeled (corrected and uncorrected) and observed pre-
ture, Tobs is the observed daily average temperature for cipitation and temperature. Differences between uncor-
a subbasin, and Tbasin is the modeled subbasin average rected and observed precipitation are sometimes very
temperature. In Eq. (2) an overbar denotes the average large; especially in the Alps differences of more than
over the considered period (1979–95 for the validation 500 mm yr21 exist. For all subbasins, however, the cor-
dataset; 1961–95 for the reference dataset) and s de- rection performs well: the spatial MAE (weighted to area,
notes the standard deviation. for temporal averages) drops from 161 to 9 mm yr21 for
Results of the bias correction of the validation dataset, the validation dataset and from 276 to 12 mm yr21 for the
as well as the reference dataset, can be found in Figs. 3 reference dataset. For temperature, the weighted MAE is
and 4. Figure 3 shows climatologies of corrected and also reduced significantly: for the validation dataset from
1 FEBRUARY 2010 HURKMANS ET AL. 685

FIG. 4. Results of the bias corrections in terms of spatial patterns. Each map shows the difference
between corrected or uncorrected data vs observations. Differences in mean annual temperature and
mean cumulative precipitation are shown for both the (top) validation and (bottom) reference datasets.
Positive values indicate an overestimation with respect to the observations.

0.918 to 0.068C and for the reference dataset from 0.858 described in Hurkmans et al. (2008), using 5 calibration
to 0.068C. Additional validation analyses (Terink et al. parameters and a numerically efficient algorithm, the
2008) showed that, apart from these numbers—also downhill-Simplex algorithm (Nelder and Mead 1965), to
other statistics, such as the extreme values (low ex- optimize them simultaneously. The Nash–Sutcliffe model
ceedance probabilities)—the lag-1 autocorrelation and efficiency (Nash and Sutcliffe 1970) was used as objec-
the number of wet days in the corrected datasets im- tive function. A difference with the application described
proved with respect to the uncorrected ones. An impor- in Hurkmans et al. (2008) is the fact that, in this study,
tant statistic for extreme peak flows for the Rhine basin is the model was calibrated for each of seven zones sepa-
that of 10-day precipitation sums (e.g., Shabalova et al. rately, including five of the largest tributaries. These
2003). Terink et al. (2008) analyzed those as well and zones are shown in Fig. 1. Calibration and validation
found for the reference dataset an improvement of ex- results for the entire basin and the main tributaries are
treme 10-day precipitation sums in the corrected data shown in Table 2.
compared to the uncorrected data. For the validation From Table 2 it appears that, especially for moun-
dataset, this improvement is smaller. tainous areas such as the Rhine upstream of Maxau and
the Neckar, performance indicators are relatively low.
b. Model calibration
This can have several reasons: first, the model was forced
The bias-corrected validation dataset is used to cali- by REMO model output, as was discussed previously.
brate VIC. Because the calibration process is computa- Differences between observations and this modeled
tionally quite demanding, we calibrated the model on a data are at times quite large, as was also pointed out by
relatively short period: 1994–97. For each simulation, Hurkmans et al. (2008), who compared the datasets
1993 is used to initialize the model. The remainder of the extensively. Even though the structural error or bias was
validation dataset (1998–2003) is used to validate the largely removed by the bias correction (see section 2a),
model. The model was calibrated in a similar way as the variation over time was not changed and significant
686 JOURNAL OF CLIMATE VOLUME 23

TABLE 2. Characteristics of observed streamflow and calibration and validation results for five of the main tributaries of the Rhine
(Ruhr, Lahn, Main, Mosel, and Neckar) and three locations along the main branch of the Rhine: Maxau, Andernach, and Lobith. Mean,
maximum, and mean annual maximum (Qmam) streamflow are all calculated for the period 1990–2005 (the extent of the database that is
available). For the calibration period (1994–97) and the validation period (1998–2003), the correlation coefficient (r), Nash–Sutcliffe
modeling efficiency (E), and the relative volume error (RVE) are shown.

Observed Calibration Validation

Subcatchment Station Qmean Qmax Qmam r E RVE r E RVE
— — (m3 s21) (m3 s21) (m3 s21) (2) (2) (%) (2) (2) (%)
Ruhr Hattingen 72.6 867.0 579.4 0.58 0.22 34.57 0.61 0.30 27.58
Lahn Kalkofen 46.0 587.4 373.0 0.80 0.59 213.37 0.61 0.27 27.31
Main Raunheim 175.2 1990.8 1053.3 0.81 0.65 27.25 0.69 0.42 27.95
Mosel Cochem 338.6 4008.9 2492.3 0.71 0.49 1.21 0.69 0.37 12.02
Neckar Rockenau 142.6 2105.4 1326.0 0.54 0.07 23.14 0.61 20.15 35.84
Rhine Maxau 1278.3 4427.8 3269.5 0.59 0.10 11.20 0.50 20.81 26.82
Rhine Andernach 2076.0 10460.0 7274.3 0.75 0.55 5.03 0.62 0.17 15.72
Rhine Lobith 2239.9 11774.6 7655.3 0.81 0.59 11.69 0.70 0.22 19.77

differences between observed and modeled precipitation period and the A2 scenario the 10th and 90th percentiles
time series persisted. Second, the complex topography in are also shown. Thus, Fig. 5 shows in a qualitative manner
the mountainous areas complicates hydrological model- the trends in average, minimum, and maximum stream-
ing because of the high amount of small-scale hetero- flow during the twenty-first century for all scenarios and
geneity. Third, there are some large surface reservoirs the reference period. A striking feature is the fact that
present in the Alpine part, such as the Bodensee, which there seems to be a decreasing trend both in minimum
are not included in the model but can damp peaks in the and maximum streamflow, particularly during the period
hydrograph considerably. 2052–2100. However, annual maxima are higher than
they are during the reference period throughout the
twenty-first century. From the simulated records, the
4. Results
years 1951 and 2001 are used to initialize the model for
The calibrated VIC model is now used to simulate the reference period and the scenarios, respectively. To
streamflow for the reference period and the three cli- be able to compare three periods of equal length, 2051 is
mate scenarios described in section 2c using the bias- also discarded, yielding three periods spanning 49 yr
corrected atmospheric data. Figure 5 shows annual time each. In sections 4b and 4d we investigate the differences
series of the annual mean, minimum, and maximum between the periods 2002–50 and 2052–2100 on one
streamflow at the basin outlet, Lobith. For the reference hand, and 1952–2000 on the other hand, in more detail.

FIG. 5. Time series of annual mean (middle solid lines), maximum (upper lines), and mini-
mum (lowest lines) daily streamflow at Lobith for all datasets. For the A2 scenario and the
reference situation, the shaded area represents area between the 10th and 90th percentiles.
1 FEBRUARY 2010 HURKMANS ET AL. 687

FIG. 6. Difference in annual mean precipitation between three climate scenarios for the period 2052–2100 and the reference situation
(1952–2000). Differences in (top) the winter season (DJF) and (bottom) the summer season (JJA). Positive values denote an increase
compared to the reference situation.

a. Spatial patterns of atmospheric variables Figure 6 shows precipitation increases in winter and de-
creases in summer for all scenarios. The largest increases
To help interpret the changes in streamflow that are in precipitation during the winter season are found in the
discussed later in this paper, spatial patterns of the dif- northern and western part, as well as the mountainous
ferences in precipitation, evapotranspiration, and air area of the Black Forest and the Vogues Mountains,
temperature are plotted in Figs. 6, 7, and 8, respectively. located along the southern part of the Rhine Valley. The
Temperature and precipitation are used to force the A2 scenario also shows increases in precipitation in the
VIC model and are bias corrected (section 3a), and Alpine part and in some areas in the eastern Rhine ba-
evapotranspiration is simulated by the model. Only sin, whereas the B1 scenario does not. The A1B scenario
differences between the period 2052–2100 and the ref- shows intermediate increases in those areas. In summer,
erence period are shown, both for the winter [December– the decrease in precipitation mainly occurs in the Alpine
February (DJF)] and summer season [June–August (JJA)]. part of the basin, the Neckar subbasin, and the southern
688 JOURNAL OF CLIMATE VOLUME 23

FIG. 7. As in Fig. 6, but for actual evapotranspiration.

part of the Mosel. The northern part, the Black Forest, narios, differences between those can be amplified or
and the Vogues area still show small increases in pre- reduced locally.
cipitation. Figure 6 shows the advantage of the em- The high spatial resolution of evapotranspiration as
ployed high-resolution climate scenarios: (relatively) plotted in Fig. 7 allows for a detailed analysis of regional
small-scale topographic features like the Black Forest differences. Changes in actual evapotranspiration (Fig. 7)
and the Vogues Mountains generally do not show up in are quite homogeneous in space during the winter season:
spatial analyses using a lower resolution (e.g., Shabalova the entire basin, except the upper Alpine part, experiences
et al. 2003). It should be noted that the spatial pattern of increases in evapotranspiration of 8–20 mm month21.
the applied precipitation is influenced by our bias cor- In summer, however, there are large regional differ-
rection. Because the correction was applied for each of ences. Large parts of the western part (the Mosel area),
134 subbasins separately, the overall spatial pattern is as well as the eastern part (the Main) and a small area in
‘‘drawn’’ toward that of the observations to some extent northern Switzerland, show small decreases in evapo-
(see Fig. 4). In addition, because the same exponent in transpiration. Ideally, to confirm that these decreases
Eq. (1) is applied to the reference dataset and the sce- are caused by water limitation, one would want to see a
1 FEBRUARY 2010 HURKMANS ET AL. 689

FIG. 8. As in Fig. 6, but for air temperature.

similar plot of the potential evapotranspiration. This, ilar spatial pattern as temperature, this suggests that the
however, does not belong to the standard VIC model decreasing evapotranspiration values in summer could
output. Because of the parameterizations for subgrid indeed be caused by water limitation. The latter is also
variability in the model (section 2b), it is not trivial to indicated by results of Hagemann et al. (2009), who
extract this information, therefore we used air tempera- considered climate change simulations of REMO with
ture as a proxy for potential evapotranspiration (Fig. 8). about 50-km horizontal resolution.
Radiation would be another option to use as a proxy but
b. Mean streamflow
we do not expect the spatial pattern of incoming radia-
tion to be very different from that of air temperature. Figure 9 shows the climatologies of relative differ-
The spatial pattern for temperature also looks quite ences in streamflow between each of the two future
homogeneous for both seasons, albeit the warming is periods and the reference period, at four locations in the
somewhat higher in the southern part, the valley of the Rhine basin, representing a small (Ruhr) and a large
upper Rhine, and the higher areas on either side. As- (Main) tributary, the Alpine part (Maxau), and the en-
suming that potential evapotranspiration follows a sim- tire basin (Lobith). The lower panel of Fig. 9 shows the
690 JOURNAL OF CLIMATE VOLUME 23

FIG. 9. Climatologies of the difference in streamflow between three climate scenarios (A2, A1B, and B1) and the reference situation at
four locations in the Rhine basin: Hattingen (Ruhr), Raunheim (Main), Maxau (upper Rhine), and Lobith (entire Rhine). Overall
averages are indicated with crosses and triangles at the y axes. Dash–dotted lines and triangles indicate changes between the period 2002–50
and the reference period, whereas solid lines and crosses indicate changes between the period 2052–2100 and the reference period.
Positive values denote an increase in streamflow. (bottom) The (absolute) climatology of streamflow at Maxau, where the black line is the
reference climatology.

absolute streamflow climatology for Maxau, which is dom- causes increased discharge in these seasons. The in-
inated by snowmelt. From this plot, it appears that during crease in winter and spring nearly offsets the decrease in
the period 2002–50 the maximum monthly discharge at summer: annual streamflow slightly (,10%) increases
Maxau, which occurs in June, does not change or slightly for the A2 and B1 scenarios. For the A1B scenario the
increases (for the A1B scenario). The start of the melt- annual streamflow slightly decreases (;2%). The pro-
ing season is slightly shifted to occur earlier in the year. nounced seasonal cycle that is visible at Maxau also
During the second half of the twenty-first century, how- dominates the signal at Lobith, albeit slightly damped.
ever, the melting season starts about 6–8 weeks earlier During the period 2052–2100, the maximum decrease in
and maximum discharge is significantly lower than during streamflow (in August) is about 30% for the A1B sce-
the reference period. In late summer (July to September), nario, which is again the most extreme, whereas the A2
therefore, discharge is reduced considerably. The largest scenario shows the strongest increase in April (also about
difference with the reference period is reached in August 30%). Similar to Maxau, there is an increase in stream-
and amounts to about 40% less discharge at Maxau. The flow during the period 2002–50 compared to the reference
A1B scenario shows the most severe decrease in dis- period, even in the summer months. In the tributaries, the
charge. The earlier snowmelt in combination with in- Ruhr (Hattingen) and the Main (Raunheim), the sea-
creased precipitation in winter and spring (Figs. 10 and 11) sonal cycle is less pronounced. In the Ruhr, streamflow
1 FEBRUARY 2010 HURKMANS ET AL. 691

FIG. 10. Climatologies of eight water balance components and air temperature, averaged over the entire basin. The reference period and
three climate scenarios (A2, A1B, and B1) are shown for the periods 2002–50 (dash–dotted lines) and 2052–2100 (solid lines).

increases throughout the year with 10%–20% for both c. Water balance components
the first and second half of the twenty-first century. In the
Main, finally, also an overall increase in streamflow can Figure 10 shows average monthly values of some of
be seen for the period 2002–50. The B1 scenario espe- the main water balance terms, averaged over the entire
cially shows a strong increase (up to 30%) in late sum- basin. To show the differences in more detail, the dif-
mer. In the period 2052–2100, however, also the Main ferences between the scenarios and the reference period
shows increasing streamflow in winter and spring, except are shown in Fig. 11. In addition, Table 3 shows annual
February and March, and decreasing streamflow in values of all water balance terms. As can be expected,
summer and autumn. The overall increase in streamflow precipitation shows a similar pattern as streamflow
in the Ruhr and Main subbasins is in accordance with (Fig. 9): overall increases during the period 2002–50, and
Fig. 6: the northern part of the Rhine basin, including the a decrease in the summer season during the period 2052–
Main and Ruhr, is dominated by increasing precipitation 2100 with about 15%. This decrease in precipitation is
throughout the year, although very little in summer. The amplified to the 30% decrease in streamflow (Fig. 9) by
severe decrease in precipitation in the summer season, increased evapotranspiration. This increases with about
therefore, only affects the southern tributaries, espe- 8% for the A1B and A2 scenarios during the second half
cially the Neckar and the Alpine part of the Rhine, as of the century. During the first half, this increase is still
can be seen in the lower two panels of Fig. 9. relatively modest with annual increases of about 2%. In
692 JOURNAL OF CLIMATE VOLUME 23

FIG. 11. Climatology differences between the scenarios and the reference period for eight water balance components and temperature,
averaged over the entire basin. Positive values denote increases. Differences between the period 2002–50 and the reference period are
indicated by dash–dotted lines and triangles (for annual means); differences between the period 2052–2100 and the reference period are
indicated by solid lines and crosses (for annual means).

both cases, evapotranspiration changes throughout the of the century. Whereas there is still an increase in an-
year, except for a short period in March and April. This nual values of baseflow during the first half of the cen-
pattern is similar to that of air temperature, indicating tury, during the period 2052–2100 baseflow decreases
that on average evapotranspiration remains mainly en- throughout the year and most severely in (late) summer.
ergy limited, which is not surprising given the overall This is related to the increase in evapotranspiration
increases in streamflow and precipitation. The seasonal during the second half of the century, which reduces the
effect in precipitation seems to disappear to a large ex- amount of water infiltrating and contributing to base-
tent. Whereas during the reference period the highest flow. Snowmelt is not a standard output variable of the
monthly precipitation sums occur in summer and the VIC model. It is, therefore, estimated by taking the daily
lowest in March, precipitation is distributed more equally difference of the snow water equivalent (SWE). If this
over the months in all scenarios. This causes the large difference is positive, accumulation has occurred and
seasonal differences (increase in spring, decrease in melt is assumed to be zero. We realize that this as-
summer) that are observed in Fig. 11. sumption is not always valid (melt and accumulation can
Surface runoff, as expected, closely follows the cli- occur simultaneously), but the resulting melt will be
matology of precipitation and baseflow illustrates the approximately right. Note that in the VIC model snow-
transition to dryer circumstances during the second half melt is added to precipitation before infiltration and runoff
1 FEBRUARY 2010 HURKMANS ET AL. 693

TABLE 3. Annual values of water balance fluxes (mm yr21) and storages (mm) for the reference situation and three climate scenarios
(A2, A1B, B1) for the first half of the twenty-first century (2002–50) and the second half (2052–2100).

1952–2000 2002–50 2052–2100

Ref A2 A1B B1 A2 A1B B1
Precipitation 968 1015 1082 1071 1073 1021 1050
Air temperature 8.06 8.89 8.99 8.75 11.34 11.52 10.21
Evapotranspiration 411 422 432 426 494 498 463
Interception evaporation 16.8 12.4 12.5 16.3 12.2 12.5 11.0
Bare soil evaporation 150 153 155 147 174 177 164
Transpiration 284 289 299 295 327 325 314
Surface runoff 277 313 373 349 339 327 332
Baseflow 271 274 269 290 238 197 253
Soil moisture 900 909 940 943 907 917 922
Snow water equivalent 61.1 55.7 60.7 56.6 44.5 40.3 50.6
Snowmelt 113 100 107 110 69 62 80

are calculated. It is, therefore, included in the baseflow the twentieth century are displayed that each contain
and surface runoff fluxes. For the period 2002–50, snow- 49 yr (1901–49 and 1950–98), similar to the other curves
melt increases in early summer compared with the ref- in the graph. For both the periods 2002–50 and 2052–
erence period. SWE also increases in spring and early 2100, annual maxima are higher than the reference sit-
summer, but only for the A1B scenario. This is in ac- uation for all three scenarios at nearly all return periods.
cordance with higher precipitation during the spring During the first half of the century, the A1B and B1
season (Fig. 11), as a portion of this precipitation falls as scenarios are both higher than the A2 scenarios, whereas
snow. Because the difference in air temperature is larger this reversed during the second half. During both pe-
for summer than for spring during the period 2002–50, riods the most extreme events are simulated under the
the excess snow from the spring season melts in summer. A1B scenario, with the highest peak reaching nearly
The annual values shown in Table 3 further illustrate 16 000 m3 s21. The most extreme event in the reference
the changes described above. In all scenarios precipita- simulation, about 11 000 m3 s21 (comparable to the
tion increases quite drastically already during the period near floods in 1993 and 1995, which appear as the most
2002–50 (on average 88 mm, which is similar to the in- extreme events in the ‘‘1950–98’’ curve), will occur on
crease in the period 2052–2100), whereas evapotrans- average every 5–6 yr during both halves of the twenty-
piration increases mainly during the second half of the first century for the most extreme scenarios. An event of
century. This is correlated with air temperature, which the magnitude of 1926 (the most extreme event in the
also shows a modest increase during the first half of the ‘‘1901–49’’ curve) would occur every 10–20 yr according
century and a significantly larger increase during the to the climate scenarios. During the first half of the
second half. This can also be seen in the annual values century the most extreme events seem to be larger than
for snowmelt and SWE, which show a larger decrease during the second half of the century, which is also
during the second half of the century. suggested by Fig. 5. In addition, it shows that either the
hydrological model or the precipitation in the reference
d. Extreme streamflow
dataset structurally overestimates events with low to
While the average changes in streamflow discussed in medium return periods. For the extreme events, models
section 4b provide useful information, extreme stream- and observations agree well.
flow events are more relevant from a water management A similar analysis is carried out for streamflow
perspective. In this section, both peak flows and low droughts. Several approaches exist to define a drought
flows are discussed. Figure 12 shows annual maxima as a (e.g., Tallaksen and van Lanen 2004; Fleig et al. 2006).
function of their return times for streamflow at Lobith. Because not all annual low-flow events reflect a drought,
The parameters of the generalized extreme value (GEV) we define a streamflow drought as a cumulative deficit
distribution are estimated using the maximum likeli- of streamflow below a threshold. At the moment the
hood method. To put these results in some perspective, threshold is exceeded, the event ends (Hisdal et al. 2004).
also the annual maxima from the observed record at The threshold was selected to be the 75th percentile of
Lobith [available from the Royal Netherlands Meteoro- streamflow (meaning the streamflow that is equaled or
logical Institute (KNMI) Web site: http://www.knmi.nl] exceeded for 75% of the time) during the reference
are shown. For comparison, two different periods from period (Hisdal et al. 2004). This value is a compromise
694 JOURNAL OF CLIMATE VOLUME 23

FIG. 12. Annual maximum discharge at Lobith vs its return period for the reference period (1952–2000) and three climate scenarios.
Both the periods (left) 2002–50 and (right) 2052–2100 are displayed. A GEV distribution is adjusted to the data. In addition, two periods of
49 yr from the observed record are shown (1901–49 and 1950–98).

between the number of years without any events and the every 3–10 yr depending on the scenario. The most ex-
number of multiyear events, both of which can affect the treme events during the period 2052–2100 occur in the
results of the analysis (Fleig et al. 2006). Figure 13 shows A1B scenario and have a drought deficit volume that is
the annual maximum deficit volume of such events versus more than double the volume during the reference pe-
their return periods. A generalized Pareto (GP) distri- riod. Table 4 shows the streamflow drought events in
bution is a suitable distribution to model such cumulative more detail. Because statistics are shown for different
deficits (Tallaksen et al. 2004; Fleig et al. 2006). There- locations in the basin (Fig. 1), it also shows which regions
fore, such a distribution is adjusted to the data in Fig. 13, are most susceptible to streamflow droughts. In Table 4,
again using maximum likelihood estimation. Additional it can be seen that there are indeed significant regional
streamflow drought statistics are shown in Table 4. Here, differences within the basin. While the overall trend
three statistics are selected: the average number of events confirms the trends shown in Fig. 13, there are some
per year, the average annual maximum duration, and exceptions. For example, in the Mosel catchment, num-
the average annual maximum intensity. The intensity is ber, duration, and intensity all keep decreasing during
defined as the deficit volume (of which the annual max- the twenty-first century. This seems in contradiction with
ima are plotted in Fig. 13), divided by the duration in Fig. 7. Here it can be seen that a large part of the Mosel
seconds. The intensity (shown in m3 s21) can thus be seen catchment experiences a decrease in summer evapo-
as the average streamflow deficit during the event. transpiration, supposedly because of water limitation. In
Figure 13 shows that, during the first half of the cen- winter, however, there is also an increase in precipita-
tury, streamflow drought deficits are lower than the tion. An explanation for this contradiction is the fact that
reference in all scenarios for small return periods. There the winter precipitation rapidly percolates to the ground-
are, however, some extreme events that are consider- water, where the water is out of reach for plants. Soil
ably higher than the extreme events during the reference moisture droughts thus occur because evapotranspira-
period, especially in the A2 scenario. For the second half tion is water limited. However, the groundwater sustains
of the century the picture is entirely different: for nearly relatively high baseflow levels, preventing streamflow
all return periods and all three scenarios streamflow droughts (Tallaksen and van Lanen 2004).
drought deficits are larger than in the reference simu- From Table 4, the Neckar basin appears to be the most
lation. The most extreme streamflow drought deficit in susceptible to streamflow droughts: especially for the A1B
the reference simulation, about 40 3 108 m3, occurs scenario, the average maximum duration increases from
1 FEBRUARY 2010 HURKMANS ET AL. 695

FIG. 13. Annual maximum cumulative deficit of streamflow with respect to a threshold vs its return period for the reference period
(1952–2000) and three climate scenarios. The periods (left) 2002–50 and (right) 2052–2100 are displayed. The threshold is selected to be
the 75th percentile at Lobith. A GP distribution is adjusted to the data.

29 to 81 days (the overall highest value) and the intensity structurally differs between GCMs. For example, the
rises from 29 to 60 m3 s21, which is also the highest value GCM used in this study, ECHAM5–MPI-OM, is known
relative to the threshold streamflow (which is only to give relatively strong temperature increases in the
108 m3 s21). The A1B scenario shows the highest values second half of the twenty-first century. As a model com-
of drought statistics for most subbasins and the entire parison in Adger et al. (2007) shows, ECHAM5–MPI-
basin, which corresponds to Fig. 13. Figures 6 and 7 OM is relatively ‘‘cool’’ compared to the multimodel
explain the fact that the Neckar subbasin appears to mean in the first half of the century and relatively warm in
be extremely vulnerable to streamflow drought events the second half. The reason is a relatively high sensitivity
through the combination of strongly decreasing summer of ECHAM5–MPI-OM to aerosol concentrations: in the
precipitation, only weakly increasing winter precipita- first half of the century aerosol concentrations are high
tion, and increasing summer evapotranspiration. The and their dimming effect is relatively strong. In the sec-
same holds for the Alpine area, represented by Maxau. ond half of the century, the model reacts strongly to the
Because a large part of the Rhine discharge originates decline of aerosol concentrations, causing a strong tem-
upstream of Maxau, this also has a strong influence at perature increase. This feature partly explains the contrast
Lobith, the basin outlet. in our results between the periods 2001–50 and 2051–2100
(relatively wet versus relatively dry) and further illustrates
the need for a similar analysis with an ensemble of GCMs.
5. Discussion
The fact that in CMIP the ECHAM4 model is quite
Data from a single GCM, RCM, and hydrological close to observed values for global mean temperature
model were used, whereas ideally one would use an and precipitation (Fig. 1 of Covey et al. 2003) is en-
ensemble for each model type. Model comparisons in- couraging. Also, in model comparisons regarding other
dicate a large range in model projections between GCMs. climate variables, ECHAM is ranked among the best
For example, the Coupled Model Intercomparison Pro- performing GCMs (van Ulden and van Oldenborgh
ject (CMIP) compared globally averaged air tempera- 2006). In a recent model comparison, Reichler and Kim
tures for the period 1961–90s simulated by 18 GCMs and (2008) showed that ECHAM5–MPI-OM (which we use)
found a range of more than 48C around the observed 148C performs very well compared to most other GCMS in
(Covey et al. 2003). For precipitation, a similar range was simulating the current climate in terms of many different
found. Also the projected trend in global temperature climate variables. Furthermore, the fact that our results
696 JOURNAL OF CLIMATE VOLUME 23

TABLE 4. Statistics of streamflow drought events during the reference period and the first and second half of the twenty-first century for
the three climate scenarios (A2, A1B, and B1) for seven streamflow gauges in the Rhine basin. For locations and corresponding subbasins
see Fig. 1. Three statistics are shown, indicated by S1, S2, and S3. Here S1 is the average number of events (i.e., nonexceedance of the
threshold) per year; S2 is the average of annual maximum durations in days; S3 is the average of annual maximum intensities, where the
intensity is defined as deficit volume (plotted in Fig. 13) divided by duration, in m3 s21. For each station, the threshold value (75th
percentile of the reference period) is shown in parentheses.

Reference A2 scenario A1B scenario B1 scenario

Period S1 S2 S3 S1 S2 S3 S1 S2 S3 S1 S2 S3
3 21
Hattingen (37.8 m s )
2002–50 6.5 33 17.6 5.0 27 15.6 4.6 26 12.9
5.7 33 16.8
2052–2100 7.1 35 19.7 6.7 39 19.9 6.0 33 18.9
3 21
Kalkofen (17.0 m s )
2002–50 8.1 30 5.8 7.3 25 5.5 6.0 22 4.6
7.5 33 5.6
2052–2100 8.7 32 7.1 9.0 34 7.1 7.3 28 6.2
3 21
Raunheim (92.9 m s )
2002–50 5.4 31 17.7 5.0 25 16.5 4.0 23 14.7
6.5 34 22.5
2052–2100 6.2 37 19.2 5.7 40 17.8 6.0 35 17.3
3 21
Cochem (165.9 m s )
2002–50 6.1 32 32.1 6.1 40 38.8 5.9 26 33.5
6.4 37 37.9
2052–2100 4.2 34 24.3 3.9 35 22.1 4.0 23 20.4
Rockenau (107.8 m3 s21)
2002–50 5.5 26 23.5 7.4 25 27.2 4.1 24 21.6
7.0 29 29.1
2052–2100 9.5 44 40.0 11.8 81 60.1 8.4 42 36.8
3 21
Maxau (986.9 m s )
2002–50 4.1 30 217 4.1 20 204 4.1 23 173
4.8 33 209
2052–2100 7.2 33 334 8.7 38 420 5.8 28 259
3 21
Lobith (1801 m s )
2002–50 5.8 27 424 5.4 21 400 5.0 21 363
6.8 32 475
2052–2100 7.7 38 614 8.8 46 722 6.7 33 510

qualitatively confirm results of previous studies for the Fig. 4, the spatial pattern from the RCM is ‘‘drawn’’
Rhine basin using other or multiple GCMs suggests that toward that of the observations. In addition, the differ-
the GCM we used is broadly representative. The range ence between the reference dataset and the scenarios
of temperature changes that is projected by GCMs is can be amplified or reduced somewhat because the same
further extended by the RCMs that are used for down- exponent [Eq. (1)] is applied to both. When the basin-
scaling (e.g., van den Hurk et al. 2005; Jacob et al. 2007; wide spatial average is considered, this effect largely
Déqué et al. 2007). In addition to the uncertainty in at- cancels out. This feature, however, is inherent to our
mospheric models, there is also uncertainty involved in correction for the model bias, which can be seasonally
hydrological models and their parameterizations. How- and geographically varying (e.g., Christensen et al. 2008;
ever, given the fact that each GCM and RCM requires a Leander and Buishand 2007). The above should be kept
specific bias correction (associated with different struc- in mind when interpreting the spatial patterns as pre-
tural model errors), and each combination of a hydro- sented in this paper or comparing them with uncorrected
logical model with an atmospheric dataset would require climate scenarios. Furthermore, while we assume that
a specific calibration, it was not feasible in the present the model biases are similar for the current climate and
study to use multiple models. This is a common problem the changing climate, Christensen et al. (2008) found
in hydrological climate change impact studies: most that this is not always valid and biases tend to grow
studies involving a similar model chain as the present under warming conditions and depend on the amount of
study also take into account only one (e.g., Dankers and precipitation, suggesting an overestimation of the tem-
Feyen 2008; Akhtar et al. 2009) or very few (e.g., de Wit perature increase.
et al. 2007; Booij 2005) GCMs. Finally, it is important to mention that the presented
The bias correction that was applied to the meteoro- analysis of hydrological extremes is based on relatively
logical data altered the spatial pattern of precipitation short time series (3 3 100 yr), whereas the Rhine dis-
and temperature because the correction was applied to charge at Lobith that is relevant for safety purposes has
each of 134 subbasins for which observations were a return period of 1250 yr (de Wit and Buishand 2007).
available. As was noted in section 4a and can be seen in The highest occurring streamflow in this study (about
1 FEBRUARY 2010 HURKMANS ET AL. 697

16 000 m3 s21) is associated with much shorter return more extreme droughts. Partly, this contrast originates
periods and therefore does not imply that higher values from our choice of the GCM (see section 5). Ideally, to
are not possible. Moreover, in this study river hydraulics verify this it would be necessary to use an ensemble of
and flooding are not taken into account, whereas flooding GCMs. The fact that the employed GCM simulates the
in Germany would have a significant impact on peak current climate realistically and its results are broadly
discharge at Lobith (de Wit and Buishand 2007). comparable to results from other studies, however, sug-
gests that the ECHAM5–MPI-OM model is represen-
tative, and results from a multimodel analysis would not
6. Conclusions
be completely different. Nevertheless, it would be very
The present study differs from previous climate change interesting to repeat this analysis with another, or mul-
impact assessments in the Rhine basin regarding the fol- tiple, GCMs.
lowing points: 1) the spatial resolution of the employed
climate scenarios is higher, allowing for a detailed spatial Acknowledgments. This research was financially sup-
analysis of the results and a better representation of ported by the European Commission through the FP6
extremes; 2) the hydrological model is more physically integrated Project NeWater and the BSIK ACER pro-
based; and 3) the entire twenty-first century was in- ject of the Dutch Climate Changes Spatial Planning
vestigated as opposed to a time slice of typically 30 yr. Programme. Rita Lammersen and Hendrik Buiteveld
As far as streamflow at Lobith is concerned, this study from Rijkwaterstaat Waterdienst, Lelystad, Nether-
confirms previous climate change impact assessments lands, are kindly acknowledged for providing observed
in the Rhine basin. By the end of the century, average streamflow data, and the international Commission for
streamflow at basin outlet is projected to increase by the Hydrology of the Rhine basin (CHR) is acknowl-
about 30% in winter and spring and to decrease by a edged for making available the observed precipitation
similar percentage in summer and autumn. In addition, and temperature. Finally, we thank the two anonymous
considerable increases in the frequency and magnitude reviewers for their helpful and constructive comments.
of both peak flows and streamflow droughts are pro-
jected. The majority of this effect originates in the Al- REFERENCES
pine part of the catchment. It is caused largely by a Adger, N., and Coauthors, 2007: Summary for policy makers. Cli-
decrease of the contribution of snowmelt to streamflow mate Change 2007: Impacts, Adaptation and Vulnerability,
and by a shift of the timing of the snowmelt season to M. L. Parry et al., Eds., Cambridge University Press, 7–22.
Akhtar, M., N. Ahmad, and M. J. Booij, 2009: Use of regional
earlier in the year.
climate model simulations as input for hydrological models for
The high-resolution climate scenarios and the (dis- the Hindukush–Karakorum–Himalaya region. Hydrol. Earth
tributed) hydrological model allow for a detailed analysis Syst. Sci., 13, 1075–1089.
of spatial patterns. Results indicate that the projected Barnett, T. P., J. C. Adam, and D. P. Lettenmaier, 2005: Potential
decrease in summer discharge is concentrated in the impacts of a warming climate on water availability in snow-
southern part of the basin. Over the (relatively small) dominated regions. Nature, 438, 303–309, doi:10.1038/
nature04141.
Black Forest and the Vogues Mountains, however, pre- Booij, M. J., 2005: Impact of climate change on river flooding as-
cipitation increases during the entire year. The northern sessed with different spatial model resolutions. J. Hydrol., 303,
tributaries, where there is hardly any decrease in sum- 176–198.
mer precipitation, are dominated by wetter conditions Christensen, J. H., T. R. Carter, M. Rummukainen, and
G. Amanatidis, 2007: Evaluating the performance and utility
throughout the year.
of regional climate models: The PRUDENCE project. Cli-
The availability of data for the entire century allows matic Change, 81, 1–6, doi:10.1007/s10584-006-9211-6.
the analysis of trends throughout the century, of which ——, F. Boberg, O. B. Christensen, and P. Lucas-Picher, 2008: On
an overview is displayed in Fig. 5. Toward the end of the need for bias correction of regional climate change pro-
the twenty-first century, a gradually decreasing trend in jections of temperature and precipitation. Geophys. Res. Lett.,
annual maximum, minimum, and mean streamflow ap- 35, L20709, doi:10.1029/2008GL035694.
Covey, C., K. M. AchutaRao, U. Cubasch, P. Jones, S. J. Lambert,
pears. This trend causes an interesting contrast between M. E. Mann, T. J. Philips, and K. E. Taylor, 2003: An overview
the first half and the second half of the century. Because of results from the Coupled Model Intercomparison Project.
precipitation significantly increases in all three scenar- Global Planet. Change, 37, 103–133.
ios, the first half is dominated by relatively wet condi- Dankers, R., and L. Feyen, 2008: Climate change impact on flood
tions. This also shows up in our analysis of peak flows. hazard in Europe: An assessment based on high-resolution
climate simulations. J. Geophys. Res., 113, D19105, doi:10.1029/
During the second half of the century, temperature and 2007JD009719.
evapotranspiration increase more drastically than dur- Déqué, M., and Coauthors, 2007: An intercomparison of regional
ing the first half, causing generally dryer conditions and climate simulations for Europe: Assessing uncertainties in
698 JOURNAL OF CLIMATE VOLUME 23

model projections. Climatic Change, 81, 53–70, doi:10.1007/ Leander, R., and T. A. Buishand, 2007: Resampling of regional
S10584-006-9228-x. climate model output for the simulation of extreme river flows.
de Wit, M. J. M., and T. A. Buishand, 2007: Generator of rainfall J. Hydrol., 332, 487–496, doi:10.1016/j.jhydrol.2006.08.006.
and discharge extremes (GRADE) for the Rhine and Meuse Lenderink, G., T. A. Buishand, and W. P. van Deursen, 2007: Es-
basins. Tech. Rep., Rijkswaterstaat RIZA Rep. 2007.027/ timates of future discharges of the river Rhine using two sce-
KNMI publication 218, Lelystad, Netherlands, 77 pp. nario methodologies: Direct versus delta approach. Hydrol.
——, B. van den Hurk, P. M. M. Warmerdam, P. J. J. F. Torfs, Earth Syst. Sci., 11, 1145–1159.
E. Roulin, and W. P. A. van Deursen, 2007: Impact of climate Liang, X., D. P. Lettenmaier, E. F. Wood, and S. J. Burges, 1994: A
change on low-flows in the river Meuse. Climatic Change, 82, simple hydrologically based model of land surface water and
351–372, doi:10.1007/S10584-006-9195-2. energy fluxes for general circulation models. J. Geophys. Res.,
Fleig, A. K., L. M. Tallaksen, H. Hisdal, and S. Demuth, 2006: 99 (D7), 14 415–14 458.
A global evaluation of streamflow drought characteristics. ——, and Coauthors, 1998: The Project for Intercomparison of
Hydrol. Earth Syst. Sci., 10, 535–552. Land-surface Parameterization Schemes (PILPS) phase 2(c)
Graham, L. P., S. Hagemann, S. Jaun, and M. Beniston, 2007: On Red-Arkansas River basin experiment: 2. Spatial and tem-
interpreting hydrological change from regional climate models. poral analysis of energy fluxes. Global Planet. Change, 19,
Climatic Change, 81, 97–122, doi:10.1007/s10584-006-9217-0. 137–159, doi:10.1016/S0921-8181(98)00045-9.
Hagemann, S., H. Göttel, D. Jacob, P. Lorenz, and E. Roeckner, Lohmann, D., R. Nolte-Holube, and E. Raschke, 1996: A large-
2009: Improved regional scale processes reflected in projected scale horizontal routing model to be coupled to land surface
hydrological changes over large European catchments. Cli- parameterization schemes. Tellus, 48A, 708–721.
mate Dyn., 32, 767–781, doi:10.1007/s00382-008-0403-9. ——, E. Raschke, B. Nijssen, and D. P. Lettenmaier, 1998a: Re-
Hamlet, A. F., P. W. Mote, M. P. Clark, and D. P. Lettenmaier, gional scale hydrology: I. Application of the VIC-2L model
2007: Twentieth-century trends in runoff, evapotranspiration coupled to a routing model. Hydrol. Sci. J., 43, 131–141.
and soil moisture in the western United States. J. Climate, 20, ——, ——, ——, and ——, 1998b: Regional scale hydrology: II.
1468–1486. Application of the VIC-2L model to the Weser River, Ger-
Hay, L. E., M. P. Clark, R. L. Wilby, W. J. Gutowski Jr., many. Hydrol. Sci. J., 43, 143–158.
G. H. Leavesly, Z. Pan, R. W. Arritt, and E. S. Takle, 2002: ——, and Coauthors, 1998c: The Project for Intercomparison
Use of regional climate model output for hydrologic simula- of Land-surface Parameterization Schemes (PILPS) phase
tions. J. Hydrometeor., 3, 571–590. 2(c) Red-Arkansas River basin experiment: 3. Spatial and
Hisdal, H., L. M. Tallaksen, B. Clausen, E. Peters, and A. Gustard, temporal analysis of water fluxes. Global Planet. Change, 19,
2004: Hydrological drought characteristics. Hydrological 161–179.
Drought: Processes and Estimation Methods for Streamflow and Matheussen, B., R. L. Kirschbaum, I. A. Goodman, G. M. O’Donnell,
Groundwater, Elsevier, 139–198. and D. P. Lettenmaier, 2000: Effects of land cover change on
Hurkmans, R. T. W. L., H. de Moel, J. C. J. H. Aerts, and streamflow in the interior Columbia River basin (U.S.A. and
P. A. Troch, 2008: Water balance versus land surface model in Canada). Hydrol. Processes, 14, 867–885.
the simulation of Rhine river discharges. Water Resour. Res., Middelkoop, H., and Coauthors, 2001: Impact of climate change on
44, W01418, doi:10.1029/2007WR006168. hydrological regimes and water resources management in the
——, W. Terink, R. Uijlenhoet, E. J. Moors, P. A. Troch, and Rhine basin. Climatic Change, 49, 105–128.
P. H. Verburg, 2009: Effects of land use changes on streamflow Monteith, J. L., 1965: Evaporation and environment. Proceedings
generation in the Rhine basin. Water Resour. Res., 45, of the 19th Symposium of the Society for Experimental Bi-
W06405, doi:10.1029/2008WR007574. ology, Cambridge University Press, 205–233.
Jacob, D., 2001: A note to the simulation of the annual and inter- Mücher, S., K. Steinnocher, J.-L. Champeaux, S. Griguolo,
annual variability of the water budget over the Baltic Sea K. Wester, C. Heunks, and V. van Katwijk, 2000: Establish-
drainage basin. Meteor. Atmos. Phys., 77, 61–73. ment of a 1-km Pan-European Land Cover database for en-
——, and Coauthors, 2007: An inter-comparison of regional climate vironmental monitoring. Proceedings of the Geoinformation
models for Europe: Model performance in present-day climate. for All XIXth Congress of the International Society for Pho-
Climatic Change, 81, 31–52, doi:10.1007/S10584-006-9213-4. togrammetry and Remote Sensing (ISPRS), K. J. Beek and
——, H. Göttel, S. Kotlarski, P. Lorenz, and K. Sieck, 2008: Kli- M. Molenaar, Eds., International Archives of Photogramme-
maauswirkungen und Anpassung in Deutschland—Phase 1: try and Remote Sensing, Vol. 33, GITC, 702–709.
Erstellung regionaler Klimaszenarien für Deutschland (Cli- Nakićenović, N., and R. Swart, Eds., 2000: Special Report on
mate impact and adaptation in Germany—Phase 1: Produc- Emissions Scenarios. Cambridge University Press, 612 pp.
tion of regional climatic scenarios for Germany). Research [Available online at http://www.grida.no/publications/other/
Rep. 204 41 138 UBA-FB 000969, Final Rep. to the UFO plan ipcc_sr/?src5/climate/ipcc/emission/index.htm.]
project, Federal Office for Environment Protection, Dessau, Nash, J. E., and I. V. Sutcliffe, 1970: River flow forecasting through
Germany, 154 pp. conceptual models. Part I—A discussion of principles. J. Hy-
Kleinn, J., C. Frei, J. Gurtz, D. Lüthi, P. L. Vidale, and C. Schär, drol., 10, 282–290.
2005: Hydrologic simulations in the Rhine basin driven by Nelder, J. A., and R. Mead, 1965: A simplex method for function
a regional climate model. J. Geophys. Res., 110, D04102, minimalization. Comput. J., 7, 308–313.
doi:10.1029/2004JD005143. Penman, H. L., 1948: Natural evaporation from open water, bare
Kwadijk, J., and H. Middelkoop, 1994: Estimation of impact of soil and grass. Proc. Roy. Soc. London, A193, 120–146.
climate change on the peak discharge probability of the river Pfister, L., J. Kwadijk, A. Musy, A. Bronstert, and L. Hoffmann,
Rhine. Climatic Change, 27, 199–224. 2004: Climate change, land use change and runoff prediction
——, and J. Rotmans, 1995: The impact of climate change on the in the Rhine–Meuse basins. River Res. Appl., 20, 229–241,
river Rhine: A scenario study. Climatic Change, 30, 397–425. doi:10.1002/rra/775.
1 FEBRUARY 2010 HURKMANS ET AL. 699

Pitman, A. J., and Coauthors, 1999: Key results and implica- ——, H. Madsen, and H. Hisdal, 2004: Frequency analysis. Hydro-
tions from Phase 1(c) of the Project for Intercomparison of logical Drought: Processes and Estimation Methods for Stream-
Land-surface Parameterization Schemes. Climate Dyn., 15, flow and Groundwater, L. M. Tallakesen and H. A. J. van Lanen,
673–684. Eds., Developments in Water Science, Vol. 48, Elsevier, 199–
Reichler, T., and J. Kim, 2008: How well do coupled models sim- 271.
ulate today’s climate? Bull. Amer. Meteor. Soc., 89, 303–311. Terink, W., R. T. W. L. Hurkmans, R. Uijlenhoet, P. M. M.
Reynolds, C. A., T. J. Jackson, and W. J. Rawls, 2000: Estimating Warmerdam, and P. J. J. F. Torfs, 2008: Bias correction of
water-holding capacities by linking the Food and Agriculture temperature and precipitation data for regional climate model
Organization soil map of the world with global pedon data- application to the Rhine basin. Tech. Rep., Hydrology and
Quantitative Water Management group, Wageningen Uni-
bases and continuous pedotransfer functions. Water Resour.
versity, Wageningen, Netherlands, 42 pp.
Res., 36, 3653–3662.
Trenberth, K. E., A. Dai, R. M. Rasmussen, and D. B. Parsons,
Shabalova, M. V., W. P. A. van Deursen, and T. A. Buishand, 2003:
2003: The changing character of precipitation. Bull. Amer.
Assessing future discharge of the river Rhine using regional
Meteor. Soc., 84, 1205–1217.
climate model integrations and a hydrological model. Climate
van den Hurk, B., and Coauthors, 2005: Soil control on runoff re-
Res., 23, 233–246. sponse to climate change in regional climate model simula-
Sprokkereef, E., 2001: Eine hydrologische Datenbank für das tions. J. Climate, 18, 3536–3551.
Rheingebiet (A hydrological database for the Rhine catch- van Ulden, A. P., and G. J. van Oldenborgh, 2006: Large-scale
ment). International Commission on Hydrology of the Rhine atmospheric circulation biases and changes in global climate
Basin (CHR) Rep., 22 pp. model simulations and their importance for climate change in
Tallaksen, L. M., and H. A. J. van Lanen, Eds., 2004: Hydrological Central Europe. Atmos. Chem. Phys., 6, 863–881.
Drought: Processes and Estimation Methods for Streamflow Zierl, B., and H. Bugmann, 2005: Global change impacts on hy-
and Groundwater. Developments in Water Science, Vol. 48, drological processes in Alpine catchments. Water Resour. Res.,
Elsevier, 579 pp. 41, W02028, doi:10.1029/2004WR003447.