a. The PRUDENCE simulations

In the framework of the European project PRUDENCE (see Christensen et al. 2002) a number of RCM systems were used to downscale global climate simulations from a range of GCMs and greenhouse gas emission scenarios. The RCMs differed with respect to the physical and dynamical formulations, land-use characteristics, and the grid and domain in which the models were integrated, although in all cases the models covered the major part of Europe at approximately 50-km resolution. Daily and monthly averaged output from the RCM integrations was available in a well-maintained and accessible central PRUDENCE database. From the available RCM integrations a selection was made of RCMs that (i) all simulated both a control climate (1961–90) and an A2-scenario time slice (2071–2100) derived from a specific member of the ensemble of simulations with the Hadley Centre’s high-resolution atmospheric model (named HadAM3H); (ii) reported all variables that were needed to calculate the regional-scale hydrological budgets (precipitation, evaporation, runoff, snow, and soil water content); and (iii) produced a closed water balance (within 0.5% of annual precipitation) in the control simulation. From the nine RCMs meeting the first criterion, seven were used for further analysis: the Royal Netherlands Meteorological Institute (KNMI), the Danmarks Meteorologiske Institut (DMI), the Erdgenössische Technische Hochschule (ETH), the Max Planck Institute (MPI), the Swedish Meteorological and Hydrological Institute (SMHI), the GKSS Forschungszentrum (GKSS), and the Universidad Castilla la Mancha (UCM). Table 1 gives a brief overview of the specific properties of each of the RCMs.

All RCM model results are processed as time series of monthly averaged/accumulated hydrological budget terms, averaged for the whole Rhine basin. The budget terms are

View Expanded

where P is total precipitation, R is total runoff, E is evaporation, and S is terrestial storage (consisting of soil and snow water content). Storage changes with time t are calculated from daily soil water and snow fields at the start of each month, whereas the flux terms in (1) are monthly averaged quantities. All terms are expressed in millimeters per day.

The depth over which S is calculated varies across the RCMs. Some models (like MPI) have a relatively deep soil of which the lower part is usually saturated and which could be considered to represent the saturated groundwater zone. Other models (like UCM) have a shallow soil that is unsaturated most of the time. In all cases, however, the water budget represented by (1) is closed. Runoff is defined as the sum of the water flux that does not infiltrate into the soil (surface runoff) plus the net flux of water leaving the soil volume via its lower boundary (the drainage component). Here ∂S/∂t is calculated over the entire soil volume between the surface and the deep boundary.

Owing to the differences in the grid orientation and resolution of the models the number of grid points used to calculate a domain-averaged quantity varies between the models (see Table 1), but this difference is considered to be of minor importance in the analysis.

The RCMs were driven by data produced by HadAM3H (Jones et al. 2001). This model has a fairly high spatial resolution of approximately 150 km and includes various changes compared to its parent coupled GCM, the Third Hadley Centre Coupled Model (HadCM3), which improves its simulation of the surface climate over many land areas (including Europe). It the context of the experiments reported here, it was used to simulate the climate of the recent past driven by observed sea surface temperatures (SSTs) and sea ice (SI) for the period 1960–90 and observed concentrations of greenhouse gases and emissions of sulphate aerosols. It was also used to provide a projection of the climate of 2070–2100 driven by emissions defined in the Intergovernmental Panel on Climate Change (IPCC) A2 Synthesis Report on Emission Scenarios (SRES; Nakicenovic and Swart 2000). For this experiment the sea surface forcing was derived by adding changes of SST/SI from HadCM3 driven by the same emission scenario to the observed values. Some RCM groups used the first year of the HadAM3H runs to spin up their model and reported only the last 30 yr, whereas others simply discarded the first simulation year.

b. Analysis of the terrestrial water storage from observed large-scale quantities

Derivation of the storage term in (1) from direct observations of soil moisture and snow amounts is not feasible when spatial scales of the order of a large river basin are considered. However, S can be estimated when the remaining terms are known. Atmospheric analysis archives have been used in the past (e.g., Trenberth and Guillemot 1995) to estimate the total atmospheric water convergence ∇_HQ in an area, which is formulated as

View Expanded

where Q is the total water flux (in vapor, liquid, and frozen phases) and W is the total water column at a given time. At every analysis time step, Q is calculated from the total water vapor content q and the horizontal wind speed vector V from

View Expanded

where p₀ is the surface pressure and g is the gravity acceleration. Residual terms resulting from increments in the data assimilation process, as discussed by Roads et al. (2003), are included in W. By combining Eqs. (1) and (2), P − E can be eliminated. Estimation of Q from meteorological analysis fields is considered to be a better representation of the atmospheric budget than the indirect estimates of P and E from the analysis system, which are plain model forecasts only indirectly affected by the corrections that arise from the assimilation of observations. When the area of consideration matches the basin of a large river, the observed discharge from the river may be used in combination with the total convergence to infer the change of the terrestrial water storage. Seneviratne et al. (2004) demonstrated the success of this method for catchments half the size of the Rhine basin using estimates of ∇Q and ∂W/∂t derived from ERA-40. Here ∂S/∂t derived from (1) compared very well with observations from a high-resolution soil moisture, groundwater, and snow monitoring network in Illinois. Similar networks are lacking in the Rhine basin.

Hirschi et al. (2004, manuscript submitted to J. Hydrometeor., hereafter HSS) expanded on the work by Seneviratne et al. (2004) by calculating the terrestrial storage component for a number of large river basins across the world, including the Rhine. They used discharge data retrieved from the Global Runoff Data Centre (available online at http://grdc.bafg.de/servlet/is/Entry.987.Display/), which reports measured discharge amounts, not corrected for artificial reservoirs and waterworks. Kwadijk (1993) argues that water buffering in Swiss lakes has changed the mean annual cycle of the Rhine discharge noticeably, and refers to a study by Schädler (1985) to calculate the effect of lake storage on the Rhine discharge near Lobith, Netherlands. Average summertime discharge is increased by up to 9% owing to lake retention, whereas wintertime discharge is reduced by 3%–6%. The average annual cycle reported by Kwadijk (1993) has been used to correct the observed Rhine discharge data prior to calculating the terrestrial storage from the budget analysis.

Figure 2a shows a time series of the change of the terrestrial water storage in the Rhine basin, derived from (1) and using ∇Q + ∂W/∂t from ERA-40 and observed Rhine discharge. As pointed out by Seneviratne et al. (2004), the large-scale budget analysis suffers from errors that are introduced from imperfect atmospheric observation systems (affecting the ERA-40 convergence estimate) and observation errors in the discharge itself. Especially in steep orographic terrain, lack of overlap of the atmospheric convergence domain and the basin producing the discharge may lead to errors in the budget calculation owing to the relatively coarse resolution of the atmospheric analysis. Also, the atmospheric assimilation system that provides estimates of the atmospheric water convergence does not conserve water, which may also affect the water balance closure. Moreover, multiyear (re)analysis time series suffer from discontinuities in the observation network, and part of the interannual variability of ∂S/∂t shown in Fig. 2a is attributed to the introduction of satellite observations since the early 1970s and insufficient spatial coverage by the rawinsonde network prior to that period. Therefore, at least part of the interdecadal variability is likely to be artificial, and is removed by subtracting a running mean with a 3-yr window. The selection of this time scale is somewhat arbitrary, but it is situated in between the seasonal time scale of soil moisture evolution and the decadal time scale at which soil moisture ranges are considered to be relatively stationary. Use of a 5-yr window instead did not have a noticeable effect on the results reported in this study. Also using only data from 1970 onward did not significantly affect the mean annual cycles of the budget terms (Fig. 3). Subtracting a running mean can be considered as a practical means to removing apparent artificial components of the interdecadal variability while preserving as many data as possible.

The integration of the (filtered) time series in Fig. 2a results in a time series of the terrestrial storage range, shown in Fig. 2b. The position on the vertical scale is arbitrary, and defined by setting the initial soil water content at 0 mm.

An average annual cycle of the flux terms in Eq. (1) is shown in Fig. 3. In spite of a clear annual cycle of the atmospheric moisture convergence, river discharge from the Rhine shows a fairly gradual evolution. The timing of the springtime snowmelt peak varies between years, and this peak is smoothed in the 40-yr average shown in Fig. 3. The major part of the average annual cycle in moisture convergence is buffered in the terrestrial storage reservoir (mainly the soil and groundwater reservoirs), which as a consequence displays a pronounced cycle of drying and wetting.

Also shown in Fig. 3 is the net moisture supply to the soil (P − E) for the same area, deduced from the 30-yr time slice of the HadAM3H control simulation representative for the 1961–90 period. Wintertime values of modeled P − E are overestimated, whereas during summer a slight underestimation of net moisture supply to the Rhine basin is simulated, in particular later in the summer season. This bias is at least partly related to systematic overestimation of the frequency of blocked circulation patterns (van Ulden et al. 2004, manuscript submitted to Climate Dyn.).

a. Comparison to precipitation observations

A 35-yr high density precipitation database has been compiled by the International Rhine Commission (CHR). The available precipitation database contains daily values for the period 1960–95 for 136 subbasins in the Rhine catchment area. The daily values were spatially averaged weighted by the area of each of the subbasins. Figure 4 shows a comparison between the modeled and observed average annual cycle of precipitation over the Rhine basin for each of the RCMs involved.

Most models show a considerably larger than observed range in the annual cycle, with either a peak that is too pronounced in winter or early summer precipitation or a value that is too low by the end of the summer. This variability of RCM results (driven by equal lateral boundaries) is consistent with findings of other studies (e.g., Jacob et al. 2001; Frei et al. 2003). Räisänen et al. (2004) point at a positive bias in the wintertime north–south pressure gradient over Europe in the driving GCM HadAM3H in the control climate simulation. This bias is also present in most RCM simulations, and results in an increased eastward advection of water vapor, that probably is partly responsible for the positive wintertime precipitation bias.

On the other hand, the interannual variability in the observations (shown by the error bars representing one standard deviation of the monthly means) is significant, and the average RCM results generally fall within this range.

b. Net convergence and evaporation

Large-scale observations of evaporation are not available. In principle, an estimate could be deduced by combining the ERA-40-derived estimate of P − E with the CHR database shown before, but the accumulation of errors makes this estimate fairly unreliable. A direct comparison of P − E between the models and the estimates derived from ERA-40 is presented here in the left panel in Fig. 5. The ERA-40-derived estimate of P − E is calculated according to Eq. (2) (total convergence minus ∂W/∂t), the closest possible estimate of the water flux at the land–atmosphere interface. The model-only evaporation is plotted in the right panel in Fig. 5.

The models generally have the tendency to create an annual cycle of P − E that is too strong, with more convergence than observed in winter and less convergence (or more divergence) in summer. The wintertime overestimation of convergence is closely related to an overestimation of precipitation (Fig. 4). For some models the fairly good match with the observations is a consequence of compensating errors. For instance, the high summertime precipitation in MPI (Fig. 4) is effectively compensated by high evaporative loss. The lack of divergence in summertime in the UCM model is associated to fairly low evaporation amounts.

c. Discharge and runoff

The runoff generated by the RCMs is linked to the net water flux into the soil (P – E) and the amount of this flux that is buffered by the soil water range. The processes affecting the soil hydrological balance are mutually coupled since, in general, runoff generation depends on the soil water content. The partitioning of the net water influx over the storage change and runoff is an important property of the hydrological system, as it describes the fate of the water entering the soil: the water put into runoff is lost and cannot be re-evaporated locally, while the soil water content is a storage buffer for later evaporation or runoff generation.

Figure 6 shows the average annual cycle of the runoff generation in the Rhine basin. Also shown is the observed discharge at Lobith (the gauging station downstream of the catchment for which the modeled runoff is displayed).

Similar to the results shown in Fig. 5, the annual cycle of the modeled runoff is generally larger than the observations. The RCMs produce a fairly good wintertime runoff. Except for UCM (late) summertime runoff is too low and for some models well outside one standard deviation of the interannual variability. Also, the rate at which the runoff recovers from the summer dryness is generally much faster in the RCMs than in the observations. The ordering of models in terms of summertime runoff underestimation is roughly the same as the ordering in P − E but quite different from the signature in the precipitation bias (Fig. 4), indicating a strong soil control (via evaporation or storage) on the runoff characteristics.

Many RCMs have systematic deficiencies in the parameterization of summertime convective precipitation over mountains. Also, the representation of glacier melt is not included. Since more than 50% of the Rhine discharge at Lobith originates from the mountainous area upstream of Basel, Netherlands, summertime runoff in the simulations is affected significantly by these deficiencies. Comparison between the observed change of the river discharge between Lobith and Basel and model generated runoff in the same catchment area indeed shows that in general the shape of the annual runoff cycle of the models is improved (results not shown). This issue will be addressed further later on.

d. Terrestrial storage

The terrestrial storage capacity in a climate model is—apart from the storage as snow—determined by the amount of water that can be stored in the soil. Many of the model parameters that have a strong influence on the actual evolution of the terrestrial water storage are difficult to specify from objective ancillary information, owing to large spatial variability, strong interaction between parameters, and the local climate dependence of the model sensitivity to the parameter values. Yet, their impact on the simulated hydrological cycle is large, and an independent evaluation of the effective (soil) storage capacity is valuable.

Figure 7 shows the storage change as estimated by HSS and the RCM model output. The interannual variability of the estimated value is considerable, and all models (except UCM) fall within this variability and show more or less a similar behavior. Except for UCM, the amplitude of the annual cycle is a bit larger than observed, in particular, in the KNMI model. However, the results are in fact surprisingly good, and in any case better than the estimate of ∂S/∂t produced by ERA-40. Over Europe, the soil water assimilation damps the annual cycle of soil water considerably (van den Hurk et al. 2004).

This “spaghetti plot” shown in Fig. 7 is not very informative on the role of the soil hydrological memory in the hydrological partitioning process, mainly because the interannual variability in the storage component is not considered. Moreover, where biases in the mean annual cycle of P − E, ∂S/∂t, and runoff may be at least partially explained from biases in the forcing boundary conditions, the treatment of anomalously wet or dry conditions is probably a better indicator of the behavior of the individual models. Budget closure demands that anomalies in P − E must be transferred into anomalies in either runoff or in the storage range. Figure 8 shows this partitioning of P − E anomalies over these terms separately as deduced from the observations. The sum of discharge and storage is equal to the total P − E (slope = 1.00). In this figure each symbol represents an anomaly value averaged per hydrological summer [July–August–September (JAS)] or winter [January–February–March (JFM)] season, by subtracting the average annual cycle from each monthly data point. In contrast to monthly values, anomaly correlations of summertime averages are not strongly mutually affected by the persistence of droughts or wet seasons and may be considered statistically uncorrelated.

Given the fairly small average annual cycle in the discharge observations (Fig. 4), it is not surprising that anomalous water supplies to the basin area are primarily buffered in the soil and snow components, and that only ∼13% of the anomalies are transferred as discharge within the same season.

In summertime the storage range has on average a large uptake capacity, and anomalies in convergence are rapidly absorbed in the soil. In wintertime, this buffer capacity is less and a stronger preference for discharge is displayed. However, the discharge response is still the smaller component in the partitioning of P − E anomalies.

A similar analysis was carried out for the area downstream of Basel, where complex precipitation or glacier processes in mountainous areas are less significant. The correspondence between annual total P − E derived from ERA-40 and runoff was indeed improved (thus, the bias in the terrestrial storage reservoir shown in Fig. 2a was smaller). The correlation between P − E anomalies and runoff was reduced from 13% to 4%, consistent with the elimination of the high Alpine mountain area with small storage capacity from the domain. In winter the correlation reduced from 42% to 30%.

Do the RCMs explored here reproduce this basin response under present-day climate conditions? Figure 9 shows similar plots as in Fig. 8 from two participating models as an example, and Fig. 10 summarizes the wintertime and summertime regression slopes of anomaly partitioning components. An uncertainty estimate of these regression slopes is calculated using the bootstrapping method, by calculating the standard deviation of the regression coefficient of 10 000 sets of 30 pairs.

The comparison between the anomaly partitioning in the observed data and the RCMs reveals three groups of models: the KNMI model with a higher-than-observed portion of the anomalies put in the storage and less in the runoff, the UCM model that has a very small storage capacity and consequently a strong runoff response to convergence anomalies, and the remaining models that are in between these two extremes but have a lower-than-observed contribution of the storage term. Except for UCM, the relative differences between all models and observed records are preserved when the analysis is carried out for separate season periods. During the summer months, the KNMI model reproduces about the right partitioning when compared to the Hirschi analysis, but in winter the significance of the storage range is overestimated. All other models have a correlation between runoff and P − E that is too strong both in summer and in winter. The difference in behavior between the KNMI model on one hand and ETH as example for the others is well discerned in Fig. 9. The impact of this difference in correlations on the sensitivity of the modeled regional hydrological cycle to climate change is explored in the next section.

The analysis was also carried out for the domain downstream Basel, to eliminate the possible effects of deficiencies in precipitation or glacier melt in the high mountain area. For all models, the summertime correlation between P − E anomalies and runoff decreased by 5%–10%, similar to the reduction of the equivalent correlation in Hirschi’s analysis. The wintertime correlation coefficients changed approximately 5%, but not into the same direction for all models.

a. Response of runoff to climate change

All participating models also performed a 30-yr simulation (2071–2100) driven by the HadAM3H model in which the A2 greenhouse gas emission scenario was imposed. Among other effects, the selected scenario leads to increased winter precipitation and decreased summer precipitation in the Rhine basin (Fig. 11). The range of simulated changes in precipitation across the models is about 0.5 mm day⁻¹, which is considerably less than the >2 mm day⁻¹ range displayed in the simulated present-day precipitation climate (Fig. 4). The range in simulated precipitation rates in each of the control and A2 scenario experiments is significantly larger than the range of the simulated responses to the A2 greenhouse gas scenario.

Intermodel differences in evaporation are most pronounced in summer, where parameterized soil processes control the water loss to the atmosphere (Fig. 11c). The KNMI model with the relatively large storage range (Fig. 11d) can sustain a longer period where E > P, even when both E increases and P decreases. A high evaporation rate is also sustained in the simulation by UCM, which compensates the low evaporation in the control simulation (Fig. 5) under conditions of higher air temperatures. In the UCM simulation the water is not provided from a terrestrial reservoir (which has a very small buffer capacity) but from a smaller-than-average precipitation reduction and larger-than-average runoff reduction (Figs. 11b,e, respectively). A rapid response owing to a limited soil water range is particularly evident in the simulation by SMHI, where the evaporation response to the A2 scenario rapidly changes sign between June and August.

The translation of the precipitation change into the runoff response is highly variable, in particular in the winter season. In general, the seasonal variation of runoff increases in the A2 scenario (more runoff in winter, less in summer), but the magnitude of this increase is much smaller in the KNMI model than in the others.

Figure 12 shows a summary of the change in the seasonal cycle of runoff in relation to the depth of the soil water range in the models. The change in seasonal cycle of runoff is indicated by plotting the fraction of total runoff that occurs in the summer months (JAS; left panel): higher values of this fraction indicate smaller differences between summertime and annual mean runoff. When runoff is identical in all months, the fraction is 0.25 (3 out of 12 months) per definition. The depth of the soil water range, denoted by D hereafter, is defined as the difference between the 95% and 5% percentiles in the 30-yr time series of monthly stored water. As indicated before, this definition is different from the (effective) soil storage capacity, which is usually indicated by a difference between model-dependent field capacity and wilting point multiplied by the total soil depth. Similar to the high-pass filtering in the observed time series (Fig. 2a), a 3-yr running mean was subtracted from the soil water time series prior to calculating the storage range to remove slow trends in the budget calculation. The interannual variability is estimated by calculating the standard deviation of the yearly difference between maximum and minimum amount of stored water. The value of the “observed” soil water range, also indicated in Fig. 12, must be interpreted with some caution, since it is calculated as a residual term of the water budget equation.

Consistent with Fig. 11, Fig. 12 shows that in all models the fraction of annual runoff generated during summer decreases in all models when changing from the control to the A2 simulation. Simultaneously, the winter fraction increases. Also shown is an increase of the annual storage range in all models when an A2 scenario is imposed. Wetter winters and dryer summers both expand the range of D: wetter winters partially due to increased maximum snow depth and higher soil saturation degrees when rainfall and snowmelt rapidly recharge the drained soil water, and dryer summers due to increased atmospheric evaporative demands that is attenuated by bringing the level where soil water stress inhibits evaporation to lower moisture levels.

However, in contrast to Fig. 11, the display in Fig. 12 enables us to relate the degree of the change of the runoff seasonality to the model-specific value of D. Models with a small terrestrial water buffer (extreme case: UCM; see also Fig. 7) have a smaller change of D in case of an A2 scenario, and a relatively strong increase of the runoff seasonality. Models with larger soil water buffers (extreme case: KNMI) show a smaller response of the runoff seasonality to a change in P − E, as shown clearly in Fig. 12. The relative summertime runoff for the area downstream of Basel is approximately 5% lower than the results plotted in Fig. 12, and in winter it is 5% higher, but the grouping of models is not clearly affected by the choice of the catchment domain (results not shown).

The response of runoff anomalies to anomalies in net convergence, as displayed in Fig. 10 for present-day climate conditions, may be indicative of the way the models respond to a change in the greenhouse gas scenario. To first order, the transition to an A2 scenario generates a different net convergence, which can be considered as an anomaly compared to the present-day climate reference situation. Figure 13 displays a similar plot as in Figs. 8 and 9, but now showing the multiyear mean values of R and P − E for each 30-yr simulation, grouped per model. For reference, also the slope of the response of runoff anomalies to anomalies in net convergence, as displayed in Fig. 10 for the control climate simulations, is displayed. For the summer period, the slopes shown in Fig. 10 are very similar for the A2 scenario simulations for nearly all models (not shown). This is due to the fairly straightforward coupling between net water supply (P − E) to the soil and partitioning this supply over storage and runoff.

The response of the runoff anomaly to P − E anomalies in the control simulation in each model is indicative for the response of the model to a change in greenhouse gas scenario. Although a solid correlation between these two responses is not consistently present in Fig. 13, a small anomaly response can clearly be associated with a smaller greenhouse gas response. Thus, the anomaly of P − E translates similarly into a summertime runoff anomaly, as does the multiyear averaged runoff respond to multiyear averaged net convergence between two climate change scenarios. This implies that the interannual variability of P − E versus runoff, derived from the observations (Figs. 8 and 10), may be used as a proxy of the response of mean runoff to climate change. The climate response is gradually oriented steeper when models are situated in the bottom right portions of the plot. Again, the KNMI and UCM models are relative outliers in the transition of induced changes of P − E into runoff changes: KNMI putting smaller portions of P − E anomalies into runoff (resulting in a small runoff response to climate change shown in Fig. 10), whereas UCM rapidly transfers P − E anomalies into runoff. All other models group around the same slope of ΔR/Δ(P − E), although the absolute values of the runoff in the control simulation vary widely across models.

In wintertime (Fig. 13, bottom panel) P − E increases in all RCM simulations, but runoff averaged over JFM decreases (slightly) for the KNMI model, in contrast with the general increase of wintertime runoff in response to climate change in the other models. However, the deduction of the RCM response to a change in P − E from the interannual variability in simulations of either climate forcing is less straightforward than in winter. In general, the simulated runoff in wintertime is much less correlated to the supply of water to soil alone, but is highly dependent on the accumulation and melt of snow, existence of frozen ground, and direct response to precipitation in case of saturated soils. However, from Fig. 13 it can be deduced that the depth of the annual cycle of the terrestrial storage is larger than average in the KNMI model, and this buffer capacity helps keeping the runoff response to increases in wintertime precipitation relatively low (see also Fig. 12, right panel).

b. Response of summertime temperature

Simulated temperature changes under climate change conditions are dependent on many interacting processes embedded in the RCM codes. Summertime temperature responses over land are considered to be partially sensitive to the representation of the hydrological cycle (e.g., Schär et al. 2004). Under present-day climate conditions, many RCMs suffer from an excessive continental drying, which is related to a chain of processes linking evaporation to stored soil water, local precipitation to local evaporation, interaction between the hydrological cycle and large-scale dynamics, and the dependence of the precipitation efficiency on the thermodynamic structure of the atmosphere (Schär et al. 1999). Clearly, the depth of the soil water range plays an important role in the annual cycle of evaporation and its response to climate change (cf. Fig. 11), and it is worthwhile to explore whether this effect on evaporation has a clear impact on the response of the near-surface temperature to climate change.

High values of net radiation and low values of evaporation are generally both associated with high near-surface temperatures. The relation between evaporation and temperature is therefore more ambivalent than the relation between evaporative fraction Λ (latent heat flux divided by net radiation, ignoring soil heat flux) and temperature. Here T and Λ are negatively correlated, with values of ∂T/∂Λ ranging between −4 and −20 K for the model simulations.

Figure 14 shows the multiyear mean summertime 2-m temperature and evaporative fraction for all 30-yr simulations, again grouped per RCM. In general, the calculated summertime evaporative fraction is reduced by approximately 6%–8% when changing the climate scenario from control to A2, and simultaneously the summertime temperature increases between 4 and 6 K. For each of the models, the average value of ∂T/∂Λ is calculated using both the control and A2 30-yr simulations. This slope is used to estimate the contribution of the change of Λ to the change of T, naively assuming that the temperature response is a linear superposition of a hydrologically controlled component and an “autonomous” increase due to remaining processes (radiative absorption, change of circulation statistics, etc.). Suppose that all models keep the summertime evaporative fraction in the A2 simulations similar to the control simulations, simple extrapolation of ∂T/∂Λ results in a reduction of the summertime temperature increase with 0 (KNMI) to up to nearly 2 K (SMHI).