Preparing local climate change scenarios for the Netherlands using resampling of climate model output

A method to prepare a set of four climate scenarios for the Netherlands is presented. These scenarios for climate change in 2050 and 2085 (compared to present-day) are intended for general use in climate change adaptation in the Netherlands. An ensemble of eight simulations with the global model EC-Earth and the regional climate model RACMO2 (run at 12 km resolution) is used. For each scenario time horizon, two target values of the global mean temperature rise are chosen based on the spread in the CMIP5 simulations. Next, the corresponding time periods in the EC-Earth/RACMO2 simulations are selected in which these target values of the global temperature rise are reached. The model output for these periods is then resampled using blocks of 5 yr periods. The rationale of resampling is that natural variations in the EC-Earth/RACMO2 ensemble are used to represent (part of the) uncertainty in the CMIP5 projections. Samples are then chosen with the aim of reconstructing the spread in seasonal temperature and precipitation changes in CMIP5 for the Netherlands. These selected samples form the basis of the scenarios. The resulting four scenarios represent 50–80% of the CMIP5 spread for summer and winter changes in seasonal means as well as a limited number of monthly statistics (warm, cold, wet and dry months). The strong point of the method—also in relation to the previous set of the climate scenarios for the Netherlands issued in 2006—is that it preserves nearly all physical inter-variable consistencies as they exist in the original model output in both space and time.


Introduction
Scenarios of future climate change are important tools to assist society to prepare for future climate conditions (Dessai et al 2005, Wilby et al 2009. Climate scenarios are plausible future climate states, containing sets of climatological variables that are (to a certain degree) physically consistent (van den Hurk et al 2013a, Huard et al 2014).
There is no generally accepted method to prepare climate change scenarios. Different ways to treat climate data produced by different models, observations and process understanding of the climate system, and different user requirements stemming from different intentions of using the scenarios, lead to various methods and outcomes. For instance, for the United Kingdom a large ensemble of climate simulations with physics perturbations of the global climate model (GCM) HadCM3, employed in a Bayesian framework, are used to produce probabilistic climate change scenarios Environmental Research Letters Environ. Res. Lett. 9 (2014) 115008 (13pp) doi: 10.1088/1748-9326/9/11/115008 Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI. (Murphy et al 2007). For Switzerland, Fischer et al (2012) also use Bayesian statistics applied to an ensemble of regional climate simulations from the ENSEMBLES project (van der Linden and Mitchell 2009). These scenarios start from modeling activities taking different sources of uncertainty-forcing uncertainty from uncertainty in emissions and future land-use, model uncertainty and uncertainty from natural variations of the climate system (see e.g. Deser et al 2010)into account. In other countries more emphasis on requirements of individual users is put. For instance, Whetton et al (2012) describe for Australia an iterative methodology in which the most relevant GCM results are selected in consultation with the end users. A similar approach is followed in Canada (Huard et al 2014). While many of these scenarios share similar approaches, different choices are made on a detailed level, for instance concerning the actual model data that is used. These decisions often depend on expert judgment (Murphy et al 2007, Lenderink et al 2007, Huard et al 2014. In this respect, it is important to note that information on regional climate change is in many aspects still sparse, despite the existence of abundant literature on global climate change as summarized in the latest report of the Intergovernmental Panel on Climate Change (IPCC 2013). The Coupled Model Intercomparison Project Phase 5, CMIP5  contains more than 250 GCM simulations, but high resolution, downscaled information is only available for a fraction of these simulations. For this reason, the Coordinated Regional Climate Downscaling Experiment (Giorgi et al 2009) within the World Climate Research Programme has been set up. Also, information from models and observations is often incomplete and/or conflicting. For instance, observed trends over the past decades do not match with modeled trends for Western Europe .
For the Netherlands, a relatively long tradition of producing climate scenarios exists. The first generation of scenarios, issued in 1995 with updates in 1997 and 2001, was based on a simple delta change approach. Observed time series of temperature were perturbed with a temperature change obtained in a set of GCM simulations, while precipitation was adjusted according to the observed relation between precipitation and temperature (Klein Tank et al 1995). In 2006 a set of climate scenarios derived from regional climate model results was released (van den Hurk et al 2007). With four scenarios approximately 70-80% of the spread in local seasonal mean temperature and precipitation change derived from a selection of CMIP3 model results was reproduced. To achieve this goal, the regional climate model results were statistically interpolated (and partly extrapolated) in a framework using two steering variables. The first steering variable was the global mean temperature rise and the second the strength of the circulation change over Western Europe van Oldenborgh 2006, Lenderink et al 2007). This procedure led to the loss of spatial detail and compromised the consistency between different variables.
In this paper, we describe a new method that has been used to produce the latest set of climate scenarios for the Netherlands released in 2014. It is based on the resampling of large ensemble of modeling results from a single global and regional climate model. As previously, a considerable fraction -at least 50%, but aiming at 70-80%-of the CMIP5 model spread for a number of seasonal temperature and precipitation statistics is represented. We retain the framework using global mean temperature change and circulation change as primary steering variables that distinguish the different scenarios. The new procedure provides much more spatial detail and preserves physical consistency better, albeit at the cost of some reduction in representing model spread from the CMIP5.
The paper is organized as follows. Section 2 gives a description of the GCM EC-Earth and the regional climate model RACMO2. In section 3 a short overview of the method is given, with detailed descriptions in sections 4 and 5. Finally, section 6 gives a summary and discussion.

EC-Earth and RACMO2
The basis for the climate scenarios is an ensemble of eight integrations with the GCM EC-Earth (Hazeleger et al 2010). This ensemble will be referred to as ENS-EC. All simulations use the RCP8.5 emission scenario in order to obtain a sufficiently large signal to span the desired range of global mean temperature rise. For all ensemble members the model configuration is identical. The ensemble is produced by perturbing the initial state of the atmosphere in EC-Earth in 1850. After ∼130 yr, the start of the reference period in 1981, the different members diverged also in the slowly varying components of the climate system, and they can be considered as independent realizations.
For later reference, we show in figure 1 the response of winter and summer temperature and precipitation as a function of the global temperature rise at/near the Netherlands (see figure caption for the exact area) averaged over all ensemble members in EC-Earth. The seasonal mean temperature rises almost in concert with the global mean temperature. Temperatures in cold months in winter and warm months in summer, however, rise faster than the global mean temperature. We think that this behavior is caused by local feedback through snow cover and arctic amplification in winter and soil feedbacks in summer (Screen 2014). Mean winter precipitation increases quasi-linearly with the global temperature rise at a rate of approximately 6% per degree. For low values of global temperature rise the rate of increase appears faster than for high values. Also, the rate of increase for wet months appears slightly lower and for dry months slightly higher than the mean changes. For summer, the response in precipitation is very nonlinear. For low values of the global temperature rise an increase in precipitation is still obtained, but for higher values mean summer precipitation decreases. The decrease in summer precipitation is in particular marked in dry months, which show a decrease of almost 30%.
All EC-Earth integrations have been downscaled with the regional climate model RACMO2 (van Meijgaard et al 2012). RACMO2 is run on a relatively small domain of 222 × 216 longitude-latitude grid points at 12 km resolution. The domain is centered on the Netherlands. Embedded in RACMO2 is a slab ocean model for the North Sea as described in Attema and Lenderink (2013). The small domain of RACMO ensures that the large-scale circulation statistics provided by EC-Earth are very well reproduced in RACMO2.

CMIP5 data for evaluation
The scenarios are evaluated against the results of the CMIP5 model ensemble in support of the fifth IPCC assessment report . An analysis of their spread for the Netherlands (and surrounding area) in relation to circulation changes and global temperature rise is described in van den Hurk et al (2013b). Here, we neglect the low emission scenario RCP2.6, which is mainly an emission scenario developed for mitigation, and we only take simulations driven by RCP4.5, RCP6, and RCP8.5 into account.
For evaluation we use the seasonal mean changes in precipitation and temperature for two 30 yr future scenario periods-centered around 2050 and 2085-compared to the reference period 1981-2010. In addition, we consider the spread in different percentiles of the monthly mean temperature and precipitation. For instance, the 5th and 10th percentiles of monthly mean temperature in winter are measures of cold months in winter, whereas the 90th and 95th percentiles for monthly mean summer temperature represent hot months. Changes are determined at a point in the Southeast of the Netherlands (6.25°E, 51.25°N) and all CMIP5 results are interpolated to this location. For each CMIP5 model simulation, the change in a statistic is first computed, and the cumulative distribution over all model simulations (almost 200) is subsequently constructed giving all simulations equal weights. With this approach higher weights are assigned to models that have multiple members in the CMIP5 ensemble, which contrast Response in temperature (mean, and 5 and 95th percentile of monthly means) and precipitation (mean, and 10 and 90th percentiles) as a function of the global mean temperature rise in the EC-Earth ensemble ENS-EC. For temperature the average for a rectangular longitude/ latitude box between 4.5°E/50.5°N (Southwestern corner) and 8°E/53°N (Northeastern corner) is taken. For precipitation the box is larger in order to suppress noise due to natural variability, and extends from 4°E/47°N to 14°E/53°N. For guidance, a 1-1 line for temperature, and 4 and 8% per degree lines for precipitation in winter, and 2 and −8% lines for summer (brown lines) are shown. Results are for 30 yr moving windows. The shaded area represents the 90% range due to natural variability derived from the deviations of the ensemble members from the ensemble mean.
with the approach used for the CMIP5 atlas whereby all models have equal weights (IPCC 2013: annex I). For the Netherlands and in view of the other uncertainties the differences between both approaches are quite small. From the distribution derived from the CMIP5 model simulations, we consider the spread given by the 25th to 75th, and the 10th to 90th percentile range, the latter two representing the 50% and 80% range of the CMIP5 ensemble, respectively. Our aim is to represent with our scenarios the 10th to 90th percentile range, where we refer to the 10th and 90th percentile as the low and upper target from CMIP5, respectively.

Overview of the method
Here we briefly explain the procedure to construct the scenarios, while more detailed information is given in the next two sections. The number of four scenarios has been decided after consultation with end users. The procedure is as follows (see figure 2): 1. In the first step we cover the spread in global mean temperature rise ΔT glob in CMIP5 by choosing two values for each scenario period (2050 and 2085). Subsequently we select the time periods where the global mean temperature rise in our model ensemble ENS-EC reaches these values of ΔT glob (see section 4.1). 2. In the second step (see sections 4.2 and 4.3) we resample the results of ENS-EC for the selected time periods (from the previous step). This is done by recombining 5 yr periods of the eight members of ENS-EC into new resampled climates, and selecting combinations that match with the spread in CMIP5. This provides eight resampled EC-Earth time series for each of the scenarios. 3. In the third step (see section 5) we use the corresponding downscaling with RACMO2 of the selected EC-Earth time series from the previous step to construct the final scenarios.

Selection based on the global mean temperature rise
The choice of the steering parameter ΔT glob is based on our selection of CMIP5 simulations forced by RCP4.5, RCP6 and RCP8.5 (see figure 3). The range of these model outcomes is partly the result of the different emission scenarios as illustrated by the left-hand figure-a difference in the forcingand partly the result of differences in the global mean temperature response by different climate models for a given emission scenario-a difference in climate sensitivity (Andrews et al 2012, Otto et al 2013. For 2050 we chose values of the steering parameter ΔT glob of 1 and 2°C, approximately representing the 10th to 90th percentiles of CMIP5. Likewise, for 2085 we picked 1.5 and 3.5°C. We used numbers rounded to 0.5°C in order to avoid suggesting too much confidence in these numbers, and to emphasize that these are scenarios based on the assumption of a certain rise in global mean temperature. For the EC-Earth runs the 30 yr time periods are selected in which these target values of the global temperature rise are reached (table 1). As the global mean temperature rise in the eight members is almost identical, we take the same periods for all ensemble members. For practical reasons we used time periods rounded at 5 yr, which led to small differences of at most 0.1°C from the target value of ΔT glob . We consider this insignificant compared to other uncertainties. The resulting periods are shown in table 1. Despite the EC-Earth runs being driven by the high RCP8.5 emission scenario, they unfortunately do not reach a 3.5°C global mean warming because EC-Earth has a relatively moderate climate sensitivity. Therefore we used the period of 3.0°C warming, and applied pattern scaling (Mitchell 2003, Fischer et al 2012 by multiplying all the changes obtained for this period with a factor 1.16 (3.5/3.0 degrees).
Turning to local changes of seasonal mean temperature and precipitation for the Netherlands, a comparison between the CMIP5 range for 2050 and ENS-EC results for the periods with 1 and 2 degrees global mean warming is shown in figure 4. A large fraction of the envelope provided by CMIP5 is already covered by the eight ensemble members. Note that the difference between the eight members is entirely due to natural variations that are, by construction, unrelated to differences in greenhouse gas forcing or model formulation. For instance, for the change in summer precipitation, the eight EC-Earth members approximately span the 25th to 90th percentile range of CMIP5. The simulations miss the dry tail of CMIP5 for the mean change and for the dry months (the 5th and 10th percentiles). For winter precipitation, the changes in the EC-Earth simulations are biased towards the wet side, in particular for changes in the dry months. The temperature in winter in the ensemble covers the full spread of CMIP5, which is a consequence of the large natural variability for that season. For summer, however, the upper 25% range of CMIP5 is not covered by any of the ensemble members.
Although the EC-Earth ensemble spans a reasonable part of the CMIP5 range, it is not very practical to use these simulation directly as scenarios. Besides that for a number of the statistics the model simulations are rather biased compared to CMIP5, it also not possible to select a subset of four simulations that covers the CMIP5 range. The model simulations that do well for one statistic, for instance dry months is summer (P05 and P10), do not perform necessarily well for another statistic, like wet months in winter. We therefore continue with resampling of the EC-Earth data in the next section.

Resampling
From the previously selected time periods, which are set by ΔT glob , the output of the different EC-Earth members are recombined into new, resampled climates. The resampling is done by recombining 5 yr periods from the eight members into new 30 yr climates. So, if we consider the resampled climate of 1981-2010 it is constructed by choosing member i 1 for the period 1981-1985, with i 1 є {1, 2,…,8}; member i 2 for the period 1986-1990, with i 2 є {1, 2,…,8}, etc. In this manner, all different re-combinations for each 30 yr period are constructed, which gives 8 6 (=262 144) resampled climates. This is done for both the control as well as the future period.
In a traditional bootstrap method one would draw each 5 yr period randomly out of the eight ensemble members. This is equivalent to drawing randomly from the 8 6 resampled climates constructed above as each of these drawings has the same probability by construction.
Here, we take an approach that is different than the random bootstrap approach because we are interested in obtaining a climate change signal that matches with the CMIP5 range. Therefore, instead of randomly choosing, we select samples conditionally in a procedure consisting of three steps as described in short below (with detailed information in the supplementary material). With resampling we construct for each scenario period and choice of ΔT glob two scenarios: a scenario characterized by relatively low, denoted by subscript L, changes in seasonal mean precipitation and temperature (relative to ΔT glob ) and one based on relatively high changes, denoted by a subscript H. This gives rise to four scenarios for each scenario time horizon, labeled by G L , G H , W L and W H (see table 2). For each scenario eight resampled 30 yr time series, called re-samples, are selected for both the control as well as the future period with the following procedure: 1. We first select 1000 re-samples (out of the total 8 6 ) that have a change in mean winter precipitation of 4% (for the 'L' scenarios) and 8% (for the 'H' scenarios) per  degree rise in global mean temperature. These two rates of the precipitation change are primarily based results with EC-Earth shown in figure 1. They are also consistent with expectations related to the increased moisture of the atmosphere following from the Clausius-Clapeyron relation (Trenberth et al 2003). This would imply an increase of 6-7% per degree temperature rise, considering that the local temperature rise is approximately equal to global temperature rise and that changes in relative humidity are small (see also Lenderink and Attema, this issue). 2. From the subset of 1000 re-samples, we select approximately 50 re-samples based on changes in summer precipitation, and changes in summer and winter temperature (i.e. targets 3-5 in table 2). This is not done by imposing an absolute constraint (like in step 1), but by selecting a percentage range from the remaining re-samples (see supplementary material). These percentile ranges are chosen in an iterative way doing the evaluation with the CMIP5 range (as discussed in the next section) a number of times. 3. From approximately 50 re-samples obtained above, we finally sub-select eight re-samples that have a minimal re-use of the same model data. It is e.g. not allowed that a 5 yr period from an ensemble member is re-used more than three times in the set of eight re-samples that form a scenario.
Thus, for each of the four scenarios belonging to a scenario time period, either 2050 or 2085, eight resampled EC-Earth runs of 30 yr for both control as well as the future period are obtained. These re-samples are labeled with EC-X, with X the scenarios name (G L , G H , W L and W H ), and the average change from these eight re-samples are compared with CMIP5 in the next section. For simplicity we will refer . Change in mean, and 5th to 95th percentiles (labeled 'ave' and P05-P95 on the x-axis) of monthly mean precipitation (upper panels) and monthly mean temperature (lower panels). The 10th to 90th percentile range of the CMIP5 ensemble is shown by the light shaded area, whereas the 25th to 75th percentile range is shown by the darker gray shaded area. Results from the eight (un-resampled) EC-Earth simulations (ENS-EC) are shown by the dark (W)/light (G) green dots and lines. All results are for the location 6.25°E, 51.25°N in the Southeastern part of the Netherlands (see section 2.2). to these as scenarios, although we still have to discuss the downscaling step with RACMO2 in section 5.

Evaluation against CMIP5
Changes derived from the resampled EC-Earth data, EC-G L , EC-G H , EC-W L and EC-W H , constructed for the scenario period 2050 are shown in figure 5, in comparison with the spread in CMIP5 (section 2.2). The light shaded area represents the 80% CMIP5 range, which is the target range for our scenarios. This is rather ambitious considering that we only develop four scenarios and we evaluate the spread for a range in statistics. Therefore, not for all statistics we could reach our goal of representing the target range, and we also compare to the central 50% CMIP5 range (dark shaded area).
The scenarios perform well for seasonal winter temperatures in 2050 (figure 5). For seasonal mean temperatures the 80% range in CMIP5 spans 0.8-2.4°C, while the scenarios cover the range between 1.0°C to 2.4°C. For the different percentiles of monthly mean temperatures the upper Table 2. Overview of the four scenarios for 2050. The three columns labeled with 'Target 2-5' give an overview of the constraints applied for the selection of the re-samples. For targets 3, 4 and 5 these are relative constraints, with 'o' denoting small changes, '−' negative changes and '+' positive changes (more symbols means larger changes).  target (the 90th percentile of the range) in CMIP5 is well covered with the W H scenario (red line), but the lower target (the 10th percentile) of CMIP5 is not captured as well, in particular for cold months, given by the 5th and 10th percentiles (P05 and P10). For summer temperature, the G L scenario captures the lower target of CMIP5, yet the warmest scenario, W H , only reaches the 75th percentile of CMIP5 range. For instance, the seasonal mean temperature change in the W H scenario is 2.5°C, whereas the upper target from CMIP5 is 3.2°C.

Scenario period Scenario label
The target range from CMIP5 is less well represented for precipitation. For the seasonal mean changes the CMIP5 spread is reasonably well covered with a scenario range between +4 to +17% and the target CMIP5 80% range between +0% and +16%. However, increases in precipitation are too strong in dry months (P05 and P10) whereas changes are underestimated for wet months (P90 and P95). This behavior is already present in ENS-EC, and the scenarios inherit much of this behavior. Inspection of the spatial fields of precipitation changes (not shown) reveals that there is a relatively small area with relatively high values of the change in P05 and P10 near the Netherlands, which also exists in ENS-EC with changes of 20-25%.
Summer precipitation covers only the central 50% CMIP5 range with changes between almost zero and −15% in average precipitation. However, the characteristic broadening of the range in CMIP5 for relatively dry months is represented well with the scenarios. Precipitation decreases by almost 40% in the driest months in the W H scenario.
For the end of this century, 2085, results are generally similar (figure 6). In summer, the scenarios cover slightly more of the CMIP5 range than those for 2050. In winter the change in mean precipitation in the W H scenario of 28% is beyond the 90% target of CMIP5. Analysis of the CMIP5 model results driven by RCP8.5 (RCP6.0) show that approximately 10% (5%) of the model simulations have an increase in mean winter time precipitation of approximately 30% or more. For dry as well as wet month the W H scenario approximately follows the 90th percentile, and we consider this (in particular the behavior of the wet months, P90 and P95, which is relevant for flooding) to be more important than the seasonal mean behavior.
Finally, the spatial patterns of the change in the G L and W H scenarios and in CMIP5 are shown in figure 7 for summer precipitation. For CMIP5, the model integrations were taken that were used to produce the CMIP5 atlas (IPCC 2013: annex I; taken from climexp.knmi.nl/atlas). Formally this dataset differs from the one used in the rest of the analysis, because individual ensembles members are down-weighted to give each model equal weight rather than each ensemble member. In practice the differences, however, are quite small and much smaller than other sources of uncertainty. This is shown by comparing figures 7 and 5. Similar as in figure 5, the G L scenario for the Netherlands, in figure 7, coincides roughly with the 75th percentile from CMIP5, while the W H scenarios coincides with the 25th percentile. While the patterns of changes in JJA precipitation are broadly similar, the scenarios display a more pronounced spatial structure than CMIP5. Note that for the statistics shown here spatial structures are smoother in CMIP5 than in the resampled scenarios as the natural variability has been averaged out over far more ensemble members in CMIP5. The broad agreement strengthens our argument that these discrete scenarios are fairly representative of the CMIP5 25-75% spread in summer precipitation.
Although the method has not been specifically devised to represent changes for spring and autumn, the method is still used to construct the scenarios for spring an autumn. An evaluation of the resampling results for spring and autumn is given in the supplementary material.

Downscaling with RACMO2
The ensemble of EC-Earth runs, ENS-EC, has been downscaled with the regional climate model RACMO2, which shares most (but not all) of the physical parameterizations with EC-Earth. This ensemble of simulations, referred to as ENS-RA, is used to provide high-resolution information for the Netherlands and surrounding areas for each scenario. For instance, daily statistics are derived from the RACMO2 results, while results shown in the previous section were limited to monthly mean changes. We used exactly the same resampling (for each scenario the same members for the same periods) as for EC-Earth. The simulations with RACMO2 have been performed on a relatively small domain (see section 2.1), which guaranteed that the large-scale response in EC-Earth are retained in the RACMO2 downscaling.
Results of RACMO2 for an area extending over the Southeastern part of the Netherlands, roughly South of 51°N and East of 5.2°E, are in most aspects very close to the EC-Earth results. Figure 8 shows results for 2085 derived from RACMO2 that are comparable to the results obtained by EC-Earth in figure 6.
There are two aspects for which RACMO2 clearly modifies the response of EC-Earth. The RACMO2 results predict slightly higher values of future summer time precipitation, or less drying in the G H and W H scenarios. For instance, in the W H scenario the mean precipitation response for 2085 is −21% in RACMO2 versus −26% in EC-Earth.
However, the most outspoken difference is the temperature response for cold months in winter, with changes in the 5th and 10th percentile that are approximately 1°C higher in the RACMO2 simulations in W H and W L . This is the result of resolving the land-sea contrast much better in RACMO2 (see figure 9), possibly enhanced by the more advanced boundary layer scheme in RACMO2 which better retains a shallow and cold boundary layer in winter (Lenderink andHoltslag 2004, van Meijgaard et al 2012).
Finally, a limited number of resulting changes averaged over the Netherlands are given in table 3. For instance, we obtain that increases in cold days (P01) in winter are much larger than the changes in relatively mild days (P99). In summer, the situation is reversed and changes in warm days (P99) are much larger than those on relatively cold days (P01).

Circulation response and resampling
By resampling we explicitly make use of natural variability in the climate system that is responsible for a considerable fraction of the spread in CMIP5. A large part of the natural variability is due to long term variations in the atmospheric circulation. van den Hurk et al (2013b) analyze the atmospheric response in CMIP5 and its relation to projected climate change in the Netherlands. In the supplementary material we show the difference in mean sea level pressure response between the different scenarios in comparison with the spread within the CMIP5 ensemble. It is shown that for the Western European area these pressure patterns are reasonably similar. The difference between the W H and W L scenario in winter can therefore be regarded as a difference mainly in the strength of the westerly circulation over the Netherlands, similar to the distinction between scenarios made in the previous set of climate scenarios (Lenderink et al 2007).

Variability and resampling
The procedure of resampling does affect the temporal variability structure of the original time series as samples are regrouped in an order that does not occur in the native data set. By using a block resampling with periods of 5 yr this effect is minimized for time scales up to 5 yr. Hardly any effect of the resampling exists on the inter-annual variability, which is to be expected due to the length of the block resampling. This is shown by an example in the supplementary material, where the inter-annual variability in W H is compared to the results of the original (un-resampled) model data. Note that apart from the trends virtually all variability in the Netherlands is described by the inter-annual variability, so that these short intervals already contain all natural variability (see e.g. figure 2 in van Oldenborgh et al 2012).
Resampling could affect climate variability at longer time scales. However, even inspection of the variability between the eight resampled members (computed from the 30 yr climate periods) and the variability between the eight members in the un-resampled ensemble only revealed small differences. Only for the variables that have been explicitly used in the selection procedure (such as the seasonal mean precipitation) the variability between the members is-by construction-much smaller. For instance, although the seasonal mean change in winter precipitation is prescribed for all resampled ensemble members, implying no variability between the ensemble members, the variability of monthly precipitation is hardly affected (see supplementary material).

Strength of the approach
By resampling of model fields the full physical consistency of the climate model output is retained. This applies to temporal and spatial consistence, and between all variables that are provided by the regional model. The time series are only discontinuous every fifth year between 31 December and 1 January. The effect of these cuts depend on the auto-correlation and therefore on the variable and season. For winter temperature the one-day auto-correlation at De Bilt of r = 0.86 points to a de-correlation time scale of about one week, so the first week of every 5 yr block is affected by the cuts. The auto-correlation for precipitation is so much lower (r = 0.2) that this variable is not affected by the resampling.
For all scenarios, eight members are available, thereby allowing an assessment of the variability at different time scales, changes in rare events, and compounding occurrences of correlated characteristics.

Limitations of the approach
The main weakness of the approach is that it makes use of a single model setup with EC-Earth and RACMO2. Uncertainty in the global temperature response in CMIP5-expressing uncertainty in emission and climate sensitivity-can be regained rather well by selecting different EC-Earth time periods. However, a full recovery of the (80%) CMIP5 range in the regional scale climate change response has not been achieved for all variables and statistics. Differences in the response of the atmospheric circulation can be reproduced rather well by resampling. But other sources of uncertainty at regional levels, in particular related to the representation of small-scale physics, are less well covered with this method.
As an example, a small number of CMIP5 models project a much stronger drying out over continental Europe in summer, also leading to higher temperatures. This is likely due to the representation in the model of soil processes and landatmosphere feedbacks, through clouds, precipitation and radiation (Rowell and Jones 2006). The difference with EC-Earth and RACMO2 is the apparent reason that our scenarios do not cover the upper target of the CMIP5 range for summer   temperature, and the lower target for precipitation. Recent research has shown that CMIP5 models with the tendency for a strong future drying also appear to display a warm and dry bias in the present-day climate (F Selten, manuscript in preparation). Therefore, we consider the possibility of a strong continental drying less realistic. However, an additional scenario representing this strong drying trend could be valuable for specific usage.
The uncertainty due to downscaling is also not captured with these scenarios. However, the RCM downscaling skill has been assessed by performing simulations using re-analysis boundaries. In a recent large inter-comparison between 17 regional climate models, RACMO2 performed among the top models (Christensen et al 2010) with results often close to the average model behavior (e.g. Lenderink 2010).
Yet, the use of present-day regional climate models has clear limitations when it comes to scales of tens of km's. One limitation is the representation of showers that bring extreme precipitation at local scales in summer (Kendon et al 2014). The convective processes leading to these showers are not resolved in current regional climate model such as RACMO2, but are parameterized. This probably limits the applicability of the results for these types of extremes. Changes in extreme precipitation (intensities) in the climate scenarios are therefore not solely based on RACMO2 results (Lenderink and Attema, this issue).
Finally, the problem of how to merge the information from climate models with observed trends up to now has not yet been addressed in these scenarios. The discrepancies between observed and CMIP3 modeled warming trends noted in van Oldenborgh et al (2009)

Conclusion
An ensemble of global (EC-Earth) and regional (RACMO2) climate model simulations is used to prepare climate scenarios for the Netherlands. The approach is based on scaling responses with the global temperature rise in combination with a resampling technique. It is shown that the four scenarios produced by our approach sample 50-80% of the CMIP5 uncertainty range for various statistics including changes in mean winter and summer temperature and precipitation.
The main strength of the approach is that the full physical consistence of the model fields is retained by the procedure, with only a minor effect on the variability on long time scales. As such, all variables available in the model output can be made available to users. This makes the method very suitable to produce scenarios. The obvious downside of the approach is that (in its current form) it is based on a single modeling system, which poses inherent limitations on covering the full spread in the CMIP5 simulations. Table 3. Changes in a number of statistics derived for the Netherlands in the scenarios for 2050, with PXX the change in the 'XX'th percentile, and 'w' denoting a wet-day percentile; 'wdf' is the wet-day frequency.