Projections of extreme precipitation events in regional climate simulations for Europe and the Alpine Region

Authors


Abstract

[1] Regional climate models (RCMs) from the ENSEMBLES project are analyzed to assess projected changes in 21st century heavy and extreme precipitation events over Europe. A set of 10 RCMs with horizontal grid spacing of 25 km is considered, which are driven by six GCMs under an A1B greenhouse gas scenario. The diagnostics include basic precipitation indices (including mean, wet-day frequency, intensity, and percentile exceedance) and application of generalized extreme value theory for return periods up to 100 years. Changes in precipitation climate between present (1970–1999) and future (2070–2099) conditions are presented on a European scale and in more detail for 11 European regions (mostly in supplemental figures). On the European scale, projections show increases (decreases) in mean amounts and wet-day frequency in northern (southern) Europe. This pattern is oscillating with the seasonal cycle. Changes in extremes exhibit a similar pattern, but increases in heavy events reach much further south. For instance, during spring and fall, much of the Mediterranean is projected to experience decreases in mean precipitation but increases in heavy events. Thus, projected changes in mean and extremes may show different signals. The inter-model spread is partly attributable to a GCM-dependent clustering of the climate change signal, but also affected by RCM uncertainties, in particular in summer. Despite these uncertainties, many of the projected changes are statistically significant and consistent across models. For instance, for the Alps, all models project an intensification of heavy events during fall, and these changes are statistically significant for a majority of the models considered.

1 Introduction

[2] Due to their potential to cause severe societal, economic and environmental impacts, there is increasing interest in extreme weather and climate phenomena. Scientific interest in extreme events is also increasing in response to concerns that climate change might negatively affect the frequency and intensity of such events. In relation to heavy precipitation events and floods, adaptation bears a particular urgency, as many water resource structures (such as dams, bridges, storm drains, or sewer systems) are planned for lifetimes exceeding 50 years. Thus, there is an important need to better understand heavy precipitation events in a changing climate and to develop climate change information for hydrological planning processes. The current study presents a detailed assessment of heavy and extreme precipitation events in regional climate model projections.

[3] Climate simulations project that the Earth's climate faces significant changes as a result of anthropogenic greenhouse gas and aerosol emissions [Solomon et al., 2007]. The most prominent expression is found in a global temperature increase that is evident in both observations and model simulations, but there is also overwhelming evidence that the hydrological cycle will significantly be affected as well. This evidence stems from observational studies, theoretical arguments, and modeling results.

[4] Large-scale observations show an increase in mean precipitation in the tropics and high latitudes and a decrease in the subtropics [Alexander et al., 2006; Frich et al., 2002], and there are indications for increases in precipitation intensity even in regions where mean precipitation decreases [Alexander et al., 2006]. Also on a European scale, a significant number of observational studies have addressed trends in mean and heavy precipitation [Klein Tank and Können, 2003]. In Switzerland, for example, cold-season precipitation has significantly increased in the last century [Widmann and Schär, 1997; Schmidli et al., 2002] and heavy precipitation events have intensified [Schmidli and Frei, 2005], consistent with an observed increase in large damages from floods [MeteoSchweiz, 2006]. While these trends are consistent with climate scenarios, this is not generally the case for all regions, as the externally forced signal has a similar or smaller magnitude as natural variability. Indeed, long-term reconstructions of European precipitation [Pauling et al., 2006; Pauling and Paeth, 2007; Casty et al., 2005] and flood frequency [Mudelsee et al., 2006; Pfister, 1999, 2009; Schmocker-Fackel and Naef, 2010a, 2010b] demonstrate that the natural climate is affected by strong decadal variations that have led to extended periods with unusual precipitation and flood frequency.

[5] From a more theoretical viewpoint, it is evident that both dynamic and thermodynamic effects will affect future precipitation. The dynamic effects on the one hand are associated with large-scale circulation changes. Such changes can be particularly significant if the mean flow changes relative to some precipitation-extracting mountain range or coast line. For instance, Southern and Central Europe is often affected by Mediterranean floods during fall that are associated with pronounced southerly flows that advect warm and moist air masses. Changes in circulation statistics may thus affect the frequency of heavy precipitation. There is evidence that such effects have been active in the past [e.g., Pauling et al., 2006] and are sometimes linked to large-scale teleconnection patterns [Bichet et al., 2013]. The thermodynamic effect on the other hand is associated with changes in atmospheric humidity and stratification. According to the Clausius-Clapeyron relation, the atmosphere's ability to take up moisture increases by approximately 7% per Kelvin warming. This moistening changes the atmospheric circulation and thermodynamic structure in fundamental ways [Held and Soden, 2006; O'Gorman and Schneider, 2009]. Climate model simulations confirm that in the global mean, this moistening actually takes place, while precipitation increases occur at a slower rate of 1–3% per Kelvin warming [Allen and Ingram, 2002; Emori and Brown, 2005]. Several studies also find that changes in extremes will not scale with changes in mean precipitation but possibly be controlled by the Clausius-Clapeyron relationship [Christensen and Christensen, 2003; Frei et al., 1998; Kharin and Zwiers, 2005; Pall et al., 2007].

[6] Future projections suggest that the observed trends will continue [Kharin et al., 2007; Solomon et al., 2007]. Yet, GCM simulations reveal limited estimates on subcontinental patterns due to their coarse resolution. Investigating the impact of climate change upon the hydrological cycle requires higher resolution, often with using RCMs [Giorgi, 2008, Leung et al., 2004]. Many studies focusing on European precipitation changes agree on a gradient like pattern with mean precipitation increasing in northern Europe and decreasing signals in the Mediterranean area [Beniston et al., 2007; Boberg et al. 2010, Frei et al., 2006; Giorgi and Coppola, 2007; Heinrich and Gobiet, 2011; Kundzewicz et al., 2005; Nikulin et al. 2011, Smiatek et al., 2009; Tomassini and Jacob, 2009].

[7] With this study, we analyze the output of 10 state-of-the-art RCM projections from the EU-ENSEMBLES project [van der Linden and Mitchell, 2009]. This project has provided a unique set of transient GCM and RCM simulations. We will address a centennial time scale and consider changes between a control period (1970–1999) and a scenario period (2070–2099) for an A1B scenario [Nakicenovic and Swart, 2000]. The choice of the scenario period is driven by the expectation that the climate change signal exceeds the natural variability [Hawkins and Sutton, 2009]. The analyses will closely follow a previous study of Frei et al. [2006] that has analyzed a previous generation of simulations from the EU-PRUDENCE project. Beside general model improvements, the main improvements since PRUDENCE are transient rather than time slice simulations, increased resolution (25 versus 50 km), and many different GCM simulations (rather than one single GCM) to better represent the model uncertainty. Following Frei et al. [2006], we analyze several indices for extremes and exploit generalized extreme value analysis. We account for climate variability by applying bootstrapping methods to discriminate real, statistically significant trends from internal variability.

[8] We systematically analyzed the results on a European scale and in addition conducted a detailed analysis covering 11 European subregions. However, due to space constraints, we present results primarily in terms of geographical maps covering Europe and choose the three Alpine subregions for analysis in the paper. Results for the other eight subregions are presented in the supplemental information. The Alps are particularly well suited for exemplary presentation, as this region has a high incidence of precipitation extremes, is affected by both the Atlantic and the Mediterranean Seas, and invokes both convective and stratiform precipitation processes. We will also make use of a unique precipitation data set for analysis [Frei and Schär, 1998].

[9] The objectives of the presented study are the assessment of 21st century changes in precipitation (mean and extremes) on European scale and the exploration of inter-model differences in the projected changes in the set of different RCM-GCM chains. The structure of the paper is as follows: Data and methodology are introduced in section 2. Detailed validation results for the Alpine and European regions then follow in section 3. The assessment of simulated climate changes is then presented in section 4. This presentation addresses European scale changes for a set of indices, and a detailed assessment in three Alpine subregions. Detailed results for other European subregions are presented in a large number of figures in the auxiliary material. The study is concluded in section 5.

2 Data and Methods

2.1 Indices

[10] Four descriptive measures are considered, intending to capture the (seasonal) climatology of precipitation and changes therein (see Table 1 for abbreviations and units). Measures of consideration are the mean precipitation (mea), wet-day frequency (fre), wet-day intensity (int), and the 90% quantile of the frequency distribution of wet days (q90). The 90% quantile serves as a measure for moderate extremes, which compared to estimates obtained by extreme value analysis generally features a higher detection probability for changes [Frei and Schär, 2001]. Under some circumstances, empirical measures can be related to extreme measures [Benestad et al., 2012]. All measures are directly derived from the empirical frequency distribution, using the same methodologies as in Frei et al. [2006]. A wet day is defined as a day on which the accumulated precipitation amount is larger than 1 mm. The threshold is introduced to overcome the drizzling effect that is present in many RCMs and to be consistent with previous studies [e.g., Frei et al., 2006]. In addition to the four climatological indices discussed above, return values are estimated for 1 day and 5 day accumulated precipitation (x1d and x5d; see section 2.3).

Table 1. Statistical Diagnostics Used in This Study
AbbreviationDefinitionUnit
  1. aA wet day is a day with precipitation ≥ 1mm.
  2. bReturn period: 2, 5, 10, 20, 50, 100 years.
freWet daya frequencyfraction
meaClimatological mean precipitationmm/d
intWet daya intensity, mean precipitation on days with precipitation ≥ 1 mmmm/d
q90Empirical 90% quantile of precipitation during wet daysamm/d
x1d.TTReturn value of 1 day precipitation intensity with a return period of TTbmm/d
x5d.TTReturn value of 5 day precipitation intensity with a return period of TTbmm/d

2.2 Data

2.2.1 Regional Climate Models

[11] The current analyses are based on daily precipitation totals from a set of 10 transient regional climate simulations (Table 2) from the EU-ENSEMBLES project [van der Linden and Mitchell, 2009]. The respective simulation strategy included a range of RCM experiments driven by transient GCM simulations [Christensen et al., 2010]. In addition, the RCMs are also driven by the ERA40 reanalysis [Uppala et al., 2005] for the period 1961–2000. All used RCMs are transiently run from the year 1951 (except SMHI-BCM, which starts in 1961) to 2099/2100 and feature a horizontal grid resolution of 0.22° (25 km). The data covers the entire European continent on a rotated latitude-longitude grid. All simulations are forced by the SRES A1B greenhouse gas scenario [Nakicenovic and Swart, 2000].

Table 2. ENSEMBLES Regional Climate Models Used in This Study
AcronymRCM and InstitutionGCM
DMI-ARPEGEHIRHAM5ARPEGE
Danish Meteorological Institute (DMI)
   
ETHZ-HadCM3Q0CLM2.4.6HadCM3Q0 standard sensitivity version
Swiss Federal Institute of Technology Zürich, (ETHZ)
  
HC-HadCM3Q0HadRM3.0
Hadley Centre (HC)
   
SMHI-HadCM3Q3RCA3.0HadCM3Q3 low sensitivity version
Swedish Meteorological and Hydrological Institute (SMHI)
  
HC-HadCM3Q3HadRM3.0
Hadley Centre (HC)
   
HC-HadCM3Q16HadRM3.0HadCM3Q16 high sensitivity version
Hadley Centre (HC)
   
DMI-ECHAM5HIRHAM5ECHAM5
Danish Meteorological Institute (DMI)
  
KNMI-ECHAM5RACMO2.1
Royal Netherlands Meteorological Institute (KNMI)
  
MPI-ECHAM5REMO5.7
Max Planck Institute of Meteorology (MPI)
   
SMHI-BCMRCA3.0BCM
Swedish Meteorological and Hydrological Institute (SMHI)

[12] For the purpose of the current study, a subset of the available models has been selected (Table 2), applying the following criteria: First, attention is restricted to RCMs that include the full scenario period 1951 to 2099/2100, with output at daily resolution (available at the time when the analysis was conducted). Second, we assure that as many GCMs are represented as possible. These are ECHAM5, BCM, Arpege, and the HadCM3. The last of these models is available in three versions representing different climate sensitivities: Q0 (low sensitivity), Q3 (standard sensitivity), and Q16 (high sensitivity).

2.2.2 Observational Data Sets

[13] We provide some validation for both European and Alpine regions. For Europe, validation will be conducted for precipitation intensity, while a more detailed validation for precipitation extremes is presented for the Alpine region.

[14] For the Alpine region, we use an observational reference dataset that spans the period 1971–1998 at daily resolution (in the following FS98). It was assembled by Frei and Schär [1998] and is based on a high-density observational rain gauge network. For the period 1971–1990, an average of about 5000 quality-controlled stations are available per day. Here we use an updated version. The data are available in the form of daily analysis grids with a resolution of about 25 km. FS98 is used to address the performance of the used RCMs in simulating extreme precipitation events (x1d.5) and climatological diagnostics (mea, int and fre).

[15] On a European scale, we validate the RCMs using the gridded E-OBS precipitation data set [Haylock et al., 2008] and refer to the period 1970–1999 (CTRL). We constrain the European validation to climatological indices (mea, int, and fre), as the E-OBS precipitation data set is known to systematically underestimate heavy precipitation [Lenderink, 2010; see also auxiliary material Figure S4] and is thus not suitable for the validation of extreme precipitation.

[16] For both, Europe and the Alps, validation is performed for RCM simulations driven by the ERA40 reanalysis as well as GCM experiments.

[17] The Alpine focus in terms of validating extreme precipitation is motivated by the fact that the available data set (FS98) is based on an unusually dense station network. In addition, the region is associated with a wide variety of climates ranging from Mediterranean-dominated in the south to Atlantic dominated in the north. Nevertheless, there are considerable uncertainties regarding the quality of the observational records. The two most significant factors relate to the systematic rain gauge undercatch (particularly relevant for winter conditions with snow fall) and the biased distribution of observing stations (with many valley and few mountain gauges). Associated uncertainties are also discussed in Frei and Schär [1998], Lenderink and van Meijgaard [2008], and Bellprat et al. [2012]. It should also be kept in mind that gridded (interpolated and/or spatially degraded) data sets do not capture the range of values (extremes) as well as the spatial and temporal variability that is covered by single-point measurement.

2.3 Statistical Methods

[18] In order to capture the climatological character of precipitation and its extremes, a set of several statistical diagnostics is taken into account. Along with descriptive statistical measures (section 2.1), extreme value analysis is applied to estimate rare and extreme precipitation events. All analyses are based on daily output and are carried out seasonally in order to capture the climatological characteristics and avoid compensation effects by averaging over the year. Seasonal subdivision is undertaken according to meteorological definitions (e.g., DJF = December, January, and February, winter). The focus of the analyses is on two 30 year periods (in the case of DJF 29). Period 1970–1999 serves as control period (CTRL) representing present conditions and time slice 2070–2099 as scenario period (SCEN) standing for future conditions. For validation against FS98, we employ the period 1971–1998 to be consistent with the observational reference data set FS98. Change signals refer to SCEN with respect to CTRL and are generally expressed as ratio (SCEN/CTRL). Moreover, all calculations are performed on grid-point scale. Further aspects are discussed in the following subsections.

2.3.1 Extreme Value Analysis

[19] Return values are estimated for 1 day and 5 day accumulated precipitation events (x1d and x5d) for return periods ranging from 2 to 100 years. Such events lie in the very far tail of the frequency distribution, or even exceed the highest record of an underlying data set (e.g., 30 years). For their estimation, we adopt methods of generalized extreme value theory [Coles, 2001; Katz et al., 2002]. A recent WMO report [Klein Tank and Zwiers, 2009] recommends using upper percentiles (e.g., q90) and extreme value theory for the assessment of extremes. Generalized extreme value theory, as used in our study, has also been applied in several previous climate-projection studies [Fowler et al., 2007; Frei et al., 2006; Hanel and Buishand, 2011].

[20] Generalized extreme value theory belongs to the block-maximum techniques and requires the definition of blocks from which only the highest record of each block is further considered in the process of extreme value analysis. We define blocks based on seasons within the 30 year time slices (29 for DJF). After seasonally subdividing the data, we separately extract the largest daily record within each season at each RCM grid cell. The resulting data set consists of 30 annual maxima for each season at grid-point scale. Following the “Extremal Types Theorem,” the distribution of a series of block maxima is distributed according to one of the three asymptotic limit distributions: namely the Frechet, Gumbel, or Weibull distribution. These three limit distributions can be characterized by one single distribution, the generalized extreme value distribution (GEV). The GEVs cumulative distribution function (CDF) is given by inline image with inline image. The GEV depends on three parameters: the location (μ), the scale (σ), and the shape (ξ) parameters. We estimate these three parameters using a maximum-likelihood estimation procedure and fit a GEV to all series of seasonal maxima at each grid point. The adopted estimation incorporates a Bayesian prior distribution for the shape parameter, which was proposed by Martins and Stedinger [2000] for the application to geophysical data in order to avoid unrealistic estimates. It was also used by Frei et al. [2006], who additionally illustrated the enhanced skill of implementing the prior, above all for hydrological data sets with small sample sizes.

[21] Having estimated the GEVs parameters, one is able to determine return values and associated return periods after inverting the GEVs cumulative distribution function (CDF). The resulting function is the quantile function: inline image where Xp relates to the p year return value, which is the daily precipitation rate (intensity) exceeded once in a period of p years. Respectively, Xp refers to the threshold exceeded once in any year of interest with a probability of 1/p.

[22] We address 1 day and 5 day precipitation events, intending to capture both short-term (1 day) and long-lived multi-day precipitation (5 days) events. The former are primarily related to convective events, and the latter to synoptic systems and persistent flow conditions. The estimation of 5 day events is based upon daily time series to which a 5 day running mean is applied before extreme value analysis.

2.3.2 Alpine Subregions

[23] Aiming at a detailed analysis of projected changes and inter-model differences in the Alpine region, we define three climatologically characteristic subregions. These are mapped in Figure 1. Their definition is similar as in Frei et al. [2006]. The Southern Alpine region (S) features a Mediterranean character. Other than that in Frei et al. [2006], the Northern region is split up into two subregions: The Northwestern Alpine region (NW), which features a maritime character and is mainly being influenced by Atlantic circulation regimes, and the Northeastern Alpine region (NE), which features a more continental character with the annual cycle of precipitation peaking in summer due to convective events. While the definition of these regions is subjective, it can also be motivated by a rotated EOF analysis [Schmidli et al., 2001].

Figure 1.

Alpine subregions used in this study: Northwestern (NW), Northeastern (NE) and Southern (S) Alps. Bold lines indicate the 700 m isolines as presented by the E-OBS Topography.

[24] Some of the analyses will be applied to domain-averaged daily precipitation. Compared to extreme value estimates as derived from single grid cells, spatial averaging over large areas leads to an increased probability of detecting temporal trends in return value estimates [Frei and Schär, 2001].

[25] Additionally, in the auxiliary material, we show results for the same analyses but applied to European subregions as defined by Christensen and Christensen [2007] and shown in Figure S4.

2.3.3 Uncertainty, Confidence, and Significance

[26] To identify significant change signals, resampling methods are taken into account to estimate confidence intervals [Wilks, 2006]. Uncertainty especially arises from the internal variability of the climate system. To adequately incorporate climate variability in our future and present-day projections, we apply a non-parametric bootstrapping approach.

[27] Confidence internals for estimates of domain-mean changes in diagnostics are estimated based on the following procedure (see also in Frei et al. [2006]). First, 50 bootstrap samples of absolute domain-mean values are generated for each time slice, CTRL (1970–1999) and SCEN (2070–2099). These non-parametric bootstrap samples are based on time series obtained by resampling of time steps (with replacement) from the original 30 year data set of regional averages. This choice assures that spatial correlations are accounted for. Change estimates are generated by resampling between the 50 pairs of SCEN and CTRL absolute domain-mean values by calculating 50 ratio estimates. To obtain 95% confidence intervals, quantiles are applied to the 50 estimates of change. The best estimate is defined as the median. A change signal is rated significant, if the ratio 1.0 (no change) does not fall into the confidence intervals range.

[28] Uncertainty ranges of absolute domain-mean values are also estimated by means of resampling. Like in the procedure above, 50 non-parametric bootstrap samples are estimated by resampling in a period of interest. Such estimates are used for validation purposes and in a multi-model approach, which is further explained in the respective section of the paper (section 4.3).

3 Validation

[29] Before discussing and interpreting the main results, we present the individual RCMs skill in simulating extreme precipitation in the Alpine region in terms of the 5 year return values for 1 day precipitation events in period 1971–1998 in Figure 2.

Figure 2.

Validation of 1 day heavy precipitation events in terms of seasonal 5 year return values (x1d.5) for three Alpine sub-regions (Figure 1). Left-hand panels are for ERA40-driven RCMs, right-hand panels for GCM-driven RCMs. Individual simulations are shown by vertical lines and denoted by symbols showing best estimates (symbols) and 95% bootstrap confidence intervals. FS98 observations are presented by green shaded areas and horizontal lines. All panels are based on the period 1971–1998.

[30] In addition, the skill of RCMs in simulating climatological precipitation indices for eight European subdomains in the period 1970–1999 (CTRL) is also presented. We show and discuss the results for France in Figure 3, and the other European subregions are shown in the supplemental information (Figures S5 to S11).

Figure 3.

Validation of mean precipitation (mea, top row), wet-day intensity (int, middle row), and wet-day frequency (fre, bottom row) for France (continental grid boxes between 44°N–50°N and 5°W–5°E; see auxiliary material Figure S4). Left-hand panels are for ERA40-driven RCMs, right-hand panels for GCM-driven RCMs. Individual simulations are shown by vertical lines and denoted by symbols showing best estimates (symbols) and 95% bootstrap confidence intervals. E-OBS observations are presented by blue shaded areas (95% confidence intervals) and horizontal lines (best estimates). All panels are based on the period 1970–1999 (CTRL).

[31] Figure 2 illustrates corresponding GEV estimates from both ERA40- and GCM-driven RCM simulations (left- and right-hand panels), with a focus on the four seasons and the three defined Alpine regions (Figure 1). Additionally to simulated values, estimates derived from observations (FS98) are depicted as qualitative reference. Supplementary Figures show the validation of mean (mea), frequency (fre) and intensity (int). However, these are not discussed in the following.

[32] Some of the RCMs reasonably reproduce the character and climatology of heavy precipitation in the Alps, but there are also models with significant biases. As expected, the reanalysis-driven simulations perform better than the GCM-driven simulations. To a certain degree, however, the obvious deficiencies (e.g., overestimation or underestimations in particular seasons) seen in the GCM-driven simulations are also apparent in ERA40-nested realizations. For instance, all DMI simulations (except those driven by Arpege) systematically overestimate heavy precipitation events, particularly in winter and fall, and in the Northwestern (NW) and Southern (S) Alpine subdomains. Thus, it appears that some variations of simulated values away from observations can be attributed to individual RCM parameterizations and formulations and not to driving GCMs.

[33] The most pronounced biases, primarily seen in the GCM-driven realizations, relate to an overestimation of heavy precipitation in the Northwestern (NW) and Northeastern (NE) Alps. These overestimations are most obvious in winter and spring. In the Southern Alps (S), models have a tendency to underestimate heavy precipitation in the warm seasons, particularly in summer. However, most models qualitatively capture the observed seasonal cycle in the three regions. An obvious exception is DMI-Arpege, which simulates a reversed seasonal cycle with very dry summer conditions. Based on observations, the most distinct seasonal cycle appears in NE with most severe events occurring in summer. In S we see a clear precipitation peak in fall, while in NW there is only a small seasonal cycle. The widths of the uncertainty ranges additionally give inference about inter-annual (year-to-year) variability of heavy events, being largest in NE and NW in summer and in S in winter. RCMs reproduce the observed variability qualitatively well in all seasons and regions.

[34] Overall, there is no evidence for a GCM-dependent clustering of RCM-GCM integrations in terms of simulated absolute domain-mean return values. As the corresponding ERA40-driven simulations show similarities to the GCM-driven simulations, one can rather suggest that individual RCM formulations highly govern the absolute values. It is remarkable and not fully understood that extreme indices (Figure 2) appear to be better captured by the RCMs than empirical indices such as int (see auxiliary material Figures S1S3).

[35] Figure 3 illustrates in the same format the performance of RCMs with respect to basic precipitation measures (mea, int and fre) for the European region France. France represents a case with median skill compared to the other seven analyzed European subregions (Figures S5S11). For instance, the validation in the regions British Isles, Iberian Peninsula and the Mediterranean is better, and validation is worse in the regions Scandinavia and Central Europe. Intercomparison between reference data sets E-OBS and FS98 is shown for the Alpine Region (see auxiliary material Figure S9). It shows that there are significant differences between the two data sets in terms of precipitation intensity in summer and spring, with too weak intensities observed in E-OBS.

[36] For France (Figure 3), a majority of models overestimate mean precipitation but qualitatively capture the character of the seasonal cycle. Overestimation (e.g., DMI-ECHAM5) and inter-model spread are more pronounced in GCM-driven realizations than in ERA40-driven realizations, this result is similarly seen for wet-day intensity and frequency. Contrary to the large overestimations seen in all diagnostics and seasons by the DMI-ECHAM5 model run, the Arpege-driven realization of the DMI model distinctly underestimates diagnostics in summer and, less pronounced, in fall. Most models perform reasonably well in simulating the intensity of precipitation throughout the year. However, some models tend to underestimation in summer and fall (above all HadCM driven RCMs). In winter, many models overestimate intensity. In terms of wet-day frequency the majority of models simulate too large values, except for GCM-driven runs in summer.

4 Results

4.1 European Patterns of Changes in Heavy Precipitation Events

[37] We proceed by discussing the 10-member-ensemble-median seasonal changes in precipitation diagnostics on a European scale for 2070–2099 (SCEN) versus 1970–1999 (CTRL). Figure 4 presents the 10-member-ensemble median change signals in fre, mea, int, q90, x1d5 and x5d5 (from left to right) for the four seasons (from top to bottom). Changes are shown in terms of ratio (SCEN/CTRL), and areas where 90% of the models (9 out of 10) agree on the sign of change are emphasized by a stippled overlay. Note that parts of the Mediterranean region are masked out in summer projections for extreme value estimates as they experience too dry conditions for reliable estimates on future changes. We additionally show (but not discuss) near term changes for period 2020–2049 versus 1970–1999 in the supplementary material to this paper (Figure S12).

Figure 4.

European projections of the ensemble-median climate change signals in precipitation frequency (fre), mean (mea), intensity (int), the 90% quantile (q90) and the 5 year return values of 1 day (x1d.5) and 5 day (x5d.5) precipitation intensities (left to right) for the four climatological seasons (top to bottom). The results are based on 10 ENSEMBLES RCMs (Table 2). Changes are expressed as ratio between SCEN (2070–2099) and CTRL (1970–1999), with blue (red) representing increases (decreases). Stippling denotes grid points where 90% of the RCMs agree on the sign of change. Regions experiencing too dry conditions for reliable return value estimates are masked out (grey). Bold black lines indicate the 700 m isolines as presented by the E-OBS Topography.

[38] The panels showing 2070–2099 changes suggest, that remarkable changes in the precipitations character are to be expected in the course of the 21st century. The projections imply a rising risk of more frequent extreme precipitation conditions, both in terms of extended dry spells (see mea and fre) and heavy precipitation events (see other indices), and depending upon the season.

[39] Further, the maps markedly illustrate that changes in basic diagnostics (mea and fre) scale disproportionate or even opposite to changes in intense (int and q90) and extreme (x1d5 and x5d5) diagnostics. Thus, it can be assumed that different processes govern the response of the hydrological cycle to climate change.

4.1.1 Basic Diagnostics: Mean and Frequency

[40] Projected patterns in basic diagnostics, namely, mean precipitation (mea) and wet-day frequency (fre), reveal that central Europe lies within a transition zone showing a sharp gradient between substantially decreasing precipitation in southern Europe (the Mediterranean, subtropics) and stationary (fre) or increasing (mea) signals in northern Europe (Scandinavia, high latitudes). The location of this transition zone experiences a seasonal shift: lying further north in summer (at 60°N, across southern Scandinavia) and further south in winter (at 40°N, across the southern Mediterranean Sea). Especially in summer, distinct decreases in the amount and occurrence of precipitation—with reductions larger than −30% (ratio < 0.7)—are projected across broad areas of continental Europe. The models agree on the aforementioned pattern (see stippling).

[41] Robust increases in the number of wet days (fre) are only projected in winter north of about 50°N and along the southern Alpine rim. Contrary to the projections for fre, large parts of Europe are affected by agreement on increases in mea in winter, fall, and across some regions in spring. Especially in winter, there is model agreement on distinct increases in mea north of the Alpine rim and along the southern and eastern slope of the Alps. In fall and spring, large-scale agreement on an increase in mea is only seen in proximity to the North Sea and across Eastern Europe.

4.1.2 Intense Diagnostics: Intensity and the 90% Quantile

[42] Oppositional to the variable spatial pattern regarding the direction of change in mea, intensity (int) and the 90% quantile (q90) show an obvious tendency towards widespread increases in all seasons, except summer. Positive signals are even found at some locations where mea and fre are concurrently projected to significantly reduce, like for example over France and the Balkans in spring or across large parts of Germany in summer. Widespread model agreement is seen in winter and fall, where int and q90 are at large scale projected to increase by around +15% in the ensemble median. In spring, slight increases are projected north of 45°N, whereas a peculiar pattern of amplified intensifications is evident along the northern Alpine foreland. Agreement on an intensification of summer events is only visible for Scandinavia. Other regions experience a more or less stable signal in summer with a leaning towards slight increases. In general, there is high correlation between the projected patterns for int and q90.

4.1.3 Extreme Diagnostics: 5 Year Return Values for 1 Day and 5 Day Events

[43] To address more extreme conditions than represented by int and q90, consideration is given to changes in 5 year return values, both for 1 day (x1d.5) and 5 day (x5d.5) accumulated precipitation.

[44] The spatial patterns of change in extreme diagnostics are similar to the projected patterns for intense diagnostics (int and q90) but more variable in magnitude. Despite their highly similar patterns, amplifications (from q90 to x1d.5 and x5d.5) are stronger for 1 day than for 5 day precipitation episodes. Especially in fall and winter, model agreement on increases is visible across most areas of central Europe. In comparison to changes in mea, the characteristic signal of intensification extends much further south. For instance, for France in spring and fall, mean precipitation is projected to change little, while x1d.5 exhibits ensemble median increases by about +10% to +20%.

[45] Apart from widespread and prominent increases, decreases in return values are a common feature in the projections along Mediterranean coastal areas in spring, fall and above all in summer. In summer, projected decreases even spread out north and affect France and the western parts of central Europe. However, the reductions in return values are not supported by model agreement and much less distinct than projected change in intense diagnostics (int and q90). In this regard, one should generally note that uncertainty in extreme value estimates increases at dry locations (dry grid points, respectively) with a small number of wet days. For this reason, grid points experiencing an ensemble-median wet-day frequency smaller than 5% (less than 5 days per season) under future conditions are masked as grey in the return value maps.

4.2 Alpine Projections

[46] Here we present a detailed inter-model comparison of projected domain-mean changes in several precipitation statistics for the three Alpine subdomains defined in Figure 1. Results are presented in Figure 5 in a similar format as in Figures 2 and 3. In addition, Table 3 presents full results for the Southern Alpine (S) subdomain. Full results for the Northwestern (NW) and Northeastern (NE) region are given in the supplementary material.

Figure 5.

Changes in domain-mean precipitation diagnostics (Table 1) for the three Alpine subregions Northwestern (NW),Northeastern (NE), and Southern (S) Alps (Figure 1). Changes are expressed as ratio between SCEN (2070–2099) and CTRL (1970–1999). Results are presented for precipitation frequency (fre), mean (mea), intensity (int), the 90% quantile (q90), and return values for 1 day events with return periods of 5-(x1d.5), 10-(x1d.10), and 50 years (x1d.50) (from left to right in each panel). Blocks of panels relate to DJF (top left), MAM (top right), JJA (bottom left), and SON (bottom right). Vertical lines depict 95% bootstrap confidence intervals, and symbols depict the best estimate.

Table 3. Changes in Precipitation Indices (Table 1) for the Southern Alpine Subregion (S) for the Four Climatological Seasons (Top to Bottom)a
RCM-GCMSouthern Alps
 fremeaintq90x1d.5x1d.50
  1. aChanges are expressed as ratio between SCEN (2070–2099) and CTRL (1970–1999). The numbers indicate the best estimate as obtained by non-parametric bootstrapping; bold numbers indicate significant change signals (Section 2).
DJF
DMI-Arpege1.031.001.061.091.081.06
ETH-HC01.030.971.151.151.161.08
HadRM-HC01.081.131.121.131.151.13
HadRM-HC30.961.381.171.201.131.09
HadRM-HC160.981.321.221.261.331.34
SMHI-HC31.051.151.151.121.261.24
DMI-ECHAM51.051.011.051.051.081.02
KNMI-ECHAM51.100.941.041.061.071.04
MPI-ECHAM51.091.001.051.061.071.08
SMHI-BCM0.960.890.980.991.031.03
MAM
DMI-Arpege0.830.610.960.980.930.94
ETH-HC00.950.991.081.081.131.16
HadRM-HC00.890.901.041.051.121.15
HadRM-HC30.980.951.051.061.051.09
HadRM-HC160.840.881.081.071.091.11
SMHI-HC31.051.101.091.121.111.06
DMI-ECHAM50.860.681.001.041.031.04
KNMI-ECHAM50.830.631.000.980.991.01
MPI-ECHAM50.840.571.021.021.011.05
SMHI-BCM0.960.751.011.031.001.03
JJA
DMI-Arpege0.760.911.121.151.091.25
ETH-HC00.700.520.970.960.830.91
HadRM-HC00.760.590.960.980.900.99
HadRM-HC30.800.720.920.890.901.00
HadRM-HC160.740.490.900.920.800.84
SMHI-HC30.890.870.991.011.021.04
DMI-ECHAM50.840.631.020.980.981.02
KNMI-ECHAM50.720.621.021.020.941.07
MPI-ECHAM50.710.611.011.030.991.02
SMHI-BCM0.890.750.970.980.991.05
SON
DMI-Arpege0.870.821.031.031.061.07
ETH-HC00.901.001.201.221.231.24
HadRM-HC00.881.021.121.171.211.22
HadRM-HC30.940.851.131.171.171.12
HadRM-HC160.710.741.101.131.091.20
SMHI-HC30.970.881.101.151.151.09
DMI-ECHAM50.930.861.081.111.171.22
KNMI-ECHAM50.930.871.071.101.131.10
MPI-ECHAM50.950.751.081.071.121.17
SMHI-BCM0.900.720.910.910.971.03

[47] Additionally, we also present the same set of figures for eight European subdomains, covering the entire European continent, in the supplemental material to this paper (Figures S13S20).

[48] The results demonstrate that, for most indices and seasons, the RCMs qualitatively agree concerning the structure of changes, i.e., regarding the sign of change in single diagnostics and the inter-relation of projected changes between the different diagnostics. Nonetheless, the projections are afflicted with large inter-model spread. Inter-model spread is largest for mean precipitation (mea), for which even variable signs of change are found in some seasons and regions, especially in fall. The spread is probably due to the location of the Alps near the transition zone between latitudinal bands of decreasing (in the south) and increasing (in the north) precipitation. Smallest spread is seen for intensity (int) and the 90% quantile (q90). Projections of change in return values correlate with projections for int and q90. However, inter-model spread is more pronounced and uncertainty, expressed by the width of the confidence intervals, is markedly larger.

[49] In terms of qualitative statements, a general intensification of precipitation events is robustly projected by a majority of models. Merely in summer projections and S rather uncertain signals appear for intense and extreme precipitation. In summer, articulate reductions in mea and fre are evident in all integrations and regions. Interestingly, there is a simultaneous leaning towards slightly intensifying events in NW and NE. In spring and fall, some models show summer-like reductions regarding mea and fre, above all in S.

4.2.1 Winter

[50] In winter (Figure 5, top-left panels), many models simulate increases in mean precipitation, which is mainly due to an increase in intensity (int), while changes in frequency (fre) are projected to be small and for many models not statistically significant. In NW and NE, DJF increases in mea amount to typically +10% (ensemble mean, ratio 1.1), but some models project insignificant or even negative changes. Especially in NW, the ECHAM5-driven integrations project significant and strong increases in mean (DMI: +25%, MPI: +19%, KNMI: +23%). The most pronounced signals in mea are however found in S for HadCM3-driven models, showing increases up to +38% (HC-HadCM3Q3), while the ECHAM5-driven models (represented in blue) show changes in mea that are not statistically significant. In general, increases in mea are accompanied by increases in int and q90. Significant increases in int and q90 are projected across all regions and by most models by typically around +10%. Similar increases are seen for the extreme diagnostics (x1d.5, x1d.10 and x1d.50) with return periods of 5, 10, and 50 years, but they are mainly not statistically significant.

4.2.2 Spring

[51] A robust feature in RCM projections for spring (Figure 5, top-right block of panels) is an intensification of precipitation across the northern Alpine regions in similar manner like in fall (compare to Figure 5, bottom-right panels). This almost model-independent statistically significant increase in int is not evident in fre or mea but accompanied by increases in q90 and 1 day extremes (x1d5, −10 and −50). In S, a highly GCM-dependent picture arises. On the one hand, the ECHAM5-driven (as well as the BCM and Arpege driven) models exhibit strong reductions in fre and mea, along with insignificant changes in int, while HadCM3-driven models project slight increase in intense and extreme diagnostics. Also exceptional in S is the DMI-Arpege, which projects decreases in all diagnostics, disclosing a summer-like structure of change (see also Table 3). The GCM-dependent inter-model spread in S is likely due to differences in the synoptic climatology and highlights the uncertain location of the seasonally shifting transition zone between drying (moistening) regions, discriminating increasing (decreasing) precipitation in the north (south).

4.2.3 Summer

[52] Substantial reductions are consistently projected for the drought-relevant indices mea and fre in summer (Figure 5, bottom-left block of panels). With only few exceptions, mea is projected to significantly and substantially reduce in all three regions and in all RCM projections. In S and NW, models suggest very severe reductions in mea. The majority of models project reductions exceeding −30% (ratio < 0.7), and some models even simulate reductions of −50%. (in S and NW for ETHZ-HadCM3Q0 and HC-HadCM3Q16). Projected reductions in mea are accompanied by decreases in fre, which are statistically significant throughout all model chains and regions. In NE, reductions, in mea are not as prominent as in NW and S. Despite significant reductions in mea and fre, the frequency of intense and extreme precipitation events (int, q90 and x1d.5, −10 and −50) exhibits increases (in particular in NW and NE) or little changes (in S). This effect is evident for all model chains, but particularly strong for ECHAM5-driven RCMs in NE and NW. To provide an impressive example, consider the KNMI-ECHAM5 integration in NE, which shows significant signals in all diagnostics. These are decreasing for mea (−16%) and fre (−17%) and increasing for intense (int: +12%, q90: +14%) and extreme diagnostics (x1d.5: +12%, x1d.50: +20%). Similarly, the ECHAM5 projections in NW show a significant reduction in mea of about −40%, while the 50 year return value of 1 day events is projected to significantly increase with around +25%. In summary, when considering the ensemble-mean change signal of x1d.50 one finds an increase of +8% in NW, +10% in NE and +2% in S (see also Figure 4). On the other side, the ensemble-mean change signal for mea is −31% in NW, −12% in NE and −33% in S. Ensemble-mean reductions in fre fall into a similar value range. The results pronouncedly show that changes in mean and extreme precipitation do not simply follow each other. Rather most feature change in opposite directions.

4.2.4 Fall

[53] Strong changes concerning extreme precipitation events are projected in fall (Figure 5, bottom-right panels). A majority of RCMs projects significant amplifications in intense (int and q90) and extreme (x1d.5, −10 and −50) precipitation diagnostics in all three regions. Strongest amplifications are projected in NE and NW, where increases for intense and extreme diagnostics mostly lie in a range between +10% and +30%. Concurrently to the projected intensifications, many models project reductions in fre across all regions by about −10%. The ECHAM5-driven RCMs project strongest increases in extremes (e.g., x1d.50 in NW for KNMI-ECHAM5: +36%), while HadCM3-driven RCMs feature statistically significant reductions in fre (HC-HadCM3Q16: −26%, −25% and −29% in NW, NE and S respectively). Despite the pronounced increases in the strength of precipitation, one sees a summer-like (Figure 5, bottom-left panel) behavior of changes in mea and fre in S, where 8 out of 10 models project significant decreases in fre between −3% and −29%. While most models distinctly project different signs for changes in heavy precipitation measures and fre, there is a large spread in the projections for mea. Taking NE as an example, one sees projections for mea in a range between −15% and +18%, whereas 3 models project significant reductions and 4 out of 10 models significant increases.

4.3 Assessment of Alpine Return Values and Uncertainties

[54] Return values of heavy precipitation events are commonly used for a wide range on planning steps, for instance, in the dimension of traffic infrastructures, bridges, sewer systems, dams, and for many additional water management purposes. In practice, the respective assessment is currently exclusively based on past climate records, i.e., return values are estimated from long precipitation series. With climate change, the past will lose some of its value as a guide to the future, and the assessment of heavy precipitation events will have to take into account climate model results.

[55] Here we provide some preliminary assessment of how such a change strategy might affect the assessment of extremes. To this end, we compare observation-based and model-based estimates for the past climate and consider model-based projections for the future.

[56] Figure 6 depicts return values and associated uncertainties for future (SCEN: 2070–2099, in red) and present-day (CTRL, 1970–1999, in blue) conditions, as a function of return period (between 2 and 100 years). The estimates are based on a multi-model ensemble for each climatological season (top to bottom) and for the three defined Alpine subregions (left to right; see Figure 1). Estimates based on observational records are depicted (FS98, 1971–1998, in black) to serve as quantitative validation for CTRL.

Figure 6.

Return values (mm/d) of 1 day extreme precipitation events as a function of return period (years), estimated by generalized extreme value analysis in three Alpine sub-regions (from left to right; see Figure 1) and the four climatological seasons (top to bottom). Each panel contains estimates and associated uncertainty ranges according to past observations (FS98, 1971–1998, black line and grey shadings), and model-based projections for present (CTRL, 1970–1999, blue line and light-blue shadings) and future (SCEN, 2070–2099, red line and light-red shadings) climatic conditions. Shadings denote the 95% bootstrap uncertainty ranges and bold lines the best estimate for a return value. Projections are based on a multi-model ensemble including 10 RCMs (Table 2). Best-estimate relative changes (in %) between CTRL and SCEN are labeled at the top of each panel, with blue (red) color denoting increases (decreases) exceeding ±5%, black color denotes stable conditions with changes in the range below ±5%.

[57] For all data sets considered, the estimates are derived using generalized extreme value theory (section 2). The uncertainty assessment is based on the aggregation of 50 bootstrap resampling estimates (section 2) of absolute domain-mean return values from each one of the 10 considered RCMs (Table 2). The 50 estimates from each model are then aggregated to a multi-model ensemble of 500 model-based estimates, for a return value associated with each return period (2, 5, 10, 20, 50, and 100 years) for each season and subregion. In the case of FS98, we apply the same techniques but resample 500 times to establish consistency with the multi-model ensemble. We use percentiles to define the uncertainty ranges and the best estimates. The 97.5% percentile presents the upper percentile, and the 2.5% percentile the lower bound of the 95% uncertainty range. The median of the entire set of estimates represents the best estimate for a return value.

[58] It is important to note that this uncertainty assessment based on 10 RCM-GCM chains has considerable limitations. First, limited model spread may lead to an underestimation of the associated uncertainty, as the models considered bear some structural similarity, and are not designed to span an uncertainty range [Knutti et al., 2008]. In addition, they are only driven by one single greenhouse gas emission scenario (A1B). Second, there is also the possibility that for the particular parameter considered (extreme daily precipitation) the model spread might overestimate the uncertainty, as the RCM-accomplished downscaling step may include outliers or erroneous models [Frei et al., 2003]. Finally, even the uncertainty assessment using observations likely underestimates the uncertainty, as it is based on one single (28 year) realization that underestimates decadal variations. Indeed, analysis of long-term observations and proxy data suggest that there is considerable decadal variability that cannot be captured within a few decades of data [Schmocker-Fackel and Naef, 2010a; 2010b]. The uncertainty estimates could thus be considered as an “estimate of opportunity,” which may be criticized on well-founded scientific grounds. However, such estimates regularly end up being used in water resource management, for instance, if a dam needs to be dimensioned today.

[59] Comparison of black (observations, FS98) and blue (RCMs, CTRL) curves in Figure 6 shows that the skill of the multi-model ensemble is generally good, as for most cases the best estimates of CTRL lie within the observational (FS98) uncertainty range, therewith confirming that our definition of the best estimate results in a realistic and robust measure for a return value. Best skill is seen for summer-time events in NE and NW and in fall in S. Overestimation is notably for heavy events in fall in NW, while the intensity of summer-time events is underestimated in S (see also Figure 2).

[60] In terms of future changes, the main results of Figure 6 are as follows: First, except for MAM and JJA in S, the models consistently project relative increases in best-estimate return values between CTRL and SCEN. The range of these increases depends upon season, subregion, and return period. Strongest intensifications are projected in fall: in NW, there is a uniform increase of return values across the event spectrum of about +14%. In S, one sees projected increases between +10.1% (2 year event) and +15.3% (100 year event). The largest changes are projected for NE, where relative changes lie in a range between +17.7% (2 year event) and +21.4% (100 year event). These changes are very substantial. Taking the 50 year return value in NE as an illustrative example, the change of +21.3% corresponds to a shift of the 50 year return value from 66.7 mm/d (CTRL) to 88.9 mm/d (SCEN). Such a large return value would comply with a return period larger than 100 years under present-day (CTRL) conditions (the 100 year return value amounts to 75.6 mm/d for CTRL).

[61] A peculiar effect is evident for summer and subregion S. Here, the 2 year return value is projected to decrease by −17.8%, whereas the 100 year return value is projected to increase by +4.8%. This result is consistent with the diverging behavior between mea and int previously discussed along Figures 5 and 4. A similar effect is evident for summer and subregions NW and NE, where the relative changes in return values increase with return period.

[62] Second, projected relative changes in return values are smaller than suggested by the Clausius-Clapeyron relationship (i.e., 7%/K warming, see introduction). This suggests that the simulated changes cannot simply be explained by changes in water-holding capacity with the simulated warming, but generally involve more complexity such as changes in atmospheric circulation and stratification.

[63] Third, in all regions and seasons, the observed changes in terms of best estimates are not statistically significant at the 97.5% level; i.e., the median of SCEN falls within the 95% bootstrap uncertainty range of CTRL. For certain regions and seasons, single models do strongly influence the sensitive lower and upper bounds of the uncertainty range. Exemplary is the DMI-Arpege integration in MAM in NE and S (Figures 5 and 2). Indeed, choosing an 80% (rather than 95%) range would considerable narrow uncertainty as extremely underestimating and overestimating models would not further be considered. However, we stick to our definition as it illustrates the entire spread of different model projections. Further noticeable is the systematically narrower uncertainty range of FS98 compared to CTRL and SCEN. This is partly because FS98 acts as one single realization compared to 10 individual RCM-GCM integrations that contribute to CTRL and SCEN (see above).

[64] Fourth, from a practical point of view, i.e., for an engineer that needs to dimension a planned dam or other structure, the projected changes are very substantial and exhibit a serious dilemma. Assume that this engineer has the task to protect some water resource structure against a 100 year precipitation event, and to account for the uncertainties present in this type of analysis. Consider as a specific example subregion NW in MAM, with the intent to protect against a 100 year precipitation event with a probability of 97.5% (i.e., taking the upper bounds of our analysis). This task implies the following protection levels: FS98 (observations, 1971–1998): 71.5 mm/d, CTRL (models, 1970 –1999): 87.2 mm/d and SCEN (models, 2070 –2099): 103.7 mm/d. These figures illustrate the dilemma of water management engineers, who often need to dimension a structure (i.e., dam or bridge) with a lifetime of > 100 years (without the luxury to wait for a more detailed assessment). Switching the assessment from FS98 to SCEN (in order to account for climate change) would imply to increase the protection level by almost 50% (from 71.5 mm/d to 103.7 mm/d). Yet almost half of this increase is not due to climate change, but rather due to the choice of another methodology (switching from observations to model-based assessment). This point illustrates the urgent need to develop improved assessment methodologies for heavy precipitation events.

5 Conclusion

[65] We have analyzed 10 transient regional climate projections from the ENSEMBLES project in terms of precipitation and its extremes across Europe. The assessment was based on statistical measures as derived from the empirical frequency distribution, i.e., on empirical indices of extremes and general extreme value analysis. Our results rely on the comparison of future (2070–2099) with present-day (1970–1999) climatic conditions. In the paper we have presented European scale results as well as a detailed analysis for the Alps (corresponding figures for 8 other European regions can be found in the supplemental information).

[66] The analyzed RCM-GCM chains feature a resolution of 25km and are based on 10 combinations from six different RCMs and four driving GCMs. Two of the RCMs provided two realizations driven by different GCMs and one RCM was available in three realizations, driven by three GCM integrations using three different climate sensitivities.

[67] A validation of GCM-driven and reanalysis ERA40-driven realizations of the used RCMs under present conditions shows that the models generally perform reasonable in simulating the climatology of mean precipitation and heavy events over the complex topography of the Alpine region. Most RCM simulations capture the characteristic seasonal cycle of precipitation extremes in different Alpine regions, irrespective of whether they are driven by ERA-40 reanalysis data or by a free running coupled GCMs. Validating the set of models against European scale observations using climatological indices, also shows reasonable results. However, the skill of the RCMs seems better in southern and western Europe, compared to northern and eastern parts of Europe.

[68] The presented projections suggest that remarkable changes in the character of European precipitation are to be expected by the end of the 21st century. The results imply a rising probability of more frequent extreme precipitation events and extended dry spells.

[69] The models agree on characteristic spatial dipole patterns in terms of changes in all considered diagnostics with partially sharp gradients separating increases in the north from decreases in the south of Europe. The position of the transition zone, between increasing and decreasing signals, depends on the diagnostic regarded and is dependent on the climatological season. In winter (summer), the position of the transition is generally more to the south (north). Decreasing change signals extend much more north as one regards mean precipitation and wet-day frequency. For instance, in summer models agree on substantial reductions in mean and frequency that reach up to 60°N with widespread decreases to only two thirds of present conditions. In spite of large areas of Europe affected by decreases in mean and frequency, there is model agreement on an intensification of precipitation events across all of Europe, except for very southern parts of the Mediterranean and the Iberian Peninsula. Strongest, robust and at large-scale significant increases are projected in fall and winter and in general across northern Europe. The intensification of precipitation events is largely independent of the changes in mean precipitation. Some areas show different signs of changes in the two diagnostics, with intensifying but less frequent events accompanied by decreasing precipitation amounts, especially in summer and across central Europe. These patterns accentuate that different physical processes govern the complex response of the hydrological cycle to climate change.

[70] For the Alpine region, we assessed changes in high detail and presented an inter-model comparison. In general, all models robustly project an intensification of precipitation, with the partial exception of summer and the southern Alpine region. In summer reductions in mean precipitation and precipitation frequency are evident in all projections and across the entire Alpine arch. The RCMs qualitatively agree on the structure of changes, i.e., the change signals in different diagnostics and regions. But there is still obvious inter-model spread in the quantitative statements. In some regions and seasons, the spread appears GCM dependent, in others RCM dependent. The large spatial variability of climates in the Alpine region is also reflected in terms of the magnitudes of climate changes signals in different diagnostics. In some regions, projected changes are larger than model biases (e.g., compare Northeastern Alps in Figures 2 and 5). In other regions, model biases may be substantially larger than projected changes.

[71] The methodology of our study closely followed the previous study of Frei et al. [2006] (F06) that analyzed 6 PRUDENCE RCMs with a grid resolution of 50 km and driven by one single GCM. As the current study is using the same diagnostics for similar regions, a detailed comparison between the two sets of simulations (PRUDENCE and ENSEMBLES) is feasible. Qualitatively, the results of the present study are largely in line with F06. The spatial patterns of changes in diagnostics are similar, showing dipole patterns with increases (decreases) in northern (southern) Europe. However, for fall, the ENSEMBLES simulations exhibit significant and robust intensifications of precipitation events in all Alpine subdomains, which were not projected in the PRUDENCE simulations analyzed in F06. Overall, however, differences are smaller than expected, given the fact that F06 considered RCMs driven by one single GCM. Also, both studies show a very complex structure of changes across the spectrum of diagnostics considered, with disproportionately different changes (and sometimes even reversed signs) for basic (mea, fre) and intense (int, q90, x1d) diagnostics. In both studies, most obviously these peculiar changes appear in summer and across large parts of Europe.

[72] Given the overall satisfactory skill of the models under consideration, our results provide a qualitative estimate of future precipitation changes in Europe and the Alps on regional to continental scales. The results are largely in line with previous studies based on coarser-resolved models (e.g., F06, IPCC), and present an extensive overview on currently available state-of-the-art high-resolution projections from the ENSEMBLES project. The presented results motivate to investigate and better understand the processes that govern the response of the hydrological cycle to climate change, and to assess the potential impacts upon natural and anthropogenic systems.

Acknowledgments

[73] We acknowledge the RCM data sets and E-OBS data set from the EU-FP6 project ENSEMBLES (http://ensembles-eu.metoffice.com). This research was partly funded by the Swiss National Science Foundation through the SNSF Sinergia project CRSII2_136279 “The Evolution of Mountain Permafrost in Switzerland” (TEMPS). Logistical support in assessing the ENSEMBLES data set was provided by Erich Fischer, Sven Kotlarski and the Center for Climate Systems Modeling (C2SM) at ETH Zurich. We also acknowledge Christoph Frei for very valuable comments and support during the preparation of the analyses.