The estimate of future climate changes is a great challenge for scientists. Very useful instruments for this purpose are climate models, numerical simulations of the climate system. However, because of chaotic nature of the climate system, it is not possible to capture all the processes and to simulate them precisely (Räisänen, 2007). Therefore, the simulations of climate models for any future time period can only be taken as projections, i.e. as possible trajectories related to a certain set of initial and boundary conditions. To increase the usefulness of these projections for impact studies it is necessary to investigate and quantitatively estimate the uncertainty (Fowler et al., 2007, Blenkinsop et al., 2009). The sources of uncertainties in both global and regional climate model (RCM) outputs include inaccuracies of initial and boundary conditions, the parameterizations of small-scale processes and the structure of the model, e.g. the choice of numerical schemes or spatial resolution (Tebaldi and Knutti, 2007). In case of RCMs, some additional factors come into play, e.g. problems connected to the limited integration domain (Laprise et al., 2008) or possible inconsistencies of parameterization schemes between RCM and the driving model (Denis et al., 2002).
The uncertainty connected to the structure of the model, especially the spatial and temporal resolution, type of grid and numerical methods can be evaluated using the multi-model ensembles (MMEs) (Tebaldi and Knutti, 2007) and can be divided into the uncertainty of first and second kind (Palmer, 2008). The estimate of the uncertainty of the first kind can be based on the variance of a MME. But the uncertainty of the second kind arises from the errors and simplifications common to all models, which is why it cannot be easily evaluated (Palmer, 2008). It is practically impossible to create a MME that would sample all uncertain features; therefore, it is not possible to estimate the whole range of structural uncertainty (Collins, 2007).
When the climate model outputs are used to estimate future climate change, an additional uncertainty arises owing to the unknown future development of forcings in climate system, both natural (solar and volcanic activity) and anthropogenic (greenhouse gas and aerosol emissions, changes in land-use). Moreover, due to uncertainties in biogeochemical cycles, climate sensitivity and internal variability of climate system we are not capable of predicting actual greenhouse gas concentrations based on estimated emissions (Giorgi et al., 2008). According to Bertrand et al. (2002), the changes in natural forcings are extremely unlikely to offset the anthropogenic increase in greenhouse gas concentrations as far as century-scale changes in the global mean temperature are considered. The emission scenarios created by the Intergovernmental Panel on Climate Change are alternative images of how the future might unfold and they represent an appropriate tool to analyse how driving forces may influence future emission outcomes and to assess the associated uncertainties (Nakićenović and Swart, 2000). Recently, a revisited set of four ‘representative concentration pathways’ (Moss et al., 2008, 2010; Meinshausen et al., 2011) has started to be used for climate change projections. Additionally, we must account for uncertainties in the reaction of natural ecosystems and human society to estimated climate change when we want to create regional climate change projections and assessment of climate change impacts in various sectors (Giorgi et al., 2008).
As already said, the methods for evaluation of uncertainties in climate model outputs are often based on the analysis of MMEs (e.g. Buonomo et al., 2007; Déqué et al., 2007; Holtanová et al., 2010; Kjellström et al., 2011). In case of RCMs, the ensembles usually include simulations driven by various global climate models (GCMs), and sometimes also model runs for various emission scenarios or representative concentration pathways. A very important step when analysing the uncertainty in climate model outputs is to identify the relative contribution of individual uncertainty sources. The simplest way of such analysis is to compare the variances of subsets of the MME. For example, Plavcová and Kyselý (2011) compared the span of a suite of RCM simulations driven by one GCM and the differences between the simulations of one RCM driven by several GCMs. Similarly, Kjellström et al. (2011) studied the differences between simulations of one RCM with various boundary conditions (GCMs and reanalysis) and under different emission scenarios. A more quantitative approach is the analysis of variance (Déqué et al., 2007; Holtanová et al., 2010; Déqué et al., 2012). This method divides the variance of an MME into contributions of individual sources (Section 'Data and method').
In Holtanová et al. (2010) we investigated the uncertainties in surface air temperature and precipitation changes projected by a suite of RCMs from the project PRUDENCE (see Section 'Data and method' for more information about the project) for the area of the Czech Republic. We compared relative contributions of RCM, driving GCM and emission scenario to the variance of the MME. We found out that the major influence on the variance of changes in mean seasonal surface air temperature between 2071–2100 and 1961–1990 has the driving GCM. In case of changes in seasonal precipitation, the RCM plays the major role in all seasons except for spring, when its importance is comparable to the GCM component.
In this study, we apply the same method of assessment of contributions from RCM and driving GCM on uncertainty in RCM outputs. Unlike Holtanová et al. (2010) we analyse not only the changes in surface air temperature and precipitation between reference and future period, but also the actual mean values in respective time periods, which might provide more insight into the mechanisms of how the climatic signal from driving GCM is modulated by the RCM. Furthermore, we compare the results for PRUDENCE RCMs with a newer set of model simulations from the project ENSEMBLES (see Section 'Data and method' for more information about the project), to test how the relative contributions of RCM and driving GCM are influenced, e.g. by the differences in RCM structure, horizontal resolution and the number of driving GCMs.
In the next section we describe the data and the method used for the analysis. Furthermore, we describe the results for RCM outputs from the project ENSEMBLES and compare them with the outputs from the project PRUDENCE. The last section summarizes and discusses the results.
2. Data and method
Our analysis concentrates on RCM simulations from two international European projects, PRUDENCE (http://prudence.dmi.dk) and ENSEMBLES (http://ensembles-eu.metoffice.com). Both of these projects aimed at creating a state-of-the-art simulations performed with several RCMs driven by several GCMs that would enable evaluation of uncertainties in RCM outputs and provide data for climate change studies over Europe. A brief description of the simulations used in this study is given in Tables 1 and 2.
Table 1. Regional climate models and their driving global climate models from the PRUDENCE project used for this analysis
Regional climate model
Table 2. Regional climate models and their driving simulations from the ENSEMBLES project used for this analysis
Regional climate model
The PRUDENCE project was accomplished in 2005, information about the project and model simulations completed can be found in Christensen and Christensen (2007), Jacob et al. (2007) and Déqué et al. (2007). We used the RCM simulations driven by two GCMs, HadAM3H (atmospheric general circulation model developed in the Met Office Hadley Centre, details can be found in Buonomo et al., 2007) and ECHAM4/OPYC3 (coupled atmosphere–ocean general circulation model developed in the Max Planck Institute for Meteorology, Roeckner et al., 1999). All simulations were run in the time-slice mode, for the control period 1961–1990 and future time period 2071–2100. The simulations of future climate considered in this study incorporated the emission scenarios A2 (Nakićenović and Swart, 2000). However, the RCM–GCM-scenario matrix is not complete (Table 1). There are eight RCM simulations driven by HadAM3H, but only two model runs driven by ECHAM4/OPYC3. All RCM runs used the horizontal resolution of approximately 50 km and the integration area covered most of Europe and Mediterranean. Physical parameterizations used in the RCMs are not the same (for details see Jacob et al. (2007) and their Table I), and in some cases there might be inconsistencies between the schemes used in a particular RCM and in the driving GCM. However, as the area used for the present analysis (see below) is located quite far from the borders of the integration domain, we assume that the model simulations are not influenced by these inconsistencies. Similar assumption is used in case of the ENSEMBLES RCMs.
The ENSEMBLES project was completed in December 2009. Detailed information about the project and its results, including the RCM simulations, can be found in van der Linden and Mitchell (2009). We analysed RCM runs driven by reanalysis ERA40 (Uppala et al., 2005) and three GCMs, namely HadCM3 (coupled atmosphere–ocean general circulation model developed in the Met Office Hadley Centre, details can be found in Gordon et al., 2000), ECHAM5/MPIOM (coupled atmosphere–ocean general circulation model developed in the Max Planck Institute for Meteorology, Keenlyside et al., 2008) and ARPEGE/OPA (sometimes referred to as CNRM-CM3, coupled atmosphere–ocean general circulation model developed in Centre National de Recherches Météorologiques, Salas-Mélia et al., unpublished work) (Table 2). In case of HadCM3 there were simulations with a standard version (denoted as HadCM3) and with two sets of modified parameterization schemes, HadCM3 (low) and HadCM3 (high). These simulations differ from each other, and can be considered as different driving data for the RCMs (van der Linden and Mitchell, 2009). All of the ENSEMBLES model simulations were transient runs. ERA40-driven simulations were conducted for the period of 1961–2000, whereas the GCM-driven runs for 1951–2050 or 1951–2100, only the simulations of HadRM model were ended in the year of 2099. After the year 2000 model simulations incorporated the emission scenario A1B (Nakićenović and Swart, 2000). Similar to the PRUDENCE models, the RCM–GCM matrix is not complete (Table 2). All ENSEMBLES RCM runs considered in this study used the horizontal resolution of approximately 25 km, except for the CLM/GKSS driven by ERA40, which was run in the horizontal resolution of 50 km. Similar to PRUDENCE, the integration area covered most of Europe and Mediterranean (see Fig. 5.1 in van der Linden and Mitchell, 2009).
To assess the uncertainties in both MMEs, the analysis of variance described by Déqué et al. (2007) was employed. The variance of an MME, which is taken as an estimate of the uncertainty, is divided into the contributions coming from the RCM, driving GCM and the interaction of these factors. Details about the procedure can be found in Déqué et al. (2007), here we show only a brief summary. The variance V of the MME is given by
where i identifies RCM and j refers to GCM, Xij denotes the individual ensemble members, N the number of the ensemble members and the ensemble mean. Let us denote Xi. the value of the characteristic of interest simulated by the RCM with index i averaged over all driving GCMs. Analogically we use the notation X. j. Then the part of the variance V attributed to the RCM is
where I is the number of RCMs considered. The contribution of the driving GCM to the variance is
where J is the number of driving GCMs considered. The contribution to the variance due to the interaction between GCM and RCM is
However, as the GCM–RCM matrix is not complete, it is not possible to divide the variance unambiguously into the contributions of mentioned factors [GCM, RCM and interaction term (INT)]. Their sum is significantly lower or larger than the estimated variance. Therefore, Déqué et al. (2007) suggested a data reconstruction algorithm to fill in the gaps in the matrix. This algorithm is iterative and based on minimizing the INTs. It is used in this study without any modifications. In the first step of the procedure we calculate averages Xi., X. j and as well as the variance and contributions of individual influences from available data. In each of the next iterations we calculate missing Xij using following equation:
The reconstructed values of originally missing Xij are used to recalculate the variance and its components. Equation (5) results from putting the members of sum in Equation (4) equal to zero. In this study the procedure was repeated in ten iterations. Differences in Xij between the two last iterations were always lower than 0.5% and the sum of contributions of GCM, RCM and INT are equal to the variance V (or the difference is negligible) after the last iteration.
We applied the method of analysis of variance described above on simulated 30-year mean seasonal surface air temperature and precipitation for the reference period (1961–1990) and future time periods (2071–2100 for PRUDENCE models, 2021–2050 and 2069–2098 for ENSEMBLES), and the changes of these quantities in the future period in comparison to the reference period. The changes were calculated as differences for air temperature and as ratios for precipitation. In case of the ENSEMBLES models, we distinguish two cases for the reference period 1961–1990; the MME including and excluding the simulations driven by ERA40.
The analysis of variance was applied to simulated characteristics averaged over the Czech Republic (i.e. over the 48.25°–51.75°N/11.25°–19.75°E area) to infer information useful for the national research project SP/1a6/108/07 that aimed at developing climate change scenarios for the Czech Republic. This study was conducted within this project. For the sake of simplicity we did not calculate the values for the Czech Republic only, but slightly larger area with simple rectangular borders.
3. Results for the ENSEMBLES RCMs
The results of the analysis of variance for MME from the project ENSEMBLES are shown in Figure 1 (air temperature) and Figure 2 (precipitation).
3.1. Climatic means
Concerning mean seasonal surface air temperature in the reference period, the results show that if we consider all simulations available, the contributions of RCM and driving field to the variance are almost equal (Figure 1(a)). Only in DJF the influence of driving field plays the dominant role. If we omit the model runs driven by ERA40, the influence of regional model on the variance of the ensemble becomes dominant, except for winter season when GCM has the main influence (Figure 1(b)).
In the period of 2021–2050 most of the variance of the MME is caused again by RCM, except for DJF (Figure 1(c)).
Considering the end of 21st century (2069–2098), the influence of RCM remains most important in MAM and JJA only; however, its dominance is not as large as in previous periods (Figure 1(e)). In SON and DJF the driving GCM causes larger part of the variance.
The variance of mean seasonal precipitation in the reference period derived from the RCM simulations driven by both GCMs and ERA40 is more influenced by RCM than by GCM, except for MAM, when RCM component is a bit lower than in other seasons and the INT is twice as large as in other seasons reaching 13% of the variance (Figure 2(a)). When RCM simulations driven by ERA40 are excluded, the RCM influence increases in JJA and decreases in SON. Except in SON, the variance remains therefore largely modulated by the RCMs (Figure 2(b)). Compared to the air temperature, the largest difference appears during winter. RCMs contribute to approximately 70% of the precipitation variance in DJF, while more than 90% of the air temperature variance responds to the driving GCMs.
The results of the analysis of variance for mean seasonal precipitation in the period of 2021–2050 are similar to the reference period, and also to the results for surface air temperature (except for DJF). The influence of RCM is prevailing; only in DJF the contribution of RCM is as high as from GCM (Figure 2(c)).
Considering the period of 2069–2098, the influence of GCM on the variance of mean seasonal precipitation is larger than in both previous periods (Figure 2(e)). This factor is prevailing in all seasons except for DJF, when it is almost equal to RCM. In comparison to the air temperature in this period, the influence of GCM is also larger, except for DJF, when the contrary applies.
Concerning the interaction of the two factors (driving data and RCM) and its contribution to the variance of the MME, it is lower than 1% in most cases, and therefore we consider its influence as negligible. The only exception is the reference period of 1961–1990, when we include also the simulations driven by the reanalysis ERA40. In that case the value of the INT lies between 3 and 13% (depending on season and variable). However, the statistical significance of the INTs was not tested, so we cannot say precisely whether its influence can be neglected or not in the reference period.
3.2. Simulated changes
Concerning the changes of surface air temperature in the period of 2021–2050 relative to the reference period, their variance is mainly influenced by driving GCM in MAM and DJF, while in JJA and SON the contribution from RCM and GCM is almost equal (Figure 1(d)).
In case of simulated changes of air temperature in 2069–2098, the influence of RCM slightly decreases in all seasons, comparing to the changes in the period of 2021–2050. Driving GCM plays the major role in DJF and MAM again, and also in SON. In JJA the contributions of both factors remain almost equal (Figure 1(f)).
If we consider the changes of precipitation between the period of 2021–2050 and the reference period, the part of the variance caused by RCM is prevailing in MAM and JJA. In SON and DJF, on the other hand, GCM plays the major role (Figure 2(d)). It can be seen that in DJF the influence of GCM on the precipitation change in 2021–2050 is similar to that for air temperature change, but in other seasons the results for the changes of the two quantities differ.
For changes of mean seasonal precipitation in 2069–2098, the results of analysis of variance of the MME are different from the period of 2021–2050. The influence of GCM and RCM is comparable (Figure 2(f)). The exception is DJF, when GCM plays the major role. The influence of GCM on precipitation changes is smaller than on the air temperature changes in the end of the 21st century, except for JJA, when the results for the two quantities do not differ.
4. Comparison of the results for PRUDENCE and ENSEMBLES RCMs
As described in the Introduction, in Holtanová et al. (2010) we applied the analysis of variance to the changes in mean seasonal air temperature and precipitation simulated by the PRUDENCE models for the area of the Czech Republic. In this section, we add the results for mean values in each of the time periods and compare them with results obtained for the ENSEMBLES models. The results for both MMEs are summarized in Table 3.
Table 3. Comparison of the results of the analysis of variance of the multi-model ensembles from the projects PRUDENCE (PRUD) and ENSEMBLES (ENS)
2-m air temperature – prevailing influence (%)
Precipitation – prevailing influence (%)
Grey colour indicates prevailing influence of RCM, white denotes major role of driving GCM. The numbers in brackets show the percentage of the variance coming from the source indicated. ‘Change’ denotes the results for the changes of surface air temperature and precipitation in the late 21st century period of 2071–2100 (2069–2098 for ENS) in comparison with the reference period of 1961–1990.
(c) 2071–2100 (2069–2098)
(d) 2071–2100 (2069–2098)
4.1. Climatic means
For the MME from the project PRUDENCE we found out that the influence of RCM on simulated mean seasonal air temperature in both reference and future time periods is dominant. For the ENSEMBLES models the role of RCM is less dominant than for PRUDENCE MME, the only exception is summer season in the reference period. In DJF, we even see prevailing influence of GCM (Table 3(a) and (c)).
The variance of mean seasonal precipitation simulated by the PRUDENCE models in both time periods (2071–2100 and 1961–1990) is mainly influenced by RCM in MAM, JJA and DJF. In SON, in the reference period the role of GCM and RCM are rather comparable, in the future time period the influence of driving GCM is slightly more pronounced. In case of the ENSEMBLES models, the results of the analysis of variance for mean seasonal precipitation in the reference period are quite similar to the PRUDENCE models. In the period of 2069–2098 the ratio of influences of RCM and GCM is also similar in SON and DJF. In MAM and JJA, the role of GCM is more important than the RCM, unlike in the PRUDENCE MME, where the influence of RCM increases in the late 21st century in comparison to the reference period by approximately 20% (Table 3(b) and (d)).
4.2. Simulated changes
The variance of late 21st century changes of mean seasonal air temperature simulated by both MMEs is mainly influenced by driving GCM. In SON and DJF the influence of GCM is more prevailing for ENSEMBLES models by approximately 20% than for PRUDENCE ensemble. In MAM and JJA the contributions of the driving GCM differ less than by 10% between the two MMEs (Table 3(e)).
Also for precipitation changes we see rather different results for both MMEs. In MAM for PRUDENCE models both factors have similar influences (49 and 51%), but for ENSEMBLES RCMs the GCM plays more important role (63%). In the other three seasons, the RCM has prevailing influence in case of PRUDENCE models. On the other hand, the role of GCM is more important for the ENSEMBLES MME (Table 3(f)).
5. Discussion and conclusions
We have presented results of the analysis of variance of simulated mean seasonal surface air temperature and precipitation for two MMEs and several time periods (1961–1990, 2021–2050 and the end of 21st century). The method was previously applied by Déqué et al. (2007) to the changes in precipitation and surface air temperature simulated by the PRUDENCE models over large regions of Europe (PRUDENCE regions defined by Christensen and Christensen, 2007). More recently Déqué et al. (2012) applied similarly the analysis of variance on changes of seasonal mean air temperature and precipitation simulated by ENSEMBLES RCMs in the period of 2021–2050. We apply the analysis of variance method on smaller spatial scale than these previous studies and not only to changes in climatic variables, but also to their climatic means. Generally our results for the changes of summer and winter mean air temperature and precipitation in the period of 2021–2050 for ENSEMBLES models agree well with the results of Déqué et al. (2012) for eastern and middle Europe, despite the fact that they used different procedure for missing data reconstruction.
For the ENSEMBLES regional models, the influence of RCM on mean seasonal surface air temperature in MAM is prevailing in all time periods (1961–1990, 2021–2050 and 2069–2098). Furthermore, we found out that in JJA and SON the influence of RCM prevails in 1961–1990 and 2021–2050 and decreases in 2069–2098. In DJF, on the other hand, driving GCM is the main source of the ensemble variance in all time periods. This is in agreement with results of Plavcová and Kyselý (2011) who compared the influence of RCM and driving GCM on biases of maximum and minimum surface air temperature simulated by ENSEMBLES regional models in the area of the Czech Republic. They also found larger influence of GCM in winter than in summer (they did not analyse data for SON and MAM).
For the mean surface air temperature in the 1961–1990 period we have found out that when the model runs driven by reanalysis ERA40 are excluded, the influence of the driving field decreases (except for DJF). This might be connected to the different nature of ERA40 and GCM-simulated fields and the discrepancies between them as GCMs generate their own climate and share systematic biases to a certain extent while ERA40 is based on observations and is much closer to reality. Another reason can be the procedure of missing data reconstruction. We have simulations driven by ERA40 available for all incorporated RCMs (Table 2), and therefore they are not influenced by the reconstruction like the GCM driven data, that must be reconstructed for many GCM–RCM pairs. In case of precipitation the results do not change largely if we include or exclude simulations driven by ERA40. This is probably because of the fact that simulation of precipitation is more influenced by RCM, which is connected to better representation of small-scale processes such as convection.
Considering changes in mean seasonal surface air temperature, the influence of GCM is larger than for mean values. Only in DJF, the contribution of GCM to the variance of the ensemble is comparable like for the mean values. This is probably because of the fact that for the mean values the large-scale forcing represented by lateral boundary conditions taken from GCM is stronger in DJF than in other seasons.
Concerning mean seasonal precipitation, the influence of RCM is larger than the contribution from driving field in 1961–1990 and 2021–2050, in the last time period the influence of GCM increases, similar to that for surface air temperature. This can be interpreted like that the uncertainty of simulated precipitation connected to GCM (that is simulation of large-scale features) grows during the 21st century more than the uncertainty connected to simulation of small-scale processes in RCM. The results for the changes in mean seasonal precipitation do not differ much from the results for the mean values. In some cases the influence of GCM increases slightly, but it is not as pronounced as in case of surface air temperature.
The results of the analysis of variance for the ENSEMBLES models are different from the results obtained for the PRUDENCE models in some aspects, especially for mean seasonal surface air temperature and precipitation in the end of the 21st century (for details see Section 'Climatic means' and Table 3(c) and (d)). There are significant differences between the MMEs themselves. They differ not only in the horizontal resolution used for the simulations (25 and 50 km), but also in the number of models, the driving GCMs and the formulation of RCMs. Further, for the simulations of future climate the emission scenarios considered were different. The PRUDENCE models were run for A2 scenario. On the other hand, the ENSEMBLES RCMs used A1B scenario. Another difference is the methodology employed. In the PRUDENCE project, the RCM simulations were run for two 30-year periods, while in the ENSEMBLES the transient simulations starting in 1951 were accomplished. It is not possible to simply attribute the differences in the results to one aspect, for example, different horizontal resolution or emission scenario.
One of the results that remain the same for both MMEs is that the influence of driving GCM is larger in case of changes of mean seasonal surface air temperature than for the mean values in corresponding time periods. We hypothesize that this fact can be explained as follows. The RCM modulates the information passed from the driving GCM (e.g. adapts it to more detailed orography), but the temperature changes are dominated by large-scale forcings including emission scenario, which is often adopted from the driving simulation. When considering precipitation changes no similar conclusion can be made.
Another point related to driving GCMs is that our results are based on RCM simulations driven by a limited number of GCMs (and reanalysis ERA40 in some cases). If we incorporated runs driven by more different GCMs, the variance of the MMEs would probably grow, as well as the influence of driving GCM. This statement is in accordance with the results obtained by Kalvová et al. (2009) and Kalvová et al. (2010). These studies showed that even if we choose GCMs that give realistic values of surface air temperature and precipitation in the reference period for the Czech Republic, the span of simulated changes is large. Driving GCMs incorporated in this study do not represent the whole range of these simulations.
The structure of RCM and the driving data are not the only sources of uncertainty in simulations of both observed climate and future changes, as described in the first section of this paper. Here we want to comment only on possible influence of unknown development of various forcings on the simulations of future climate. Recent studies (e.g. Déqué et al., 2007, Hawkins and Sutton, 2009, Holtanová et al., 2010, Hawkins and Sutton, 2011) have shown that the influence of emission scenario is smaller in comparison to other sources of uncertainty. However, this does not mean that the influence of anthropogenic forcings on the climate system should be neglected when creating the climate change scenarios. Rather, it highlights the need to take into account not only various emission scenarios, but also (and maybe more importantly) projections from different RCMs driven by more GCMs, if possible.
This work has been supported by the project SP/1a6/108/07 funded by the Ministry of Environment of the Czech Republic, Research Plan MSM0021620860 and the project UNCE 204020/2012. The ENSEMBLES data used in this work was funded by the EU FP6 Integrated Project ENSEMBLES (Contract number 505539) whose support is gratefully acknowledged. PRUDENCE data have been provided through the PRUDENCE data archive, funded by the EU through contract EVK2-CT2001-00132.