Climatic Change DOI 10.1007/s10584-015-1412-4
Testing climate models using an impact model: what are the advantages? Marc Stéfanon 1 & Nicolas K. Martin-StPaul 1,6 & Paul Leadley 1 & Sophie Bastin 2 & Alessandro Dell’Aquila 3 & Philippe Drobinski 4 & Clemente Gallardo 5
Received: 8 April 2014 / Accepted: 19 April 2015 # Springer Science+Business Media Dordrecht 2015
Abstract Global and regional climate model (GCM and RCM) outputs are often used as climate forcing for ecological impact models, and this potentially results in large cumulative errors because information and error are passed sequentially along the modeling chain from GCM to RCM to impact model. There are also a growing number of Earth system modeling platforms in which climate and ecological models are dynamically coupled, and in this case error amplification due to feedbacks can lead to even more serious problems. It is essential in both cases to rethink the organization of evaluation which typically relies on independent validation at each successive step, and to rely more heavily on analyses that cover the full modeling chain and thus require stronger interactions between climate and impact modelers. In this paper, we illustrate the benefits of using impact models as an additional source of information for evaluating climate models. Four RCMs that are part of the HyMeX (Hydrological cycle in Mediterranean EXperiment) and Mediterranean CORDEX projects (MED-CORDEX) were tested with observed climatology and a process-based model of European beech (Fagus sylvatica L.) tree growth and forest ecosystem functioning that has been rigorously validated. This two part analysis i) indicates that evaluation of RCMs on climate variables alone may be insufficient to determine the suitability of RCMs for studies of climate-forest interactions and ii) points to areas of improvement in these RCMs that would improve impact studies or behavior in coupled climate-ecosystem models over the spatial domain studied.
* Marc Stéfanon
[email protected] 1
Laboratoire d’Écologie Systématique et Évolution (ESE), (UMR 8079 CNRS/Univ. Paris-Sud/ AgroParisTech), Orsay, France
2
Laboratoire Atmospheres, Milieux, Observations Spatiales (LATMOS), Institut Pierre Simon Laplace (CNRS/UVSQ/UPMC), Guyancourt, France
3
Ente per le Nuove Technologie, l’Energia e l’Ambiente (ENEA), Climate Section, Rome, Italy
4
Laboratoire de Météorologie Dynamique (LMD) - Institut Pierre Simon Laplace, (CNRS/Ecole Polytechnique/ENS/UPMC), Paris, France
5
Instituto de Ciencias Ambientales de la UCLM, Toledo, Spain
6
Ecologie des Forêts Mediterraneennes, INRA, UR629, F-84914 Avignon, France
Climatic Change
1 Introduction Global and regional climate model (GCM and RCM) outputs are often used as climate forcing for ecological impact models. For the ecological community, the assessment of the climatic information is a critical key step of the impact modeling process. However evaluation of climate models using climate data is a complex task for several reasons. First, there is a large number of output variables (>100), some of which are difficult to compare with data due to lack of good measurements. Second, most diagnostics examine only a single variable at a time and generally focus on testing for systematic bias and the capacity to reproduce temporal and spatial patterns for a wide range of scales (Brands et al. 2013; Flaounas et al. 2013). Finally, the different variables are intrinsically linked through complex non-linear relationships that can include thresholds. Substantial thought has been given to developing a framework to deal with the enormous amount of data produced in the last few decades by various international exercises of model inter-comparison, such as the Coupled Model Intercomparison Project (CMIP - Taylor et al. 2012) or the COordinated Downscaling EXperiment (CORDEX - Giorgi et al. 2009). Current methodologies rely on integrated indicators that summarize information (Taylor 2001, Teuling et al. 2011). Climate models are imperfect by construction and do not perform evenly for all simulated variables and/or evaluation indicators because of the many sources of uncertainty (e.g., physical parameterization, numerical scheme, boundary conditions) (Flaounas et al. 2013; Di Luca et al. 2014). Their evaluation is also made difficult because of the uncertainty of the observations (Flaounas et al. 2012), hence the importance given to inter-comparison exercises. Once these evaluations performed, a key question remains: how good is good enough? The answer depends on the application. One under-explored way in global and regional climate modeling is process-based evaluation rather than a ‘holistic’ variable-based evaluation (Lung et al. 2013). Along the different components of the modeling chain from GCM to RCM to impact model, information is passed and error is typically assessed sequentially. Rethinking the organization of this chain from linear to cyclical by using impact models as an additional source of information for evaluating climate models can provide many benefits, but the use of impact indicators to test climate models is infrequent. This approach is particularly insightful because many impacts integrate a broad range of climate signals and can be sensitive to small changes in climate drivers (Lung et al. 2013). We have tested four RCMs with both observed climatology and two impact models. Our analysis focuses on climate impacts on forests using European beech (Fagus sylvatica L.) as an example. This species is a dominant and a representative tree of temperate deciduous broadleaf forests in Europe, and is of high economic importance for timber (Fang and Lechowicz 2006). The long time period between planting and harvesting of trees in exploited forests means that beech trees often integrate the local climatic signals over periods exceeding a century. Moreover the low migration capacity and long life cycle of trees in non-managed forests make trees and forests particularly vulnerable to climate change impacts (Lindner et al. 2010; Cheaib et al. 2012). We argue that analyzing the differences between these two diagnostics highlights i) the non-linear error propagation along the modeling chain, ii) the weighting of climatic processes and variables of interest for a given type of impact model, and iii) the variables and processes of the climate model requiring improvement.
Climatic Change
2 Methodology 2.1 Models and data Five different climate datasets over France were used: four are provided by RCMs and one is a reference dataset based on surface meteorological observations. The RCM outputs were provided by the HyMeX (Hydrological cycle in Mediterranean EXperiment, Drobinski et al. 2014) and Mediterranean CORDEX (MED-CORDEX, Ruti et al. 2014) archives: ALADIN (Colin et al. 2010), RegCM (Pal et al. 2007), WRF (Skamarock et al. 2008), PROMES (Domínguez et al. 2010). These models were selected based on the availability of the daily data required by the tree growth and forest ecosystem model, CASTANEA (see below). The RCM simulations were performed at a 50 km resolution for the period from 1989 to 2008 by a dynamical downscaling of the ERA-INTERIM reanalysis (Dee et al. 2011). The SAFRAN analysis (Quintana-Seguí et al. 2008) was used as a reference dataset. SAFRAN is derived from the climatological network of Météo-France and provided the required variables on a 8 km grid. We used two models of tree response to climate, the process-based tree growth and forest ecosystem model CASTANEA (Dufrêne et al 2005) and the niche-based model BIOMOD (Thuiller 2003), as the basis for assessing the performance of atmospheric variables provided by the four RCMs. CASTANEA is a forest soil-vegetation-atmosphere model coupled with a tree growth module. It simulates carbon and water fluxes at a daily time step for an average tree in a homogeneous stand of forest using 5 atmospheric variables (temperature, rainfall, solar downward radiation, wind speed at 10 m, relative humidity at 2 m) provided by forcing datasets. CASTANEA simulates carbon and water fluxes, including gross and net ecosystem photosynthesis, respiration, transpiration, latent heat flux, soil water content and tree growth for several major European tree species including European beech (Davi et al. 2005, 2008; Delpierre et al. 2012). CASTANEA has been thoroughly validated for beech and several other major European tree species using ecosystem CO2 flux, water flux and tree growth data from multiple sites (Davi et al. 2005, 2008; Cheaib et al. 2012; Delpierre et al. 2012). CASTANEA can also be used to simulate the distributional range (presence/absence) of a tree species by assuming that tree growth is a proxy of its ability to persist in a given environment (Cheaib et al. 2012). It has been validated against current tree distribution using the French National Forest Inventory (NFI, for more information see www.ifn.fr). BIOMOD is a correlative species distribution model which links observed presences of a species to climate variables trough statistical relationships. This type of model is used widely in the ecological community with around 22,000 related publications over the 1999–2009 period (Thuiller et al. 2009). These models require mean annual climate variables as input (e.g., annual precipitation, temperature, potential evapotranspiration).
2.2 RCM evaluation RCMs were evaluated using two approaches: i) a conventional method using standard tools for evaluation based on climatic variables (Taylor diagrams and overall bias calculations) and ii) an ecological approach based on the simulation of tree distributional range computed with the growth model CASTANEA and a correlative species distribution model. Temperature, rainfall and relative humidity were used in the climatic evaluation based on knowledge that these variables play the most important roles in simulating tree growth and
Climatic Change
distribution in CASTANEA (Davi et al. 2005, 2008; Cheaib et al. 2012; Delpierre et al. 2012). Climate variables were computed over the period May-August since most of annual photosynthesis and radial growth of beech occurs during this time (Lebaube et al. 2000; Schmitt et al. 2000). Multi-annual lags could potentially add a confounding factor in this analysis, but observations from the French Permanent Plot Network (RENECOFOR - Lebourgeois et al. 2005) suggest that the role of weather and drought conditions during the previous season are of little relevance for growth dynamics of European beech. The climatic evaluation was performed for each climatic variable using Taylor diagrams for each model. Taylor diagrams provide a graphical summary of how closely a given simulated variable matches a reference variable and are a standard method of model evaluation in the climate modeling community (Taylor 2001). The similarity between the two patterns is quantified using correlation, centered root-meansquare difference and the magnitude of their variations (standard deviations). Taylor diagrams put a heavy emphasis on the ability of models to reproduce spatial and temporal patterns. Taylor diagrams do not provide information about overall bias because the means of the fields are subtracted out before computing the secondorder statistics used in the diagram (Taylor 2001). Spatially and temporally average overall biases were also calculated for all three climate variables. Taylor diagrams are widely used by climate community to evaluate model performance. Other performance metrics exist and could provide additional insight into the evaluation of climatic variables (Gleckler et al. 2008). For the ecological evaluation, the RCM climatic outputs as well as the SAFRAN reference dataset were used as inputs for CASTANEA to stimulate the European beech growth. CASTANEA simulations were performed for the study area (France) on a 8 km grid to match the soil water holding capacity database (Cheaib et al. 2012). Consequently, RCM climatic inputs at 50 km were resampled on an 8 km grid. Because of highly non-linear effect linked to soil moisture, we consider that the loss of information by upscaling to 50 km the soil database is more critical for the results than a rescaling of climatic data to 8 km. Tree growth (gC/m2/yr) was transformed to presence/absence by comparing the map of simulations to the binary map of presence/ absence observations. This was achieved by computing a threshold value of tree growth that maximizes the goodness-of-fit of the CASTANEA simulations forced by SAFRAN. Goodness-of-fit is calculated with the TSS (True Skill Statistics, Allouche et al. 2006), a statistic ranking that has values ranging between −1 and 1. TSS accounts for both the presences correctly predicted and the absences correctly predicted. A score of 1 means that both presence and absence are predicted perfectly whereas a score of −1 means that no points of the grid are predicted correctly. Simulations have also been performed with an ensemble average of five different statistical species distribution models implemented in the BIOMOD modeling platform (Thuiller 2014). Statistical relationships were calibrated using a sub-sample of the observed species distribution with the SAFRAN dataset and evaluated using the remaining data. BIOMOD has been extensively evaluated over France and generally provides better results than other correlative species distribution models because it is based on a multi-model ensemble method (Marmion et al. 2009; Cheaib et al. 2012). The same calibration was applied to others simulations forced by the RCMs and goodness-of-fit indices (TSS) were computed. We used SAFRAN for the calibration since it is our climatic reference.
Climatic Change
2.3 Evaluation of the individual climate variables for the WRF RCM Climate models cannot represent perfectly all aspects of climate; as a result errors in the temporal pattern, the spatial patterns as well as systematic bias are often reported. Such biases often preclude the dynamic coupling between vegetation models and climate models and the correct assessment of impacts unless some form of bias correction is performed (Ehret et al. 2012; Ruffault et al. 2014). Based on the RCM evaluations described in subsection 2.2), we performed a sensitivity analysis with the WRF model in order to identify which climate variables contribute the most to deteriorate the ecological signal produced by CASTANEA. To do so, five sensitivity tests are carried out by using different climate datasets. Each test includes one variable from WRF while other variables are provided by SAFRAN.
3 Results and discussion 3.1 Evaluation of RCMs based on climatic variables Figure 1 shows the Taylor diagrams for the temporal and spatial patterns of surface air temperature, rainfall, and surface relative humidity, with respect to the SAFRAN analysis. Table 1 shows the biases of each variable over the period May-August. The temporal variability of temperature is well reproduced by all RCMs since all points are tightly clustered around the SAFRAN reference point in the Taylor diagram. The spatial pattern of temperature is reasonably well reproduced by three of the RCMs (correlation coefficient of ca. 0.7 and standard deviations that are similar the SAFRAN reference), but RegCM has much a lower correlation coefficient (0.2) and underestimates spatial variability. The RMSD for spatial variability is near or above 2 °C for all models indicating substantial spatial heterogeneity in the relationship between the models and the reference. Systematic bias is less than 1 °C for all the models except WRF which shows a large warm bias during the summer (+2.4 °C, Table 1). Compared to temperature, temporal correlations of rainfall between RCMs and the SAFR AN reference were lower (ca. 0.7–0.9). The ability of PROMES to simulate temporal variation in rainfall is slightly lower than for the other RCMs based on correlation and RMSD. Concerning the spatial pattern of rainfall, RegCM has the lowest correlation with the SAFR AN analysis (0.4), when compared to others models (ca. 0.7). ALADIN and PROMES have higher spatial variability (standard deviation) and RegCM has lower spatial variability than the reference. In terms of total rainfall during the growing season (May–August), PROMES, RegCM and ALADIN are close to SAFRAN (differences of <20 mm), while WRF has a large rainfall deficit (95 mm, Table 1). The spread amongst models is larger for the temporal variability of relative humidity than for the precipitation and temperature. Overall ALADIN performs the poorest for temporal variability with the lowest correlation coefficient, the highest RMSD and an overestimation of temporal variability compared to the reference. The spatial pattern of relative humidity is the best simulated by ALADIN and PROMES based on correlation (0.65). WRF has a lower correlation (0.5), and RegCM the lowest (0.2). When spatially and temporally averaged, PROMES, RegCM and ALADIN are close to SAFRAN (<5 %) while WRF shows a large humidity deficit
Climatic Change
Fig. 1 Taylor diagrams for the surface air temperature (upper panels), rainfall (middle panels) and surface relative humidity (lower panels) obtained from the regional climate models during the May-August period for temporal variability (right column) and spatial variability (left column). ALADIN, PROMES, RegCM and WRF simulations are represented by ●, ♦, ▪, ▲ marker respectively. SAFRAN (✰ marker) is the reference datas. The standard deviation is a measure of temporal variance over the entire time period (left panels) or spatial variance over the entire spatial domain (right panels) and is indicated by black contour lines; the correlation coefficient is the temporal or spatial pattern correlation and is indicated by blue lines; and the centered root mean square difference (RMSD) is indicated by the green contour lines and is a measure of the variation in differences between the centered model and Bobserved^ datasets. Simulated climate is closest to Bobservations^ when they are closest to the SAFRAN point
(20 %). This result for WRF is expected since relative humidity is directly dependent on temperature and partly depends on the precipitation for surface humidity content. This climatic evaluation is only pertinent for the spatial domain covered by France, the growing season of beech and climatic variables that are important for studying climate impacts on beech: evaluation for other spatial domains, time periods or climate variables could lead to different conclusions. Given these caveats, the results of the climate analyses can be summarized as follows. ALADIN has low bias for key climate variables and generally scores well on measures spatial variability. WRF does as well as or better than all other models in terms of
Climatic Change Table 1 Difference in daily mean surface temperature, cumulated rainfall and relative humidity between the RCMs (ALADIN, PROMES, RegCM, WRF) minus the SAFRAN observations ALADIN-SAFRAN
PROMES-SAFRAN
RegCM-SAFRAN
WRF-SAFRAN
Temperature (°C)
−0.11
−0.46
−0.87
2.4
Cumulated rainfall (mm)
0.49
4.3
19
−95
Relative humidity (%)
0.35
5.4
2.9
−20
Results are computed over the May–August period and averaged over France
reproducing spatial and temporal variability in key climate variables, but ranks the lowest in terms of bias for climate variables. PROMES and RegCM are intermediate cases in terms of bias and both have some cases in which they do not reproduce spatial patterns well, with RegCM having particularly low scores for spatial patterns of all three climate variables. In summary, we find that there are no clear outliers among the RCMs and without a detailed knowledge of both climatic and impact model, we are not able to anticipate the results of the impact model. How well will ALADIN do given that it has low biases but does only moderately well, as do the other RCMs, at reproducing spatial patterns knowing that our goal is to reproduce spatial patterns of beech distribution? More importantly, many climate impact analyses are now carried out with analyses that include multiple climate model inputs as a means of estimating uncertainty. WRF, PROMES and RegCM generally performed less well than ALADIN, but does this compromise their utility for impact modeling? Among the issues highlighted by the climatic analysis which is most important, the ability to produce spatiotemporal patterns well (WRF), or to have limited bias (PROMES and RegCM)?
3.2 Evaluation of RCMs based on tree distribution European beech is known to be drought sensitive, and with this background knowledge we expect that important biases in precipitation and temperature could substantially lower the RCM suitability for modeling its distribution. Based on the purely climatic analysis from Table 1, it can be predicted a priori that ALADIN is likely to do best of the four models in an impact study where a comparison is made using SAFRAN as a reference. To verify this hypothesis, we compared the ability of the process-based model CASTANEA to simulate the presence/absence of European beech depending on the type of climatic forcing used (Fig. 2). Modeled beech distribution using CASTANEA has a TSS value of 0.45 using SAFRAN as the climatic reference dataset. This is a good performance compared with other process-based models (Cheaib et al. 2012). Discrepancies produced between the CASTANEA simulation and observations are produced by uncertainties in CASTANEA, but some of what appears to be an overprediction of range may be the result of CASTANEA simulating beech presence in areas were there is currently little forest or where management has selected against beech (Cheaib et al. 2012). Based on these considerations, we have focused on how closely the predicted RCM distributions are to SAFRAN distributions and why. In order to facilitate this comparison, the TSS of the RCMs was expressed in relative to the SAFRAN TSS (Fig. 2g). The ALADIN simulation has the highest goodness-of-fit (71 % of the SAFRAN goodness-offit). RegCM, PROMES and especially WRF (<10 %) have a lower TSS. This is explained by an overestimation of beech presence along the Atlantic coast (PROMES), the Mediterranean and Southwestern France (RegCM). The WRF model produces an overall underestimation of
Climatic Change
beech distribution over nearly all of France. BIOMOD does better than CASTANEA in predicting current beech distributions (TSS=0.62 using SAFRAN input), because it is fit to the data. Even though this is a very different type of model, BIOMOD simulations are in good agreement with CASTANEA results (Fig. 2h). The ALADIN simulation has again the highest goodness-of-fit (64 % of the SAFRAN goodness-of-fit), TSS and model ranking are very similar, with both RegCM and WRF doing particularly poorly.
Fig. 2 European beech distribution : Observed (a) and modeled by CASTANEA using atmospheric variables from SAFRAN (b), ALADIN (c), PROMES (d), RegCM (e), WRF (f). Panel g) indicates the average TSS computed with CASTANEA simulations over France (% relative to SAFRAN TSS). Panel h) indicates the average TSS with BIOMOD simulations over France (% relative to SAFRAN TSS, maps not shown)
Climatic Change
Thereafter, we examined the details of the WRF underestimation of beech distribution by assessing which of the different climatic variables affected the most the modeled beech distribution produced by CASTANEA. WRF was chosen because of its very low TSS and large potential improvement. We performed five additional simulations. They are based on using atmospheric variables provided by SAFRAN, and replacing them one-by-one with output from WRF. Simulations using the 100 % SAFRAN and WRF forcings (Fig. 2g) are also presented to show the maximum and minimum TSS possible (Fig. 3). Consistent with our expectations, the lowest TSS are found for the simulation performed with only WRF variables. Simulations with results similar to SAFRAN are found when the WRF variable is not a limiting factor for the beech growth (i.e., downward radiation) or of low relevance (i.e., wind). Temperature, relative humidity and precipitation are in ascending order the atmospheric variables that contribute the most to the deterioration of the simulated ecological signal. However, during the growing season, temperature and relative humidity also modify the water budget of trees through evapotranspiration. High temperature increases the evaporative demand while low relative humidity enhances the vapor pressure deficit (VPD). A high VPD causes the stomata closure and thus decreases carbon assimilation, which impedes the beech growth or induces a dieback on the long term. This analysis therefore pinpoints low precipitation as the most important factor explaining the low predictive ability of CASTANEA when driven by uncorrected WRF output, with low relatively humidity and high temperatures as the second and third most important factors. The underestimation of beech distribution with WRF is caused by an overall dry and warm bias that strongly decreases the soil moisture. WRF systematic biases are relatively easy to correct, but are also the main drawbacks to using WRF together with CASTANEA.
Fig. 3 Change in goodness-of-fit (% relative to SAFRAN TSS) for beech distribution produced by CASTANEA in a sensitivity test. The sensitivity study was carried out by using SAFRAN climate variables jointly with one WRF variable. The previous simulations using the SAFRAN and WRF forcings are presented as a reference for the maximum and minimum TSS possible. SDR = Solar Downward Radiation, Ws = Wind speed, RH = Relative Humidity, R = Rainfall, T = Temperature. The lower the value, the more the WRF variable degraded the simulated distribution of beech compared to SAFRAN
Climatic Change
Concerning the RegCM model, a cold bias (up to 4 °C) especially located in the Southern France allows simulated beech presence where it is known to be absent, because temperatures fall below a climatic threshold in CASTANEA. This is compensated by a warm bias in others locations (i.e. mountainous areas) when spatially averaged, however it is partly visible with the Taylor diagrams (Fig. 1b, d). A similar mechanism is at work along the Atlantic coast with PROMES, except that the cold bias is much smaller (1 °C). Nevertheless, the associated increase of tree growth is sufficient to trigger widespread beech presence in an area where it is currently absent in observations. It appears that the temporal variability of the different RCMs is fairly similar and of small relevance in explaining the difference in beech distribution between the climate models. This area-averaged diagnostic highlights that RCMs have similar performance at simulating variables over larger spatial scales. However, the spatial variability at local scale is much more critical: the response in beech distribution depends on which extent the RCMs stay -or do not stay- within the range of a sustainable climate for beech. Relatively small climate departures can lead to incorrect predictions of presence or absence, especially when close to climatic limits of beech distributions. Our analysis indicates that the limited biases and ability to reproduce spatio-temporal patterns in ALADIN may be sufficient to allow it to feed directly into impact studies for this spatio-temporal domain with no bias correction. In contrast, RegCM and WRF clearly cannot be used without corrections because their TSS indicates that they are barely better than random, and substantial degradation of projected distributions occurs with uncorrected PROMES. Fixing these issues, either for impact studies or for development of dynamically coupled models, therefore requires bias correction. The climatic and tree distribution analyses are coherent concerning the type of correction required. WRF requires the most straightforward type of bias correction because the bias is relatively constant in time and space. Indeed, we find that bias-corrected WRF (linear shift) performs as well as ALADIN for analyses of impacts on beech (data not shown). PROMES, and RegCM in particular, have difficulties reproducing some aspects of spatial patterns of both climate variables and tree distributions, requiring more complex spatial bias correction. Importantly, no model appears to require temporal bias correction for studying this type of impact, and this is very positive since this type of bias correction is particularly problematic. In general, bias correction is not without potentially serious drawbacks - since it may add a source of uncertainties (White and Toumi 2013; Ruffault et al. 2014) - so the advantages and disadvantages should be weighed carefully before using bias-corrected climate models (Ehret et al. 2012).
4 Conclusion We used a multi-model approach to simulate European beech distribution over France. We demonstrate that biases in climate models, rather than discrepancies in spatial pattern, were the primary source of error in modeling tree distribution, keeping in mind these results apply to France and for European beech (a drought-sensitive species). Most importantly, we show that it is difficult to evaluate the quality of a climate model without knowing how the impact model will respond. That is, evaluations with climate data are insufficient to provide measure of whether a climate model is Bgood enough^ or provide a ranking of models for a particular impact. Impact models can be very sensitive to changes in input variables, such that even small discrepancies are amplified and highlighted. Consequently impact models provide an
Climatic Change
important complementary point of view to identify the improvement effort required for the different climate models by pinpointing their limiting processes and variables. Such analysis could be especially important when climate models are dynamically coupled with ecological model because of error amplification. This paper also demonstrates the relevance for impact modelers to run their models with the simulated present climate before running simulations for future climates. Nevertheless the use of impact models to evaluate climate models is surprisingly rare (Lung et al. 2013). This may be explained by several factors including i) limited interactions between the climate and impact communities, ii) the high diversity of impact models that need to be tested and iii) the lack of standardized tools for carrying out analyses. Testing with impact models might also be useful in regions where climate data is of insufficient quality for reliable model benchmarking (Brands et al. 2013). The use of ecological data with an inverse model could provide a range of climatic informations in location with sparse weather station networks, as in boreal latitudes. This approach, already used in paleoclimate, was validated with modern data for most of atmospheric variables considered (Kaplan et al. 2003; Wu et al. 2013). Its wider application could also provide a Bcrossvalidation^ for the global high resolution gridded bioclimatic dataset (<1 km) which suffers from a lack of precision at fine scale (Bedia et al. 2013). Acknowledgments This work has been funded by the Humboldt project funded by the GIS (Groupement d’Intérêt Scientifique) BClimat, Environnement et Société^. This work has also been carried out in the framework of the LabEx BASC (Biodiversity, Agrosystems, Society and Climate, ANR-11-LABX-0034). This work also contributes to the HyMeX program (HYdrological cycle in The Mediterranean EXperiment – http://www.hymex. org) through INSU- MISTRALS support and the GEWEX program of WCRP. The simulations used in this work were downloaded from the Med-CORDEX database (www.medcordex.eu). We acknowledge: (i) Samuel Somot for providing one of the simulation outputs; (ii) the ESPRI/IPSL database teams and Julien Lenseigne for their technical support in data archive and simulations runs.
References Allouche O, Tsoar A, Ronen K (2006) Assessing the accuracy of species distribution models: prevalence, kappa and the true skill statistic (TSS). J Appl Ecol 43(6):1223–1232 Bedia J, Herrera S, Gutiérrez JM (2013) Dangers of using global bioclimatic datasets for ecological niche modeling. Limitations for future climate projections. Glob Planet Chang 107:1–12 Brands S, Herrera S et al (2013) How well do CMIP5 earth system models simulate present climate conditions in Europe and Africa? Clim Dyn 41(3-4):803–817 Cheaib A, Badeau V et al (2012) Climate change impacts on tree ranges: model intercomparison facilitates understanding and quantification of uncertainty. Ecol Lett 15(6):533–544 Colin J, Déqué M, Radu R, Somot S (2010) Sensitivity study of heavy precipitation in limited area model climate simulations: influence of the size of the domain and the use of the spectral nudging technique. Tellus A 62(5):591–604 Davi H, Dufrêne E, Granier A, Le Dantec V, Barbaroux C, François C, Bréda N (2005) Modelling carbon and water cycles in a beech forest: part II.: validation of the main processes from organ to stand scale. Ecol Model 185(2):387–405 Davi H, Barbaroux C, Dufrêne E, François C, Montpied P, Bréda N, Badeck F (2008) Modelling leaf mass per area in forest canopy as affected by prevailing radiation conditions. Ecol Model 211(3– 4):339–349 Dee DP, Uppala SM, Simmons AJ et al (2011) The era-interim reanalysis: configuration and performance of the data assimilation system. QJRMS 137(656):553–597 Delpierre N, Soudani K et al (2012) Quantifying the influence of climate and biological drivers on the interannual variability of carbon exchanges in European forests through process-based modelling. Agric For Meteorol 154:99–112
Climatic Change Di Luca A, Flaounas E, Drobinski P, Brossier CL (2014) The atmospheric component of the Mediterranean Sea water budget in a WRF multi-physics ensemble and observations. Clim Dyn 1–27 Domínguez M, Gaertner MA, de Rosnay P, Losada T (2010) A regional climate model simulation over West Africa: parameterization tests and analysis of land surface fields. Clim Dyn 35(1):249–265 Drobinski P, Ducrocq V, Alpert P et al (2014) HyMeX, a 10-year multidisciplinary program on the mediterranean water cycle. Bull Am Meteorol Soc. doi:10.1175/BAMS-D-12-00242.1 Dufrêne E, Davi H, François C, Maire GL, Dantec VL, Granier A (2005) Modelling carbon and water cycles in a beech forest: part I: model description and uncertainty analysis on modelled NEE. Ecol Model 185(2):407– 436 Ehret U, Zehe E, Wulfmeyer V, Warrach-Sagi K, Liebert J (2012) HESS opinions Bshould we apply bias correction to global and regional climate model data?^. Hydrol Earth Syst Sci Discuss 9(4):5355–5387 Fang J, Lechowicz M (2006) Climatic limits for the present distribution of beech (Fagus L.) species in the world. J Biogeogr 33(10):1804–1819 Flaounas E, Drobinski P, Vrac M, Bastin S, Lebeaupin-Brossier C, Stéfanon M, Borga M, Calvet JC (2013) Precipitation and temperature space-time variability and extremes in the mediterranean region: evaluation of dynamical and statistical downscaling methods. Clim Dyn 40(11–12):2687–2705 Flaounas E, Drobinski P, Borga M et al (2012) Assessment of gridded observations used for climate model validation in the Mediterranean region: the HyMeX and MED-CORDEX framework. Environ Res Lett 7(2): 024017 Giorgi F, Jones C, Asrar GR et al. (2009). Addressing climate information needs at the regionallevel: the CORDEX framework. WMO Bulletin 58 (3) - July 2009 Gleckler PJ, Taylor KE, Doutriaux C (2008) Performance metrics for climate models. J Geophys Res 113: D06104. doi:10.1029/2007JD008972 Kaplan JO, Bigelow NH, Prentice IC, et al. (2003) Climate change and arctic ecosystems: 2. Modeling, paleodata-model comparisons, and future projections. J Geophys Res 108(D19) Lebaube S, Le Goff N, Ottorini JM, Granier A (2000) Carbon balance and tree growth in a Fagus sylvatica stand. Ann For Sci 57(1):49–61 Lebourgeois F, Bréda N, Ulrich E, Granier A (2005) Climate–tree-growth relationships of European beech (Fagus sylvatica L.) in the French Permanent Plot Network (RENECOFOR). Trees 19:385–401 Lindner M, Maroschek M et al (2010) Climate change impacts, adaptive capacity, and vulnerability of European forest ecosystems. For Ecol Manag 259(4):698–709 Lung T, Dosio A, Becker W, Lavalle C, Bouwer LM (2013) Assessing the influence of climate model uncertainty on EU-wide climate change impact indicators. Clim Chang 120(1-2):211–227 Marmion M, Parviainen M, Luoto M, Heikkinen RK, Thuiller W (2009) Evaluation of consensus methods in predictive species distribution modelling. Divers Distrib 15:59–69 Pal JS, Giorgi F, Bi X et al (2007) Regional climate modeling for the developing world: the ICTP RegCM3 and RegCNET. Bull Am Meteorol Soc 88(9):1395–1409 Quintana-Seguí P, Le Moigne P, Durand Y et al (2008) Analysis of near-surface atmospheric variables: validation of the SAFRAN analysis over France. J Appl Meteorol Climatol 47:92–107 Ruffault J, Martin-StPaul NK, Duffet C, Goge F, Mouillot F (2014). Projecting future drought in Mediterranean forests: bias correction of climate models matters! Theor Appl Climatol 1–10 Ruti P, Somot S, Dubois C, et al (2014) MED-CORDEX initiative for Mediterranean Climate studies (Bulletin of the American Meteorological Society, in revision) Schmitt U, Möller R, Eckstein D (2000) Seasonal wood formation dynamics of beech (Fagus sylvatica L.) and black locust (Robinia pseudoacacia L.) as determined by the “pinning” technique. J Appl Bot 74:10–16 Skamarock WC, Klemp JB, Dudhia J, Gill DO, Barker DM, Duda MG, Huang XY, Wang W, Powers JG (2008) A description of the advanced research wrf version 3. Technical Report, NCAR Taylor KE (2001) Summarizing multiple aspects of model performance in single diagram. J Geophys Res 106(D7):7183–7192 Taylor KE, Stouffer RJ, Meehl GA (2012) An overview of CMIP5 and the experiment design. Bull Am Meteorol Soc 93(4):485–498 Teuling AJ, Stöckli R, Seneviratne SI (2011) Bivariate colour maps for visualizing climate data. Int J Climatol 31(9):1408–1412 Thuiller W (2003) BIOMOD – optimizing predictions of species distributions and projecting potential future shifts under global change. Glob Chang Biol 9:1353–1362 Thuiller W (2014) Editorial commentary on ‘patterns and uncertainties of species’ range shifts under climate change’. Glob Chang Biol 20(12):3593–3594 Thuiller W, Lafourcade B, Engler R, Araújo MB (2009) BIOMOD – a platform for ensemble forecasting of species distributions. Ecography 32:369–373
Climatic Change White RH, Toumi R (2013) The limitations of bias correcting regional climate model inputs. Geophys Res Lett 40(12):2907–2912 Wu H, Guo Z, Guiot J et al (2013) Elevation-induced climate change as a dominant factor causing the late Miocene C4 plant expansion in the Himalayan foreland. Glob Chang Biol. doi:10.1111/gcb.12426