This work uses a network of GPS stations over Europe from which a homogenized integrated water vapor (IWV) dataset has been retrieved, completed with colocated temperature and precipitation measurements over specific stations to (i) estimate the biases of six regional climate models over Europe in terms of humidity; (ii) understand their origins; and (iii) finally assess the impact of these biases on the frequency of occurrence of precipitation. The evaluated simulations have been performed in the framework of HYMEX/Med-CORDEX programs and cover the Mediterranean area and part of Europe at horizontal resolutions of 50 to 12 km.
The analysis shows that models tend to overestimate the low values of IWV
and the use of the nudging technique reduces the differences between GPS and
simulated IWV. Results suggest that physics of models mostly explain the
mean biases, while dynamics affects the variability. The land
surface–atmosphere exchanges affect the estimation of IWV over most part of
Europe, especially in summer. The limitations of the models to represent
these processes explain part of their biases in IWV. However, models
correctly simulate the dependance between IWV and temperature, and
specifically the deviation that this relationship experiences regarding the
Clausius–Clapeyron law after a critical value of temperature (
Finally, it is shown over the SIRTA observatory (near Paris) that the frequency of occurrence of light precipitation is strongly conditioned by the biases in IWV and by the precision of the models to reproduce the distribution of IWV as a function of the temperature. The results of the models indicate that a similar dependence occurs in other areas of Europe, especially where precipitation has a predominantly convective character. According to the observations, for each range of temperature, there is a critical value of IWV from which precipitation starts to increase. The critical values and the probability of exceeding them are simulated with a bias that depends on the model. Those models, which generally present light precipitation too often, show lower critical values and higher probability of exceeding them.
Humidity plays a major role in the water and energy cycles due to its strong radiative effect associated with a positive feedback on climate (Randall et al., 2007) and its importance to control precipitation and particularly extreme ones (Held and Soden, 2006; Neelin et al., 2009; Sahany et al., 2012). Trends and variability of humidity and precipitation are strongly correlated (Trenberth et al., 2003; Zhang et al., 2013) and several studies have revealed that the rate of increase in daily extreme precipitation is highly connected with the warming following the Clausius–Clapeyron (C-C) relation (Allen and Ingram, 2002; Pall et al., 2007; Kharin et al., 2007). This rate of precipitation is indeed affected by the humidity content of the atmosphere (integrated water vapor, IWV), which rises as the climate warms (e.g., Trenberth, 2011). Nevertheless, dynamical processes (O'Gorman and Schneider, 2009; Sugiyama et al., 2010; Singleton and Toumi, 2013; Muller, 2013; Drobinski et al., 2016), lack of humidity sources leading to a decrease in relative humidity (RH; Drobinski et al., 2018), or low or high precipitation efficiency (Drobinski et al., 2016; Trenberth et al., 2003) can explain the deviation from C-C rate locally. Humidity variability at regional scale – and not only at the surface – thus needs to be assessed to better anticipate the precipitation change, and more specifically the rate of heavy precipitation, which are not well estimated by global models (e.g., Allan and Soden, 2008).
Another aspect that links IWV and precipitation concerns the triggering of precipitation and thus the frequency of occurrence of precipitation: Holloway and Neelin (2009) showed that precipitation over the tropical oceans is strongly sensitive to free-tropospheric humidity even more than surface humidity, and Neelin et al. (2009) and Sahany et al. (2012) further conclude that there exists a threshold of IWV, which depends on the mean tropospheric temperature, over which precipitation starts to increase significantly. They also showed that this critical value of IWV does not correspond to the saturation value when temperature increases, i.e., that at higher temperature, deep convection occurs at a lower value of relative humidity. This means that IWV is a relevant parameter to measure over long-term periods, at high temporal resolution and at the regional scale in order to establish the relationship between IWV–precipitation and temperature and monitor its possible evolution. Models still have strong difficulties in adequately simulating the water cycle (Trenberth et al., 2003; Flato et al., 2013), and often presents the “too often too light precipitation” problem (e.g., Sun et al., 2006; Panthou et al., 2016). A better knowledge of the IWV–precipitation relationship would be a help to better constrain models.
Up to now, very few long-term (> 15 years) and homogeneous datasets of water vapor measurements exist, even less at subdaily timescales. These datasets are necessary to understand the humidity variability at regional scales at different timescales. Besides, the colocation of such measurements with independent measurements of precipitation and vertical profiles of temperature provide a strong added value for better climate understanding. Reanalyses are of course a good tool to have these three parameters colocated over long-term and at subdaily timescales; however, precipitation mostly relies on the model physics. Moreover, Flato et al. (2013) have shown that even in reanalyses, the relationship between the IWV trend and the temperature trend presents differences between reanalyses and deviates from C-C over tropical oceans.
In this study, we make use of the Global Positioning System (GPS) IWV dataset
that has been processed as done by Parracho et al. (2018) and which
provides IWV measurements from over 100 European sites covering a period
of 5 years or more. The GPS technique accurately measures IWV (accuracy
around 1–2 kg m
The paper is organized as follows: Sect. 2 presents the observational datasets and the different simulations used in this study. Section 3 describes the methodology to compare observations and models. In Sect. 4, the ability of models to reproduce the mean value of humidity and its variability over Europe at different timescales is evaluated. The influence of dynamical and physical processes is discussed, and a special focus on the scaling of IWV with temperature is developed. In Sect. 5, the issue of how much a bias in IWV can enhance the problem of “too often too light precipitation” behavior of models is raised by considering the relationship between mean tropospheric temperature, IWV and precipitation in the different models and observations over the SIRTA supersite in France. Then, by considering other stations across Europe, the generalization of this relationship is assessed. Finally, a conclusion is given in Sect. 6.
The GPS dataset used in this study is based on homogeneously reprocessed GPS delay data produced by the NASA's Jet Propulsion Laboratory in the framework of the first International GNSS Service (IGS) reprocessing campaign. The data cover the period from January 1995 to May 2011 and include more than 400 stations globally. For the present study, the delay data were screened and converted into 6-hourly IWV estimates as described in Parracho et al. (2018). Then, daily values are computed using this 6-hourly dataset. The dataset was restricted to the period from January 1995 to December 2008 and includes 95 GPS stations over Europe as shown in Fig. 1. It represents many more stations than in the study of Parracho et al. (2018) because for the purpose of their study which estimates IWV trends, they restricted their selection to stations with only small gaps over the 15-year period from 1995 to 2010. Here, only stations with less than 5 years of observations are not considered in the evaluation of model humidity bias.
Mean values of IWV in winter
This study also uses observations collected at the SIRTA atmospheric
observatory, located 20 km southwest of Paris (48.7
The Météo-France COMEPHORE (“COmbinaison en vue de la Meilleure
Estimation de la Precipitation HOraiRE”) product is used to allow a fairer
intercomparison between models and observations than the single
rain gauge (Chen and Knutson, 2008): it is an hourly reanalysis of
precipitation by merging radar data and rain gauges over France at
The Météo-France radiosoundings, launched twice a day from Trappes (near 00:00 and 12:00 UTC), 15 km to the west of SIRTA, are also used to compute the mean tropospheric temperature (more details in Sect. 3).
List of models used in the study.
The list of regional climate models (RCMs) and details about the settings are given in Table 1. All the simulations use the 6-hourly European Center for Medium-Range Weather Forecast (ECMWF) reanalyses ERA-Interim (Dee et al., 2011) as RCM boundary conditions. They cover at least the period 1989–2008 as initially recommended in the MED-CORDEX project (Ruti et al., 2015). For LMDZ, which is a global model with regional zoom capability, temperature, wind speed and specific humidity are nudged towards the ERA-Interim fields outside the MED-CORDEX domain. It must be noted that the mesh of LMDZ is not regular within the zoom region and the resolution varies between 50 and 30 km. All other RCMs are forced at the boundaries using 3-dimensional re-analyses of wind, humidity, temperature or potential temperature, and geopotential height. For CCLM, cloud ice and liquid water are additionally prescribed at the domain boundaries. The IPSL WRF simulation uses nudging at all scales within the domain for temperature, wind and humidity above the planetary boundary layer (Salameh et al., 2010; Omrani et al., 2013, 2015). The other models did not use nudging in the Med-CORDEX domain.
Simulations used here were produced from five models (ALADIN V5.2, Colin et
al., 2010; CCLM, Rockel et al., 2008; WRF V3.1.1, Skamarock et al., 2008; LMDZ
V4, Hourdin et al., 2006; PROMES, Dominguez et al., 2010, 2013) with a
horizontal resolution around 50 km (0.44
The GPS observations are supplemented by the HadISD v.2.0.1.2016 subdaily dataset of surface parameters (e.g., temperature, dew point temperature, wind, pressure; Dunn et al., 2012). It is global and based on the Integrated Surface Database (ISD) dataset from NOAA's National Climatic Data Center. Stations were selected on the basis of their length of record and reporting frequency before they are passed through a suite of quality control tests. It is a joint effort from the MetOffice Hadley Center and the National Center for Atmospheric research (NCAR).
To compare models and observations, we consider differences as the “model minus observation” results throughout the paper.
Values of linear regression slopes obtained for each
model/each month when plotting IWVmod–IWVobs as a function of the difference in
altitudes between model and GPS at each station. The first figure is for
native grid of the model and the second figure is for model regridded to
ERAI grid. It is expressed in 10
Each modeling group has provided a file containing the gridded IWV over the 1995–2008 period at daily resolution on its native grid. The IWV is either computed online or offline in the model. The offline computation can introduce some errors due to vertical integration over the discretized vertical grid. For each model, we extracted the value of IWV at the closest grid point of GPS stations. We did it using the native grid, and also after having regridded all the model outputs to ERA-Interim (ERAI) grid, to have a fair comparison with ERAI when necessary. The difference in altitude between the GPS station and the closest grid point is difficult to take into account and can introduce strong bias over complex terrain (Hagemann et al., 2003; Wang and Zhang, 2009). As a consequence, the stations where the difference in altitude is higher than 500 m were removed from the analysis. Note the number of stations that are removed depends on each model since the models do not use the same topography and the same projection. Then, a linear correction is applied on model outputs to reduce the bias due to orography: for each model and each month, we plotted the difference in the monthly averaged IWV values between the model and GPS as a function of the difference in altitude and we concluded that a linear correction can be applied to take into account the difference in altitude. For each different month, the slope of the linear regression between these two differences is computed and we apply the corresponding correction to IWV values of the model. The values of the slope of the linear regression for each model and each month are indicated in Table 2, both for the native grid of the model and for the regridded outputs. Various evaluation metrics (Table 3) have been computed with and without correction for these two grids. The correction does not impact the variability scores (e.g., interannual variability), but does affect the mean bias (not always by an improvement) and slightly reduces the standard deviation of the difference (Table 3). It does not affect the ranking of performance between models.
Mean bias and standard deviation (SD) in
kilograms per meter (kg m
Note that at SIRTA, the difference in altitude is weak and results are thus not impacted by this problem.
The following comparisons were made with SIRTA observations:
Occurrence (%) of nonprecipitating days (first number), very light precipitation (second number) and light precipitation (third number) for different datasets (columns) computed over different years (rows) for two different periods of the year: W is for winter (Julian day from 1 to 100) and S is for summer (Julian day from 151 to 251). Native grids are used.
Table 4 presents an estimate of the frequencies of occurrence of nonprecipitating days, very light precipitation and light precipitation for winter (Julian day 1 to 100) and summer (Julian day 151 to 251) when considering either the common period of the two datasets (2004–2007) or the full period of each dataset (i.e., 1997–2007 for COMEPHORE and 2003–2015 for SIRTA rain gauge). The number of dry days increases for the two estimations from COMEPHORE when considering a longer period than the common period. For ReOBS, the statistics remain similar for the two different periods. However, when considering the most recent years only, which will be used in the next section (2008–2015), the number of dry days increases. The influence of the number of years, the years considered and the products used to estimate these frequencies of occurrences is generally small but significant. Even though model errors are most of the time beyond this uncertainty, it has to be kept in mind in the following analysis.
The objective is to characterize how precipitation depends on IWV for
different ranges of mean tropospheric temperature. To do that, we divided our
datasets into four different bins of temperature: the first bin is for
temperature less than or equal to 254.5 K, the second bin is for temperature
between 254.5 and 258 K, the third one corresponds to temperature between
258 and 262.5 K and the fourth one is for temperature higher than 262.5 K.
This choice has been made to ensure a high number of samples in each bin of
temperature to proceed to the next step. In this way, the mean
tropospheric temperature of one bin is similar more or less 1 K for all the
datasets. Then, in each bin of temperature, the daily mean precipitation
rates are sorted according increasing values of daily mean IWV. IWV bins are
then defined such that they contain an equal number of pairs of
precipitation rates and IWV (40 samples) to ensure a reasonable number of days
to compute the 50th quantile of precipitation, which indicates if there are
more precipitating days than nonprecipitating days. The range of each IWV
bin is thus not constant but allows the transition
between mostly nonprecipitating days and mostly precipitating days to be identified quite easily. For
each model and for observations, a critical value of IWV (
Figure 1 indicates the mean values of IWV retrieved from GPS measurements in
winter (Fig. 1a) and in summer (Fig. 1b). In winter, higher values are
observed along the Atlantic and Mediterranean coasts while central and
eastern Europe exhibit very low values of IWV. In summer, there are two
different regimes: (i) north of 45
Table 3 shows the mean bias and the standard deviation (SD) of ERA-Interim
and RCMs' IWV on their native grid or regridded to ERAI grid, in comparison
to GPS estimates. Daily datasets are used to compute these statistics and
exactly the same sampling is used between models and observations. It shows
that the mean bias ranges from 0.5 to 1.0 kg m
The standard deviation indicates a large spread around observations for all
models (
Occurrence for 30-day periods from January to December over the
years 2004–2007 of
Percentage of simulated daily mean IWV values which overestimate
the GPS ones
To the first order, the IWV variability is dominated by the seasonal cycle (shown in Fig. 1), which is underestimated by models (not shown), and which explains part of the model standard deviation: indeed, Fig. 3 shows the percentage of simulated daily mean IWV values which overestimate GPS values at each station for the ensemble of the five models regridded to the ERAI grid, in winter and summer. Figure 3a and b present results without height correction, while Fig. 3c and d are done with height-corrected data. In winter, more than 70 % of values are overestimated over most stations, with or without height corrections. In summer, this percentage decreases appreciably in most stations and in almost half of them it reaches values below 50 % (Fig. 3b). The use of height correction homogenizes the results in summer between stations and the very low or very high percentages do not appear anymore when this correction is applied. UCLM50 simulation is the moistest, especially in summer (not shown), which explains its high standard deviation in Table 3.
The IWV variability also comes from the interannual variability. For each month, we computed its anomaly by subtracting the average value of the month over all the years. We then computed the correlation between the anomalies of the GPS IWV estimates and those from the models. The minimum and maximum values of these correlations for each model when all stations are considered are indicated in Table 3 (last column). As indicated by these numbers that range between 0.78 and 0.99, the interannual variability is captured well by the model, which is not surprising since this variability is mostly driven by large-scale advection of air masses (all models use 6-hourly ERA-Interim parameters as lateral boundary conditions). Note that the maximum correlation is very high for all models, while the minimum values are higher for ERAI, LMD50 and IPSL50 than for the three others. This is mostly due to a specific month (not shown): in January 1996, three models have an anomaly very different from the observations and, as a consequence, the interannual correlation for January goes down.
To conclude, the RCM configuration allows a reasonable representation of the large-scale advection of air masses by the models, which is an important driver of humidity within the RCM domains (see also Trenberth et al., 2005). This is further improved when the model is nudged towards reanalysis, as in the IPSL model (Table 3).
Nevertheless RCM errors are significant, with a difficulty in reproducing low
values of IWV, generating a mean positive bias for all models. Most of the
contribution of humidity to the integrated water vapor comes from the
surface, and the interaction between environment and clouds and boundary
layer. Small-scale processes are thus important to reproduce moisture
sources and sinks (precipitation, mesoscale circulations, evaporation and
evapotranspiration, clouds and microphysics). To better understand these
errors, we assess the link between surface humidity and IWV at different
timescales. Figure 4 displays the monthly mean values of IWV versus monthly
mean values of 2 m specific humidity (Q2) averaged over all stations where
and when both IWV from GPS and Q2 from HadISD are available. Monthly means
are computed if at least 60 concomitant (both IWV from GPS and Q2 from
surface station) values are available (i.e., about two values per day out of a possible four). A total of 3238 months are obtained, spread over 42 different
stations. The average number of stations per month is 19 with a maximum of
30 stations. Figure 4 shows that 2 m specific humidity is a very good proxy
for IWV at the monthly scale. All models but IPSL have a similar
relationship between the two variables to the observed one (slope of
Summertime interannual correlation between IWV and Q2 at GPS stations. In bold, the mean value for each model, followed by min and max values. The other values in the table corresponds to the standard deviation of the difference in correlation between two models.
Monthly values of IWV as a function of monthly values of 2 m specific humidity (Q2) averaged over all stations where and when both IWV from GPS and Q2 from HadISD are available. Monthly means are computed if at least 60 coexisting values exist (i.e., about two values per day over a possible four). A total of 3238 months are obtained, spread over 42 different stations. The average number of stations per month is 19 with a maximum of 30 stations). The color of circles corresponding to each model is indicated on the legend. All model outputs have been regridded to the ERAI grid. The first number indicates the slope of the regression obtained when considering the same months than observations at each station, while the second number is the slope of the regression when considering all months at all grid points (only models).
Despite the strong correlation between the annual cycle of Q2 and those of
IWV, Q2 is not necessarily a good proxy for IWV at other timescales or to
tackle model biases (for instance IPSL bias for surface humidity is strongly
negative while it is weak and slightly positive for IWV). While in the
wintertime, humidity variability mostly originates from the air mass
advection, summertime variability is mainly affected by land–surface
interactions and boundary layer processes. Several studies have shown the
existence of a large spread in the representation of the surface fluxes and
land–atmosphere coupling strength between models over Europe, due to the
fact that Europe is a zone of transition between the regime of “energy-limited” areas with low land–atmosphere coupling strength and those of “oil-moisture-limited” areas with high
land–atmosphere coupling strength. The
difficulty to represent soil conditions and surface fluxes is then increased
(Cheruy et al., 2015; Boe and Terray, 2014; Fischer et al., 2007; Knist
et al., 2017). The interannual correlations between IWV and Q2 summertime
anomalies are indicated for each model in Table 5, as averaged values over
all GPS stations and minimum and maximum values across all GPS stations (an
attempt was made for GPS IWV and HadISD Q2 but there were too many missing
values). For most stations, the correlation is higher than 0.5 with a mean
value around 0.8 for the five models. The standard deviation of the difference
between models is around 0.15, which reveals a good agreement between them.
Some stations, however, indicate higher differences, as indicated by the
minimum values that strongly differs between models. IPSL50 and LMD50 models
present strong correlation at all stations (
A way to check the behavior of models is to consider the relationship
between IWV and temperature. At global scale, the scaling of IWV with
temperature is expected to follow the Clausius–Clapeyron (C-C) law, at a
rate of about 6 %
Physically,
To determine the slope after
As observed (e.g., Ruzmaikin et al., 2014), on average the troposphere can be
separated into two layers, one being the boundary layer (BL) and the other
one the free troposphere (FT) – assuming constant humidity within the two
layers, IWV can be expressed as follows:
RH
Equation (1) can thus be approximated by the following:
Values of the left-hand side (LHS),
second term of the right-hand side (RHS) and total RHS
of Eq. (2) computed for the SIRTA site for two different temperatures
(10 and 20
At temperatures lower than
Figure 5a indicates that despite the bias at low temperature, models
generally capture the deviation from C-C and the IWV maximum value that is
reached at high temperature, except for the UCLM model for which IWV
continues to increase slightly with
The link between the IWV and RH evolutions for the model ensemble is shown
in Fig. 6a. Once again, the transition from one regime to another is smoother
than for observations (Fig. 5b), but the decrease in RH starts around the
same range of temperature as for the transition of IWV–
In conclusion to this section, models tend to overestimate low values of IWV. Although they generally capture the IWV scaling with temperature well, small-scale processes also explain part of the standard deviation (not only induced by the deviations from driving data) when considering the differences with GPS IWV data. In the next section, we consider the impact of these humidity biases in models of light precipitation occurrence.
Figure 7 displays the 50th percentile of precipitation (computed
including days without precipitation) as a function of IWV for four
different bins of mean tropospheric temperature for both observations and
the models (see Sect. 3 for the details of the methodology). The
observations are considered for the period 2008–2015 and the different
models for the period 2001–2008 (i.e., the same number of years, despite the
time shift). For all these datasets, there exists a critical value of IWV,
The 50th quantile of precipitation as a function of IWV for four
different bins of tropospheric temperature at 48.7
The second reason for the high frequency of occurrence of very light
precipitation in these models is that the probability of exceeding this
critical value of IWV is strongly overestimated in these two models in
comparison with observations (Fig. 8b). Statistically speaking, it means that
the models that are too humid have a positive bias in light precipitation.
For CMCC, the dipole between winter and summer observed in the estimate of
nonprecipitating days is also observed here: though the critical value of
IWV is correct at low and moderate
Period for which the maximum value of the temperature bin (
In addition to the period 2001–2008 discussed above, we tested two other 8-year
periods (1989–1996 and 1995–2002) and the entire period (1989–2008) to
assess the influence of the considered period on the results. Figure S1 in the Supplement and
Table 7 show that results are rather robust among models and periods, though
some uncertainty exists in both
Latitude/longitude/altitude (
To have an idea of the models' behavior over other parts of Europe, several
other stations are considered in this section. Their locations are shown in Fig. 1b by black diamonds and details are given in Table 8. Except for
Marseille located in the south of France where the COMEPHORE product
associated with GPS has been used, the model outputs are not compared with
observations for the other stations. Figure 9 displays the occurrence of
nonprecipitating days for the different models and from the COMEPHORE product
at the station location when available (over France),
To relate these characteristics to temperature and humidity, we reproduced
the analysis done at SIRTA. The value of the critical value of IWV over
which precipitation starts to increase is generally similar among models,
despite increasing dispersion with temperature. This value depends on the
stations, indicating the influence of local specificities in the estimation
of this relationship. For instance, UCLM50, which is the model with the most
important difference between southern and northern stations for the annual
cycle of dry days, also indicates strong differences in the
This work uses GPS integrated water vapor measurements associated with temperature and precipitation measurements to (i) estimate the biases of six regional climate models over Europe in terms of humidity; (ii) understand their origins; and (iii) finally assess the impact of these biases on the occurrence of precipitation.
The first part of the study aimed at evaluating the mean bias and standard deviations of IWV in models compared to GPS measurements at interannual, seasonal and daily timescales. An interesting result is that all models overestimate the lower values of IWV (nighttime, wintertime) at all stations. The spread among models is increased during summertime. Our analysis suggests that the model physics mostly explain the mean biases, while dynamics affects the variability. The use of nudging towards reanalyses thus improves the representation of the large-scale advection of air masses and reduces the standard deviation of differences between GPS retrieved IWV values and simulated ones. The land surface–atmosphere interactions are crucial in the estimation of IWV over most part of Europe, especially in summer, and explain part of the mean biases. However, the relationship between IWV and temperature, which deviates from the Clausius–Clapeyron law after a critical value of temperature, is generally well captured by models. This critical temperature presents a spatial variability since it corresponds to the value when relative humidity starts to decrease. It is thus strongly dependent on local processes which drive the local humidity sources (from evaporation and advection). This explains why the maximum values of IWV are not necessarily observed over warmer areas, which often corresponds to dry areas, where a soil-moisture-limited regime is dominant.
The improvement in humidity representation may also help in the representation of precipitation distribution. Indeed, in the second part of this study, it is shown that the biases in IWV and most importantly IWV's distributions as a function of temperature strongly impact the occurrence of light precipitation over France, and most generally over areas where convection is the main process of precipitation triggering. For each range of mean tropospheric temperature, there exists a critical value of IWV over which a pickup in precipitation occurs. This is observed and simulated by models, but the critical values and the probability of exceeding them vary between models and observations. Models which present light precipitation too often generally show lower critical values and higher probability of exceeding them. Thus, a better knowledge and representation of the triggering thresholds of precipitation and of their variability should potentially help to improve the representation of the whole precipitation distribution in models. The ensemble of simulations with implicit and explicit convection that will be performed in the framework of the Flagship Pilot Studies' convective-permitting climate simulation of the CORDEX project will allow us to assess the sensitivity of precipitation triggering and distribution to the model resolution. Issues that will be explored in more detail following this work will be the role of humidity in (i) the triggering of precipitation in simulations at different resolution, (ii) the low precipitation rates (precipitation efficiency) and (iii) the impact of too-easy triggering in the entire precipitation distribution.
The SIRTA ReOBS netCDF file is available at
The supplement related to this article is available online at:
The analysis has been performed by SB, with contribution from PD and MC and some inputs from all the other authors to improve the paper. RR also suggested some references to push the analysis forward and provided the CNRM50 dataset. MC provided the SIRTA-ReOBS dataset, OB the GPS dataset with some work from AP, LL the LMD50 simulation, PL and DC the CMCC50 dataset, and CG and MDA the UCLM dataset.
The authors declare that they have no conflict of interest.
This article is part of the special issue “Hydrological cycle in the Mediterranean (ACP/AMT/GMD/HESS/NHESS/OS inter-journal SI)”. It is not associated with a conference.
This work is a contribution to the HyMeX program (HYdrological cycle in The Mediterranean EXperiment) through INSU-MISTRALS support and the MEDCORDEX program (Coordinated Regional climate Downscaling EXperiment–Mediterranean region). This research has received funding from the French National Research Agency (ANR) project REMEMBER (grant ANR-12-SENV-001) and is a contribution to the VEGA project through LEFE/INSU support, to the EECLAT project through LEFE-INSU and CNES supports and to the GNSS4SWEC COST action ES1206 through EU support. It was supported by the IPSL group for regional climate and environmental studies, with granted access to the HPC resources of GENCI/IDRIS (under allocation i2011010227), by SIRTA Working Group “Climate studies” and by the national infrastructure ACTRIS-FR, identified on the French road map for Research Infrastructures, published by the Ministry of Research. The SIRTA-ReOBS effort also benefited from the support of the L-IPSL funded by ANR under the “Programme d'Investissements d'Avenir” (grant ANR-10-LABX-0018) and by the EUCLIPSE project funded by the European Commission under the Seventh Framework Program (grant no. 244067). To process the data, this study benefited from the IPSL mesocenter ESPRI facility which is supported by CNRS, UPMC, Labex L-IPSL, CNES and Ecole Polytechnique. We would like to acknowledge the SIRTA team for collecting data, Cindy Lebeaupin-Brossier and Marc Stefanon for providing simulation outputs, the CNES (Centre National d'Etudes Spatiales) for partially funded Marjolaine Chiriaco research, Emmanuele Lombardi and ENEA for the Med-CORDEX database, Samuel Somot for his role in Med-CORDEX coordination, and Thomas Noel for its help in post-processing data. The work to carry out the simulations of the UCLM model was funded by the Spanish Ministry of Education and Science and the European Regional Development Fund, through grant CGL2007-66440-C04-02. Edited by: Heini Wernli Reviewed by: Juan José Gómez-Navarro and one anonymous referee