Causes of future Mediterranean precipitation decline depend on the season

Future mean precipitation in the Mediterranean is projected to decrease year-round in response to global warming, threatening to aggravate water stress in the region, which can cause social and economic difficulties. We investigate possible causes of the Mediterranean drying in regional climate simulations. To test the influence of multiple large-scale drivers on the drying, we sequentially add them to the simulations. We find that the causes of the Mediterranean drying depend on the season. The summer drying results from the land-ocean warming contrast, and from lapse-rate and other thermodynamic changes, but only weakly depends on circulation changes. In contrast, to reproduce the simulated Mediterranean winter drying, additional changes in the circulation and atmospheric state have to be represented in the simulations. Since land-ocean contrast, thermodynamic and lapse-rate changes are more robust in climate simulations than circulation changes, the uncertainty associated with the projected drying should be considered smaller in summer than in winter.


Introduction
The Mediterranean is among the global 'hot-spots' of climate change, where severe consequences of climate change are expected (Giorgi 2006). One of the leading causes of the vulnerability of the Mediterranean to greenhouse gas-driven warming is the projected decline in mean precipitation (Cramer et al 2018, García-Ruiz et al 2011. The precipitation decrease is projected year-round (Giorgi andLionello 2008, Seager et al 2014), and is opposite to the atmospheric moisture content, which increases with warming (Held and Soden 2006). A drying trend in the Mediterranean was already observed in the last century (Mariotti et al 2015), and especially the southern parts of the region presently experience severe water stress (Cramer et al 2018). Multiple sectors such as agriculture, industry, infrastructure, and tourism compete for the available freshwater resources (Kovats et al 2014). A further decline in Mediterranean precipitation will likely aggravate water stress and social tensions, which can even lead to conflicts and migration (Niang et al 2014). Also, the Mediterranean drying threatens ecosystems and human health (García-Ruiz et al 2011, Kovats et al 2014, Cramer et al 2018. To attenuate the severe impacts of climate change in the Mediterranean, costly adaptation and mitigation measures are inevitable (Kovats et al 2014). Stakeholders thus rely on reliable projections of precipitation changes in the Mediterranean, which means that a detailed understanding of the responsible physical pathways and associated uncertainties is necessary. The suggested causes of the Mediterranean drying in the literature are diverse. Any further distribution of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI. circulation changes (Zappa and Shepherd 2017). Regarding thermodynamics, changes in atmospheric humidity and moisture saturation can influence the drying (Seager et al 2014). Furthermore, the summer drying has been related to the land-ocean warming contrast (Rowell and Jones 2006, Boé and Terray 2014, He and Soden 2016, lapse-rate changes (Kröner et al 2017, Brogli et al 2019, misrepresentations of largescale circulation features (Bladé et al 2012, Van Haren et al 2015 and land-atmosphere feedbacks (Rowell and Jones 2006). Some potential causes of the drying have been challenged by studies finding a merely weak influence of changes in circulation (Kendon et al 2010, He andSoden 2016), baroclinic eddy activity (Seager et al 2014), summer weather regimes (Boé et al 2009) or large-scale humidity changes (He and Soden 2016). There is thus a need to clarify the importance of the suggested causes for the Mediterranean drying and there are also indications that the causes could be different for summer and winter.
With this study, we aim for clarification and identify the most important drivers behind the Mediterranean drying throughout the year by using regional climate model (RCM) simulations. To this end, potential drivers simulated by global climate models (GCMs) are added to the boundary conditions of the RCM simulations as in Kröner et al (2017) and Brogli et al (2019). We investigate the influence of these drivers on the regional climate simulations separately for the summer and winter season using a consistent simulation design.

Materials and methods
2.1. Global and regional climate models Two global climate models provide the large-scale drivers, namely HadGEM2-ES (Martin et al 2011) and MPI-ESM-LR (Stevens et al 2013). We separately downscale both GCMs and present the mean of the two hereafter. To do so we use the RCM COSMO in Climate mode (COSMO-CLM4.8_clm17) (Steppeler et al 2003, Rockel et al 2008, Doms et al 2011 at a horizontal resolution of 0.44°, and using 40 vertical levels. The RCM is driven by lateral boundary conditions and SSTs and we simulate the 30 yr future period (2070-2099) assuming the high-end RCP8.5 (Moss et al 2010) emission scenario, and compare to the 30 yr historical period . The soilmoisture is initialized by extracting the field from a 30 yr long reanalysis-driven spinup simulation. The RCM is run with a constant aerosol climatology for the historical and scenario period, whereas the GCMs contain the large-scale changes associated with global aerosol forcing. The simulation domain covers the pan-European continent. At least one year is used as a spinup in the regional simulations before the analysis. Simulations using COSMO-CLM4.8_clm17 are a part of the EURO-CORDEX ensemble , and the performance concerning the historical climate is compared with the majority of the RCMs in the EURO-CORDEX ensemble in CH2018 (2018) (see their figures 13.19 and 13.21). Generally, the biases in COSMO-CLM4.8_clm17 driven by the two GCMs we are using here are among the smallest in the model ensemble, yet in the Mediterranean, the downscaled HadGEM2-ES simulation exhibits a dry bias in summer and the downscaled MPI-ESM-LR simulation a wet bias in winter. The simulated circulation patterns from the original GCM simulations are reproduced in the RCM, albeit slightly weaker (figure S7 in the supplementary material available online at stacks.iop. org/ERL/14/114017/mmedia).

Simulation strategy
The current simulation strategy follows the separation methodology used in Kröner et al (2017) and Brogli et al (2019), but is specifically adapted to the purpose of the current study. To identify the main causes of the Mediterranean summer and winter drying, we distinguish between 'drivers' and 'feedbacks'. Drivers refer to large-scale changes in the atmospheric environment that are communicated to the continental scale as changes in large-scale temperature, humidity, and circulation. Feedbacks refer to changes in continental and sub-continental-scale processes, such as related to land-surface, radiative and cloud feedbacks. The current study attempts a separation of the climate change response in terms of its drivers, without specifically addressing the contributions of the associated feedbacks and uncertainties connected to the feedbacks. The drivers are communicated to the RCM simulations by lateral-boundary forcing, while the feedbacks result from physical processes within the RCM domain. Thus, to assess uncertainties associated with feedbacks, it would be necessary to compare multiple RCMs, which is not feasible in this study.
To investigate climate change drivers, in addition to the historical and future simulations described above, we perform three further RCM experiments for both GCMs. In each experiment, we modify the initial and boundary conditions of the historical simulation with large-scale drivers that could cause the Mediterranean drying. With this framework, we can compare the experiments to the unmodified historical simulation, and assess whether the imposed drivers induce precipitation changes. The three additional experiments are named thermodynamics and lapse rate (TDLR), sea surface temperature (SSTE) and mean state & circulation (MEA). Additionally, we assess the full climate change signal (FCC) by comparing the historical and future simulation. The changes we include as initial and boundary conditions in each experiment are listed in table 1. The most important properties of all experiments will be graphically presented in section 3.
From a technical standpoint, we change either only the temperature (TDLR, SSTE) or the temperature and wind (MEA) in the experiments. When we change the temperature at a boundary grid-point, we change the humidity assuming constant relative humidity and All processes provided by GCM, e.g. high-frequency circulation changes and local deviations from constant relative humidity changes ✓ adapt the pressure according to the hydrostatic balance. All changes are based on the 30 yr mean annual cycle of the differences in temperature or wind between the future and historical simulation. The mean annual cycle is further smoothed using a spectral filter (Bosshard et al 2011, Kröner et al 2017, which removes high-frequency variability in temperature or wind changes. Only a timemean annual cycle remains ( figure S4). Each year in the experiments is modified with the same annual cycle. The smoothed annual cycle is imposed onto the boundary fields of the historical simulation, retaining the full historical variability, meaning that the historical sequence of weather systems at the lateral boundaries is unchanged. The CO 2 concentrations are increased in all experiments.
The employed simulation strategy, quantifying different drivers by modifying the boundary conditions, is often referred to as the pseudo-global warming (PGW) approach (

Thermodynamic and circulation changes
Here we present changes in atmospheric properties, for all the experiments shown in table 1, which are relevant for precipitation changes. Figure 1 shows the simulation properties for the summer season and figure 2 for winter. We present the near-surface temperature (2 m warming), relative humidity (2 m ΔRH), low-level specific humidity (850 hPa Δq), mid-atmospheric absolute wind speed (500 hPa ΔU), geopotential height anomaly (500 hPa ΔZ g ) and eddy kinetic energy (500 hPa ΔEKE). EKE is a eulerian proxy for the strength of storm-tracks Hodges 2002, Shaw et al 2016). EKE is computed here based on the wind on 500 hPa, filtered using a ten-day high-pass Butterworth filter of fifth order. Subsequently, we compute EKE as -

Thermodynamics and lapse rate
The TDLR simulation is the most simple experiment, where the temperature at the lateral boundaries is changed by a horizontally uniform increment, and the specific humidity by assuming unchanged relative humidity. The uniform temperature increment is derived from the domain mean change between the future and historical simulation at different model levels ( figure S5). The prescribed SST is raised by the same temperature increment as the lowermost atmospheric level, which limits the magnitude of the landocean warming contrast (figures 1(a) and 2(a)). As given by the experimental design, a strong nearsurface warming and an increase in specific humidity are seen for both summer (figures 1(a), (i)) and winter (figures 2(a), (i)) over land and ocean. However, the magnitude and spatial pattern of the humidity and temperature changes are not matching the full climate change signal, in particular for the increase in specific humidity, which is strong in comparison (figures 1(l) and 2(l)). The TDLR experiment also features an increase in static-stability (figure S5), which is caused by the temperature dependence of the moist-adiabatic lapse rate (Schneider 2007). Spatial differences in land temperature and relative humidity changes are dynamically induced by the RCM, for example, the summer warming in the Mediterranean is amplified (figure 1(a)) in response to lapse-rate changes (Kröner et al 2017, Brogli et al 2019. The amplified summer warming is connected to a moderate decrease in 2 m relative humidity (figure 1(e)), since the locally warmer air features an increased saturation vapor pressure (Sherwood and Fu 2014, Byrne and O'Gorman 2016). In winter, changes in near-surface relative humidity are small (figure 2(e)). Dynamic changes are largely absent in TDLR, as there is virtually no change in wind speed and geopotential height (figures 1(m), (q) and 2(m), (q)). Also, there is no relevant change in eddy kinetic energy (figures 1(u), 2(u)).

Sea surface temperature
The SSTE experiment includes the changes described for TDLR, but SST changes projected by the GCMs, both the mean and patters, are additionally included. This reflects the fact that the ocean warms less than the atmosphere, which can be seen for our simulations in figures 1(b) and 2(b). We observe a weaker increase in atmospheric humidity in SSTE compared to TDLR since the oceanic humidity source regions experience less warming (figures 1(f), (j) and 2(f), (j)). Yet, lowering the SST has a stronger effect on the specific humidity compared to the relative humidity, especially over land (figures 1(f), (j) and 2(f), (j)). As in TDLR, dynamic changes are largely absent in SSTE (figures 1(n), (r), (v) and 2(n), (r), (v)). Previous studies have found that the land-ocean warming contrast originates from differences in the near-surface moisture availability, i.e. from the fact that the marine boundary layer is moister than the boundary layer over land

Mean state and circulation
In MEA, the time-mean atmospheric state and circulation are changed, this means, that MEA includes spatially varying thermodynamic and dynamic changes. As expected, changes in the seasonal mean geopotential 2 Note that even though this term is used in a large number of studies, it can potentially be misunderstood (Somerville and Hassol 2011). height anomaly and wind speed are seen (figures 1(o), (s) and 2(o), (s)). Even though we observe changes in the seasonal mean circulation, the respective simulations are fed with the baroclinic eddies of the historical period and therefore no substantial changes in eddy kinetic energy arise (figures 1(w) and 2(w)). The increase of specific humidity is even lower than in SSTE, especially in winter (figures 1(k) and 2(k)). The decreased moisture availability could be caused by mean flow changes, for example by increased subsidence. Another possibility is connected to spatial differences in the temperature increase present in MEA. If an atmospheric moisture source region locally warms less in MEA than in SSTE, the connected moisture content will then also increase less. In summer, a strong decrease in surface relative humidity in the Mediterranean can be seen in MEA ( Figure 1(g)), connected to the amplified land-ocean contrast ( figure 1(c)).

Full climate change
In most aspects, the FCC simulation is very similar to MEA, the main difference being high-frequency changes in the atmospheric circulation, such as changes in the baroclinic eddies fed at the boundaries, and the resulting changes in eddy kinetic energy that are only noteworthy in FCC (figures 1(x) and 2(x)). However, as also shown in Brogli et al (2019), the high-frequency changes are not important for the temperature change (figures 1(d) and 2(d)). Furthermore, relative humidity can vary according to the GCM in FCC (instead of being constant at the model boundary), but this only has a small influence (figures 1(h), (l) and 2(h), (l)).

Precipitation changes
Here we present how the increasingly complex drivers included in the experiments influence precipitation changes over Europe in summer (figure 3) and winter (figure 4). Precipitation changes are shown in terms of mean, frequency (amount of days with precipitation 1 mm), and intensity (amount of precipitation on days with 1 mm), following Rajczak and Schär (2017). In both seasons, whenever the mean precipitation decreases, the decrease in precipitation is connected to decreasing precipitation frequency, while the precipitation intensity generally increases due to the increased moisture availability, which is consistent with Rajczak and Schär (2017) and Polade et al (2017). The patterns of seasonal mean precipitation change simulated by our FCC experiments (figures 3(d) and 4(d)) closely resemble the simulation results of larger ensembles of regional (CH2018 2018) and global (Christensen et al 2013) climate simulations.

Summer season
A decrease in mean summer precipitation over the Mediterranean occurs in both the TDLR and SSTE experiments (figures 3(a), (b)), even in the absence of dynamic changes. Especially the land-ocean warming contrast is important, as the precipitation decrease is substantially stronger in SSTE compared to TDLR (∼−40% versus ∼−25%). The reduced ocean warming in SSTE compared to TDLR weakens the moisture increase ( figure 1(j)). Together with the amplified Mediterranean land warming, decreasing relative humidity, and higher static stability, a substantial decrease in mean precipitation and precipitation frequency occurs (figures 3(b), (f)). The precipitation decrease in the TDLR experiment (Figure 3(a)) is weaker than in SSTE, due to the weaker land-ocean warming contrast ( figure 1(a)).
The changes in the seasonal mean atmospheric state and circulation further reduce summer precipitation ( figure 3(c)). The additional reduction of precipitation will depend upon the circulation changes of the driving GCM ( figure S2). However, for the two models considered, this only modifies the pattern and intensity of the drying. The comparably low humidity increase, which can be connected to circulation changes or local temperature change anomalies (figures 1(g), (k)), may play a role in MEA too. We regionally find a further reduction of precipitation in FCC compared to MEA (figures 3(c), (d)). This suggests that large-scale changes in baroclinic eddies, relative humidity, and other high-frequency processes contribute to the reduction of future Mediterranean mean summer precipitation, but are not necessary for the drying to occur. Both MEA and FCC have a more important influence in areas north of the Mediterranean.
In absolute numbers, our FCC simulations project a summer precipitation decline of ∼−30 mm/season, of which ∼−20 mm/season are readily explained by SSTE (figure S1). The absolute numbers however strongly depend on the Mediterranean sub-region and decrease from north-west to south-east (figure S1).

Winter season
The same drivers leading to a precipitation decline in summer, do not generally induce decreasing Mediterranean precipitation in winter (figure 4). Thermodynamics alone lead to a domain-wide mean precipitation increase ( figure 4(a)). A smaller precipitation increase results when the land-ocean warming contrast and ocean warming patterns are considered ( figure 4(b)). Yet, only by considering changes in the seasonal mean atmospheric state and circulation, the Mediterranean drying is generated (figure 4(c)). As for summer, the changes baroclinic eddy activity in FCC over the Mediterranean as seen in the EKE (figure 2(x)) contribute to the drying but are not necessary for the drying to occur (figures 4(c), (d)). Yet, for land areas north of the Mediterranean sea, FCC has substantial influence, for example over the Iberian Peninsula the precipitation only decreases in FFC. The absolute precipitation decline over land for FCC is around 20-30 mm/season (figure S1), and thus is on the same order of magnitude as in summer.
The MEA precipitation changes are dynamically consistent with the increasing geopotential height anomaly over the Mediterranean (figure 2(s)). This agrees with studies suggesting an influence of increasing mean atmospheric pressure and subsidence in that region (Giorgi and Lionello 2008, Seager et al 2014, Zappa et al 2015b. The higher atmospheric pressure has been suggested to be connected to a northward shift of the Hadley cell edge and jet stream (figure 2(o)) (Giorgi andLionello 2008, Seager et al 2014). As in summer, the dynamically or thermodynamically induced weaker humidity changes in MEA compared to SSTE likely contribute to the drying (figure 2(k)). Furthermore, the strong increase in geopotential height over the Mediterranean is also connected to locally stronger mid-atmospheric warming and a decrease in relative humidity (not shown), which can also influence precipitation.

Comparison of seasons
As shown above, the summer Mediterranean drying is projected when only a few simple drivers are considered (thermodynamics, lapse-rate changes and landocean contrast), which are connected to well-understood properties of the climate system (see section 3, or Joshi et al (2008), Byrne and O'Gorman (2013b), Hall (2014), Sherwood and Fu (2014), Brogli et al (2019)). Additional changes in atmospheric state and circulation only act to enhance the summer drying. In contrast, to project the winter drying, changes in the mean atmospheric state and circulation have to be included, which depend on the downscaled GCM ( figure S3). In general, circulation changes in state-ofthe-art GCMs are considered to come with a high uncertainty (Shepherd 2014, Fereday et al 2018. The results presented here, therefore, suggest higher confidence in projections of the summer Mediterranean drying compared to the winter drying. The summer drying is supported by physical understanding, a variety of model simulations (Christensen et al 2013) and also observations (Mariotti 2010, Mariotti et al 2015. Thus, if greenhouse gas emissions continue to rise unabated, we have to expect an enhanced Mediterranean summer drying. The situation in winter is different. While some observations support Mediterranean winter drying (Mariotti 2010, Mariotti et al 2015, Seager et al 2019, large-scale circulation changes are poorly understood and models disagree on the projected drying, both in large ensembles (Christensen et al 2013) and our regional simulations (figure S3). These seasonal differences in the causes of the Mediterranean drying and the different associated uncertainties are relevant for studying the impact of climate change on freshwater resources, as the absolute decrease in precipitation is similar in summer and winter (figure S1).
Why do the causes of Mediterranean precipitation changes follow different physical pathways depending on the season? The reasons behind this are likely connected to the seasonally dominant precipitation formation mechanisms. Locally formed convective precipitation is more abundant in summer, thus the local thermodynamic state, including tropospheric humidity and temperature, should play a central role. Winter precipitation, on the other hand, is predominantly caused by synoptic processes and largescale ocean to land transport, and thus we should expect changes in the large-scale circulation and mean state of the atmosphere to be relevant.

Conclusions
We investigated the causes of the projected Mediterranean precipitation decline for the summer and winter season, which will increase the water stress for humans and ecosystems if climate change continues at the current rate (Cramer et al 2018). By adding various large-scale drivers to the boundary conditions of regional climate simulations, we found that the causes of the precipitation decline differ depending on the season. In summer, drying is a more direct consequence of greenhouse-gas driven climate warming than in winter. In summer, the atmosphere will be warmer, further away from saturation, and more stably stratified. We showed that, in combination with the land-ocean warming contrast, these changes suffice to induce the Mediterranean summer drying. Changes in the seasonal mean circulation, the connected changes in the seasonal mean atmospheric state, and changes in the baroclinic eddies entering the domain further reduce the summer precipitation but are not the primary driver for the precipitation decline. The opposite result is found for winter, where mean state and circulation changes are necessary to induce declining Mediterranean precipitation, and the decline is partially reinforced by changes in baroclinic eddies. The changes in the mean atmospheric state and circulation in winter are connected to an elevated geopotential height anomaly over the Mediterranean stemming from the driving GCMs. Seager et al (2014Seager et al ( , 2019 suggested a combination of time-mean circulation and specific humidity changes as causes of the Mediterranean winter drying, consistent with our findings.
Our results strengthen the confidence in the projected decline in summer freshwater availability since declining summer precipitation is induced by basic thermodynamic properties of the changing climate system and is supported by multiple lines of evidence. In winter, evidence also points towards a precipitation decrease, but it depends on more uncertain dynamic changes. More reliable projections of winter precipitation will thus require further improvement of the representation of circulation changes in climate models.