The latter decades of the twentieth century and the early years of the twenty-first century have seen many extreme weather events, among which heat waves and extreme precipitation in particular have become increasingly frequent in many global areas (IPCC 2014). Compared to other natural disasters such as floods or storms, detecting and quantifying droughts is more complex, since droughts are characterized by a slow onset and a high resilience to their effects, while long-term impacts may emerge months or even years after the drought peak (Vogt and Somma 2000; Wilhite 2000; Wilhite et al. 2007). Another level of complexity arises from the many different definitions of drought, including meteorological, agricultural, hydrological, socioeconomic, and ecological droughts (Mishra and Singh 2010; Crausbay et al. 2017). Different types of droughts can lead to different, often cascading impacts, affecting various economic sectors such as agriculture (Schmidhuber and Tubiello 2007; Li et al. 2009), hydroelectric and thermal power generation (Bartos and Chester 2015), public water supply (Iglesias et al. 2009), waterborne transport, and tourism (Thomas et al. 2013). Environmental and social impacts include, for example, vegetation stress (Vicente-Serrano et al. 2013), wetland, soil and land degradation (Bai et al. 2008), and links with migration (Kelley et al. 2015). Consequently, the recognition of drought as a climate hazard, as well as a better understanding of its manifold aspects, is becoming an urgent priority in a warming world (Dai 2011), with the result that drought is becoming a “hot topic” in climatology (Trenberth et al. 2014).
In this study, we focus on meteorological drought, which is caused by a prolonged rainfall deficit, often enhanced by other meteorological conditions, such as high temperatures, high evapotranspiration rates, and desiccating winds (Palmer 1965; Wilhite and Glantz 1985). In recent decades, many studies have reported an overall global tendency toward more frequent and severe meteorological drought events (e.g., Dai 2011, 2013; Spinoni et al. 2014; Osborn et al. 2016), even though the consensus about the extent and magnitude of the change is not universal (Seneviratne 2012; Sheffield et al. 2012; Hauser et al. 2017). Although most studies agree on the location of recent past drought hot spots—namely, the Mediterranean region, western North America, southern South America, large parts of Africa, and northeastern China (Trenberth et al. 2014; Spinoni et al. 2015a; Coelho et al. 2016; Cook et al. 2016; Dai and Zhao 2017; Zittis 2018)—other regions have also been hit by megadroughts in recent years. Examples are western North America including Mexico from 1999 to 2007 (Stahle et al. 2009), Australia from 2001 to 2009 (van Dijk et al. 2013), Russia in 2010 (Wegren 2011), California in 2013–14 (S. Wang et al. 2014), Europe over the last two decades (Hanel et al. 2018), South Africa in 2015–18 (Masante et al. 2018), and Kenya in 2014–19 (Reliefweb 2019).
In contrast with past events, the overall picture for meteorological drought projections is still incomplete. A number of studies investigated multimodel hydrological and meteorological drought projections based on global climate models (GCMs) of previous (e.g., CMIP3; Seager et al. 2007; Sheffield and Wood 2008; Dai 2011; Orlowsky and Seneviratne 2012) and current generations (CMIP5; Prudhomme et al. 2014; Touma et al. 2015; Ukkola et al. 2018). However, such projections are often presented with medium spatial resolution (i.e., not better than 1°) and sometimes using only a limited number of simulations. Due to such limitations, most projections suffer from large uncertainties (Burke and Brown 2008; Dai 2013; Orlowsky and Seneviratne 2013; Zhao and Dai 2017; Lu et al. 2019). On the other hand, some studies investigated drought hazard projections on selected countries or regions (Cook et al. 2015; Spinoni et al. 2018) by means of regional climate models (RCMs).
This study aims at improving the available meteorological drought projections by using—for the first time, to our knowledge—a large number of simulations (103) based on a combination of GCMs and RCMs, and producing high spatial resolution (0.44° or ~50 km) global projections of drought frequency, severity, and peak events (i.e., historically undocumented extreme droughts) for the twenty-first century. The RCMs are guided by the parent GCMs but, being able to represent small-scale processes and features (Rummukainen 2010), they have been shown to simulate more accurately present-day, observed precipitation characteristics and higher-order statistics, and in turn to “add value” to the performances of GCMs (Feser et al. 2011; Di Luca et al. 2012; Giorgi et al. 2014; Dosio et al. 2015; Torma et al. 2015; Kendon et al. 2017; Dosio et al. 2019). Consequently, as we discuss in section 3, the use of RCMs, coupled with GCMs, can help showing drought-related spatial patterns that the use of GCMs alone cannot provide.
We considered two climate scenarios: the moderate representative concentration pathway (RCP) 4.5 and the more extreme RCP8.5 (van Vuuren et al. 2011). The RCM simulations were produced in the framework of the Coordinated Regional Climate Downscaling Experiment (CORDEX; www.cordex.org) at a spatial resolution of 0.44°. Although some CORDEX simulations have previously been used in drought-related studies at regional scale (e.g., Meresa et al. 2016; Zahradníček et al. 2016; Diasso and Abiodun 2017; Um et al. 2017; Spinoni et al. 2018; Tabari and Willems 2018), they have never been applied for global-scale drought analyses.
This study also aims to answer the following question: where and to what extent will the projected temperature rise (IPCC 2014) play a crucial role in increased drought frequency and severity? Similar to a previous study focused on Europe (Spinoni et al. 2018), here we separately investigated drought projections based on both standardized precipitation index (SPI; McKee et al. 1993) and standardized precipitation-evapotranspiration index (SPEI; Vicente-Serrano et al. 2010), in order to evaluate the importance of including temperature in drought projections. SPI uses only precipitation as input (Spinoni et al. 2014), while SPEI uses both precipitation and potential evapotranspiration, which incorporates the effects of temperature (Beguería et al. 2014).
This is not the first attempt to investigate global drought projections using different drought indicators: Touma et al. (2015) compared four indicators, including the SPI and the SPEI, using the results of 15 GCMs. They regridded the outputs at common 1° resolution, unavoidably introducing an interpolation bias because only 2 of the 15 GCMs used have a spatial resolution comparable to 1°. In our study, the use of RCMs—over the native common 0.44° grid—allows a higher resolution without the need to regrid the outputs. Moreover, the larger number of simulations, especially over some regions, allows deeper evaluation of the uncertainties and more robust analysis of statistical significance of projected changes.
The remainder of this paper is structured in three main sections. In section 2, the data and methods are described, with a focus on the CORDEX dataset, the drought indicators, and the definition of drought-related variables. In section 3, the increase or decrease in drought frequency and severity from 1981–2010 to 2071–2100, both at global and macroregional spatial scale, are analyzed. The relative importance of temperature and/or precipitation as meteorological drivers for future droughts, is also discussed in section 3, focusing on areas where the two drought indicators result in diverging projections. Section 4 summarizes the results of the study and anticipates possible further steps.
2. Data and methods
a. Input data: Gathering macroregional CORDEX simulations
The CORDEX initiative is a World Climate Research Programme (WCRP) core project (Giorgi et al. 2009; Giorgi and Gutowski 2015), which has promoted the provision of climate information at regional scale by means of coordinated regional climate downscaling (RCD) techniques (Hewitson and Crane 1996), over several continental regions of the world. Different institutions and research groups (for the complete list, see www.cordex.org) have contributed to producing climate outputs based on a variety of RCMs over 14 geographical domains, covering the main continental areas of the world (see Table 1).
CORDEX domains and combinations of GCMs and RCMs available for each region. The letter C before the acronym is introduced to avoid confusion with the macroregions shown in Fig. 2 and used for regional statistics. CSIRO-CCAM is a global model with a stretched grid.
The CORDEX outputs consist of multivariable time series at different spatial and time resolutions and climate scenarios. For each CORDEX domain (i.e., region), a set of simulations is available, depending on the combinations of GCMs and RCMs. For our purposes, only the simulations including data for daily precipitation (P) and minimum and maximum temperatures (TN and TX) from 1981–2100, for both RCP4.5 and RCP8.5, were considered. Unfortunately, we could not extend our analyses to RCP2.6 because, at the time of inquiry, the corresponding simulations did not cover the entire world. The selected spatial resolution is 0.44° (~50 km), as higher-resolution data (0.11° or 0.22°) are not available for all domains.
The primary sources of CORDEX data are the Earth System Grid Federation (ESGF) web portals. However, not every CORDEX simulation used in this study was available on the ESGF portal at the beginning of our data search. In the meantime, new simulations have been added to the ESGF data catalogue, but some areas are still covered by only a few simulations. Consequently, we obtained as yet unpublished data directly from the contact points for each domain. In total, we collected 103 GCM–RCM simulations. Those provided for the CORDEX domain Australia (AUS) are included in those provided at global scale by the Commonwealth Scientific and Industrial Research Organization (CSIRO). Unfortunately, no simulations for Southeast Asia (SEA) could be obtained. However, those belonging to East Asia (EAS) also include SEA in the domain.
The complete list of GCMs and RCMs related to the 103 simulations is reported in Table 1 and the responsible institute and key references in Table 2. The exceptionally large number of simulations used makes this study unique, and particularly valuable. Nonetheless, it must be pointed out that, as the number of simulations varies from region to region, the robustness of the results may be affected over regions where the number of simulations is limited (see section 3). Figure 1 shows the number of simulations per region, with the smallest number (<10) over Australia and southern Siberia, and the largest (>60) over the eastern Mediterranean region, where many CORDEX domains overlap.
GCMs and the RCMs used in this study, institute(s) owning the intellectual property, and key reference(s).
In general, any procedure of subselecting models potentially introduces bias, so the use of varying number of simulations in different regions needs validation. Unfortunately (see Table 1), no simulations generated by the same combination of GCMs and RCMs are available over all the CORDEX domains, with the exception of the six GCMs coupled with the CSIRO-CCAM. Thus, performing a validation by comparing drought projections obtained using all 103 simulations versus those obtained using a combination of models based on a single RCM is likely to be depending too much on that single RCM. However, as discussed in section 3, the absence of clear discontinuities over the bordering areas between CORDEX domains suggests that the spatial distribution of the future drought conditions is not biased by using of different simulations over different regions.
Independently for each simulation, we computed drought indicators and derived drought variables, based on temperature and precipitation data, and (only at a later stage) the 103 outputs were merged over a common grid. Generally, the use of a large number of simulations avoided pronounced discrepancies along the borders of CORDEX domains, where simulations for different domains overlap. Over all domain borders, we tested whether the use of simulations from only one domain would give substantially different results compared with simulations from another domain. In only two cases, nonnegligible discrepancies were found: along the Urals (borders between the Europe, central Asia, and Arctic domains) and, to a minor extent, southeastern China (borders between the central Asia, South Asia, and East Asia domains). Eventually, over these areas—as in all other areas—we elected to use the ensemble median of all simulations, in order to maintain methodological homogeneity with the rest of the global areas. See section 3 for details.
b. Meteorological drought indicators: SPI and SPEI
For each simulation, we converted the daily data into monthly averages for minimum and maximum temperature, and monthly sums for precipitation. As all simulations are compliant with CORDEX standards, no gaps or spurious data were found, except for very few cases of unrealistically extremely low winter minimum temperature over northeastern Siberia in one of the simulations—which we nevertheless decided to retain. In fact, as our analyses is based on median values from all simulations available for a certain grid point, large outliers are excluded.
Estimating potential evapotranspiration (PET) in an environment with a changing atmospheric CO2 concentration is not straightforward (Roderick et al. 2015; Milly and Dunne 2016). We used the Hargreaves–Samani equation (H-S; Hargreaves and Samani 1985), which derives PET by estimating solar radiation from minimum and maximum temperature and is frequently used in drought studies (e.g., Vangelis et al. 2011; Vicente-Serrano et al. 2011). The use of both minimum and maximum temperature avoids the large overestimation of droughts in dry and hot periods by models based on mean temperature only, as the Thornthwaite’s model (Th; Thornthwaite 1948; Weiß and Menzel 2008; Shahidian et al. 2012). On the other hand, the H-S method tends to overestimate PET in humid regions and underestimate it in regions with high wind speed (Temesgen et al. 1999). The H-S method uses extraterrestrial radiation rather than solar radiation and neglects atmospheric transmissivity, which is influenced by high moisture content in the atmosphere in humid regions. Moreover, the H-S method does not consider atmospheric moisture, which is particularly important in humid regions, where PET tends to decrease as atmospheric moisture increases (McKenney and Rosenberg 1993; Tabari 2010).
More realistic estimations of PET, suitable for drought-related studies (Sheffield et al. 2012; Trenberth et al. 2014; Dai and Zhao 2017), could be obtained with the Penman–Monteith method (P-M; Allen et al. 2006). For example, Hosseinzadehtalaei et al. (2017) found smaller bias in future PET changes with the P-M method compared to the H-S method. In spite of this shortcoming, P-M is considered more realistic because is based on sunshine duration, temperature, vapor pressure, humidity, and wind speed data. However, it makes use of questionable physical assumptions, as its parameterization refers to a surface of grass with a sufficient amount of water; therefore, in very or extremely dry periods P-M tends to overestimate PET (Brutsaert and Parlange 1998). Unfortunately, such variables are available only for a limited number of CORDEX simulations, which is why we opted for H-S to compute PET, as was done in Spinoni et al. (2018) for Europe.
For each simulation, climate scenario, and grid point, we computed time series of SPI and SPEI values. Following McKee et al. (1993) for SPI and Vicente-Serrano et al. (2010) for SPEI, we fitted precipitation data on a gamma distribution to obtain SPI, and the difference between precipitation and PET on a log-logistic distribution to obtain SPEI.
The time scale of the drought indicator is sometimes used to define the type of drought, especially when the study focuses on drought impacts; that is, short scales (up to 3 months) refer to meteorological droughts, medium scales (6 months) to agricultural droughts, and longer scales (12 months or more) to hydrological droughts (Heim 2002). Rather than this definition, in this study we investigated meteorological droughts as driven by meteorological variables (Mishra and Singh 2010, 2011) using two meteorological indicators (SPI-12 and SPEI-12), similar to Spinoni et al. (2018) for Europe. We used a 12-month accumulation period when computing drought indicators (SPI-12 and SPEI-12), this being a compromise between short time scales suitable to detect the specific time when a drought event occurs and long time scales suitable for multiannual cycles. The analyses on seasonal drought projections at different warming levels using shorter time scales (in particular the SPI-3 and the SPEI-3) are left to future research.
As discussed in previous studies using a similar methodology (Spinoni et al. 2015b, 2018), we selected the entire period (1981–2100) as a baseline period to fit the underlying distribution of the drought indicators. In fact, the choice of a shorter period, possibly characterized by frequent and severe droughts, could influence the indicator over the entire period, leading to underestimation of droughts in other periods, or vice versa. Moreover, the longer the baseline period, the more robust the standardized drought indicators (Wu et al. 2005). In contrast, if only the past decades are chosen as a reference period, the possible local acclimatization as the century progresses cannot be taken into account. In particular, using past data as a reference period to investigate future drought events might introduce bias, since “normal” conditions in the past may become anomalous in the future, so that events at the end of the twenty-first century could be unrealistically extreme. Note that the baseline period described above should not be confused with the reference period (1981–2010) used in comparing the projected drought quantities.
The SPI and SPEI results have been analyzed separately, as we specifically wanted to isolate the effect of temperature on meteorological drought projections. The role of temperature, which is often incorporated in drought studies as PET, is critical and much debated in the scientific literature (Dai et al. 2018), due to the fact that, in the context of progressive warming, an increase of precipitation can be outbalanced by a larger atmospheric evaporative demand forced by higher temperatures. By separating the projected changes in drought variables according to the SPI and SPEI indicators, we have been able to analyze whether or not, and where, projected changes in precipitation and/or temperature drive future changes in drought frequency and severity. Furthermore, meteorological drought impacts can be better correlated with SPI or SPEI, depending on the socioeconomic sector involved (Naumann et al. 2015), and therefore different users can benefit from this study if the results for both indicators are presented separately.
c. Drought frequency, severity, and extreme events
Once we had computed time series, from 1981 to 2100, for the SPI-12 and SPEI-12 values at gridpoint scale (0.44°), for all simulations and both RCPs, we applied the same methodology as described in Spinoni et al. (2014) to detect drought events, that is, using the “run theory” as proposed by Yevjevich (1967). That is to say: a drought event starts when the drought indicator falls below one negative standard deviation for at least two consecutive months and ends when the indicator turns positive.
Drought frequency (DF) is then defined as the number of events in a given period, with the two investigated 30-yr periods in this study being 1981–2010 (representing the reference period) and 2071–2100 (representing the far future). The severity of an event is estimated as the sum, in absolute values, of all the monthly indicator values between the start and the end of the event. Since our focus is on the change in severity of the average drought event between the two periods, here drought severity (DS) refers to the average severity—not the total severity—of drought events in the selected period. As the world is likely to face more extreme events during the twenty-first century (IPCC 2014), we defined a specific quantity (PK, for “peak events”) representing the number of drought events during 2071–2100 that are more severe than the most severe event that occurred during 1981–2010.
In the maps included in this paper, the drought quantities are presented as the median values over all simulations available for the corresponding grid point. Thus all the available simulations for each point were used, as we prioritized maximum possible use of information. Using median values, together with the overall large number of simulations, helps to minimize the impact on the results of individual simulations, which may be biased and thus lead to biased SPI and SPEI time series. Over the borders between two CORDEX domains (regions), some model grids do not perfectly overlap (although the shift is in most cases less than 0.05°). Therefore, we interpolated the shifted simulations over a common 0.44° grid, using an interpolation scheme based on radial Gaussian weights. The only area where we found nonnegligible discontinuity between domains is over the Urals (Europe and central Asia).
The core results of this study focus on the changes in the selected drought variables between the reference period (1981–2010) and the far future (2071–2100). If not explicitly stated otherwise, such changes are considered robust in sign if at least two-thirds of the simulations indicate a change with the same sign. For instance, for a given grid point where 10 simulations are available, an increase in drought frequency is defined as robust in sign if at least seven simulations project an increase. We evaluated the possibility of using a larger threshold (e.g., 75% of model sign agreements), but the robustness in sign of the results was sensitive to outliers in regions where a very limited number of runs (<10) is available.
The results have also been analyzed at the macroregional scale, using the regions described in the Special Report of the IPCC “Managing the Risks of Extreme Events and Disasters to Advance Climate Change Adaptation” (IPCC 2012). However, as is evident in Fig. 2, we made some minor changes. The Caribbean islands and Central America form one region, as do the north tropical Pacific and northern Australia. We discarded the west Indian Ocean and southern and eastern tropical Pacific islands, due to the small fraction of land.
As in Spinoni et al. (2014) and Spinoni et al. (2018), we excluded from our analyses extremely arid or very cold areas, such as the Sahara and Antarctica. Areas excluded are those with a 30-yr (1981–2010) average annual ratio of precipitation to PET below 0.05 (arid), and with a similar 30-yr average annual PET below 365 mm (cold). These areas are not considered when showing global or macroregional percentage changes in areas affected by drought.
3. Results and discussion
a. Validation of drought projections: CORDEX data versus observed data
Before using the ensemble of CORDEX simulations to analyze drought projections, we tested their reliability versus observed data for 1981–2010 (the recent past). In the CORDEX simulations this period is a combination of the historical experiments data (until 2005) and future projections data (2006–10, driven by the RCP). The observational datasets used for validation are the Global Precipitation Climatology Centre (GPCC, version 7) of the German Meteorological Office (DWD) (Schneider et al. 2008; Becker et al. 2013), and the Climate Research Unit Time series (CRUTS, version 4.01) of the University of East Anglia (Harris et al. 2014). We obtained precipitation data from GPCC and temperature data from CRUTS. Although CRUTS includes PET, computed based on the Penman–Monteith method, we used minimum and maximum temperature data to obtain PET based on the Hargreaves–Samani equation, and in turn, SPI and SPEI, to ensure homogeneity with the CORDEX outputs.
The drought variables selected for validation were drought frequency (DF) and drought severity (DS; averaged over events) during 1981–2010. The spatial resolution of the GPCC and CRUTS gridded data (0.5°) is slightly coarser than that of the CORDEX data (0.44°), so we interpolated DF and DS derived from the observational datasets over the CORDEX grid. We selected a kriging-based interpolation method (Cressie 1990) based on weighted Gaussian distance between points and a search radius of 75 km. The resulting error is likely to be negligible, as the difference in spatial resolution is small, and other sources of bias (e.g., low number of input stations in remote regions) can be more relevant.
Results show that CORDEX ensemble median values slightly underestimate both DF and DS, generally, when compared with the observational datasets (Fig. 3). Globally, the underestimation is larger for DF (about 12% for SPI and 11% for SPEI) and smaller for DS (about 10% for SPI and 8% for SPEI). However, for both drought variables (DF and DS) and indicators (SPI and SPEI), more than 50% of the land areas show differences smaller than 5%. Locally, regions where the underestimation is largest (on average close to 15%) are visible over the central United States, northwestern Mexico, and western Canada (for DF), as well as Angola and the mountainous regions of central Asia (for DS). In contrast, the largest overestimation (about 10%) by CORDEX simulations is visible, locally, over the Democratic Republic of Congo (for DF) and Australia (for DS based on SPI).
While the discussed discrepancies do not directly affect the results shown in the following sections, the validation exercise is useful to investigate for which regions the CORDEX simulations are more or less reliable. Although the ensemble median does not show excessively large discrepancies, individual ensemble members may have larger errors. During the first phase of testing, we tested the reliability of single-model runs by applying a bootstrapping technique to the ensemble: we defined a criterion for excluding a simulation if it showed drought frequency (over 1981–2010 and based on SPEI) with an absolute difference versus the observational datasets of more than three events per 10 years, and covering more than 66% of its domain. This never occurred for the 103 simulations used in this study, although a couple did show discrepancies above the threshold for large areas of Siberia.
b. The twenty-first century: A drying or wetting warming?
Climate simulations are in agreement regarding a warming world during the twenty-first century (Meehl et al. 2007; IPCC 2014); therefore, we can expect a global increase in PET driven by temperature. An increase in evaporative demand, however, does not per se result in an intensification of drought frequency and/or severity. Only the combined effect of changes in both rainfall and PET will determine where droughts become more or less frequent and severe.
Figure 4 shows the change in the ensemble median of mean temperature (TM) and total precipitation (P) between the end of the twenty-first century and the reference period. The upper panels show that, under both climate scenarios, the overall increase in mean temperature is robust in both magnitude and sign over the vast majority of land areas. Precipitation is projected to increase or decrease depending on the region and scenario, showing larger spatial and temporal variability and, in general, a wetting or drying tendency for RCP4.5 corresponds to a stronger wetting or drying tendency for RCP8.5. In particular, the increase is robust in both sign and magnitude over northern latitudes, central Australia, and Antarctica, whereas over the eastern United States, the Horn of Africa, India, most of China, and southeastern Asia most models agree on the sign of the change, but not necessarily on its magnitude. On the contrary, a drying that is robust in sign is projected for the Mediterranean and South Africa, parts of Mexico and southern Argentina, and eastern China, especially under RCP4.5. However, over most of central and western United States, Mexico, South America, central Europe, and Africa the change in precipitation is not robust, either in sign or magnitude (e.g., see Dosio et al. 2019). However, the Amazon subdomain is a special case, once it is a large region that includes more biomes than only Amazon tropical forest. For this reason, in further studies this subdomain may be split into minor regions, to investigate whether more robust local precipitation trends could be expected.
Although the main scope of this study is to investigate changes in drought frequency and severity, and not in temperature and precipitation per se, Fig. 4 represents the first map of its kind (to our knowledge) to show global temperature and precipitation projections based on a large ensemble of RCMs at a high spatial resolution (0.44°). Temperature projections shown in Fig. 4 agree with those reported in the latest IPCC Assessment Report (IPCC 2014), based on global simulations from the phase 5 of the Coupled Model Intercomparison Project (CMIP5) (Taylor et al. 2012) and with the previous set of simulations from CMIP3 (Knutti and Sedláček 2013). As shown for example by Dosio (2017) for Africa, RCM projections for temperature largely agree with those of the driving GCMs. For precipitation, there are numerous areas with an uncertain change, in line with previously published GCM-based precipitation projections (Power et al. 2012; Knutti and Sedláček 2013; IPCC 2014), but uncertainties may be reduced by using constrained or weighted GCM ensembles, as done in Mexico and Central America (Colorado-Ruiz et al. 2018) and in the Arctic (Knutti et al. 2017). In general both GCM- and CORDEX-based results show overall similar spatial patterns, especially the drying tendency over Chile, the Mediterranean region, and southern Africa (e.g., Dosio et al. 2019).
Table 3 (which refers to the same acronyms as in Fig. 2) summarizes the projected temperature and precipitation changes between 1981–2010 and 2071–2100 at macroregional scale. Over land, the global average temperature increase by the end of the twenty-first century is estimated at 2.6°C for RCP4.5 and at 4.8°C for RCP8.5, being most extreme over the Arctic region (ARC) and least extreme over southern South America (SSA). At global scale, annual precipitation is projected to increase, on average, by approximately 8% for RCP4.5 and 5% for RCP8.5. For both climate scenarios, approximately 73% of the lands will face an increase in precipitation by the end of the twenty-first century (Table 3). The fraction of land area projected to become wetter is particularly small for two macroregions, the Mediterranean region (MED; i.e., 5% for RCP4.5 and 1.3% for RCP8.5) and southern Africa (SAF; 23% for RCP4.5 and 19.1% for RCP8.5), in agreement with the drying tendency discussed previously. Finally, southern Australia (SAU) shows the largest difference between precipitation projection depending on the underlying RCP, with the fraction of land projected to become wetter under RCP4.5 (i.e., 92.1%) greatly reducing under RCP8.5 (i.e., 34.1%).
Average mean temperature (ΔTM) and precipitation (∆P) differences between 1981–2010 and 2071–2100 under the RCP4.5 and the RCP8.5 for 28 macroregions and at global scale (only over land). The last two columns show the percentage of areas in which precipitation is projected to increase. The regions with an increase or decrease in precipitation larger than 10% are highlighted in bold.
c. Drought frequency, severity, and extreme droughts projections
Before analyzing the drought projections for the RCMs, it is interesting to briefly discuss those obtained by the driving GCMs. Figure 5 shows the changes in drought frequency (DF; events per decade) and average severity of drought events (DS; average severity per decade) between the reference period (1981–2010) and the far future (2071–2100) under the RCP4.5 and the RCP8.5. As input data, we used the ensemble median of 16 GCMs (see Table 1), regridded at medium spatial resolution (1.8°). The choice of such common resolution depends on the optimal choice between the single resolutions (from 0.75° × 0.75° to 2.5° × 2°). The same interpolation scheme used for regridding the CORDEX simulation was used.
Figure 5 shows some clear spatial patterns: the drought frequency and severity are projected to increase in much larger areas according to the SPEI than to the SPI, although they agree about the increase of both quantities (under both scenarios) over the Amazon forest, southern South America, the Mediterranean region, southern Africa, and southern Australia. According to the SPI, most of the areas at high latitudes are projected to see a decrease in both drought frequency and severity. The areas where less than two-thirds of the simulations agree on the sign of change are different according to the SPEI (medium high latitudes in Northern Hemisphere and equatorial Africa) and the SPI (central Europe, the Middle East, and parts of Brazil).
The same analyses were repeated using the 103 CORDEX simulations, improving the spatial resolution from 1.8° to 0.44°. Figure 6 shows the changes in DF: as expected, the area projected to experience more drought events in the future is much larger according to SPEI (approximately 72% for both RCPs) than with SPI (approximately 17% for RCP4.5 and 16% for RCP8.5). The corresponding values per macroregion are reported in Table 4. The two indicators agree on the projected decrease in DF over high latitudes and southeastern Asia and on the increase over the Mediterranean region, Chile and Argentina, southern Africa, and southeastern China. Areas where the change is not robust in sign show some differences. The projected change is not robust in sign over India for SPEI, while it is robust in sign for SPI. On the contrary, over the U.S. Midwest, northwestern Mexico, central Europe, and tropical Africa the projected change is robust in sign for SPEI and not for SPI. Some regions (Table 4) show opposite tendencies, in particular under the RCP8.5. Examples are eastern North America, Amazonia (under RCP8.5 only), the Horn of Africa, central Asia, and central Australia, mainly due to the nonrobust precipitation changes over those areas.
Percentage of area in which drought frequency is projected to increase (decrease) from 1981–2010 to 2071–2100 under RCP4.5 and RCP8.5 and according to SPI-12 and SPEI-12. The change is uncertain (Unc) if less than two-thirds of simulations agree on the sign of change, otherwise (ΔDF > 0 or ΔDF < 0) more than two thirds of the model agree on the change in sign.
In general, the two scenarios agree on the sign of change between the recent past and far future for each of the indicators. For SPI, the change in DF under RCP8.5 is in general larger than that under RCP4.5. However, SPEI shows a larger increase in DF under RCP4.5 compared with RCP8.5 especially for the Mediterranean region, most of central Asia, and Africa. A possible explanation for this somewhat counterintuitive result can be given by combining the information from Fig. 6 (frequency; see also Fig. 7 for validation) and Fig. 8 (severity, which is linked to drought duration). In fact, over these areas, under RCP8.5, the length of the droughts is projected to increase enormously (with some droughts lasting for several years) with the result that their frequency is reduced. This hypothesis is confirmed by analyzing the changes in drought duration (for SPEI): over 97.3% of the mentioned areas, droughts are projected to last much longer under RCP8.5 than RCP4.5 (not shown). Under the moderate scenario the droughts are projected to be more frequent than in recent past, but the increase in severity and duration will be smaller than under the more extreme scenario.
The spatial patterns of the driving GCMs are in general similar to those of the RCMs, but with the use of RCMs some different patterns are found. First, the projected increase of drought frequency is smaller (according to CORDEX) over Australia, where the SPI projects a decrease in central territories. Second, two areas (India for the SPEI and tropical Africa for the SPI) show not robust (in sign) changes according to CORDEX simulations. Third, GCMs tend to generally overestimate the increase in DF under the RCP8.5 over Northern Hemisphere compared to the RCMs. This partly depends on the effect that a coarser resolution unavoidably introduces, but the use of RCMs (which account for regional physical features) becomes very useful to distinguish between regions with a moderate, large, or very large increase as it occurs in the western United States and central Asia. In fact, in Fig. 6 the borders between areas with progressively larger changes are better defined and the use of a larger number of simulations leads to more reliable delineation of areas with robust (in sign) changes.
Figures 6 and 8 show no clear discontinuities over borders between CORDEX domains, proving that the spatial patterns of the drought projections do not depend on the different set of simulations used in different regions. However, this is valid for the ensemble medians, while some discontinuities over bordering regions occur when computing the standard deviation of changes in drought frequency (Fig. 7), in particular over the Urals and, to lesser extent, over eastern China. Using the best sample [i.e., the eight simulations (eight is the minimum number of simulations over any area); see Fig. 1] with the smallest spread over each CORDEX domain, such discontinuities disappear.
Given the large number of simulations employed, one could expect a larger intersimulation spread (Fig. 7), but for DF it is smaller than 0.5 events per decade over most land areas. The areas with largest spread (and consequently less robust outputs in terms of magnitude) for the SPEI are eastern Canada, the Baltic republics, central Russia, India, and—for RCP8.5 only—the Horn of Africa and southeastern Asia. For the SPI, changes in DF under the RCP4.5 show no particular areas with large spread, whereas under RCP8.5 the spread is relevant for equatorial and tropical latitudes (i.e., areas with larger annual precipitation totals). However, for an ensemble of climate models, the geographical distribution of the uncertainties represented by model spread at the gridpoint scale could overestimate the projected range, leading to physically implausible patterns of change on global and regional scales, as climate change impacts will never be realized as the worst (or best) case everywhere (Madsen et al. 2017).
The changes in DS (see Fig. 8) are larger in percentage than those for DF. For SPEI, only latitudes higher than 55°N and southeastern Asia will face a decrease in DS. For SPI, on the other hand, southern Chile and Argentina, the Mediterranean region, large parts of southern Africa, and (under RCP8.5 only) southeastern China and southwestern Australia are projected to face an increase in DS. Moreover, the regions where the change is not robust in sign are larger for SPI than for SPEI, and consequently the areas with opposite robust tendencies in sign (i.e., increase for SPEI and decrease for SPI) are limited to central Asia and central Australia. The intersimulation spread for DS is spatially similar to that for DF (see Fig. 7), and thus we do not show the corresponding maps.
For DS, the overall spatial patterns using GCMs only or the combinations of GCMs and RCMs are almost identical according to the SPEI, although under the RCP4.5 the use of RCMs makes a notable difference over western United States (Figs. 5 and 6). The projections of DS according to the SPI show remarkable differences under RCP8.5 over South America, where the use of RCMs turns positive changes into not robust in sign or even slightly negative changes in tropical South America, in agreement with the analyses of Llopart et al. (2014) and Sánchez et al. (2015), who showed that the downscaling RCMs can project a positive precipitation signal even though the driving GCMs show little or negative change. Moreover, under the RCP4.5, the decrease in DS at high latitudes in the Northern Hemisphere is larger according to the use of RCMs than to GCMs only. In general, the use of RCMs helps providing a better representation of areas with robust (in sign) changes according to the SPI in both hemispheres.
Based on Figs. 6 and 8 we can highlight regions where drought events are projected to be both more frequent and severe (Fig. 9). According to SPEI, most of the regions show a large fraction of area falling in this “worst case” situation, while for SPI this happens for only a few regions, in particular the Mediterranean region and southern Africa. Over the Tibetan Plateau, the two indicators completely diverge, although results for this region may be largely influenced by its complex orography. In addition, this region contains the smallest number of grid points used for the analysis, due to the masking of very cold high-elevation areas. At a global scale, for SPI, the regions with a projected decrease of both drought variables (DF and DS) are clearly the majority; on the other hand, the net difference between regions with an increase and those with a decrease of both DF and DS is positive for SPEI.
One of the main consequences of climate change is that record-breaking (i.e., never previously recorded) extreme events are expected to happen, such as the 2010 Russian drought and heatwave (Trenberth and Fasullo 2012; Dosio et al. 2018). To estimate this possible evolution in the twenty-first century, we calculated how many events in 2071–2100 are projected to be more severe than the most severe ones that occurred in 1981–2010 (PK; Fig. 10). For SPI, under both scenarios, this will occur over approximately 33% of the unmasked lands. Moreover, for SPI under RCP8.5, only western South America, the Mediterranean region, and the Mediterranean-like southwestern parts of southern Africa will experience three or more droughts never recorded in 1981–2010 (see Table 5). For SPEI, such extreme droughts not recorded in 1981–2010 will involve approximately 75% of the unmasked lands under both scenarios and, under RCP8.5, 40% will face at least three such unrecorded events. According to Table 5, only a few regions will be hit by unrecorded events over less than 80% of their area, under both scenarios: high-latitude areas (Alaska, Canada–Greenland–Iceland, northern Europe, and northern Asia) and Southeast Asia.