Assessment of Precipitation and Hydrological Droughts in South America through Statistically Downscaled CMIP6 Pro-jections

1. Introduction

Climate change is undeniable, as is its attribution to anthropogenic greenhouse gas-emitting activities, which have unequivocally intensified global warming, evidenced by the 1.1 °C increase in global surface temperature from 2011-2020 relative to 1850-1900 [1]. Additionally, best estimates indicate that continued greenhouse gas emissions will cause a 1.5 °C increase in the near term (2021-2040) [1]. Moreover, even if the countries fully implement the commitments made in the Paris Agreement, global warming is expected to exceed 2 °C by the end of the century [2,3]. The progressive warming will intensify the global hydrological cycle so that compound heatwaves and droughts are projected to become more frequent, including concurrent events in multiple locations [1].

Droughts are a natural phenomenon characterized by a continuous persistence of precipitation deficit, which occurs in almost all climate zones, including those with high precipitation rates, such as the Amazonia [4]. Prolonged periods of drought cause innumerable damages, such as losses in agriculture and livestock, contamination of waters, reduction of water availability for daily consumption and water energy generation, and environmental risks with fires [5,6]. Due to climate change and increasing temperatures, droughts are expected to increase in frequency and duration [7,8].

Studies with climate projections indicate an increased risk of extreme droughts in different parts of the world, depending on the seasons and drought indicators analyzed. For example, considering North America, studies with climate projections of phases 5 and 6 of the Coupled Model Intercomparison Project (CMIP) show an increase in the duration and frequency of droughts in the southeast [9], central and northern [10,11,12,13,14] United States, and southwestern Canada [15]. Similarly, drought events are projected to intensify in parts of Asia [11,12,13,14,16,17], Africa [11,14,19], Europe [11,18], Oceania [11,13,14,18], and South America [11,13,14,18,20,21,22,23,24,25].

In South America (SA), climate change’s effects on droughts are evident in different sectors, such as Northeastern and Southeastern Brazil, Amazonia, and the continent’s southeast. The semi-arid Northeast of Brazil (NEB) is one of the historically most vulnerable regions to droughts [26], with several events recorded since the 16th century [26,27], and numerous socio-economic impacts such as damage to agricultural production, livestock, loss of human life from hunger, malnutrition, and disease, migrations to urban centers, and losses in regional and national economies [27]. The recent drought that began in 2012 and extended partially into 2018 (considered the most severe in recent decades) prompted public damage management policies that lessened the drought’s impacts but were insufficient to contain its effects on nearly nine million people [22,27,28]. Towards the end of the 21st century, CMIP5 climate projections under the RCP-8.5 scenario indicate an increase of up to 54% in the proportion of areas under water stress conditions in the NEB, considering warming above 4 °C [27].

Southeastern Brazil (SEB) has also experienced some of the worst droughts in recent decades, such as in 2001 [29], 2014/2015 [30,31,32,33,34], 2018 [6], and most recently, in 2020/2021 [35]. Although in the 2014/2015 event, Brazil already had an interconnected electric system, which avoided an electric crisis in the country (unlike the 2001 event), this drought presented the largest rainfall deficit since 1960 in the metropolitan region of São Paulo (one of the most populated regions of Brazil), causing a water crisis with reservoirs at extremely low levels [31,32,36]. Considering the climatology of droughts in the region, observational data from 1901-2010 indicated for the western and central (eastern) sectors of the state of São Paulo a tendency of increase (reduction) of severe drought events [5].

Several drought events have also occurred in the Amazonia (AMZ), accentuating forest fires, affecting the region’s biota, and signaling the risk of a tipping point [37,38,39,40]. Furthermore, future scenarios suggest reduced rainfall and intensified droughts in the forest [23,24,37,41,42], which may strengthen the process of savannization of the biome [37,38,43]. Additionally, in the last decade, other Brazilian regions have also experienced severe droughts, such as the 2019/2020 droughts in the Brazilian Pantanal and Midwest [44,45,46] and the 2012/2013 and 2019/2020 droughts in southern Brazil (SB) [47,48].

Another South American region affected by droughts is Southeastern South America (SESA), which covers northern Argentina, Paraguay, Uruguay, and SB. The 2008/2009 drought in the region was among the most severe in the last 50 years [49]. Analyses of precipitation and soil moisture anomalies from 1979-2018 have shown that droughts tend to last up to three months in the sector, resulting in significant losses in agricultural production [50]. Additionally, several studies have shown an increasing trend of warm days/nights [51,52,53,54] and the occurrence of dry spells [54,55] in the region in the last decades.

Droughts are a complex and multiform phenomenon [35], and different quantitative indicators allow their assessment [56]. The Standardized Precipitation Index (SPI), developed by McKee et al. [57], quantifies the rainfall deficit or excess on different time scales. SPI on time scales greater than six months is employed to identify and characterize hydrological droughts, which cause reduced soil moisture levels, river flows, groundwater recharge, and reservoir levels [56,58,59]. The SPI index proves advantageous because besides allowing the evaluation of drought impacts on different hydrological cycle components (using different time scales), it requires only rainfall data as input variables in the index computation [58]. However, as SPI does not account for the temperature component, its analysis disregards evapotranspiration processes, which play an essential role in the hydrological cycle [56,58]. Despite its limitations, several studies have employed the SPI-12 index (SPI index on a 12-month time scale) due to its simplicity of implementation to identify hydrological droughts in SA [6,56,58,60,61,62,63,64,65,66,67].

In this context of climate research, global climate models (GCMs) are a primary tool for investigating climate system elements [68]. A new generation of GCMs from CMIP has recently been available to the scientific community, comprising the sixth phase (CMIP6) of the project’s experiment. The CMIP6-GCMs present aspects of improvement over previous generations, such as higher spatial resolution and better parameterization schemes of the physical and biogeochemical processes of the climate system [69]. In addition, CMIP6 models employ the Scenario Model Intercomparison Project (ScenarioMIP), which provides climate projections based on the latest greenhouse gas emission and land use scenarios, the Shared Socio-economic Pathways (SSPs) [70]. SSPs characterize a more realistic socio-economic development by considering different social, economic, technological, and political scenarios [71].

Despite their crucial role in climate research, GCMs have limitations due to their coarse spatial resolution, which restricts their use in analyzing regional-scale processes and impacts. In this framework, dynamical and statistical downscaling techniques address these limitations of global models. Dynamical downscaling employs regional climate models (RCMs) that use initial and boundary conditions provided by GCMs, while statistical downscaling determines relationships between large-scale atmospheric circulation factors and local climate [72]. Dynamical downscaling is helpful for the analysis of local-scale climate phenomena, but such a procedure requires high computational costs [73]. On the other hand, statistical downscaling requires less computational effort and avoids the propagation of systematic errors arising from GCMs [74].

Regarding SA, one observes a literature more focused on the dynamical approach of air temperature, precipitation, and wind [75,76,77,78,79,80,81,82,83,84,85]. However, recent studies have also applied the statistical method and demonstrated its competence in representing the spatial distribution and extreme events of temperature and precipitation [74,83,86,87,88,89,90,91]. Research with statistical downscaling of precipitation projections from CMIP6 to SA showed that the Quantile Delta Mapping (QDM) bias correction technique, developed by Cannon et al. [92], performed well in correcting the systematic errors in the different quantiles of the probability distributions of the GCMs raw simulations, evidencing its ability in reproducing seasonal variability and extreme properties [91]. With the QDM technique, it is possible to use data as reanalysis to interpolate the historical projections to the reference dataset’s spatial resolution and, from transfer functions, apply the same correction to future projections. The QDM method proved to be advantageous since it preserves the model-projected relative changes and trends (i.e., if a model has a dry trend in a specific region, it will be kept after the spatial disaggregation and bias correction) and corrects the systematic biases in the quantiles of the modeled data with respect to the observations [92].

Given the background, this study aims to: (a) apply the QDM bias correction technique and statistical downscaling to historical simulations and climate projections of precipitation from a CMIP6 multi-model ensemble; (b) employ the bias-corrected estimates to present a set of precipitation projections at intermediate resolution (50 km) in SA; (c) employ the bias-corrected projections and the SPI-12 Index to identify the four types of hydrological drought events (mild, moderate, severe, and extreme), and their aspects (frequency, duration, severity, intensity, and peak) in eight subdomains of SA. There is still a shortage of studies analyzing hydrological droughts in SA with post-processed projections from the CMIP6-GCMs, and this study intends to fill such a gap. In addition, this research can assist decision-makers and energy planners in better future management of water resources on the continent.

2. Materials and Methods

2.1. Study Area

The study area comprises the SA continent (Figure 1), located at latitudes 12°N-55°S. Its extensive latitudinal coverage provides a climate heterogeneity of tropical, subtropical, and extratropical regions, as well as a diverse geography that includes particular areas such as the Andes Mountains, the Atacama Desert, the Amazon Rainforest, and the semi-arid Northeast of Brazil [93,94]. The South American Monsoon System (SAMS) primarily influences the central SA, with two well-defined seasons marked by the rainy season from November to March and the dry season from May to September [93,94,95]. On the other hand, the subtropical western portion of the continent concentrates its rainfall in the austral winter months due to the passage of cold fronts and cut-off lows [93,94]. Still, the northern portion of SA does not have a well-defined dry season, being strongly influenced by the Intertropical Convergence Zone (ITCZ), with maximum rainfall in the austral autumn and early winter [96], and presenting one of the wettest places in the globe [97].

Table 2. Geographic coordinates of the subdomains selected for hydrological drought analysis.

Subdomain	Area
1	5°N–5°S 68°W–74°W
2	2.5°S–10°S 53°W–63°W
3	4.5°S–11°S 36°W–45°W
4	11.5°S–19.5°S 40°W–47°W
5	11.5°S–19.5°S 48°W–57°W
6	20°S–24.5°S 41°W–53°W
7	25°S–35°S 48.5°W–58°W
8	30°S–40°S 65°W–73°W

Table 2. Geographic coordinates of the subdomains selected for hydrological drought analysis.

Subdomain	Area
1	5°N–5°S 68°W–74°W
2	2.5°S–10°S 53°W–63°W
3	4.5°S–11°S 36°W–45°W
4	11.5°S–19.5°S 40°W–47°W
5	11.5°S–19.5°S 48°W–57°W
6	20°S–24.5°S 41°W–53°W
7	25°S–35°S 48.5°W–58°W
8	30°S–40°S 65°W–73°W

The rectangles illustrated in Figure 1 indicate the subdomains selected to analyze hydrological drought events on the continent. We considered subdomain 8 (Patagonia) because, in this region, the Andes have a lower height, while for the other parts of the Cordillera, we did not perform analysis due to the reference data also having uncertainties [98].

2.2. CMIP6-GCMs Selection

The study employed precipitation projections from eight CMIP6-GCMs, comprehending the historical period (1995-2014) and two greenhouse gas emission scenarios (SSP2-4.5 and SSP5-8.5) in the future period (2020-2099). The SSP2-4.5 scenario denotes a moderate emission scenario, while SSP5-8.5 considers a high greenhouse gas emission context, representing a period with little effort to mitigate climate change effects [70]. The GCMs dataset comprises precipitation projections every three hours provided on the Earth System Grid Federation (ESGF) platform (available at https://esgfnode.llnl.gov/search/cmip6).

At the early stage of this study (January 2022), we selected the models that best represented the South American climate in terms of precipitation and air temperature, which is a response to atmospheric circulation. To select the GCMs, we used the methodology of Rupp et al. [99], whereby several metrics evaluate the best models based on regionally averaged properties and large-scale patterns. Thus, the identification of the best-performing models included the calculation of the following parameters with monthly data from 50 CMIP6-GCMs for different SA subdomains (figure not shown): (a) mean and standard deviation for each year (1995-2014); (b) spatial correlation calculated for each season (DJF, MAM, JJA, SON) and year (1995-2014) with Pearson’s correlation coefficient; (c) mean amplitude, defined as the difference of the variables between January and July; (d) linear trend, calculated for complete time series (rather than by seasons) using the method of least squares and angular coefficients for the indication of a positive or negative trend.

Ranking the GCMs according to their performance is not trivial, as several statistical metrics and seasonal seasons are evaluated. Therefore, we compiled all the information by standardizing the metrics (giving equal weight/importance to each metric) to rank the GCMs in terms of performance according to the methodology proposed by Rupp et al. [99]. For each model $\mathrm{i}$ and metric $\mathrm{j}$, the bias ${\mathrm{E}}_{\mathrm{i},\mathrm{j}}$ is calculated as:

$${\mathrm{E}}_{\mathrm{i},\mathrm{j}}=\left|{\mathrm{x}}_{\mathrm{obs},\mathrm{j}}-{\mathrm{x}}_{\mathrm{i},\mathrm{j}}\right|,$$

(1)

where ${\mathrm{E}}_{\mathrm{i},\mathrm{j}}$ is the absolute error (absolute bias value), and ${\mathrm{x}}_{\mathrm{obs},\mathrm{j}}$ and ${\mathrm{x}}_{\mathrm{i},\mathrm{j}}$ are the observed and simulated metrics, respectively. The next step was to calculate the relative error ${\mathrm{E}}_{\mathrm{i},\mathrm{j}}^{*}$ (which can be interpreted as a standardized time series) by Equation 2:

$${\mathrm{E}}_{\mathrm{i},\mathrm{j}}^{*}=\frac{{\mathrm{E}}_{\mathrm{i},\mathrm{j}}-\min \left({\mathrm{E}}_{\mathrm{i},\mathrm{j}}\right)}{\max \left({\mathrm{E}}_{\mathrm{i},\mathrm{j}}\right)-\min \left({\mathrm{E}}_{\mathrm{i},\mathrm{j}}\right)},$$

(2)

where $\max \left({\mathrm{E}}_{\mathrm{i},\mathrm{j}}\right)$ and $\min \left({\mathrm{E}}_{\mathrm{i},\mathrm{j}}\right)$ are the functions used to select a time series’ maximum and minimum values, respectively. If the metric is a correlation, each function $\max \left({\mathrm{E}}_{\mathrm{i},\mathrm{j}}\right)$ or $\min \left({\mathrm{E}}_{\mathrm{i},\mathrm{j}}\right)$ is reversed. In this metric, since the absolute error ${\mathrm{E}}_{\mathrm{i},\mathrm{j}}$ is divided by the amplitude of the error, ${\mathrm{E}}_{\mathrm{i},\mathrm{j}}^{*}=0$ indicates that the model has perfect performance and ${\mathrm{E}}_{\mathrm{i},\mathrm{j}}^{*}=1$ denotes poor performance. According to Rupp et al. [99], the relative error is then summed over all statistical metrics $\mathrm{m}$ of a model, which provides the total relative error ${\mathrm{E}}_{\mathrm{i},\mathrm{tot}}^{*}$:

$${\mathrm{E}}_{\mathrm{i},\mathrm{tot}}^{*}={\displaystyle \sum }_{\mathrm{j}=1}^{\mathrm{m}}{\mathrm{E}}_{\mathrm{i},\mathrm{j}}^{*},\left(3\right)$$

(3)

The final step was to rank the models according to their relative error on a scale from 0 to 1, where each ${\mathrm{E}}_{\mathrm{i},\mathrm{j}}^{*}$ is divided by the maximum value of ${\mathrm{E}}_{\mathrm{i},\mathrm{j}}^{*}$. Figure 2 illustrates that the best models are on the left-hand side (values closest to zero). We could not necessarily select the best models shown in Figure 2 due to the absence of hourly/daily data and/or projections in the ESGF database. Thus, by concurrently analyzing the availability of high-frequency data and projections, the best models (indicated with red bars) were selected for this study (Table 2). Furthermore, three selected GCMs (EC-Earth3, IPSL-CM6A-LR, and MPI-ESM1-2-LR) were previously validated and performed well in representing the SA climate [100]. Although a good simulation of the historical period does not determine more accurate climate projections for the same model, it ensures more reliable future estimates since the poor ability to simulate historical climate is likely reflected in poorer future projections [101].

Table 2. Information on each CMIP6-GCM employed in the study.

Model	Resolution (°Lat × °Lon)	Institute	Reference
CMCC-CM2-SR5	1.25 × 0.94	Fondazione Centro Euro-Mediterraneo sui Cambiamenti Climatici	Lovato and Peano [102]
CMCC-ESM2	1.25 × 0.94	Fondazione Centro Euro-Mediterraneo sui Cambiamenti Climatici	Lovato et al. [103]
EC-Earth3	0.70 × 0.70	EC-Earth Consortium	Döscher et al. [104]
GFDL-ESM4	1.25 × 1.00	Geophysical Fluid Dynamics Laboratory	Krasting et al. [105]
IPSL-CM6A-LR	2.50 × 1.26	Institut Pierre Simon Laplace	Boucher et al. [106]
MIROC6	1.41 × 1.41	Japan Agency for Marine-Earth Science and Technology	Tatebe and Watanabe [107]
MPI-ESM1-2-LR	0.94 × 0.94	Max Planck Institute for Meteorology	Wieners et al. [108]
MRI-ESM2-0	1.13 × 1.13	Meteorological Research Institute	Yukimoto et al. [109]

Table 2. Information on each CMIP6-GCM employed in the study.

Model	Resolution (°Lat × °Lon)	Institute	Reference
CMCC-CM2-SR5	1.25 × 0.94	Fondazione Centro Euro-Mediterraneo sui Cambiamenti Climatici	Lovato and Peano [102]
CMCC-ESM2	1.25 × 0.94	Fondazione Centro Euro-Mediterraneo sui Cambiamenti Climatici	Lovato et al. [103]
EC-Earth3	0.70 × 0.70	EC-Earth Consortium	Döscher et al. [104]
GFDL-ESM4	1.25 × 1.00	Geophysical Fluid Dynamics Laboratory	Krasting et al. [105]
IPSL-CM6A-LR	2.50 × 1.26	Institut Pierre Simon Laplace	Boucher et al. [106]
MIROC6	1.41 × 1.41	Japan Agency for Marine-Earth Science and Technology	Tatebe and Watanabe [107]
MPI-ESM1-2-LR	0.94 × 0.94	Max Planck Institute for Meteorology	Wieners et al. [108]
MRI-ESM2-0	1.13 × 1.13	Meteorological Research Institute	Yukimoto et al. [109]

2.3. Reference Dataset

The study used precipitation analysis from the Climate Prediction Center Gauge-Based Analysis of Global Daily Precipitation (CPC) [110] to validate the historical simulations of the CMIP6-GCMs. For this, daily data from 1995 to 2014 were used, with 0.5° horizontal resolution (available at https://ftp.cpc.ncep.noaa.gov/precip/CPC_UNI_PRCP/GAUGE_GLB/RT/). CPC data proved adept at representing average and seasonal precipitation patterns over most of SA [94] but present uncertainties in regions of complex topography such as the Andes Mountains [98].

We point out that the information in the CMIP6 tutorials does not clarify the period of daily precipitation accumulation. Precipitation data were obtained with a frequency of three hours, and the daily accumulation followed the recommendations of the World Meteorological Organization (WMO) [111] to avoid errors like comparing data with different periods for the daily accumulation. That is, the rainfall for a given day is accumulated from 1200 Z of the previous day to 1200 Z of the day in question. However, it is worth noting that the CMIP6 models provide the accumulations at 0130, 0430 Z until completing 24 hours. Thus, the accumulation was performed from 1330 Z to 1030 Z of the following day to be as close as possible to the WMO definition. Furthermore, CPC precipitation analysis also accumulates the daily rainfall in the 1200 to 1200 Z interval.

2.4. Bias Correction and Statistical Downscaling

One way to overcome the limitations imposed by the coarse resolution of GCMs is through statistical downscaling methods, which establish statistical relationships between model outputs and reference data [112]. Statistical downscaling techniques are classified into three types: transfer function or regression models, weather generators, and weather typing [110], and this study used transfer functions. This method was chosen due to its simplicity of implementation and for preserving time series trends (for more details, see Cannon et al. [92]). This methodology is also known as Bias Correction-Statistical Downscaling (BCSD).

This study used the BCSD method to downscale the simulations and projections of CMIP6-GCMs. Bias correction was performed with the QDM technique [92] to the historical simulations (1995-2014), and the transfer functions were applied to the future projections (2020-2099). According to Cannon et al. [92], the QDM technique preserves the model-projected trends and relative changes (e.g., if the GCM shows a dry trend in a given region, that trend will be maintained after bias correction) and corrects systematic biases in the quantiles of the modeled data relative to the reference one. Moreover, compared to the quantile mapping technique, the QDM technique proved advantageous because it is less susceptible to problems such as inflating relative trends in extreme values [92].

Before bias correction, spatial disaggregation was applied to the CMIP6-GCMs outputs, and the model data were downscaled to the 0.5° resolution as the CPC data with bilinear interpolation. Several studies show that bilinear interpolation provides consistent estimates by adjusting the spatially correlated behavior of the variable [91,113,114,115,116,117,118,119,120].

After spatial disaggregation, bias correction was applied with the QDM method. This method follows three steps [92]: first, the trend is removed from all projected individual quantiles. Next, the detrended quantiles are bias-corrected with the quantile mapping technique. Finally, the projected changes are superimposed on the bias-corrected outputs. Let $\mathrm{o}$ and $\mathrm{p}$ be the observed and projected data, and $\mathrm{h}$ and $\mathrm{f}$ are the historical and future periods, respectively. The definition of the non-exceedance probability of the observed (${\mathrm{x}}_{\mathrm{h},\mathrm{o}})$ and projected (${\mathrm{x}}_{\mathrm{h},\mathrm{p}})$ historical and future (${\mathrm{x}}_{\mathrm{f},\mathrm{p}})$ data is accounted for as:

$$\begin{gathered} {\mathrm{p}}_{\mathrm{f},\mathrm{p}}\left(\mathrm{t}\right)=\mathrm{F}\left({\mathrm{x}}_{\mathrm{f},\mathrm{p}}\left(\mathrm{t}\right)\right) \\ {\mathrm{p}}_{\mathrm{h},\mathrm{p}}\left(\mathrm{t}\right)=\mathrm{F}\left({\mathrm{x}}_{\mathrm{h},\mathrm{p}}\left(\mathrm{t}\right)\right) \\ {\mathrm{p}}_{\mathrm{h},\mathrm{o}}\left(\mathrm{t}\right)=\mathrm{F}\left({\mathrm{x}}_{\mathrm{h},\mathrm{o}}\left(\mathrm{t}\right)\right) \end{gathered}$$

(4)

where $\mathrm{p}$ and $\mathrm{F}$ denote the non-exceedance probability associated with a specific value in time and the cumulative distribution function (CDF), respectively. The change factor, which associates the historical simulation outputs with those of the future period, was calculated with Equation 5:

$${\Delta }^{\mathrm{M}}\left(\mathrm{t}\right)=\frac{{\mathrm{F}}_{\mathrm{f},\mathrm{p}}^{-1}\left({\mathrm{p}}_{\mathrm{f},\mathrm{p}}\left(\mathrm{t}\right)\right)}{{\mathrm{F}}_{\mathrm{h},\mathrm{p}}^{-1}\left({\mathrm{p}}_{\mathrm{f},\mathrm{p}}\left(\mathrm{t}\right)\right)}=\frac{{\mathrm{x}}_{\mathrm{f},\mathrm{p}}\left(\mathrm{t}\right)}{{\mathrm{F}}_{\mathrm{h},\mathrm{p}}^{-1}\left({\mathrm{p}}_{\mathrm{f},\mathrm{p}}\left(\mathrm{t}\right)\right)},$$

(5)

where ${\mathrm{F}}^{-1}$ denotes the inverse CDF and ${\Delta }^{\mathrm{M}}\left(\mathrm{t}\right)$ is the multiplicative factor of change between the simulated quantiles of the historical and future periods. Finally, the bias correction in the future projections was obtained by applying the multiplicative relative change ${\Delta }^{\mathrm{M}}\left(\mathrm{t}\right)$ to the historical values with the corrected bias, according to Equation 6:

$${\hat{\mathrm{x}}}_{\mathrm{f},\mathrm{p}}\left(\mathrm{t}\right)={\Delta }^{\mathrm{M}}\left(\mathrm{t}\right)·{\mathrm{F}}_{\mathrm{h},\mathrm{o}}^{-1}\left({\mathrm{p}}_{\mathrm{f},\mathrm{p}}\left(\mathrm{t}\right)\right),$$

(6)

The historical period (1995-2014) was used for the training set to adjust the future projections (2020-2099) with the QDM algorithm. This time window was chosen due to computational resources and to follow the same reference period used by the International Panel on Climate Change (IPCC). The Python-based package xclim [121] was used to perform the calculations.

2.5. Test of Statistical Significance for the Difference in Climatological Mean Values

To assess whether the differences in mean climatological values in the future period (2020-2099) of the CMIP6-GCMs are statistically significant compared to the historical period (1995-2014), we used the Student’s t-test. This test assumes the null hypothesis (H₀) of no difference between the two datasets against the alternative hypothesis of a difference between the two ensembles. The test was computed according to Equation 7:

$$\mathrm{t}=\frac{{\overline{\mathrm{X}}}_{\mathrm{f}}-{\overline{\mathrm{X}}}_{\mathrm{h}}}{\sqrt{\frac{{\mathrm{s}}_{\mathrm{f}}^2}{\mathrm{n}}+\frac{{\mathrm{s}}_{\mathrm{h}}^2}{\mathrm{n}}}},$$

(7)

where ${\mathrm{s}}_{\mathrm{f}}$ and ${\mathrm{s}}_{\mathrm{h}}$ are the standard deviation values of the future and historical datasets, respectively, and $\mathrm{n}$ comprises the number of values in each set. The associated degree of freedom $\mathrm{v}$ was estimated as:

$$\mathrm{v}=\frac{{\left(\frac{{\mathrm{s}}_{\mathrm{f}}^2}{\mathrm{n}}+\frac{{\mathrm{s}}_{\mathrm{h}}^2}{\mathrm{n}}\right)}^2}{\frac{{\mathrm{s}}_{\mathrm{f}}^4}{{\mathrm{n}}^2\left(\mathrm{n}-1\right)}+\frac{{\mathrm{s}}_{\mathrm{h}}^4}{{\mathrm{n}}^2\left(\mathrm{n}-1\right)}},$$

(8)

The test was performed using a significance level α of 5%. Thus, when the probability value (p-value) found was less than 5%, the null hypothesis of no difference between the two sets was rejected in favor of the alternative hypothesis, indicating statistical evidence of the difference between the mean values of the two periods evaluated.

2.6. Standardized Precipitation Index (SPI)

The SPI index, developed by McKee et al. [57], quantifies the rainfall deficit or excess on different time scales, evaluating the intensity of dry and wet periods. The SPI-12 index uses a time scale of 12 months. Besides identifying long-term rainfall patterns, SPI-12 can also be associated with flows, reservoir levels, and groundwater anomalies, helping to evaluate hydrological droughts [56,58].

According to Santos et al. [61], the first step in calculating the SPI is determining a probability density function that describes the time series. For this, the gamma function is a good fit for continuous variables with a lower limit equal to zero and no upper limit, widely used for studying historical precipitation series [61,122]. The gamma function is given by Equation 9:

$$\mathrm{g}\left(\mathrm{x}\right)=\frac{{\mathrm{X}}^{\mathsf{\alpha }-1}·{\mathrm{e}}^{\frac{-\mathrm{x}}{\mathsf{\beta }}}}{{\mathsf{\beta }}^{\mathsf{\alpha }}Г\left(\mathsf{\alpha }\right)}, \mathrm{para} \mathrm{X}>0,$$

(9)

where $\mathsf{\alpha }>0$ is the shape parameter, $\mathsf{\beta }>0$ is the scale parameter, $\mathrm{x}>0$ is the precipitation amount (mm) and $Г\left(\mathsf{\alpha }\right)$ is the full gamma function (for more details, see Wilks [122]). To estimate $\mathsf{\alpha }$ and $\mathsf{\beta }$ parameters of the gamma distribution, equations 10 and 12 were used, respectively:

$$\mathsf{\alpha }=\frac{1}{4\mathrm{A}}\left(1+\sqrt{1+\frac{4\mathrm{A}}{3}}\right),$$

(10)

$$\mathrm{A}=\ln \left(\overline{\mathrm{x}}\right)-\frac{\sum \ln \left(\mathrm{x}\right)}{\mathrm{n}},$$

(11)

$$\mathsf{\beta }=\frac{\overline{\mathrm{x}}}{\mathsf{\alpha }},$$

(12)

where $\overline{\mathrm{x}}$ is the mean rainfall, and $\mathrm{n}$ is the number of observations.

Thus, the cumulative distribution is then transformed into a normal probability distribution with a mean equal to zero and a standard deviation equal to one. Next, the cumulative probability of occurrence of each monthly value is estimated. This probability is applied to the inverse normal function to find the SPI value. The SPI is the difference of the observed rainfall minus the mean of the specific time interval, divided by the standard deviation, according to Equation 13:

$$\mathrm{SPI}={\mathrm{Z}}_{\mathrm{i}}=\frac{\left({\mathrm{P}}_{\mathrm{i}}-{\overline{\mathrm{P}}}_{\mathrm{i}}\right)}{{\mathsf{\sigma }}_{\mathrm{i}}},$$

(13)

where ${\mathrm{P}}_{\mathrm{i}}$ is the observed precipitation, ${\overline{\mathrm{P}}}_{\mathrm{i}}$and ${\mathsf{\sigma }}_{\mathrm{i}}$ are the mean and standard deviation of the fitted series, respectively.

As our results will show that the SSP5-8.5 scenario presents the most significant changes, the SPI-12 analysis was performed only under this scenario. Each bias-corrected CMIP6-GCM calculated SPI-12, and the BSCD ensemble (the bias-corrected CMIP6 ensemble) index was obtained by averaging the indices estimated by the individual models. Drought events were classified based on SPI values defined by McKee et al. [57]. Drought starts when the SPI falls below zero and ends when it becomes positive [57]. In this work, we used the thresholds defined by McKee et al. [57] (Table 3) to select and analyze all the drought events from mild to extreme categories. In addition, drought events were analyzed based on five characteristics: frequency (number of drought events in a period), duration (number of months between the first and last month of the event), severity (absolute sum of all SPI values during the event), intensity (ratio between severity and duration), and peak (largest absolute value of SPI recorded during the event).

3. Results and Discussion

3.1. Historical Simulations

The historical simulations (1995-2014) of precipitation obtained by the ensemble of eight CMIP6-GCMs before (raw ensemble) and after applying statistical downscaling (BCSD ensemble) are presented in Figure 3. Considering the austral summer (DJF), the raw ensemble tends to overestimate precipitation over most of Brazil and the west coast of SA (Figure 3a5). Contrarily, underestimates occur in northwest SA and north-central Argentina. In addition, the overestimation of rainfall over the Andes is notable. During summer, precipitation patterns exhibit a northwest-southeast orientation over the continent due to the action of the South Atlantic Convergence Zone (SACZ) [94]. On average, the raw ensemble represents the continental distribution of rainfall associated with the SACZ, but it amplifies and shifts the core of maximum precipitation to the southeast and northeast of SA.

The underestimation of precipitation during summer over northwestern SA and northern Brazil is also seen in other studies with CMIP6 models [123,124,125], as well as by CMIP5 models [126], associated with a less satisfactory representation of the Intertropical Convergence Zone (ITCZ), arising from the models’ oversensitivity to sea surface temperature (SST) and deficiency in simulating surface wind convergence. Although the CMIP6 models show considerable improvement in reproducing rainfall magnitudes over SA relative to the CMIP5 models, the simulation of ITCZ position and intensity is still deficient, which partially justifies the negative rainfall biases over northern Brazil and northern SA [125]. The systematic underestimation of rainfall in the Amazon Basin is due to an insufficient representation of different processes such as cumulus convection, biosphere-atmosphere interactions in the forest, soil moisture, and surface processes, as well as a low coverage of rainfall stations in the region, which influences the analysis of the magnitude and location of precipitation [127].

In addition, GCMs tend to produce overly intense precipitation over the central Andes in Bolivia, Peru, Ecuador, and southwestern Colombia due to excessive modeled convection and lack of topographic representativeness. In addition, validating the simulations in these areas includes many uncertainties due to the scarcity of rainfall stations in mountainous regions [123,128]. Historical simulations of the CMIP6 ensemble without bias correction indicate better performance in reproducing precipitation patterns in SA during winter and spring, reiterating previous results [125].

Considering the BCSD ensemble, one notices a significant reduction of biases across the continent, especially on the west coast of SA and northeastern Brazil. Despite a better representation of the intensity and location of rainfall maxima associated with the SACZ, the ensemble still overestimates precipitation at the center of the continental SACZ, which is mainly controlled by internal climate variability and has low or negligible predictability associated with SST variations [129,130].

Similarly, during austral autumn (Figure 3b4), there is a marked reduction in the ensemble systematic biases, and the errors concentrate in northern SA, portions of north and northeastern Brazil, northeastern Peru, central Brazil, and western Chaco. Considering the rainfall biases north of 10°S obtained by the raw ensemble (Figure 3b3), BCSD adjusts the spatial distribution of rainfall, providing a simulated field analogous to the observed one, although with the persistence of larger overestimates in the far north of Peru and Brazil (Figure 3b2).

In the winter and spring seasons, the reducing raw ensemble’s systematic errors in most of SA is notable, mainly on the continent’s west coast and portions of Colombia and Venezuela (Figure 3c4,3d4). In winter, rainfall overestimates concentrate on the north of the equator, partially justified by the less satisfactory representation of the ITCZ by GCMs, while in spring, the positive precipitation bias in western Amazonas persists even after correction. In summary, we conclude that BCSD efficiently reduces the systematic errors of GCMs and ensures more reliable projections about future climate conditions. In general, the biases that persist after applying the correction occur in problematic sectors for global climate modeling, such as the tropical region and continental portion of the SACZ.

3.2. BCSD Ensemble Projections of Precipitation under the SSP2-4.5 and SSP5-8.5 Forcing Scenarios

Figure 4 and Figure 5 present the precipitation climate projections obtained by the BCSD ensemble under the SSP2-4.5 and SSP5-8.5 forcing scenarios, respectively. Under the SSP2-4.5 scenario, BCSD ensemble projects for summer and fall increase by up to 10% over much of Brazil for the coming decades. From 2080, up to 20% growth is projected in Brazil’s southeastern and northeastern sectors. In contrast, up to 20% reductions are projected in the extreme north of SA, with their sign diminished by the end of the 21st century.

In the winter season (Figure 4c), the BCSD ensemble projects more expressive reductions starting in 2040, with regions of maximum decrease (up to 50%) beginning in 2080 in the central-western and northeastern Brazil sectors. In the spring (Figure 4d), the BCSD ensemble projects a significant reduction in rainfall, intensified after 2060, with reductions above 20% in large parts of central and northeastern Brazil. The results obtained here partially agree with those of other studies that used projections from the CMIP5 and CMIP6 models. Under the RCP4.5 forcing scenario, mean annual patterns from the ensemble of 26 CMIP5 models indicate decreases of up to 150 mm year-1 in the far north of SA, as well as decreases in annual rainfall over much of central SA, a slight increase over isolated portions of Northeast Brazil, and larger increases over southern Brazil [131].

Similarly, the CMIP5 ensemble indicates increased rainfall over southeastern SA and reduced rainfall over Amazonia and northern SA during the summer of 2050-2080 [23]. In winter, increased precipitation is also seen over western SA, extending from Ecuador to Argentina [23], a pattern analogous to that found here. Additionally, an ensemble composed of 38 CMIP6-GMCs projects increased precipitation (~0.3 mm day^-1) over Northeast and South Brazil sectors during the summer of 2040-2059 and a reduction of the same magnitude over nearly all of SA during the winter [123]. For the period 2080-2099, projections show even wetter (drier) conditions in southern Brazil (Amazonas and northern SA) during summer and intensified rainfall reduction across the continent during winter [123]. On the other hand, a study with the global HadGEM2-ES model nested with the Eta regional model under the RCP4.5 forcing scenario shows a projection of increased precipitation over most of the Amazon Basin, southern Brazil, and the northern portion of the coastal coast of Northeast Brazil, as well as decreased rainfall over much of the Midwest, Southeast, and central Northeast regions of Brazil [132]. We stress that the similarities and differences between the results of the studies are due to factors such as different models used, emission scenarios employed, reference periods chosen, and validation data.

Considering the SSP5-8.5 emission scenario (Figure 5), spatial patterns of projected seasonal change in precipitation are similar to those obtained for the SSP2-4.5 scenario but with the most intense sign of change. During summer (Figure 5a), an average increase of 10% is projected over most of Brazil and Argentina, and the growth intensifies after 2060, principally over portions of Northeastern Brazil and central-southern Argentina. The changes in fall (Figure 5b) are similar to the SSP2-4.5 scenario but indicate more intense precipitation increases in the Bahia state (Brazil), southern Brazil, central-eastern Argentina, and the central Andes.

In winter (Figure 5c), the BCSD ensemble projects rainfall decrease over much of central Brazil and Bolivia, extending into northern Argentina and southeastern, northeastern, and north Brazil. From 2060 onwards, the BCSD ensemble shows decreases of up to 50% in the Midwest and coastal Northeast areas. In contrast, a substantial rainfall increase for Brazil’s southeast and south coasts is observed from 2080 onward. In spring (Figure 5d), the projections indicate more drastic changes, with decreases of more than 10% over most of Brazil and northern SA, with more intense reductions (up to 50%) over the northern portion of coastal Northeast Brazil. During this season, projected increases in precipitation occur in isolated regions such as the coasts of Peru and Ecuador and northern Chile.

The results agree with those of Ruffato-Ferreira et al. [132], in which there is a trend of increasing water scarcity mainly in central Brazil, as well as a progressive increase in water availability in the southern and southeastern Atlantic basins, favoring southern Brazil. In addition, the São Francisco River Basin is the most vulnerable in a maximum emission scenario, accentuating water scarcity in Northeast Brazil. Similarly, CMIP5 projections indicate increases of about 100 mm year^-1 by the end of the 21st century in southern Brazil and parts of Peru, Ecuador, Colombia, and Venezuela. In comparison, areas between south Chile and Argentina and the far north of SA may experience reductions of up to 150 mm year^-1 [131].

The higher severity of precipitation reductions in SA under the SSP5-8.5 scenario was also obtained by CMIP5 models nested with different regional climate models [23,24,133]. Among the possible causes for the dry conditions projected for Amazonia and northern SA is the weakening of the northeast trade winds at the end of the 21st century, inducing a decrease in moisture transport from the ocean to the continent [23,73]. Additionally, studies with CMIP6-GCMs under the SSP5-8.5 forcing scenario also provide projections of expressive precipitation reduction over much of the continent, mainly in the Midwest, Southeast, Northeast, and North of Brazil and northern SA, with decreases of up to 1.2 mm day^-1 in the most affected regions [123,125]. On the other hand, SESA and southern Brazil will likely experience higher rainfall volumes in the coming decades, exposing these regions to the progressive frequency of extreme daily precipitation events and an increase in the number of consecutive wet days [23].

Analyses of projected changes in rainfall with GCMs from CMIP3, CMIP5, and CMIP6 over Brazil show that the projected signal depends on the CMIP generation considered, except for southern Brazil, where an increase is seen in all cases [134]. While CMIP3 projects an increase in rainfall in northern Brazil (especially in the western portion), CMIP5 and CMIP6 models project a reduction. In Northeast Brazil, the projections are also divergent among the CMIP generations, with CMIP5 indicating an increase in rainfall throughout the territory. At the same time, CMIP3 and CMIP6 project an increase (reduction) in rainfall in the region’s northern (southern) sector. In the Midwest and Southeast regions, the sign depends on the family of CMIP used, with increased precipitation projected by CMIP5 and decreased rainfall estimated by CMIP3 and CMIP6. In summary, multi-model ensembles show that CMIP3 most accurately represents precipitation extremes in Northeastern Brazil, while CMIP5 performs best for the Midwest, and CMIP6 provides the most accurate projections for the remaining Brazilian regions [134].

Finally, we recommend caution for energy planning with the projections analyzed here. More robust evaluations should consider the vegetation of different biomes since it plays a crucial role in the water balance and greater detail of the projected scenarios of land use and land cover changes. Furthermore, intrinsic to the process of climate modeling, the uncertainties and inaccuracies associated with different models limit a greater assertiveness and require pondering in decision-making based on the projections.

3.3. Temporal Series of the BCSD Ensemble SPI-12 Index under the SSP5-8.5 Forcing Scenario

Figure 6 shows the SPI-12 temporal series (2020-2099) provided by the BCSD ensemble under the SSP5-8.5 emission scenario for eight SA subdomains. The SPI-12 index was obtained by averaging the indices calculated for each individual projection. In R1 (northwestern Amazonia), 31 drought episodes were identified between 2020-2099, all belonging to the mild drought category. The longest episode occurs from 07/2086 to 09/2088, totaling 37 months, followed by the episode from 06/2065 to 05/2068 (36 months) and severity of 10.02. In addition, other long-lasting drought episodes occur from 11/2044 to 08/2047 and from 08/2049 to 05/2052 (34 months each).

For R2 (central Amazonia), 30 drought episodes were identified in the period 2020-2099, categorized as mild droughts, and 56% of the episodes (17 cases) have a duration of 10 months or longer. The longest episode occurs from 03/2061 to 11/2065 (57 months), with a severity of 13.89. Other longer episodes occur from 12/2085 to 02/2089 (39 months), from 04/2039 to 08/2041 (29 months), and from 10/2073 to 12/2075 (27 months).

Although different droughts have occurred in Amazonia during El Niño-Southern Oscillation (ENSO) events, SST anomalies in the tropical North Atlantic (TNA) also play an important role in the region’s rainfall regime [4,37,40,135,136]. The anomalous warming in TNA is associated with the northward displacement of the ITCZ, changes in the north-south divergent circulation, and weakening of the trade winds and moisture flux from the tropical Atlantic, inducing a reduction of rainfall in the southern, northern, and eastern sectors of Amazonia [4,37,136]. Furthermore, ENSO events are related to anomalies in the east-west Walker circulation, with convection over the central Pacific and subsidence over east and central Amazonia [4,135,136].

In general, drought events related to warm SST anomalies in TNA show a north-south gradient with drier (wetter) conditions in southern (northern) Amazonia, while droughts linked to ENSO events show a southwest-northeast gradient with drier conditions in northeastern Amazonia [136]. However, overlapping effects of both teleconnection mechanisms also affect the region, such as the 2010 drought associated with successive ENSO episodes during the austral summer and the warmer TNA during the austral autumn and winter [37]. Similarly, the severe drought of 2015-2016 was associated with intense warm anomalies in the central Pacific and TNA, with marked effects in northeastern Amazonia [137].

Considering the occurrence of drought events in 2015-2100 relative to the 1850-2014 period under the SSP5-8.5 scenario, Wang et al. [14] found an increase in the frequency of droughts in northern SA during the 21st century, as well as more prolonged droughts and more than 50% increase in the extent of areas affected. On the other hand, the variability of drought-related statistical results provided by CMIP6 models is greater in the tropics than in other latitudinal zones, implying that GCMs need improvement in capturing drought-causing patterns in equatorial regions [7]. Furthermore, models from CMIP5 and CMIP6 indicate divergence in rainfall projections over the area, and models from CMIP6 show no improvement in simulating total precipitation and consecutive dry days relative to the previous generation of CMIP [138].

In R3 (northern sector of Northeast Brazil), 31 drought episodes were identified in 2020-2099, all belonging to the mild drought category. About 61% of the episodes (19 events) present a duration equal to or longer than 10 months. The longest-lasting hydrological drought episode occurs from 11/2027 to 12/2032 (62 months), followed by the episodes from 01/2041 to 02/2044 (38 months) and 12/2034 to 12/2037 (37 months). In R4 (central sector of Northeast Brazil), 32 hydrological drought episodes were counted in 2020-2099, all classified as mild droughts. About 72% of the episodes (23 cases) are 10 months or longer. Four longer-lasting episodes were obtained from 01/2035 to 12/2040 (72 months), from 11/2026 to 11/2030 (49 months), from 01/2067 to 12/2069 (36 months), and from 12/2085 to 02/2088 (27 months). Our results corroborate previous analyses since CMIP6 projections suggest an increase in the number of dry days in Northeast Brazil (mainly in DJF and MAM), with an estimated increase of up to 8.0 and 14.7% in the near (2016-2040) and far (2076-2100) futures, respectively, under the SSP5-8.5 scenario [139].

Precipitation in Northeast Brazil is marked by interannual variability and drought events are attributed to ENSO and the anomalously northern position of the ITCZ, resulting from a warmer TNA [12,37,136,140]. However, extratropical variability modes also influence rainfall distribution in the region, as analyses from 1980-2009 concluded that drought events in this period showed annular patterns in both hemispheres (South Annular Mode and North Annular Mode) well configured during DJF (pre-rainy season in the region), both in years with and without ENSO [141].

For R5 (Midwest region of Brazil), 31 drought episodes were identified, all classified as mild droughts, with the most extended episode from 11/2026 to 02/2031 (52 months), followed by other long-lasting events from 02/2043 to 10/2046 (45 months), and from 10/20635 to 11/2038 (38 months). About 61% of the drought episodes (19 cases) are 10 months or longer.

Marengo et al. [46] report no evident direct relationship between drought events in the region and SST anomalies in the Pacific and Atlantic oceans. While the 2019-2020 drought was associated with anomalous warming in TNA, earlier events occurred with simultaneous warming of the northern tropical and equatorial Pacific and cooling of TNA. Overall, the authors conclude that droughts in the region may be triggered by warmer SSTs in the North Atlantic and North Pacific (which promote the northward displacement of the ITCZ and reduce precipitation in southern Amazonia and the Midwest), which reduce moisture transport from Amazonia to the region. However, regional factors such as water balance and soil moisture influence the sector’s interannual seasonality of droughts and floods. In this context, there is an increasing tendency in the water deficit in deforested regions due to the expansion of agriculture and cattle ranching, contributing to local warming and reduced precipitation [35].

In R6 (Southeast region of Brazil), 33 drought episodes were counted in the period 2020-2099, all belonging to the mild drought category, with the longest-lasting episode from 04/2020 to 01/2026 (70 months), followed by episodes with 61 months (from 12/2035 to 12/2040) and 23 months (11/2032 to 09/2034). Approximately 60% of projected drought episodes are 10 months or longer.

Analyses of SPI-1 and SPI-12 in the north and northwest areas of the Rio de Janeiro state for the 1967-2013 period indicated a higher occurrence of events in the moderately and extremely dry categories, as well as a higher frequency of droughts in the two regions of the state during ENSO cycles in both phases of the Pacific Decadal Oscillation [62]. Moreover, analyses of drought events in the Paraná River basin showed that hydrological droughts in the 1981-2021 period were the most severe and intense [35]. Furthermore, studies show that the severe drought of 2014-2015 was associated with anomalous warming in the western tropical Pacific that initiated a wave train along the south Pacific, which in turn resulted in anomalous anticyclonic circulation in the southwest Atlantic, expanding the west flank of the South Atlantic Subtropical Anticyclone (SASA) and restricting the entry of low-pressure systems into southeastern Brazil [30,32]. Additionally, analyses of summer droughts during 1961-2010 in the São Paulo state show a prevalence of anomalous subsidence of the Hadley cell’s descending branch over the region, inhibiting convective activity in the sector [33].

In R7 (Southern Brazil and Uruguay region), 29 drought episodes were computed, all categorized as mild droughts, of which 62% (18 episodes) have a duration of 10 months or longer. The longest drought episode refers to the period from 01/2022 to 01/2028 (73 months), followed by cases of 41 months (12/2032 to 04/2036) and 39 months (08/2036 to 10/2039). Many of the droughts that have occurred in the region are linked to the cold phase of ENSO (La Niña), but other factors also contribute to the onset and intensification of droughts in the sector, such as the development of atmospheric blockings in the south Pacific, warmer SST anomalies in TNA occurring concurrently with La Niña, as well as more regional and local aspects such as reduced moisture transport to the region caused by deforestation in Amazonia [49,50,142,143]. Attribution study infers that the rainfall deficit occurring in the Southern part of Brazil, Argentina, and Uruguay since 2019 is not only partially induced by the action of La Niña but also caused by higher temperatures that reduce water availability in the region, indicating that although the decrease in rainfall is associated with natural climate variability, the consequences of drought are becoming more severe due to increasing temperatures [143].

Finally, in R8 (western Patagonia), 40 drought episodes were identified in 2020-2099, all classified as mild droughts, of which 52.5% (21 cases) have a duration of 10 months or more. The longest-lasting episode occurs from 06/2050 to 07/2053 (38 months), followed by episodes of 32 months (01/2076 to 08/2078) and 28 months (09/2066 to 12/2068). This sector has experienced intense droughts recently [144], substantially affecting socio-economic activities in the region.

In conclusion, the BCSD ensemble shows that all SA subdomains analyzed are prone to drought episodes during the 21st century. Although the ensemble projects predominantly mild droughts, perhaps due to smoothing the most extreme projections, a considerable proportion of episodes last longer than 10 months. The significant occurrence of longer hydrological drought episodes corroborates analyses previously performed with CMIP6 models that indicated an increase in event duration during the 21st century under the SSP5-8.5 scenario in SA and a higher frequency of longer-lasting events [14].

3.4. Projections of Drought Parameters by the Bias-Corrected CMIP6-GCMs and BCSD Ensemble

Figure 7 and Figure 8 show the droughts parameters projected by the eight bias-corrected CMIP6-GCMs, as well as by CPC (for the historical period only) and BCSD ensemble under the SSP5-8.5 scenario.

For R1, of the 301 episodes identified by all datasets, 72% (218 cases) belong to the mild drought category, 16% (48 cases) correspond to the moderate drought class, 9% (28 cases) are of severe drought events, and 2% (7 cases) belong to the extreme drought category. Moreover, only three GCMs (CMCC-CM2-SR5, CMCC-ESM2, and MPI-ESM1-2-LR) indicate a slight increase in the average number of drought episodes in the 2020-2099 relative to the historical period, while the IPSL-CM6A-LR model and BCSD ensemble suggest a reduction of up to 27% and 23%, respectively.

On the other hand, seven of the nine datasets show an increase in the duration of drought episodes in 2020-2099 relative to 1996-2014. The IPSL-CM6A-LR and EC-Earth3 models indicate an increase of 47% and 32% in the duration (in months) of the events, respectively, while the BCSD ensemble provides an average increase of 25%. Similarly, most GCMs (and the BCSD ensemble) converge on increasing severity of drought episodes in the 21st century, with the IPSL-CM6A-LR model and the BCSD ensemble indicating increases of 51% and 40%, respectively. Regarding the intensity and peak parameters, GCMs show mixed signals, and the MRI-ESM2-0 and EC-Earth3 models show 12% and 18% increases in intensity and peak, respectively. In general, in this region, all GCMs overestimate the number of drought episodes over the historical period, and the GFDL-ESM4 and IPSL-CM6A-LR models show the largest range of parameter estimates for the 2020-2099 period.

In R2, of the 252 episodes identified by all datasets, 65% (163 cases) correspond to the mild drought class, 17% (43 cases) are of moderate drought events, 12% (30 cases) refer to severe drought events, and 6% (16 cases) are of extreme drought events. Additionally, about 70% of the datasets converge to a growing number of drought episodes in the coming decades relative to 1996-2014, with the MPI-ESM1-2-LR and MRI-ESM2-0 models indicating an average increase of up to 45% and 38%, respectively. In addition, half of the datasets show an increase in drought duration in the 21st century, with the CMCC-CM2-SR5 model providing an average increase of up to 50%. Similarly, this model projects an average increase of 24% in the severity of drought episodes. For the intensity and peak parameters, the GCMs show mixed signals, with the EC-Earth3 model providing an average increase of up to 19% in the magnitude of both parameters. In general, in this region, the GCMs perform better in representing drought episodes during the historical period, with the CMCC-ESM2 and GFDL-ESM4 models providing the same number of episodes obtained by CPC.

In R3, 245 drought episodes were identified by the datasets, of which 64% (156 cases) correspond to mild drought episodes, 23% (56 cases) are moderate droughts, 9% (21 cases) are severe drought events, and 5% (12 cases) are extreme drought events. In this region, half of the datasets project an increase in the frequency of drought episodes in the 2020-2099 period (relative to 1996-2014), and half show a decrease. CMCC-ESM2 model indicates an average reduction of up to 47% in the number of episodes, but the MIROC6 model shows an average increase of up to 33%. Contrarily, only two GCMs (GFDL-ESM4 and IPSL-CM6A-LR) provide a reduction in episode duration, while all the others project increase during the following decades. MRI-ESM2-0 and CMCC-ESM2 yield average increases in the duration of drought episodes of up to 44% and 100%, respectively. These same models also provide the largest average increases in severity, corresponding to 44% and 168% (by MRI-ESM2-0 and CMCC-ESM2, respectively). As for the other regions, the intensity and peak projections show mixed signals, with the CMCC-ESM2 model indicating a 44% increase in the average peak magnitude of episodes in 2020-2099, while the EC-Earth3 model shows a 34% reduction.

For R4, 267 drought episodes were identified by all datasets, of which 58% (156 cases) were classified as mild droughts, 21% (57 cases) as moderate droughts, 11% (29 cases) as severe droughts, and 9% (25 cases) as extreme droughts. In this sector, GCMs show divergent projections about the frequency of drought episodes in 2020-2099. While the IPSL-CM6A-LR model projects an average increase of up to 50% in frequency, the MPI-ESM1-2-LR model estimates an average reduction of up to 43%. However, models converge about the increasing duration over the coming decades, with only three GCMs projecting reductions (CMCC-ESM2, IPSL-CM6A-LR, and MRI-ESM2-0) and the MPI-ESM1-2-LR model indicating an average increase of up to 61%. Similarly, GCMS are more homogenous in the increasing severity, with the MPI-ESM1-2-LR model providing an average increase of up to 100%. Whilst GCMs project mixed signals about changes in the intensity, projections of peak changes are more concordant, with most models indicating an increase. In this case, the MIROC6 model projects an average increase of up to 37% in peak episodes over the coming decades.

In R5, all datasets total 277 drought episodes, of which 64% (176 cases) correspond to the mild drought class, 21% (58 cases) are moderate drought events, 11% (31 cases) are severe drought events, and 4% (12 cases) are extreme drought events. In this region, more than half of the datasets project an increasing frequency in drought episodes in 2020-2099, with the GFDL-ESM4 model indicating an average increase of up to 63%, while the MPI-ESM1-2-LR model projects an average reduction of up to 46%. Similarly, the same proportion of models projects an increase in episode duration over the coming decades, with the CMCC-ESM2 and MPI-ESM1-2-LR models providing average increases of up to 40% and 90%, respectively.

Regarding severity, only three GCMs project a reduction (EC-Earth3, MIROC6, and GFDL-ESM4), and the MPI-ESM1-2-LR model estimates an average increase of up to 106%. For the intensity and peak parameters, the signals provided are heterogeneous, with projections of change in intensity ranging from -32% (by EC-Earth3) to 34% (by CMCC-CM2-SR5) and amplitude of change in peak from -17% (by MRI-ESM2-0) to 21% (by IPSL-CM6A-LR).

For R6, of the 324 episodes identified by the datasets, 70% (227 cases) were classified as mild droughts, 16% (51 cases) as moderate droughts, 12% (38 cases) as severe droughts, and 2% (8 cases) as extreme droughts. In this sector, half of the ensembles project a reduction in the frequency of drought episodes (with the EC-Earth3 model providing an average reduction of up to 21%). Inversely, another half suggests an increase (with the CMCC-CM2-SR5 model projecting an average increase of up to 32%). Regarding episode duration in the 2020-2099 period, only two GCMs project a reduction (CMCC-CM2-SR5 and CMCC-ESM2), while the MIROC6 and EC-Earth3 models provide an average increase of 29% and 37%, respectively. Likewise, only three GCMs indicate a reduction in severity (CMCC-CM2-SR5, CMCC-ESM2, and MPI-ESM1-2-LR), while the IPSL-CM6A-LR and EC-Earth3 models project an average increase of 27% and 70%, respectively. Of the same, only the models CMCC-CM2-SR5 and MPI-ESM1-2-LR project a reduction in the intensity of drought episodes in 2020-2099, while the MRI-ESM2-0 model indicates an average increase of up to 158%. Regarding the peak of identified episodes, the outputs indicate mixed signals, with the EC-Earth3 model projecting an average increase of 26% and the MIROC6 model providing an average reduction of up to 22%.

In R7, a total of 293 drought episodes was obtained, with 70% (206 cases) being mild droughts, 19% (56 cases) of moderate droughts, 9% (27 cases) being severe droughts, and 1% (4 cases) of extreme droughts. Regarding the frequency, only two GCMs project an increase in the incidence (21% and 25% by the CMCC-ESM2 and MIROC6 models, respectively), while models such as EC-Earth3 and CMCC-CM2-SR5 indicate a reduction of 25% and 40%, respectively. Contrarily, only two GCMs project a reduction in episode duration (2% and 5% by CMCC-ESM2 and MIROC6 models, respectively), while EC-Earth3 and CMCC-CM2-SR5 models indicate a 49% and 101% increase, respectively. Alike, most models converge to a signal of increased severity of hydrological drought episodes in the region in 2020-2099, with the GFDL-ESM4, EC-Earth3, and CMCC-CM2-SR5 models projecting average increases of 65%, 96%, and 125%, respectively. The datasets indicate mixed signals for intensity, with decreases and increases ranging from 4% to 55% and 6% to 13%, respectively. Similarly, projections of change in episode peaks are also more heterogeneous, with reductions and increases ranging from 0.10% to 30% and 7% to 28%, respectively.

Finally, in R8, all datasets give a total of 331 drought episodes in 2020-2099, of which 81% (268 cases) are classified as mild droughts, 13% (43 cases) as moderate droughts, 5% (17 cases) as severe droughts, and 1% (3 cases) as extreme droughts. Half of the outputs project a reduction in the duration of episodes in the sector, while another indicates an increase. As for the other regions, most GCMs converge on a signal of increasing duration of drought episodes, with the GFDL-ESM4 model projecting an average increase of up to 34%. Likewise, projections among the models are more concordant about the increasing severity, with the IPSL-CMC6A-LR model providing an average increase of up to 41%. Additionally, most GCMs project an increase in peak episodes in 2020-2099 (only the EC-Earth3 model indicates an average reduction of up to 3%), with the BCSD ensemble providing an average increase of up to 32%.

In summary, GCMs project mixed signals about changes in drought events’ frequency, intensity, and peak magnitudes during the coming decades in the SA subdomains. On the other hand, the projections are more homogeneous regarding the duration and severity of the episodes, with most models converging to increasing magnitudes of both parameters in all sectors evaluated. Concerning the different categories of droughts (mild, moderate, severe, and extreme), results show a larger occurrence of mild droughts. However, regions like the northern and central Northeast, Midwest, and Southeast Brazil show a substantial proportion (above 20%) of moderate drought events, as well as a relevant occurrence (above 10%) of severe drought events in the Amazonia region, central Northeast, Midwest, and Southeast Brazil. In this context, we highlight that although the BCSD ensemble provides predominantly mild drought episodes, individual analyses of the CMIP6-GCMs indicate expressive frequencies of moderate and severe events in all the evaluated subdomains.

4. Conclusions

In this study, we applied statistical downscaling to CMIP6 precipitation projections in SA using the CPC data as a reference to evaluate future changes in precipitation and the occurrence of hydrological droughts in the continent. To this end, we used the QDM technique developed by Cannon et al. [92], and the method proved effective in reducing systematic biases and preserving the trends of GCMs’ projections. For the coming decades, the post-processed precipitation projections indicated reduced rainfall in sectors such as northern SA, North, Northeast, Midwest, and Southeast Brazil and increased precipitation in southern Brazil and SESA regions. Such changes are more prominent during the austral spring (SON), and their signal is more robust under the SSP5-8.5 scenario, corroborating the literature.

The SPI-12 Index analysis showed considerable variability of projections among the GCMs about drought event parameters such as frequency, intensity, and peak. On the other hand, concerning duration and severity, a more remarkable agreement was observed among the GCMs regarding the intensification of both aspects in practically all the subdomains analyzed. Considering the different categories of drought events, the results showed a substantial frequency of moderate and severe droughts in the Northeast, Midwest, and Southeast Brazil. In addition, the individual CMIP6-GCMs and the BCSD ensemble project a considerable proportion of events with duration equal to or greater than 10 months in all evaluated South American sectors.

Given the results, it is valid to highlight some limitations of the study. Firstly, we point out that although the SPI-12 analysis is easy to implement because it uses only precipitation information, it disregards crucial aspects of the droughts related to temperature and evapotranspiration processes. In this context, analyses have indicated that the frequency and duration of droughts in SA are driven mainly by climate factors such as maximum and minimum temperatures, net surface radiation, and precipitation [145]. Thus, future studies employing statistical downscaling should address these aspects for assessing droughts on the continent during the 21st century.

Furthermore, although the study employed eight GCMs and the multi-model ensemble, it is recommended that future works use a larger number of models and forcing scenarios to minimize the uncertainties associated with the projections. Moreover, the projections presented here have an intermediate spatial resolution of 50 km, which limits the spatial detail of the analyses performed. In this sense, further research should consider climate projections with a finer spatial resolution to ensure greater accuracy of the results. However, the feasibility of such approaches demands high computational costs. Overall, despite the uncertainties associated with GCMs, the identification of hydrological drought episodes, and the bias correction technique, the results presented here can provide valuable subsidies to decision-makers and energy planners for better management of water resources in the South American continent for the coming decades.