Climate Change Impacts On Evapotranspiration in Brazil: a Multi-model Assessment

A large part of Brazil is highly vulnerable to climate changes projected for the end of the 21st century. Analyzing these vulnerabilities is particularly important for agriculture, since the country is one of the largest agricultural commodity producers in the world. Changes in the reference evapotranspiration (ET o ) can impact crops and make cultivation unfeasible. However, studies on ET o patterns under climate change scenarios for Brazil have been restricted to regional scales and use too few climate models or too simpli�ed water balance models for their analysis. This can lead to uncertainties in assessing the impacts of climate change on ET o . Therefore, this study seeks to analyze ET o patterns in Brazil towards the end of the 21st century using two methods that are better at estimating regional ET o , i.e., the Turc and Abtew methods, under two radiative forcing scenarios (RCPs 4.5 and 8.5). Daily data on near surface air temperature (mean and maximum), global solar radiation, and near surface relative humidity from six General Circulation Models (GCMs) from the Coupled Model Intercomparison Project Phase 5 (CMIP5) were used to analyze the simulations and projections for climate change. The performance of climate simulations is heterogeneous among the GCMs, with overestimations (~ 2.5 mm day -1 ) in some models, and underestimations (~ 1.5 mm day -1 ) in others. In general, climate change projections indicate increases of up to 1 mm day -1 in ET o , mainly in the North, Northeast, and Center-West regions of Brazil. Both estimation methods showed similar spatial patterns, however the Turc method projected lower intensity changes compared to the Abtew method.

Since Brazil is one of the largest agricultural commodity producers in the world, it requires e cient irrigation systems to properly manage water resources and ensure sustainable commercial production (Jerszurki et al. 2019;Monteiro et al. 2021).Studies on climate change impacts and vulnerabilities have been increasingly targeted towards different socioeconomic sectors (Guimarães et al. 2016;Lyra et al. 2017), especially the agricultural sector, which is highly dependent on climate conditions (Santos et al. 2017;Por rio et al. 2018).
The reference evapotranspiration (ET o ) is the main agrometeorological variable used by irrigation projects, and also considered in the assessment of agricultural impacts through indicators of drought, and agricultural crop and forestry growth and yield (Fan et  This study seeks to contribute to scienti c literature, relative to studies that have already been carried out, by analyzing the in uence of climate projections to the end of the 21st century (2071-2100) on daily evapotranspiration in Brazil using six climate models under two radiative forcing scenarios based and using two methods that best estimate ET o .
2 Materials And Methods

Methods for estimating the reference evapotranspiration
According to Monteiro et al. (2021), the Turc (Valipour et al. 2017) and Abtew (Abtew, 1996) methods are the most appropriate for estimating ET o for Brazil under current climatic conditions.Therefore, these two methods will be used in this study to analyze the ET o under future climate scenarios.
The ET o in the Turc (Tu) and Abtew (Ab) methods is calculated, respectively, using equations 1 and 2:

Where
is the reference evapotranspiration (mm day -1 ); is global solar radiation (MJ m -2 day -1 ); is the mean daily near surface air temperature (°C); is the maximum daily near surface air temperature (°C); and is the latent heat for vaporization (2.4418 MJ kg -1 ).Furthermore, in Equation 1, when near surface relative air humidity or when RH < 50%.

Analysis of the performance of ET o estimation methods using data from climate models
The performance analysis for the ET o values calculated from data from six General Circulation Models  This subset of CMIP5 models (Table 1) was chosen given the availability of daily data, given that the necessary input variables for the Tu and Ab methods (Equations 1 and 2) were contained in their databases, and since they are widely used in literature.
Validation was performed on a seasonal and annual scales for each GCM and considering the average from the six models (ensemble mean  3 Results And Discussion

Climate models performance
The performance of GCMs in simulating ET o show under/overestimations that varied according to region and model (Figures 1 and 2).By contrast, the results using the Tu method (Figure 1) and the Ab method (Figure 2) were quite similar.The ensemble mean of GCMs smoothened the individual biases of the models, but still showed overestimations for some locations throughout the year, mainly in the North of Brazil (Figures 1 and 2).
There was divergence in ET o estimation among the GCMs when using Tu (Figure 1) and Ab (Figure 2) methods, different from results from Llopart et al. (2020) when using a combination of 2 global and 8 regional models for South America.The CanESM2, CNRM-CM5, IPSL-CM5A-MR and MIROC-ESM models showed overestimations (up to 2.5 mm day -1 ) in both methods, mainly in the North of Brazil, and in parts of the Center-West.The HadGEM2-CC and MRI-CGCM3 models showed underestimations (up to 1.5 mm day -1 ) in the Southern and Northeastern regions of Brazil, respectively.In general, both methods showed the same under/overestimates per region and per model.The only difference was that under/overestimates were more intense in the Ab method (Figure 2) than in the Tu method (Figure 1).
When individually analyzing model performance in simulating the variables that are used to calculate ET o , we identi ed that the temperature patterns (mean and maximum) were similar among the GCMs, except for CanESM2, which had overestimations up to 5°C for the northern region of Brazil during the austral spring (Supplementary Material 1 and 2).For RH (SM.3), the CanESM2, CNRM-CM5 and IPSL-CM5A-MR models gave underestimates above 10%, mainly in the North, and overestimates from the Northeast to South of Brazil.The HadGEM2-CC, MIROC-ESM and MRI-CGCM3 models showed an opposite trend for under/overestimated, i.e., the models could not adequately simulate RH, showing discrepancies among the GCMs with respect to the data.The GCMs patterns diverged from each other mainly in the R s simulations (SM.4), with most overestimates (greater than 3 MJ m -2 day -1 ) for all of Brazil throughout the year.In general, the MRI-CGCM3 better represented the climate variables (in magnitude and spatial pattern), and consequently better represented ET o estimates (Figures 1 and 2).
This result is different from Guimarães et al. ( 2016) performed for the Northeast of Brazil, where the HadGEM2-CC climate model performed the best for ET o (correlation = 0.6 to 0.8) of all the GCMs studied.Both methods project a general increase (0.6 to 1 mm day -1 ) for ET o , mainly in the North, Northeast, and Center-West of Brazil for the CanESM2, HadGEM2-CC and MIROC-ESM models.The CNRM-CM5 and MRI-CGCM3 models showed less intense increases (0.2 to 0.4 mm day -1 ) for all of Brazil.The IPSL-CM5A-MR model had the smallest projected increases (0 to 0.2 mm day -1 ), mainly using the Tu method.Almost all of Brazil will be affected by increases greater than or equal to 1 mm day -1 in the ET o rate under greater radiative forcing scenarios (Figures 4 and 6).The climate projections by the ensemble mean for the six GCMs using the different estimation methods (Tu and Ab) showed a tendency for increased ET o for all of Brazil, but the magnitude of this increase was smoother, mainly in the North, Northeast, and Center-West of Brazil (Figures 3-6).These results corroborate the results of Cardoso and Justino (2014), who calculated ET o using the Penman-Monteith method for a regional climate model coupled with a potential vegetation model, and Llopart et al. (2020), who considered a simpli ed water balance with regional climate models.Both authors obtained an increase of up to 3 mm day -1 for the Northern region of Brazil.However, our results differ from Andrade et al. (2020), who used soil water assessment tools and regional climate models and obtained a reduction of 0.36 mm day -1 for a part of Northeastern Brazil.

Climate changes on ET o
The climate change projections for T, T max , RH and R s for RCP 4.5 show similar spatial patterns, but with lower intensities than RCP 8.5.For brevity's sake, the supplementary material shows projections for only these variables for scenario RCP 8.5 (SM.5-8).The GCMs show good agreement among each other for T (SM.5) and T max (SM.6) projections, with increases towards the end of the 21st century up to 6°C for RCP 8.5.The most intense temperature changes (from 4 to 6°C) were projected in the CanESM2, HadGEM2-CC, IPSL-CM5A-MR and MIROC-ESM models, mainly for the North and Center-West regions of Brazil.The CNRM-CM5 and MRI-CGCM3 models projected less intense increases (from 2 to 3°C) for southern Brazil.
RH projections showed variable spatial patterns and magnitudes among the GCMs for all of Brazil (SM.7).Nonetheless, generally these tended to decrease by 6% (RCP 4.5) to 10% (RCP 8.5) towards the end of the 21st century.This RH reduction can be explained by decreased precipitation across most of Brazil day -1 in the extreme North of Brazil, and inland in the Northeast.This pattern was also observed by Cardoso and Justino (2014), who explained these divergences as changes in surface albedo and in the heterogeneity of precipitation projections from the individual regional climate models.
In The six climate models showed different simulations for global solar radiation and relative humidity (largest discrepancy), except for mean and maximum air temperature input variables.Additionally, the models showed some divergence for ET o simulations using the Tu and Ab methods, with overestimates (up to 2.5 mm day -1 ) in some climate models, and underestimates (up to 1.5 mm day -1 ) in others.Therefore, since the Tu method was slightly superior to the Ab method when comparing to the observed data, the Turc method was more reliable in estimating ET o for future climate chance scenarios.
Despite the divergences in the climate models for some input variables used to calculate ET o , climate projections indicated similar patterns for the analyzed climate models, and for the two ET o estimation methods used, with projected increases (1 mm day -1 ) mainly in the North, Northeast, and Center-West of Brazil.The results were more intense when using the Ab method.
The assessment of the impacts of climate change on evapotranspiration performed by this study can be useful in outlining adaptation measures to cope with damages caused by changes to various sectors of the economy, e.g., agriculture, forestry, and hydroelectric generation in Brazil.Additionally, future studies seeking to verify climate change evapotranspiration trends should also consider both above and below ground variables, as well as they should be performed with recent state-of-the art GCMs (CMIP6), that have been evaluated in the IPCC Sixth Assessment Report (IPCC AR6), using different socioeconomic pathways.

1
Introduction Extreme weather and climate event changes have been recorded and projected for different regions of South America (Natividade et al. 2017; Avila-Diaz et al. 2020a,b; Cerón et al. 2021; Regoto et al. 2021.Generally, warmer climates are projected for all of Brazil by the end of the 21st century, possibly increasing by 5°C in the south of the Amazon, in the Center-West, and the western part of Minas Gerais state (Torres and Marengo 2014; IPCC 2021).Furthermore, models project heterogeneous rainfall trends in the form of reduced rainfall at lower latitudes and increased rainfall at higher latitudes (IPCC, 2013, 2021; Llopart et al. 2020), making much of Brazil vulnerable to climate change (Torres et al. 2012; Darela et al. 2016; Silva et al. 2019; Lapola et al. 2020; Torres et al. 2021).

2. 3 nd
Projected changes to the reference evapotranspirationThe projections were based on seasonal and annual analysis on the six GCMs and the ensemble mean, using the Tu and Ab methods and two radiative forcing scenarios, called Representative Concentration Pathways (RCPs) (4.5 and 8.5), projected to the end of the 21st century (2071-2100).RCPs 4.5 and 8.5 represent intermediate radiative forcing scenarios (4.5 W m -2 ) and more intense radiative forcing scenarios (8.5 W m -2 ) and correspond to equivalent CO2 concentrations at 650 and 1370 ppm, respectively(Moss et al. 2010;Van Vuuren et al. 2011).Projected future climate changes were calculated by taking the difference between the climatological average for future period (2071-2100) from the climatological average from the historical period(1980-  2005), considering the two analyzed RCPs, and the different ET o estimation methods (Tu and Ab).

Figures 3
Figures 3 to 6 show the seasonal and annual climate changes projected for ET o using the Tu and Ab methods under different radiative forcing scenarios for the end of the 21st century (2071-2100) for all of Brazil.In general, climate change projections for ET o relative to RCP 4.5 (Figures3 and 5) show similar spatial patterns with lower intensity compared to RCP 8.5 (Figures4 and 6).Additionally, the projected climate changes for ET o show similar spatial patterns and magnitudes between the different estimation methods.The Tu method gave lower intensity results compared to the Ab method.

(
Llopart et al. 2020;Sousa et al. 2019).In the South of Brazil, where precipitation increases are projected, there was no signi cant projection for increased/reduced RH.In the CanESM2, CNRM-CM5, HadGEM2-CC and MIROC-ESM models, the greatest RH reduction (~ 10%) occurred in the North and Center-West of Brazil projected throughout the year, and in the Northeast of Brazil during JJA and SON in the MRI-CGCM3 model.By contrast, the IPSL-CM5A-MR model did not show signi cant trends towards increased or decreased RH towards the end of the 21st century.The R s projections were different among the six GCMs (SM.8).The CanESM2, HadGEM2-CC, MIROC-ESM and MRI-CGCM3 models projected increased R s at around 3 MJ m -2 day -1 , mainly in the North and Northeast of Brazil.The CNRM-CM5 model did not show any signi cant increase (or reduction) for the end of the 21st century.The IPSL-CM5A-MR model, on the other hand, showed a reduction at 1.5 MJ m-2

Figure 4 Seasonal 5
Figure 4 al. 2016; Dewes et al. 2017; Jerszurki et al. 2019; Monteiro et al. 2021).Plants dissipate heat into the atmosphere via ET o to keep their plant tissue temperatures at appropriate levels for their metabolisms (Devi and Reddy 2018; Abreu et al. 2022).The higher temperatures and irregular precipitation patterns that are projected for Brazil (IPCC, 2013, 2021) are expected to affect ET o (Valipour et al. 2017; Wang et al. 2018), since increases in air temperature tend to increase evapotranspiration, and because precipitation controls the amount of soil water available.Plants lose water to the atmosphere at higher rates with increased evapotranspiration (Santos et al. 2017), and this can impact certain crops and make cultivating them unfeasible if water availability is not adequate (Ramirez-Cabral et al. 2017; Tavares et al. 2018; Elli et al. 2020).It can also reduce crop productivity and quality (Heinemann et al. 2017; Tironi et al. 2017; Fraga et al. 2019).Authors like Wang et al. (2007) and Zhang et al. (2015) reported opposite trends in the so-called 'evaporation paradox', where increases in air temperature reduced evapotranspiration.Given the 'evaporation paradox', global increases in air temperature may not result in increased evapotranspiration (Liu et al. 2018), if there are combined in uences from variations (increases/decreases) in other metrological variables like wind speed, relative humidity, and precipitation (Fan et al. 2016).Since there are uncertainties as to the aforementioned trends, the contribution of these evapotranspiration-altering meteorological variables needs to be studied and analyzed individually (Zhang et al. 2015; Liu et al. 2018; Monteiro et al. 2021), to better understand evapotranspiration patterns under climate change scenarios (Gondim et al. 2018; Moses and Hambira 2018).Although the ET o variable is sensitive to projected climate changes (Dewes et al. 2017; Valipour et al. 2017; Liu et al. 2018; Jerszurki et al. 2019; Llopart et al. 2020), it is not directly available in climate model databases, e.g., those belonging to the Coupled Model Intercomparison Project (CMIP), given its nature and complexity, making it di cult to study under climate change scenarios (Valipour et al. 2017).Therefore, from ET o estimation methods that best represent current climate conditions (Monteiro et al. 2021), it is necessary to verify the spatiotemporal ET o patterns under future climate conditions.Furthermore, there are few studies on projected evapotranspiration changes for Brazil, and the ones that do exist are limited in that i) they are restricted to a single regional (or local) scale(Lyra etal.2017; Gondim et al. 2018; Santos et al. 2019; Sousa et al. 2019); and/or ii) they use too few climate models (Pan et al. 2015; Guimarães et al. 2016; Jerszurki et al. 2019) or too simpli ed water balance models (Llopart et al. 2020) in their analyses.Therefore, these studies may contain inconsistencies in assessing climate change impacts on ET o and do not represent all of Brazil.
Xavier et al. (2016)and Ab methods.The data provided byXavier et al. (2016)have horizontal resolution equal to 0.25° latitude/longitude covering all of Brazil.The six GCMs used in this study (speci ed in Table1) belong to the Coupled Model Intercomparison Project Phase 5 (CMIP5), from the World Climate Research Program, made available via the Earth System Grid data portal (https://esgf-data.dkrz.from/search/cmip5-dkrz/).The GCMs have horizontal resolution ranging from 1.1° to 2.8° latitude/longitude, which were later interpolated to the 0.25° grid to compare with GWD (Table1 (GCMs) from 1980-2005 was performed by comparing the ET o calculated from the observed data spatialized to grid points (GWD), provided by Xavier et al. (2016) (https://utexas.app.box.com/v/Xavieretal-IJOC- ).
). Statistical bias was used to quantify how well the GCMs simulate ET o , as per: 3 Where are the ET o values obtained from the GCMs data; are the observed ET o values obtained from gridded weather dataset provided by Xavier et al. (2016); and is the number of daily observations (from 1980 to 2005).The bias was calculated considering Tu and Ab methods.
Fan et al. 2016;Gao et al. 2017;Lin et al. 2018d temporal patterns expected in the signal change (increase), since the projected increases in air temperature and reductions in relative humidity should lead to increased ET o (Lemos Filho et al. 2010; Santos et al. 2017; Jerszurki et al. 2019).With respect to spatial and temporal patterns, results released by the IPCC (2013, 2021), Torres and Marengo (2014) and Torres et al. (2021) proved that increases in temperature will be more intense in the North, Northeast and Center-West of Brazil, and there will be different precipitation pattern changes, which will be negative (positive) in the Northeast (South).This indicates greater ET o increases in the North, Northeast, and Center-West of Brazil, and possibly lesser increases in the South of Brazil, as demonstrated using both methods (Figures3-6).Some authors (e.g.,Fan et al. 2016;Gao et al. 2017;Lin et al. 2018) emphasize that impacts to evapotranspiration arise from interactions between climatic factors and local conditions, e.g., type of vegetation cover, and the impacts of human activities.Such factors increase uncertainties with respect to the contribution that each variable, both above and below ground(Ruosteenoja etal.2018; Monteiro et al.The analyses were carried out with the ensemble mean and individually to analyze the response of each model, despite MRI-CGCM3 being slightly superior in simulating ET o for historical period. This study analyzed the reference evapotranspiration at the end of the 21st century, using two ET o estimation methods (Tu and Ab), two radiative forcing scenarios (RCPs 4.5 and 8.5), and six climate models from the CMIP5 (CanESM2; CNRM-CM5; HadGEM2-CC; IPSL-CM5A-MR; MIROC-ESM; MRI-CGCM3).