Evapotranspiration under Drought Conditions: The Case Study of a Seasonally Dry Atlantic Forest

Information on evapotranspiration (ET) has the potential to clarify drought’s effects on the water balance of natural ecosystems. Here, we use a 6-year dataset to present daily ET trends under different drought conditions in a seasonally dry Atlantic Forest in southeast Brazil as well as environmental and biophysical controls. Generalized linear models (GLMs) were applied to highlight the main controls on ET. Significant differences for ET were not found under different drought conditions (near normal, moderately dry, and severely dry). ET responded positively to net radiation (Rn), bulk surface (gs), and aerodynamic (ga) conductance. Drought severity and soil water storage (SWS) did not significantly affect ET. We attributed the regularization of ET to (i) the stability in the SWS observed in the study site (dystrophic red latosol); (ii) the tree species’ adaptations to cope with water stress; (iii) the alternation between droughts and near-normal conditions (which have increased the water in the system on an annual basis); and (iv) the monthly variations in water inputs within the hydrological years. Finally, our study sheds light on the ecosystem characteristics that may represent sources of resilience when facing the droughts predicted in climate change scenarios.


Introduction
Evapotranspiration (ET) is the integration of transpiration and surface evaporation. Transpiration is the water exchange with the atmosphere when the vegetation takes up carbon dioxide for photosynthesis [1]. Evaporation occurs from the soil surface and the wet canopy following the energy gradient [1]. ET is a key process relating the surface to the atmosphere by coupling the water and energy exchanges. Therefore, local and regional climates are driven by these exchanges of mass and energy [2]. Changes in rainfall pattern, energy and soil moisture availability, temperature increase, and other aspects related to climate change [3] can misbalance these water and energy fluxes in natural ecosystems. Therefore, understanding the impact of stressful weather conditions (e.g., droughts) is essential for improving the management of natural resources in a changing climate.
Biomes respond differently whether droughts happen in humid, semi-arid, or arid climates, i.e., evapotranspiration can remain the same, decline, or cease depending on whether the biome has previously faced dry conditions. Biomes that naturally face dry periods have developed mechanisms to overcome drought effects. For instance, the Brazilian savanna was less sensitive to droughts than the humid Amazonia [4]. The humid Amazonia had a negative response because it is a normally wet ecosystem, and therefore, it

Data Monitoring and Study Period
There are two meteorological stations in the SDAF (Figure 1). One placed on the top of a 22-m-high meteorological tower (i.e., above the canopy) and the other inside the remnant (Figure 1). Both meteorological stations are equipped with a 0.20-mm-resolution tipping bucket rain gauge (Campbell Scientific CR10X), Wind Speed and Direction Set Smart Sensor (S-WSET-A), Silicon Pyranometer (S-LIB-M003), and 12-bit Temperature/Relative Humidity Smart Sensor (S-THB-M002) to measure rainfall, wind speed and direction, net radiation, and air temperature/relative humidity, respectively. These meteorological variables were monitored in a 10 min time step, integrated to one hour, and stored in a Hobbo datalogger.
The study period spans from 2013 to 2019, encompassing six hydrological years. The The studied region is characterized by well-defined seasons (Cwa Köppen-type climate) with rainfall concentrated in the summer (December to March) [35]. The long-term average annual rainfall considering the period between 1991 and 2020 is 1383.4 mm, with 85% occurring in the wet period between October and March [36]. The annual mean air temperature is 20.6 • C, ranging from 17.2 • C (June and July) to 23 • C in February.

Data Monitoring and Study Period
There are two meteorological stations in the SDAF (Figure 1). One placed on the top of a 22-m-high meteorological tower (i.e., above the canopy) and the other inside the remnant ( Figure 1). Both meteorological stations are equipped with a 0.20-mm-resolution tipping bucket rain gauge (Campbell Scientific CR10X), Wind Speed and Direction Set Smart Sensor (S-WSET-A), Silicon Pyranometer (S-LIB-M003), and 12-bit Temperature/Relative Humidity Smart Sensor (S-THB-M002) to measure rainfall, wind speed and direction, net radiation, and air temperature/relative humidity, respectively. These meteorological variables were monitored in a 10 min time step, integrated to one hour, and stored in a Hobbo datalogger.
The study period spans from 2013 to 2019, encompassing six hydrological years. The hydrological year in the studied region starts in October and ends in September of the next year (2013-2014, 2014-2015, 2015-2016, 2016-2017, 2017-2018, and 2018-2019). The authors of [37] applied the Standard Precipitation Index (SPI) to assess the meteorological conditions throughout these years. They observed anomalous meteorological conditions, which affected the canopy interception [6], water balance [38], and rainfall redistribution [37] of the SDAF. According to the World Meteorological Organization [39] and SPI values calculated by [37], the hydrological years were classified as near normal (NN), moderately dry (MD), and severely dry (MD) (Figure 2). These three meteorological conditions were the bases for the assessment of the behavior of evapotranspiration (ET) and how the SDAF responds to droughts.

Daily Evapotranspiration
The latent heat flux (LE) was calculated on an hourly step using the energy balance equation: . G was set to zero because of its small contribution observed in tall and complex forests [6,40]. H was calculated considering the temperature gradient between the AFR and the air above the canopy [4]: where ρ is the dry air density [kg m −3 where the resultant ET is in mm d −1 .

Aerodynamic Conductance
There are different forms to calculate ga for tropical forests [41]. In this study, we used the Monin-Obukhov similarity theory [20] considering neutral stability conditions of the atmosphere [1]. This approach is recommended for tropical forests that have only wind speed monitoring [41]: Graphical representation of the droughts' classification of the hydrological years according to SPI [37] in the SDAF. SD = severely dry; MD = moderately dry; NN = near normal.
These three meteorological conditions were the bases for the assessment of the behavior of evapotranspiration (ET) and how the SDAF responds to droughts.

Daily Evapotranspiration
The latent heat flux (LE) was calculated on an hourly step using the energy balance equation: . G was set to zero because of its small contribution observed in tall and complex forests [6,40]. H was calculated considering the temperature gradient between the AFR and the air above the canopy [4]: where is the dry air density [kg m −3 where the resultant ET is in mm d −1 .

Aerodynamic Conductance
There are different forms to calculate g a for tropical forests [41]. In this study, we used the Monin-Obukhov similarity theory [20] considering neutral stability conditions of the atmosphere [1]. This approach is recommended for tropical forests that have only wind speed monitoring [41]: where k is the von Kárman's constant set as 0.41, u is the wind velocity above the canopy [m s −1 ], z is the height of wind velocity monitoring [22.0 m], Z 0M and Z 0H are the roughness length for momentum and heat transfer, respectively, and d is the zero-plane displacement height [m]. The calculation of Z 0M , Z 0H , and d followed the recommendations of [42].

Bulk Surface Conductance
Besides the aerodynamic conductance (g a ), another parameter defining the rate of water transferring from the forest to the atmosphere is the bulk surface conductance (g s ). This parameter is closely related to stomatal dynamics and characteristics and describes the physiological control on ET [9,41]. The g s parameter can be calculated by inverting the Penman-Monteith equation as detailed by [43] where ∆ is the slope of the saturation vapor pressure curve [kPa , VPD is the vapor pressure deficit [kPa], and g s is computed in m s −1 .

Soil Water Storage
Soil moisture was monitored in 32 points across the SDAF ( Figure 1) using a multisensor capacitance probe (Profile Probe PR2/6 model, Delta-T Devices, Cambridge, UK). This monitoring was performed at five depths (0.10, 0.20, 0.30, 0.40, and 1.00 m) either on a biweekly or monthly frequency, depending upon weather conditions. The depth information was integrated to account for soil water storage (SWS) in the 1.0 m profile: where i is the depth of soil moisture monitoring, n is the number of depths, θ is the soil moisture [cm 3 cm −3 ], h is the layer thickness between i and i + 1 (mm), and SWS is in mm. The SWS of the AFR was taken as the average across the 32 points.

Statistical Analysis
In order to verify the individual effects of the weather conditions (near normal, moderately dry, and severely dry), stomatal conductance (g s ), aerodynamic conductance (g a ), net radiation (R n ), vapor pressure deficit (VPD), and soil water storage (SWS) on ET (objective ii), we organized the data in average monthly values. The data aggregation in the monthly average was an alternative approach to deal with the data gaps in daily ET and to match the soil moisture monitoring scale. Then, we conducted exploratory analyses with (i) descriptive statics (mean, standard deviation, and coefficient of variation) of ET and explanatory variables; (ii) Pearson correlation matrix to assess linear correlation among the variables; and (iii) histogram and Shapiro-Wilk test for checking the normality of ET ( Figure S1). The effects of each explanatory variable on ET were assessed by means of a generalized linear model (GLM). Therefore, we fitted a GLM (family = Gaussian) to assess the effects of the above-mentioned explanatory variables on ET (global model).
ET~Period + g a + g s + R n + VPD + SWS Additionally, we conducted a multimodel analysis from the global model based on Akaike information criterion (AIC). The best models select according to the AIC were the models with ∆AICc < 2 considered equally supported [44]. The final model corresponded to the average of the coefficients of the set of pre-selected models [45]. The models were fitted using the function glm and the MuMIn package [46] for R [47].

Priestley-Taylor and Decoupling Coefficients
The evapotranspiration is driven by both environmental and physiological conditions [9]. Assessing these controls is important to point the drivers of ET and highlight limiting factors, mainly during stressful weather conditions (such as droughts). Two coefficients that are usually applied [5,11,17,19] are (i) the Priestley-Taylor coefficient (ET/ET eq ) and (ii) the decoupling coefficient (Ω). ET/ET eq highlights whether energy availability or soil moisture controls ET. ET ET eq is known as the equilibrium evapotranspiration because only the available energy and air temperature drive it [11]. For ET/ET eq ≥ 1, ET is controlled by the available energy (R n ), i.e., soil moisture is not limiting ET. On the other hand, ET/ET eq < 1 indicates the soil moisture as the limiting factor of ET [48].
The decoupling coefficient (Ω) was calculated according to [49] where ∆ is the slope of the saturation vapor pressure curve [kPa . Ω ranges from 0 to 1: When Ω → 1, the canopy is decoupled from the atmosphere above it, and ET is controlled by R n and g s . When Ω → 0, the canopy is strongly coupled to the atmosphere. To be coupled means that the vapor pressure within the stomata equals that in the atmosphere [11], i.e., ET is controlled by how g s responds to VPD.

Gross Rainfall Patterns
The long-term time series used by [37] to calculate the SPI was applied to appraise the patterns of monthly gross rainfall (GR) in the study period ( Although the meteorological drought was properly classified by [37], the deficit observed in each hydrological year was due to different contributions throughout the months ( The long-term time series used by [37] to calculate the SPI was applied to appraise the patterns of monthly gross rainfall (GR) in the study period (  Although the meteorological drought was properly classified by [37], the deficit observed in each hydrological year was due to different contributions throughout the

Energy Balance
Although gaps were observed in the meteorological time series (Figure 4), they did not preclude the energy balance analyses because the energy components were evenly distributed in the near normal, moderately dry, and severely dry periods.

Daily Evapotranspiration (ET)
The seasonal behavior of ET followed the energy availability (Figures 4 and 5). ET reached up to 14 mm d −1 in the wet period and decreased to 0.01 mm d −1 in the dry period. ET values were more grouped during the dry periods up to October, when the  Daily R n ranged from minimal values (~0 W m −2 ) in the dry period (April to September) to values up to 480 W m −2 in the wet period (October to November) regardless of the drought condition ( Figure 4). Minimal values of R n were also observed in the wet period due to cloudy days. LE ranged from 0 to 400 W m −2 following the temporal behavior of R n . This highlights the importance of energy availability for forest evapotranspiration in the study period. On the other hand, most of the H was between 0 and 50 W m −2 , with some peaks up to 150 W m −2 .

Daily Evapotranspiration (ET)
The seasonal behavior of ET followed the energy availability (Figures 4 and 5). ET reached up to 14 mm d −1 in the wet period and decreased to 0.01 mm d −1 in the dry period. ET values were more grouped during the dry periods up to October, when the rainfall begins. After that, cloudy and sunny days increased the variability of ET in the wet period ( Figure 5). ET ranged from 0.12 to 14 mm d −1 in the near normal and from 0.01 to 12 mm d −1 in the moderately and severely dry conditions ( Figure 6). On average, ET was 5.4 ± 2, 4.6 ± 2.1, and 5.6 ± 3.0 mm d −1 in the near normal, moderately dry, and severely dry, respectively. The ET pattern was almost the same between the periods with smaller amplitude  ET ranged from 0.12 to 14 mm d −1 in the near normal and from 0.01 to 12 mm d −1 in the moderately and severely dry conditions ( Figure 6). On average, ET was 5.4 ± 2, 4.6 ± 2.1, and 5.6 ± 3.0 mm d −1 in the near normal, moderately dry, and severely dry, respectively. The ET pattern was almost the same between the periods with smaller amplitude in the moderately dry year. Despite the difference in the rainfall patterns between the droughts [37], there was no difference in the daily VPD ( Figure S2). Daily VPD ranged from 0.06 to 1.82 kPa in the near-normal conditions (Figure 7) (average of 0.73 ± 0.32 kPa). It ranged from 0.04 to 1.87 kPa in the moderately dry conditions (average of 0.61 ± 0.27 kPa); and from 0.05 to 1.94 kPa in the severely dry conditions (average of 0.77 ± 0.34 kPa). The highest values of daily VPD were observed during the wet period and in the end of the dry period.
The temporal pattern of daily Tair was the same throughout the study period ( Figure  7). Tair followed the energy availability ( Figure 4) and attained its maximum (28.4 °C) and minimum (11.9 °C) values in the wet and dry periods, respectively. Daily Tair averaged 22.0 ± 1.1 °C in the wet period and 19.1 ± 1.8 °C in the dry period. The same behavior was not observed for RH since air saturation and dryness occurred in both the wet and dry periods (Figure 7). This pattern is a response to the rainfall occurrence and the variability of Tair. RH ranged from 41.8% to 98.3% and averaged 71.2% ± 10% throughout the study period. Despite the difference in the rainfall patterns between the droughts [37], there was no difference in the daily VPD ( Figure S2 Values of ga ranged from 3.4 to 40.9 m s −1 and averaged 16.9 ± 6.3 m s −1 in the study period. There was no difference in the temporal distribution of ga between the drought conditions and the wet and dry periods. Minimum and maximum gs were ~0 and ~100 m s −1 , respectively, regardless of the drought condition. These values were observed in both The temporal pattern of daily T air was the same throughout the study period (Figure 7). T air followed the energy availability ( Figure 4) and attained its maximum (28.4 • C) and minimum (11.9 • C) values in the wet and dry periods, respectively. Daily T air averaged 22.0 ± 1.1 • C in the wet period and 19.1 ± 1.8 • C in the dry period. The same behavior was not observed for RH since air saturation and dryness occurred in both the wet and dry periods (Figure 7). This pattern is a response to the rainfall occurrence and the variability of T air . RH ranged from 41.8% to 98.3% and averaged 71.2% ± 10% throughout the study period.
Values of g a ranged from 3.4 to 40.9 m s −1 and averaged 16.9 ± 6.3 m s −1 in the study period. There was no difference in the temporal distribution of g a between the drought conditions and the wet and dry periods. Minimum and maximum g s were~0 and 100 m s −1 , respectively, regardless of the drought condition. These values were observed in both the dry and wet periods of every hydrological year (October to September) (Figure 7). There was a scattering of g s in the wet period due to the temporal variability of water and energy availability. On the other hand, g s values were more grouped in the end of the dry season as a consequence of few rainfall events. This indicated some seasonality of g s in the SDAF. On average, g s was 24.41 ± 20.45 m s −1 .
SWS ranged from 178 to 319 mm, from 179 to 314 mm, and from 178 to 340 mm in the near normal, moderately dry, and severely dry, respectively ( Figure 6). On average, SWS was 242 ± 41, 229 ± 35, and 256 ± 38 mm in the near normal, moderately dry, and severely dry, respectively. Although wetter soil conditions were observed in the severely dry years (Figure 6), the minimum values were almost the same between the periods.

Environmental and Biophysical Controls of ET
The descriptive analyses of the monthly daily average data pointed to g a and g s as the variables with the highest and the lowest coefficient of variation in our dataset, respectively (Table 1). Moreover, a higher Pearson coefficient of correlation was observed between ET and R n , and the lowest correlation was observed between ET and SWS ( Figure 8). Table 1. Descriptive analyses of the monthly daily average of evapotranspiration (ET), bulk surface conductance (g s ), aerodynamic conductance (g a ), net radiation (R n ), vapor pressure deficit (VPD), and soil water storage (SWS) for the 2013-2019 period in a seasonally dry Atlantic tropical forest. Dataset used in the statistical modeling. The global model pointed to g s , g a , R n , and VPD as significant explanatory variables of ET (Table 2). There was no difference in ET between the weather conditions (near normal, moderately dry, and severely dry). There was also no significant effect of SWS on ET variability.

Mean Standard Deviation Coefficient of Variation
In accordance with the global model, multimodel inference selected only one model (AIC < 2) containing g s , g a , Rn, and VPD as significant explanatory variables (Figure 9) (see the rank of models and the coefficients of the final model in Tables S1 and S2). All explanatory variables had positive effects on ET, notably R n , which had a relatively higher influence on ET.  The global model pointed to gs, ga, Rn, and VPD as significant explanatory variables of ET (Table 2). There was no difference in ET between the weather conditions (near normal, moderately dry, and severely dry). There was also no significant effect of SWS on ET variability.

Priestley-Taylor and Decoupling Coefficients
The Priestley-Taylor coefficient (ET/ET eq ) was basically constant throughout the study period ( Figure 5), with an average of 1.20 ± 0.14. ET/ET eq > 1 highlighted the importance of energy availability and air temperature for the daily ET regardless of the weather condition. The soil moisture became a limiting factor only in a few cases when ET/ET eq < 1. In such cases, the energy radiation was basically converted to sensible heat (H) (Figure 4).
The decoupling coefficient (Ω) presented some seasonality with greater values in the wet period ( Figure 5). However, the average Ω of 0.77 ± 0.02 indicated a small temporal variability. In this sense, the canopy was decoupled from the atmosphere (2013-2019) because it ranged between 0.72 and 0.82 regardless of the weather condition.

SWS
0.011348 0.072357 0.157 0.877338 In accordance with the global model, multimodel inference selected only one model (AIC < 2) containing gs, ga, Rn, and VPD as significant explanatory variables (Figure 9) (see the rank of models and the coefficients of the final model in Tables S1 and S2). All explanatory variables had positive effects on ET, notably Rn, which had a relatively higher influence on ET.

Priestley-Taylor and Decoupling Coefficients
The Priestley-Taylor coefficient (ET/ETeq) was basically constant throughout the study period ( Figure 5), with an average of 1.20 ± 0.14. ET/ETeq > 1 highlighted the importance of energy availability and air temperature for the daily ET regardless of the weather condition. The soil moisture became a limiting factor only in a few cases when ET/ETeq < 1. In such cases, the energy radiation was basically converted to sensible heat (H) (Figure 4).
The decoupling coefficient (Ω) presented some seasonality with greater values in the wet period ( Figure 5). However, the average Ω of 0.77 ± 0.02 indicated a small temporal

Energy Balance and Daily Evapotranspiration
The temporal dynamics of LE followed R n indicating that most of the energy availability is converted to evapotranspiration (Figure 4). This pattern is mainly observed in biomes in which soil water is not a limiting factor [50,51]. The contribution of H increases when the system changes from energy limited to water limited [5]. This was observed in the Brazilian savanna during the dry season, when the variability of LE was less correlated to R n as soil water storage became a limiting factor [4]. The importance of H increased in a semi-evergreen tropical forest in Vietnam during a hot-dry year [50]. The greater importance of H was observed in a few days in the SDAF throughout the study period ( Figure 4) and is likely related to rainless, dry, and hot days.
The daily evapotranspiration of the SDAF was in accordance with the values observed by [52] in an Atlantic Forest in the Mantiqueira range in southeastern Brazil. ET ranged from 1.53 mm d −1 (dry season) to 5.1 mm d −1 (wet season) in the Mantiqueira range, with an average of 3.26 mm d −1 . The authors of [50] observed values of ET ranging from 3.64 ± 1.02 mm d −1 in the dry season to 4.65 ± 1.02 mm d −1 in the wet season in a semi-evergreen tropical forest in Vietnam. Moreover, the authors of [53] observed ET values ranging from 0.3 to 6.7 mm d −1 in the Amazon River basin. The lowest values are associated with cold weather and sparse vegetation in the Andes Cordillera, whereas the greatest ET occurred in the lowland regions characterized by abundant rainfall and tropical moist forest [53]. On the other hand, ET was lower than 2.1 mm d −1 in a semi-arid alpine step in the Tibetan Plateau [17]. Moreover, ET ranged from 0.8 o 1.3 mm d −1 in the semi-arid Caatinga in Brazil [5]. These results highlight the importance of water and energy availability for daily evapotranspiration since tropical and wet climates present greater ET.

Environmental and Biophysical Controls
ET is controlled either by the environment or by physiology and is defined by interand intra-annual weather variability [4], energy availability [30,54], soil moisture [4,5], soil temperature [21], vapor pressure deficit [18], stomatal dynamics [17], root depth [14], vegetation adaptation [17,19,38,50], and canopy aerodynamics [41]. Energy availability (R n ) was the main driver of ET in the seasonally dry Atlantic Forest (SDAF) (Figure 8). R n was also the main driver of ET in arid ecosystems in China [30] and Brazil [54]; tropical evergreen forest [54], savanna [4], and wetland [51] in Brazil; and tropical rainforest in French Guiana [21], when SWS was not a limiting factor. SWS and weather condition (NN, MD, and SD) were not explanatory variables of ET in the SDAF (Figure 8). This highlights that R n was the limiting factor of ET regardless of the drought severity.
This pattern was confirmed by the values of ET/ET eq , which indicated that ET was controlled by R n (ET/ET eq > 1) and not by soil moisture (ET/ET eq < 1) (Figure 4). The importance of VPD and g a (Figure 9) also highlights the SWS availability, i.e., the greater the atmospheric demand (VPD) or canopy aerodynamics (g a ), the greater ET. Values of Ω close to one highlighted the importance of R n and g s to drive ET variability. g s was neither controlled by SWS nor VPD (Figure 8) since the canopy was decoupled from the atmosphere (Ω~1), i.e., ET was not limited by high atmosphere demand because the soil (latosol) was able to provide water even during dry years ( Figure 6). The importance of g s (Figure 9) was due to the stomatal response to the temporal variability of rainfall and energy availability (cloudy/sunny days and dry/wet periods).
The canopy coupled with the atmosphere in the dry season in the arid Caatinga in Brazil [5] highlights the control of stomatal closure (indicated by g s ) in ET due to high VPD and low SWS. This was also observed in the Brazilian savanna [4], in the semi-arid alpine ecosystem in the Tibetan Plateau [17], and in a semi-evergreen tropical forest in Vietnam [50]. The stomatal closure is the first mechanism activated by vegetation to avoid water loss to the atmosphere [55]. This is followed by leaf loss and tree mortality [56] in a delayed response to persistent droughts [57]. Other mechanisms have been developed by some species to overcome droughts such as root water uptake in deep layers [14], small leaves [58], and deciduousness [59]. Such adaptations are more observed in vegetation that has naturally dry periods because increased tree mortality and a greater recovery time were observed after the 2010 drought in the humid Amazon Forest compared to a semiarid biome in Brazil [54], i.e., humid forests suffer more with droughts because water is normally abundant. However, other ecosystems worldwide, which are somehow adapted to seasonally dry periods, had difficulties in maintaining ET in drought conditions [11,50]. These difficulties were not observed in the SDAF because of the latosol's role in this ecosystem.
The latosol of the SDAF is a weathered soil with granular structure and micro-porosity, which confers to it a high storage capacity and infiltrability [34,38]. These capacities are boosted in the upper layers by organic matter, which was crucial for infiltration and water redistribution in the SDAF [34]. Local topography is another important factor driving soil moisture in the SDAF [34,60]. The gentle slope together with the high infiltration capacity decreased the runoff generation, which was not observed in fields (even under intense events) nor in simulations of the water balance [38]. These characteristics were crucial for maintaining ET values because SWS was stable throughout the study years (2013-2019) ( Figure 6). In contrast, the Rhodic Cambisol was unable to support ET in a semi-evergreen tropical forest in Vietnam, where up to 50% of the trees lose their leaves to overcome dry periods [50]. Cambisol is shallow and linked to step topography [52,61], and therefore, it is unable to provide water as the studied latosol does.
Besides soil characteristics, the temporal distribution of rainfall and the adaptations of trees can be highlighted as other factors regularizing ET between the weather conditions. The SPI accounts for meteorological drought, which is not promptly transferred to the soil. ET and soil water redistribution are relatively slow processes [54], and it takes time for the SWS to respond to the meteorological drought. Moreover, the rainfall of some months was close to the long-term average in the moderately and severely dry years (Figure 3), which provided water inputs to the SDAF, recharging the soil. The near-normal years also contributed to maintaining soil moisture in droughts, as they represent eventual recharges of the system. This resulted in the same SWS between NN, MD, and SD years ( Figure 6). Regarding the forest adaptations, more than 50% of the trees in our forest lose their leaves during the dry season [62], which naturally reduces ET on a seasonal basis and potentially make the system more resilient to drought. Moreover, the SDAF's most abundant tree species, Xylopia brasiliensis Spreng. and Copaifera langsdorffii Desf., hold dry-affiliated traits, such as small leaves and deciduousness [24], also favoring drought-coping. Summing up, these adaptations of some tree species, the holding water capacity of the latosol, the presence of months with enough rainfall to recharge the soil, and near-normal years among droughts seem to be responsible for keeping ET only limited to energy availability ( Figure 9).

Shortcomings and Implications
This study was performed considering the temporal availability of the data. Therefore, different time steps were taking depending on the data used. For instance, the coefficients (ET/ET eq and Ω) and the energy balance could be assessed in a daily step. However, the GLM was applied on the monthly averaged daily values of ET, g s , g a , and VPD because of the availability of SWS. Soil moisture was available on a monthly step because it was manually monitored. This precluded the analyses of the temporal dynamics of soil moisture and ET. However, soil moisture has been a driver of ET only when it becomes a limitation for root water uptake [4,17]. This was not the case in the SDAF since soil moisture became a limitation for ET (ET/ET eq < 1) only in a few moments throughout the study period ( Figure 5). This reinforces the results of the GLM since daily ET follows the temporal behavior of R n (Figures 4 and 5). It may be interesting to decrease the temporal monitoring of soil moisture as droughts become more frequent, intense, and lasting. SWS can start being a limiting factor of ET in these conditions. Moreover, the drought was classified considering the rainfall deficit on a hydrological-year scale. The temporal distribution of rainfall within the hydrological year is another important driver of ET as one can observe different behaviors of ET and its attributes in the same drought categories (Figures 5 and 7). In this sense, future studies should take into consideration the temporal distribution of GR since it is likely the cause of the different behavior of ET and its attributes in the same drought categories (NN, MD, and SD). For instance, the differences in ET ( Figure 5) in the SD years (2013-2014 and 2017-2018) are likely due to the temporal distribution of GR (Figure 3), which caused different dynamics of soil moisture. However, it does not preclude the results of this study because we aimed to assess the overall condition of ET in different drought severities.
Despite the shortcomings, the importance of the water holding capacity of the latosol, monthly rainfalls to recharge the soil, and the adapted mechanism of some trees (such as small leaves and deciduousness) for regulating ET during droughts could be assessed with the available dataset. This was possible due to the time series used (six hydrological years) of daily information and a robust methodology (GLM and coefficients) to extract the most information from the dataset.
In this way, our study sheds light on the ecosystem's characteristics that may represent sources of resilience when facing the droughts predicted in climate change scenarios. We could observe that energy availability is the great limiting factor for ET in the SDAF regardless of the drought severity. This means that the SDAF can cope with some stressful weather conditions expected in climate change scenarios. This is likely due to the soil (latosol) capacity to store water and the physiological adaptation of some trees.

Conclusions
ET is driven by different environmental and physiological characteristics, which can become limiting factors depending on the weather condition. Droughts can change the dynamics of ET by limiting water availability. On the other hand, energy availability (R n ) can limit ET when SWS is enough. Both the holding water capacity and the gentle topography of the dystrophic red latosol were crucial to regularizing ET in the SDAF. The values of ET/ET eq > 1 and Ω~1 highlighted R n as the limiting factor of ET regardless of the drought severity. These results were confirmed by the observed values of SWS and the results of the GLM, which did not differ between the weather conditions. Moreover, some tree species can cope with different drought severities due to adapted mechanisms (such as deciduousness, leaf size, stomatal control, and deep water uptake). These mechanisms were developed to deal with the naturally dry periods faced by these forests. However, this ability to maintain ET in droughts was not observed in other seasonally dry forests where the soil was shallow, poorly structured, and step (e.g., Cambisol). Therefore, although physiological mechanisms are crucial for regulating water loss in dry conditions, the soil's characteristics (such as holding water capacity and infiltrability) are fundamental to maintaining the ET even during severe droughts.
This study sheds light on the importance of the soil's characteristics in a seasonally dry Atlantic Forest to deal with dry weather. Therefore, the decisions about the areas of the Atlantic Forest biome to be preserved (and regenerated) should take the soil into consideration, because it may define the success of the conservation area under climate change scenarios.
The main conclusions of this study can be summarized in four characteristics that drive ET dynamics: • Soil: the holding water and the infiltration capacities of the dystrophic red latosol were crucial to sustaining ET even in severe drought conditions. • Weather: R n was the limiting factor of ET in the near-normal, moderately dry, and severely dry hydrological years. • Physiological mechanisms: deciduousness and small leaves were crucial for regulating water loss in dry conditions (favoring drought-coping).

•
Rainfall distribution: the inter-and intra-annual variability of gross rainfall favored soil recharge and ET.
Supplementary Materials: The following supporting information can be downloaded at https: //www.mdpi.com/article/10.3390/atmos13060871/s1: Figure S1: Histogram of ET and Shapiro-Wilk normality test attesting normality. Figure S2: Box plots of daily R n , g a , g s , and VPD in a seasonally dry Atlantic Forest under different drought periods (NN = near normal; MD = moderately dry; SD = severely dry). Table S1: GLM models of ET and the likely explanatory variables R n , g a , g s , SWS, VPD, and period. Table S2: Explanatory variables of ET in the final GLM model. Informed Consent Statement: Not applicable.

Data Availability Statement:
The data presented in this study are available on request from the corresponding author.

Conflicts of Interest:
The authors declare no conflict of interest.