Assessment of an Automated Calibration of the SEBAL Algorithm to Estimate Dry-Season Surface-Energy Partitioning in a Forest–Savanna Transition in Brazil

: Evapotranspiration ( ET ) provides a strong connection between surface energy and hydrological cycles. Advancements in remote sensing techniques have increased our understanding of energy and terrestrial water balances as well as the interaction between surface and atmosphere over large areas. In this study, we computed surface energy ﬂuxes using the Surface Energy Balance Algorithm for Land (SEBAL) algorithm and a simpliﬁed adaptation of the CIMEC (Calibration using Inverse Modeling at Extreme Conditions) process for automated endmember selection. Our main purpose was to assess and compare the accuracy of the automated calibration of the SEBAL algorithm using two di ﬀ erent sources of meteorological input data (ground measurements from an eddy covariance ﬂux tower and reanalysis data from Modern-Era Reanalysis for Research and Applications version 2 (MERRA-2)) to estimate the dry season partitioning of surface energy and water ﬂuxes in a transitional area between tropical rainforest and savanna. The area is located in Brazil and is subject to deforestation and cropland expansion. The SEBAL estimates were validated using eddy covariance measurements (2004 to 2006) from the Large-Scale Biosphere-Atmosphere Experiment in the Amazon (LBA) at the Bananal Java é s (JAV) site. Results indicated a high accuracy for daily ET , using both ground measurements and MERRA-2 reanalysis, suggesting a low sensitivity to meteorological inputs. For daily ET estimates, we found a root mean square error (RMSE) of 0.35 mm day − 1 for both observed and reanalysis meteorology using accurate quantiles for endmembers selection, yielding an error lower than 9% (RMSE compared to the average daily ET ). Overall, the ET rates in forest areas were 4.2 mm day − 1 , while in grassland / pasture and agricultural areas we found average rates between 2.0 and 3.2 mm day − 1 , with signiﬁcant changes in energy partitioning according to land cover. Thus, results are promising for the use of reanalysis data to estimate regional scale patterns of sensible heat ( H ) and latent heat ( LE ) ﬂuxes, especially in areas subject to deforestation.

to manual calibration [55], especially for time-series analysis [54]. However, there is still a need to refine the endmembers selection process, which will further improve the model accuracy since ET estimations are sensitive to changes in endmembers [55][56][57][58].
Regional applications of the SEBAL are usually based on local meteorological measurements [52,55,59,60], including wind speed (u), relative humidity (RH), and shortwave incident radiation (R s ). On the other hand, large-scale ET estimations can potentially benefit from the use of global, gridded meteorological variables (e.g., reanalysis methods) due to their global spatial coverage and long time-series [17,21,24,[61][62][63]. In the case of reanalysis, data assimilation techniques are employed to combine global climatic models with observed data to improve the accuracy of global meteorological estimates [64][65][66][67]. For instance, Modern-Era Reanalysis for Research and Applications version 2 (MERRA-2) reanalysis was developed by the Global Modeling and Assimilation Office Project, providing consistent long-term meteorological datasets [68].
In the context of assessing techniques of automated ET time-series mapping and using global meteorological and remote sensing datasets, the goals of this study were to investigate how (1) the hot and cold endmembers selection and (2) the use of global reanalysis data as meteorological input impact the accuracy of the SEBAL in estimating the dry-season surface energy and water fluxes, when compared to eddy covariance measurements. We selected a transitional area between tropical rainforest and savanna in the southeastern part of the Amazon as the study area, with an experimental site located in a seasonal floodplain forest-savanna ecosystem. This area is characterized by a climatological 5-month dry season [69,70], and is strongly affected by land cover conversion from forest and savanna to pasture and cropland [71].

Research Overview
To estimate surface energy fluxes, we first ran the SEBAL algorithm using ground measurements (SEBAL-T) and MERRA-2 reanalysis data as meteorological inputs (SEBAL-M). Landsat 5 Thematic Mapper (TM) data were used in both cases as remote sensing inputs. To assess uncertainties related to MERRA-2, we also compared MERRA-2 estimates to meteorological measurements. For validation, SEBAL estimations were assessed using measured radiative and turbulent fluxes. Subsequently, the averages of surface fluxes for each land cover class of the study area were quantified. A flowchart of the methodology is presented in Figure 1. Detailed information about the methodology is presented in the next sections.
Remote Sens. 2020, 12, x FOR PEER REVIEW 3 of 24 when compared to manual calibration [55], especially for time-series analysis [54]. However, there is still a need to refine the endmembers selection process, which will further improve the model accuracy since estimations are sensitive to changes in endmembers [55][56][57][58]. Regional applications of the SEBAL are usually based on local meteorological measurements [52,55,59,60], including wind speed ( ), relative humidity ( ), and shortwave incident radiation ( ). On the other hand, large-scale estimations can potentially benefit from the use of global, gridded meteorological variables (e.g., reanalysis methods) due to their global spatial coverage and long time-series [17,21,24,[61][62][63]. In the case of reanalysis, data assimilation techniques are employed to combine global climatic models with observed data to improve the accuracy of global meteorological estimates [64][65][66][67]. For instance, Modern-Era Reanalysis for Research and Applications version 2 (MERRA-2) reanalysis was developed by the Global Modeling and Assimilation Office Project, providing consistent long-term meteorological datasets [68].
In the context of assessing techniques of automated time-series mapping and using global meteorological and remote sensing datasets, the goals of this study were to investigate how (1) the hot and cold endmembers selection and (2) the use of global reanalysis data as meteorological input impact the accuracy of the SEBAL in estimating the dry-season surface energy and water fluxes, when compared to eddy covariance measurements. We selected a transitional area between tropical rainforest and savanna in the southeastern part of the Amazon as the study area, with an experimental site located in a seasonal floodplain forest-savanna ecosystem. This area is characterized by a climatological 5-month dry season [69,70], and is strongly affected by land cover conversion from forest and savanna to pasture and cropland [71].

Research Overview
To estimate surface energy fluxes, we first ran the SEBAL algorithm using ground measurements (SEBAL-T) and MERRA-2 reanalysis data as meteorological inputs (SEBAL-M). Landsat 5 Thematic Mapper (TM) data were used in both cases as remote sensing inputs. To assess uncertainties related to MERRA-2, we also compared MERRA-2 estimates to meteorological measurements. For validation, SEBAL estimations were assessed using measured radiative and turbulent fluxes. Subsequently, the averages of surface fluxes for each land cover class of the study area were quantified. A flowchart of the methodology is presented in Figure 1. Detailed information about the methodology is presented in the next sections.

Micrometeorological and Eddy Covariance Measurements
The micrometeorological and eddy covariance tower is located at 9 • 49 16.1"S and 50 • 08 55.3"W, in a transitional area between the Amazon and the Cerrado biomes (Figure 2), at the northern boundary of the Bananal island, between the Araguaia and Javaés rivers, over a seasonal forest-savanna ecosystem floodplain. Natural vegetation within the tower footprint is composed mostly of woodland savanna (with trees up to 18 m and sparse bushes) and fragments of wooded savanna (with trees up to 5 m) and grassland, with a 5-month dry season (May to September) and a 3-month flood season (January to March) [69,70]. The climate is defined as tropical with a dry winter (Aw Koppen's climate classification) [72], with annual precipitation around 1755 mm and 90% of the precipitation during the wet season, causing seasonal flooding [73].

Micrometeorological and Eddy Covariance Measurements
The micrometeorological and eddy covariance tower is located at 9°49'16.1"S and 50°08'55.3"W, in a transitional area between the Amazon and the Cerrado biomes (Figure 2), at the northern boundary of the Bananal island, between the Araguaia and Javaés rivers, over a seasonal forest-savanna ecosystem floodplain. Natural vegetation within the tower footprint is composed mostly of woodland savanna (with trees up to 18 m and sparse bushes) and fragments of wooded savanna (with trees up to 5 m) and grassland, with a 5-month dry season (May to September) and a 3-month flood season (January to March) [69,70]. The climate is defined as tropical with a dry winter (Aw Koppen's climate classification) [72], with annual precipitation around 1755 mm and 90% of the precipitation during the wet season, causing seasonal flooding [73]. Micrometeorological and eddy covariance measurements at the Bananal Javaés site (JAV), available from 10/2003 to 12/2006, are part of the Large-Scale Biosphere and Atmosphere Experiment in the Amazon (LBA) [74], reported in da Rocha et al. [70]. LBA is an international research program focused on understanding the interactions between terrestrial ecosystems and the atmosphere within a region of great human and natural complexity (encompassing a wide range of land-use and land-cover types and climate conditions) on the Amazon biome [75]. Data from the JAV site have been described [69,70,76] in terms of their environmental characteristics, instrumentation (Table 1), measurement procedures, and results. We used quality-controlled data from Saleska et al. [77], recorded above the canopy at 30 min intervals, to perform our analysis. We did not gap fill the data. Micrometeorological and eddy covariance measurements at the Bananal Javaés site (JAV), available from 10/2003 to 12/2006, are part of the Large-Scale Biosphere and Atmosphere Experiment in the Amazon (LBA) [74], reported in da Rocha et al. [70]. LBA is an international research program focused on understanding the interactions between terrestrial ecosystems and the atmosphere within a region of great human and natural complexity (encompassing a wide range of land-use and land-cover types and climate conditions) on the Amazon biome [75]. Data from the JAV site have been described [69,70,76] in terms of their environmental characteristics, instrumentation (Table 1), measurement procedures, and results. We used quality-controlled data from Saleska et al. [77], recorded above the canopy at 30 min intervals, to perform our analysis. We did not gap fill the data.
The energy balance closure (defined as R n − G = H + LE, where G is the ground heat flux and R n is the net radiation) at several LBA sites presented an imbalance of 70%-105% [70] due to systematic biases in instruments, energy sources not considered, as well as losses of turbulent fluxes at high and low frequencies [78]. For the JAV site, during the dry season, the energy balance closure presented a negative imbalance of 15% [70]. Therefore, the energy balance closure was forced using the Bowen ratio (β) technique (Equation (1)). The residual available energy was then distributed between H and LE, and the energy imbalance was closed to conserve the energy partitioning, according to Equations (2) and (3) [79]. This simple and reliable approach, used in a wide range of climates [80], was adopted since it can perform similarly to other methods, despite remaining uncertainties about energy balance closure [81].

MERRA-2 Meteorological Reanalysis Data
As global gridded reanalysis inputs, we used reanalysis data from MERRA-2, with spatial resolution of 0.5 • × 0.625 • and hourly temporal resolution [68]. Meteorological reanalysis inputs include daily average R s and hourly T a (average between 10:00 and 11:00 am local time to coincide with the satellite overpass), u, specific humidity (q), and atmospheric pressure (P). MERRA-2 RH was computed from q, according to Equations (4)-(6) [82]. To combine meteorological reanalysis and remote sensing inputs, MERRA-2 data were interpolated from coarse spatial resolution to match the spatial resolution of Landsat images (30 m) using a simple bicubic interpolation method (MATLAB®R2012b).
where e and e s are vapor pressure and saturated vapor pressure, respectively, while P is the atmospheric pressure.

Landsat Data
Landsat 5 TM images were selected periodically during the dry season (June to September), with consideration given to cloud-cover condition and micrometeorological and eddy covariance measurements availability. After excluding images with greater than 10% cloud cover, a total of 12 images were obtained ( Table 2). Landsat images have a temporal resolution of 16 days and a spatial resolution of 30 m for multispectral data and 120 m for thermal data. We used Landsat data because of the high spatial resolution and data availability during the eddy covariance measurements, encompassing the 2004-2006 dry seasons. To compute SEBAL variables we used bands 1 to 7 and to select high confidence pixels we used the quality attributes band.

Surface Energy Balance Algorithm for Land (SEBAL)
SEBAL calculates instantaneous LE at the moment of satellite overpass (around 10:30 am local time) as the residual in the surface energy balance Equation (Equation (7)).
where ω λ is a coefficient related to the ratio of the solar radiation at the surface occurring within a given spectral band and ρ λ is the surface reflectance for this band.
To estimate T s based on Landsat images, we used the modified Planck Equation (Equation (10)), following Allen et al. [27], with atmospheric and surface emissivity corrections.
where ε NB is the surface emissivity corresponding to the thermal sensor wavelength, R c is the corrected thermal radiance from the surface based on the spectral radiance (Landsat band 6), while K 1 and K 2 are constants [27].
To estimate H (Equation (11)) it is necessary to select endmembers to represent extreme conditions for internal calibration [26]. Based on the hot and cold endmembers selection, SEBAL considers a linear relation between T s and dT between two heights z 1 and z 2 . H is assumed to be zero for the cold pixel and maximum for the hot pixel, while LE is assumed to be zero for the hot pixel and maximum for the cold pixel.
where ρ is the air density, Cp is the specific heat capacity of air, and r ah is the aerodynamic resistance to turbulent heat transport from the evaporating surface at height z 1 to the air above the evaporating surface z 2 (Equation (12)), considering the atmospheric stability correction based on the Monin-Obhukov similarity hypothesis [26].
where k is the von Karman constant (0.41) and u * the friction velocity, estimated for each pixel (Equation (13)).
where u 200 is wind speed at 200 m and z om is the surface roughness for each pixel, estimated according to Bastiaanssen [83].
To select the hot and cold endmembers, we used a simplified version adapted from the automated methodology from METRIC [27] based on the Calibration using Inverse Modeling at Extreme Conditions (CIMEC) process [54]. The CIMEC process employs a population of candidate members based on quantiles of remote sensing estimations of T s and NDVI to select the hot (dry) and cold (wet) ends of the ET spectrum, aiming at consistent internal calibration for time-series applications. Allen et al. [54] defined a subset of endmembers within the highest 5% of NDVI and the lowest 20% of T s to select the cold extreme condition, while endmembers within the lowest 10% NDVI and with the highest 20% T s were used to select the hot extreme condition. After [54], several additional criteria were proposed to improve and refine the endmember selection, sometimes with an increase in complexity and time processing [17,31,55,56,85]. In this research, we proposed an assessment of the automated calibration of the endmembers based on the Allen et al. [54] methodology to accurately select the quantiles to estimate the surface energy components. We tested for the effect of nine different automated endmember selection quantiles (Table 3) over an area of 100 × 100 km, based solely on T s and NDVI, without additional filters. The first quantile group was proposed by Allen et al. [54]. For the other eight groups we used different quantiles to select the endmembers. Table 3. Quantiles used for the automated selection of cold and hot endmembers.

Endmember Group
NDVI Cold T s Cold NDVI Hot T s Hot To estimate the daily average net radiation (R n24h ) we used the methodology proposed by de Bruin [86] (Equation (14)).
where R s 24h is the shortwave incident daily radiation, R a 24h is the exoatmospheric shortwave radiation, while C s is an empirical constant, calibrated according to local climate conditions. We used a generalized reduced gradient nonlinear method for calibration [87], yielding a value of 115 for clear-sky days and atmospheric conditions similar to the cloud-free Landsat image conditions, reducing the influence of low atmospheric transmissivity during all sky conditions. Remote Sens. 2020, 12, 1108 8 of 23 The evaporative fraction (Λ) (Equation (15)) was used to upscale instantaneous estimates of LE to daily ET (ET 24h ) (Equation (16)), assuming that Λ is constant during the daytime [26], when soil moisture does not change significantly and advection does not occur [40].
where λ is the latent heat of vaporization of water [82].

MapBiomas Land Cover Dataset
To assess the spatial variations of surface energy and water fluxes over the main land cover conditions in the study area, we used a land cover classification from MapBiomas version 3.1. MapBiomas is a collaborative network to generate annual land-use and land-cover maps in Brazil with spatial resolution of 30 m through processing of the Landsat collection (1985-2018) with a random forest algorithm within Google Earth Engine cloud computing [88], available at http://mapbiomas.org.

MERRA-2 and SEBAL Assessment
To evaluate MERRA-2 meteorological reanalysis data and to validate SEBAL surface fluxes, we used the performance metrics of mean absolute error (MAE) and root mean square error (RMSE) between observed (O i ) and predicted (P i ) values (Equations (17) and (18)). Ground measurements were used to validate a single MERRA-2 coarse pixel (for the location of the JAV site) and to validate SEBAL, considering an average estimation for an area centered in the JAV site with a radius of 1 km. This average was adopted based on the tower footprint, since in both wet and dry seasons, up to 75% of the footprint flux originates within 1 km of the flux tower radius [69].
where n is the number of samples.

Validation of MERRA-2 Reanalysis Data
To assess MERRA-2 meteorological reanalysis accuracy and uncertainties in SEBAL surface flux estimations, we compared T a , R s , RH, and u to ground measurements ( Figure 3). For R s , we found an overestimation (MAE) of 35.8 W m −2 and an error (RMSE) of 40.8 W m −2 , representing 17.0% and 19.4% of the average measured R s , respectively. Overall, MERRA-2 R s errors are similar in range during both wet and dry seasons. For u and RH we also found an overestimation. Whereas for u, MAE was 1.3 m s −1 and RMSE was 1.6 m s −1 , for RH we found 15.7% and 18.3%, respectively, with higher errors during the dry season and an overestimation lag of 2 to 3 months since the beginning of the dry season. On the other hand, for T a we found an underestimation (MAE) of 3.9 • C and an RMSE of 4.7 • C, with higher underestimation during the dry season, yielding RMSE up to 6.2 • C. During the wet (flooded) season, MAE and RMSE decreased to 2.4 and 2.8 • C, respectively. higher errors during the dry season and an overestimation lag of 2 to 3 months since the beginning of the dry season. On the other hand, for we found an underestimation (MAE) of 3.9°C and an RMSE of 4.7°C, with higher underestimation during the dry season, yielding RMSE up to 6.2°C. During the wet (flooded) season, MAE and RMSE decreased to 2.4 and 2.8°C, respectively.

Validation of SEBAL Instantaneous Surface Flux Estimates
For performance assessment, the instantaneous surface fluxes ( , , , and ) estimated by SEBAL-T and SEBAL-M averaged over the 1 km radius area (tower footprint) were compared against radiative and eddy covariance measured data (at 10:30 am local time) at the JAV site. Results of nine different quantiles of endmembers (combinations of and quantiles) are shown in Figure 4.

Validation of SEBAL Instantaneous Surface Flux Estimates
For performance assessment, the instantaneous surface fluxes (R n , G, H, and LE) estimated by SEBAL-T and SEBAL-M averaged over the 1 km radius area (tower footprint) were compared against radiative and eddy covariance measured data (at 10:30 am local time) at the JAV site. Results of nine different quantiles of endmembers (combinations of T s and NDVI quantiles) are shown in Figure 4. Despite similar results for R n and G estimations using different meteorological datasets and quantiles to select the endmembers, we found notable differences in H and LE estimations, mainly related to the endmember selection criteria. The average H measurement without energy balance closure was 40.9 W m −2 , and the average H from SEBAL-T estimations measurements ranged between 19.0 and 45.8 W m −2 for quantiles g A -g Ts4 , yielding an average of 19.1 W m −2 for quantiles g IV1 − g IV4 . SEBAL-M estimations were slightly higher than SEBAL-T ones, ranging between 25.0 and 68.4 W m −2 for quantiles g A − g Ts4 , with an average of 21.91 W m −2 for quantiles g IV1 − g IV4 . When considering the energy balance closure, the average H measurement increased to 79   With an increase in H estimations (for both meteorological inputs) for quantiles g A − g Ts4 , LE estimations decreased for the same quantiles and held steady for quantiles g IV1 − g IV4 . Overall, without energy balance closure, instantaneous LE was overestimated, with MAE (RMSE) ranging between 111.6 (128.1) and 156.3 (169.9) W m −2 for SEBAL-T and between 95.8 (98.7) and 149.3 (152.0) W m −2 for SEBAL-M. When correcting instantaneous measurements of LE according to the energy balance closure, LE increased accuracy, with an underestimation, with MAE (RMSE) ranging between 52.3 (66.0) and 97.1 (113.1) W m −2 for SEBAL-T and between 89.5 (93.9) and 143.0 (144.5) W m −2 for SEBAL-M. As a residual of the energy balance, instantaneous estimates of LE for quantiles g A and g IV1 − g IV4 yielded the most accurate estimates. MAE and RMSE of all estimated quantiles for Rn, G, H, and LE are presented in Tables S1 and S2.

Assessment of SEBAL Daily Evapotranspiration Estimates
SEBAL upscaling from instantaneous fluxes of LE to daily ET (ET 24h ) integrates the daily net radiation (R n 24h ), assuming Λ as a constant indicator of energy partitioning throughout the day (from sunrise to sunset). To investigate SEBAL upscaling, the accuracy of the estimated Λ was first assessed in relation to the measured data for instantaneous and daily measurements (considering or not the energy balance closure). Then, R n 24h , Λ, and ET 24h estimates were evaluated ( Figure 5). SEBAL based Λ led to overestimation for the nine quantile groups when compared to instantaneous Λ computed from measured data (at 10:30 am local time). However, when it was compared to the daily Λ computed from measured daily averages, the accuracy increased significantly for both meteorological inputs. For instantaneous Λ we found an average of 0.44 ± 0.14 (mean ± standard deviation), while for daily average Λ (measured) we found 0.75 ± 0.16 (without energy balance closure) and 0.82 ± 0.05 (forcing the energy balance closure). For SEBAL-T estimates, Λ values ranged between 0.76 and 0.96 and for SEBAL-M, Λ ranged between 0.74 and 0.95. For the most accurate SEBAL estimates of Λ, quantile g Ts3 , we found an RMSE of 0.08 (0.19) and 0.10 (0.20), with (without) daily energy balance closure, for SEBAL-M and SEBAL-T estimations, respectively. Overall, SEBAL estimates of instantaneous Λ were relatively close to daily ground measurements when considering the energy balance closure. Furthermore, R n 24h estimates yielded high accuracy when driven by both meteorological inputs. Overall, an underestimation was obtained, with an RMSE of 13.5 W m −2 for SEBAL-T and 9.8 W m −2 for SEBAL-M. MAE and RMSE of all quantiles for Λ and ET 24h are presented in Tables S3 and S4. Our results indicate that daily estimates of using the SEBAL algorithm and both meteorological inputs were consistently accurate when compared to eddy covariance measurements. Without the energy balance closure, SEBAL-T led to an average RMSE of 0.87 mm day −1 (for SEBAL-M we found an RMSE of 0.85 mm day −1 ). On the other hand, when using the energy balance closure technique to compute daily , SEBAL accuracy increased significantly, yielding an average RMSE of 0.35 mm day −1 for both meteorological inputs (an improvement of 60%). The most accurate quantile used for the endmember selection was , with of 5% and 10% and of 0.01%, for cold and hot endmembers, respectively, representing an error of 8.6% (RMSE compared to the average daily ), with errors up to 15% and 23% for the less accurate quantile of endmembers selection, for SEBAL-T and SEBAL-M, respectively.

Spatial Assessment of Surface Energy and Water Fluxes
Based on the combination of MERRA-2 reanalysis and Landsat images SEBAL estimations, we computed the average of each surface energy ( and ) and water ( ) flux component according to land cover, using data from MapBiomas [88]. The study area presents a large heterogeneity in vegetation physiognomies, dominated by savanna vegetation, with an significant increase in croplands [9,89,90]. To illustrate the spatial pattern of vegetation characteristics and surface processes, we computed the dry season average and and daily for 2004 ( Figure 6). values larger than zero were found over forested areas (flooded or non-flooded), which are associated to values lower (close to the minimum 24.5 °C values) than savanna and croplands, which reached up to 32 °C and also yielded a higher proportion in the energy partition. The largest values were obtained for croplands, while natural savanna regions reached intermediate values.

Figure 5.
Comparison between ground measurements (where in situ inst and day stand for instantaneous and daily measurements without energy balance closure, respectively, and in situ EBC for daily measurements with energy balance closure) and SEBAL estimates of evaporative fraction (Λ) (a), daily net radiation (R n 24h ) (b), and daily evapotranspiration (ET 24h ) (c) driven by in situ meteorological measurements (SEBAL-T) and MERRA-2 (SEBAL-M).
Our results indicate that daily estimates of ET using the SEBAL algorithm and both meteorological inputs were consistently accurate when compared to eddy covariance measurements. Without the energy balance closure, SEBAL-T led to an average RMSE of 0.87 mm day −1 (for SEBAL-M we found an RMSE of 0.85 mm day −1 ). On the other hand, when using the energy balance closure technique to compute daily ET, SEBAL accuracy increased significantly, yielding an average RMSE of 0.35 mm day −1 for both meteorological inputs (an improvement of 60%). The most accurate quantile used for the endmember selection was g Ts4 , with NDVI of 5% and 10% and T s of 0.01%, for cold and hot endmembers, respectively, representing an error of 8.6% (RMSE compared to the average daily ET), with errors up to 15% and 23% for the less accurate quantile of endmembers selection, for SEBAL-T and SEBAL-M, respectively.

Spatial Assessment of Surface Energy and Water Fluxes
Based on the combination of MERRA-2 reanalysis and Landsat images SEBAL estimations, we computed the average of each surface energy (H and LE) and water (ET) flux component according to land cover, using data from MapBiomas [88]. The study area presents a large heterogeneity in vegetation physiognomies, dominated by savanna vegetation, with an significant increase in croplands [9,89,90]. To illustrate the spatial pattern of vegetation characteristics and surface processes, we computed the dry season average NDVI and T s and daily ET for 2004 ( Figure 6). NDVI values larger than zero were found over forested areas (flooded or non-flooded), which are associated to T s values lower (close to the minimum 24.5 • C values) than savanna and croplands, which reached up to 32 • C and also yielded a higher H proportion in the energy partition. The largest T s values were obtained for croplands, while natural savanna regions reached intermediate values. There are marked differences between tropical forest, savanna vegetation, and cropland areas with respect to land-atmosphere exchanges. Converting forest vegetation to cropland and pasture substantially changes surface energy (H and LE) and water (ET) fluxes (Figure 7). Taking into consideration the average surface fluxes according to land cover classes for all analyzed images and using SEBAL-T results from g Ts4 quantile, our estimates of daily ET ranged between 2.0 and 4.2 mm day −1 for pasture and tropical forest, while grassland, savanna vegetation, and cropland presented rates of 2.9, 3.1, and 3.2 mm day −1 , respectively. Finally, Figure 8

Uncertainties in the SEBAL Model Structure
Our estimates of instantaneous are consistent with other studies [91], with errors lower than

Uncertainties in the SEBAL Model Structure
Our estimates of instantaneous R n are consistent with other studies [91], with errors lower than 5%, despite overestimations when driven by MERRA-2 and underestimations when driven by ground measurements. However, we found considerable discrepancies in instantaneous G values, with RMSE higher than 108 W m −2 . Measured G represented only a small fraction of R n , corresponding to around 3% of this component (7 W m −2 on average). Results from [92] indicate that the canopy and air-column heat storage in the Amazon during the dry season is around 80 W m −2 during mid-morning, being much higher than G. Since canopy heat storage plays an important role on a diurnal time scale, it cannot be neglected, despite its integration on a daily time scale only accounting for a small portion of the available energy [93]. On a daily basis, heat gained during the daytime is released at night through thermal radiation and evaporation [94].
The poor performance of the SEBAL in estimating G may also be related to the forest-savanna flooding ecosystem, since the equation used to calculate G was derived for short vegetation and non-flooded areas. The main components to calculate G using the SEBAL algorithm are T s , NDVI, and R n , representing surface fluxes (including canopy and ground surface), while measured G only accounts for ground surface. This is why the energy budget for tropical forests should be rewritten (Equation (19)) [92].
where S corresponds to changes in total heat content within the canopy and Q p corresponds to the energy of photosynthesis. We also note that we used globally available equations in this study for estimating the energy balance components, however, more refined parameters could be adopted in a future work to improve algorithm performance and model structure.

Uncertainties in the SEBAL Instantaneous and Daily ET Estimates
Overall, the main SEBAL uncertainties are related to H estimates and the selection of the hot and cold endmembers [35], yielding higher errors than the other surface energy balance components. However, we found accurate H estimates using g Ts1 − g Ts4 quantile groups, especially when compared to instantaneous measurements without energy balance closure. On the other hand, due to energy balance components being neglected or not computed for both eddy covariance measurements and SEBAL estimates, such as ∆S at diurnal time scale, SEBAL instantaneous LE was overestimated, with MAE higher than 95.8 W m −2 for SEBAL-T and 111.6 W m −2 for SEBAL-M, suggesting significant differences on the instantaneous surface energy fluxes when compared to ground measurements. For instance, systematic biases in instrumentation and energy sources not considered in the process (storage in the air column or biomass) could give rise to uncertainties in the energy balance as well as losses of the turbulent fluxes at high or low frequencies [78]. Considering that SEBAL instantaneous LE estimates were calculated as the residual of the energy balance, we also corrected measured LE considering the energy balance closure at an instantaneous time scale. In this case, LE was underestimated and MAE decreased significantly.
On the other hand, results indicate that the SEBAL slightly overestimated daily ET when compared to eddy covariance measurements, which is in accordance with other studies [51,55,61,[95][96][97], with errors ranging between 0.3-0.4 mm day −1 when compared to measurements forced by energy balance closure techniques, and 0.6-0.8 mm day −1 when compared to measurements without energy balance closure. A better performance of the SEBAL at a daily time scale is possibly related to the small fraction of G and ∆S on a daily time scale [92,94], closely relying on the accurate estimates of Λ and R n 24h . The use of the Λ for upscaling instantaneous to daily surface fluxes based on remote sensing estimates is reliable, with assessments indicating a difference of less than 12% when compared to measured Λ, evidencing the agreement between SEBAL estimates and eddy covariance measurements [98], in close accordance with our findings.

The SEBAL Sensitivity to Meteorological Inputs
Overall, MERRA-2 provides more self-consistent meteorological data than its previous versions at large scale [68,99]. Although some validations of the recent MERRA-2 surface and atmospheric reanalysis have been published [100,101], there is a lack of comprehensive assessment at regional scales.
Meteorological reanalysis uncertainties are mainly related to model structure and parameterization [102][103][104], since it is a theory-based calculation [105]. Therefore, the coarse spatial resolution can result in an increased likelihood of discrepancies between meteorological reanalysis and ground measurements, since they usually have limited spatial representativeness [102]. Meteorological reanalysis only provides large scale averages and does not represent the subgrid (misrepresenting local meteorological processes), therefore it suffers from biases especially over heterogeneous surfaces [106]. Thus, despite the moderate accuracy and biases at regional scales, as found at the JAV site, land data assimilation and reanalysis are becoming important sources of meteorological inputs for ET estimations [107].
Our results indicate that the use of ground measurements or global reanalysis data (MERRA-2) as meteorological inputs leads to low sensitivity on SEBAL accuracy to estimate surface energy and water fluxes. Only a few studies addressed the influence of meteorological inputs on SEBAL estimates, pointing out the small [108] to moderate sensitivity [109] of u. Considering T a and RH, used as inputs to compute downward longwave radiation and atmospheric transmissivity, respectively, the low sensitivity is related to the small fraction of these components in the surface energy balance during clear-sky days. On the other hand, Biggs et al. [17] reported a lower accuracy for the SEBAL driven by the use of MERRA-1 Land global climate data, especially for daily R n , when compared to estimates driven by ground measurements. Indeed, errors in radiation (R s or R n ) can cause large errors in LE and ET estimations, e.g., Kich [109] demonstrated that an error of 10% in R n can cause an error up to 25% in LE.
Our findings confirm that temperature-based models rely on T s and dT, with a high sensitivity of Λ and R n 24h to upscaling instantaneous to daily fluxes [108,110,111], while vegetation-based models depend on vegetation conditions and reanalysis data from global climate grids, with a higher sensitivity to meteorological inputs [17,112]. Since our estimates of Λ and R n 24h yielded a high accuracy for both meteorological inputs, the most accurate daily ET estimates were for SEBAL-T and SEBAL-M when compared to vegetation-based model estimates in the Amazon biome [112,113].

The SEBAL Sensitivity to Endmember Selection
Endmember selection was assessed by using different quantiles of T s and NDVI. Our results clearly show that the selection of both hot and cold extreme endmembers result in high sensitivity to H, LE, and ET, playing a key role in the magnitude of surface fluxes estimates. Choragudi [108] and Long et al. [35] evaluated the sensitivity of the endmember selection, reporting a higher sensitivity for the hot endmember selection, where T s itself plays a secondary role in LE and ET estimations [35,57]. In contrast, changes in NDVI quantiles showed a lower sensitivity when compared to T s , which is expected since the endmember selection is closely related to dT [108].
An example of endmember candidates based on the nine quantile groups of T s and NDVI is presented in Figure 9 for the year 2004 DOY 200. Histograms of T s used for internal calibration are also presented. It is clear that the endmember selection based on T s and NDVI for different quantiles has a low sensitivity on the cold and a high sensitivity on the hot selection. For quantiles g T s 1 − g T s 4 , with changes in T s quantiles (keeping NDVI quantiles fixed), there is a decrease of 0.65K for T s at the cold endmember and an increase of 2.4K for T s at the hot endmember. On the other hand, for quantiles g IV 1 − g IV 4 , with changes in NDVI quantiles (keeping T s quantiles fixed), we found small changes in T s for both cold and hot endmembers (increase of 0.90 K and a decrease of 0.15 K for the hot and cold endmembers, respectively). Our findings are in agreement with Long et al. [35], indicating a higher sensitivity for T s rather than NDVI on H and LE estimates. This higher sensitivity on the hot and dry endmember candidates was originally demonstrated by Bastiaanssen [53] and also observed by previous studies [58,114], where the endmembers play a key role in the SEBAL's internal calibration. The strong and accurate performance of the SEBAL and other surface energy balance models in heterogeneous areas was also demonstrated by other studies [115,116] also considering wet and dry climate conditions [61,95]. The importance of the selection of the endmembers for estimation was also highlighted [35,57,114], however, our results indicate that the influence of pixel selection is more relevant than the meteorological data used to force the model, making reanalysis data such as from MERRA-2 useful for surface energy balance models.

Spatial Assessment of Surface Energy and Water Fluxes
Our study area is located in a transitional zone between forest and savanna in Central Brazil. Our estimates of energy partition (sensible and latent heat) for different land uses are in accordance with Restrepo-Coupe et al. [117], who showed that converted areas (forest to cropland and pasture) presented a different behavior with relative reductions in dry season , consistent with increasing water limitation due to loss of deep roots that can access soil water. They are also in agreement with Oliveira et al. [118] and Dias et al. [4], who conducted studies in the southwest (Ji-Paraná River basin) and southeast (Xingu River basin) Amazon on the conversion between forest vegetation and cropland and pasture, showing substantial changes on surface energy ( and ) and water ( ) fluxes. Furthermore, changes in reflect alterations on surface energy fluxes partitioning between and . For instance, for forest areas we found a of 0.19, while pasture areas presented a of 0.60. In this context, land cover conversion leads to an increase of , with an associated increase in and decrease in , causing an increase of and decrease in , mainly due to deforestation resulting in less vegetation being available to transpire water to the atmosphere [119].

Conclusions
Recent techniques of remote sensing and reanalysis data have improved our understanding of surface energy and water fluxes in the tropics. In this study, we presented an assessment of the SEBAL algorithm to estimate , and driven by ground measurements and MERRA-2 reanalysis data. The comparison between SEBAL estimates against ground measurements yielded high accuracy for , however, we found a high bias for , suggesting a regional calibration. For This higher sensitivity on the hot and dry endmember candidates was originally demonstrated by Bastiaanssen [53] and also observed by previous studies [58,114], where the endmembers play a key role in the SEBAL's internal calibration. The strong and accurate performance of the SEBAL and other surface energy balance models in heterogeneous areas was also demonstrated by other studies [115,116] also considering wet and dry climate conditions [61,95]. The importance of the selection of the endmembers for ET estimation was also highlighted [35,57,114], however, our results indicate that the influence of pixel selection is more relevant than the meteorological data used to force the model, making reanalysis data such as from MERRA-2 useful for surface energy balance models.

Spatial Assessment of Surface Energy and Water Fluxes
Our study area is located in a transitional zone between forest and savanna in Central Brazil. Our estimates of energy partition (sensible and latent heat) for different land uses are in accordance with Restrepo-Coupe et al. [117], who showed that converted areas (forest to cropland and pasture) presented a different behavior with relative reductions in dry season ET, consistent with increasing water limitation due to loss of deep roots that can access soil water. They are also in agreement with Oliveira et al. [118] and Dias et al. [4], who conducted studies in the southwest (Ji-Paraná River basin) and southeast (Xingu River basin) Amazon on the conversion between forest vegetation and cropland and pasture, showing substantial changes on surface energy (H and LE) and water (ET) fluxes. Furthermore, changes in β reflect alterations on surface energy fluxes partitioning between H and LE. For instance, for forest areas we found a β of 0.19, while pasture areas presented a β of 0.60. In this context, land cover conversion leads to an increase of β, with an associated increase in H and decrease in LE, causing an increase of T a and decrease in ET, mainly due to deforestation resulting in less vegetation being available to transpire water to the atmosphere [119].

Conclusions
Recent techniques of remote sensing and reanalysis data have improved our understanding of surface energy and water fluxes in the tropics. In this study, we presented an assessment of the SEBAL algorithm to estimate H, LE and ET driven by ground measurements and MERRA-2 reanalysis data. The comparison between SEBAL estimates against ground measurements yielded high accuracy for R n , however, we found a high bias for G, suggesting a regional calibration. For H and LE, eddy covariance measurements corrected with energy balance closure techniques yielded more accurate results when using both meteorological inputs in comparison to the non-closure measurements. Overall, SEBAL daily ET estimates presented higher accuracy than instantaneous fluxes. Furthermore, the automatic selection of endmembers was evaluated considering different T S and NDVI quantiles. Although daily ET estimates were less sensitive to the meteorological data, the endmember selection for SEBAL's internal calibration played a key role in estimations of surface fluxes. T S quantiles showed a higher sensitivity than NDVI quantiles, indicating its high influence on H and LE estimates on tropical biomes. Indeed, the strength of the SEBAL algorithm consists in the endmembers selection, while meteorological inputs have a lower impact on ET accuracy, providing consistent surface-energy partitioning at multiple land cover conditions independent of low accuracy input datasets.
The analyses reported here suggest that the SEBAL algorithm can capture surface fluxes for different land cover conditions at a regional scale within an acceptable accuracy, with significant potential for spatial and temporal monitoring based on MERRA-2 global reanalysis data and Landsat images in data-limited areas. In future research, long-term changes of surface fluxes across different land covers will be addressed, aiming to understand the impact of deforestation on land-atmosphere exchanges.