Radiation Balance Estimates Over Southeastern Brazil: Ground Observations, Satellite and Reanalysis

Resumo Pela grande atividade econômica e densidade populacional, a região sudeste do Brasil vivencia processos acelerados de mudanças no uso e cobertura do solo, que contribuem para modificações no balanço de radiação (BR) na superfície. Neste estudo, avaliamos os componentes do BR de diferentes conjuntos de dados de alta resolução, última geração, nesta região do país em dois períodos (fev/2005-jan/2006 e mar/2015-fev/2016). Em geral, todos os conjuntos de dados representaram adequadamente a sazonalidade dos componentes do BR (exceto albedo). O ERA5-Land apresenta com o menor erro relativo médio para representar o albedo (≈ 15%), radiação de onda longa atmosférica (DLWR ≈ 4,5%) e radiação de onda longa da superfície (ULWR ≈ 3,6%). Na estimativa de radiação de ondas curtas, o GLASS foi o melhor (≈ 14%). As incertezas neste último podem estar associadas à dificuldade para representar a variabilidade de cobertura de nuvens no período chuvoso. As falhas na estimativa do albedo devem-se à incapacidade de simular as propriedades da superfície. DLWR e ULWR apresentaram os melhores desempenhos e suas incertezas estiveram relacionadas a problemas no cálculo das temperaturas do ar e da superfície, respectivamente. ERA5-Land e GLASS são adequados para estimar os componentes da BR no sudeste do Brasil.


Introduction
Brazilian southeast region (SE) is the main driver of economic development in the country, since it concentrates around 42% of the population and produces approximately 56% of the gross domestic product (IBGE, 2018).All this development lead to large changes in land cover.Urban areas, pastures and crops -such as coffee, soy and sugarcane -replaced areas of native vegetation in the Atlantic Forest and Cerrado biomes, leaving them with a low proportion of remaining native vegetation and biodiversity highly threatened (Alvarenga et al., 2016;Dias et al., 2016;Grecchi et al., 2014;Sano et al., 2010).
Microclimate studies of biomes associate deforestation with potential impacts on the climate system, primary caused by changing the components of radiation balance (RB) (Conte et al., 2019;Macdougall and Beltrami, 2017).Estimates of the amount of solar energy received, reflected and emitted by the Earth's surface and the atmosphere, plays a prominent role to the understanding of past weather, long-term features and future climate (Wild, 2016(Wild, , 2020)).Moreover, the study of the radiation balance on the earth's surface is relevant to solve some scientific questions, such as the practical viability of solar energy, the reaction of plants to the wavelengths necessary for photosynthesis, the potential consumption of water and productivity of crops/ecosystems, among others.(Green et al., 2017;Mercado et al., 2009;Wild et al., 2015).
Observations of the RB components have been carried out for some decades through net radiometers (Driemel et al., 2018;Wild et al., 2017).However, this instrumentation besides having a high cost and requiring special techniques of operation and calibration, provides measurements that are representative only for small areas that cannot be used at larger scales (Ferreira et al., 2020).In most case in Brazil meteorological stations only measures shortwave component (Xavier et al., 2016), while the longwave component is restricted to individual micrometeorological experiments or recent networks of specialized stations, such as SONDA (INPE, 2021).
Remote sensing and reanalysis appear as alternatives to solve the absence and limitations of surface measurements of the RB components.There are two types of remote sensing algorithms, those that use only information from satellite sensors (Bisht et al., 2005;Ramírez-Cuesta et al., 2018), and others that combine on-board sensors with surface station data (Amatya et al., 2015;Carmona et al., 2015;Ferreira et al., 2020;Silva et al., 2015).However, previous methodologies has been restricted to cloudless days.Bisht andBras (2010, 2011) proposed a model to estimate the RB components (instantaneous and daily) for all sky conditions, which uses only MODIS sensor information.Although useful, this methodology computed the daily RB from only two measurements, which are insufficient to capture cloud dynamics, especially for latitudes higher than 30°, where the MODIS's passage is 1-2 days (Wolfe et al., 2002).On the other hand, reanalyses offer temporal high resolution and consistency, as well as data under all sky conditions.These are frequently used for atmospheric model validation, and have contributed to clarify the relative importance of the data assimilation improvements versus observational improvements for numerical weather prediction over the last decades (Bengtsson et al., 2007).However, caveats result from their coarse resolutions (> 100 km) and assimilation of data obtained from atmospheric profiles and measurements, which cause systematic biases in the RB products (Jia et al., 2018;Slater, 2016;Zhang et al., 2016).
To overcome the problems mentioned above, studies have combined satellite and reanalysis data to determine the RB components at global (Verma et al., 2016) and regional scales (Moukomla and Blanken, 2017;Oliveira et al., 2016;Yu et al., 2014).In addition, with the recent advances in data assimilation, increment in the quality and quantity of observed data, increase in spatial resolutions (< 30 km) and improvements in the parameterization of models, blended datasets have been made available for involving different meteorological parameters, such as the RB components (Muñoz-Sabater et al., 2021;Rodell et al., 2004;Zhang et al., 2019).Nevertheless, it is crucial to assess the reliability of each dataset, and to identify the strengths and underlying biases associated, before using them (Dolinar et al., 2016).
In this context, the present study focuses on evaluating the performance of state-of-the-art high-resolution datasets (reanalyses and satellite), that can suitably fill the lack of observation in Brazil as an alternative for monitoring the RB components.The paper is structured as follows.Section 2 introduces the study area, climate datasets and methods.Section 3 details and discusses the performance in estimating the RB components based on different datasets, while the section 4 presents the concluding remarks.

Study area
The study area corresponds to the southeast region of Brazil, located between latitudes 14°S and 25°S and longitudes 54°W and 39°W (Fig. 1).The region has a territorial extension of 924,620 km 2 (approximately 10 times the Portugal surface area, or 1.6 times of the Iberian Peninsula), formed by the States of Minas Gerais (MG), São Paulo (SP), Espirito Santo (ES) and Rio de Janeiro (RJ).In addition, it has a population of approximately 88.6 million people.The region is covered by Mata Atlântica (Atlantic Forest), Cerrado (Savanna) and Caatinga (Fig. 1b).Mata Atlântica is concentrated mainly in the east of the territory, while Cerrado is present in the interior of SP and west of MG.Caatinga biome corresponds to small portion of the north of the region.The seasonal cycle of precipitation, humidity and circulation are determined by the South Atlantic Convergence Zone (SACZ) during the spring and summer seasons, while the frontal systems dominate the winter season, which is predominantly dry (Bernardino et al., 2018;Zilli et al., 2017).The relief, which varies from sea level to altitudes above 1800 m, has an influence on the region temperature, decreasing it as the elevation increases (Cavalcanti et al., 2009).

Observed data for validation
Measurements of the RB components from three different surfaces were used to validate the dataset outputs, as described below.These stations are representative of three different environment conditions, in terms of vegetation and land cover changes.The short data collection periods (one year) are because the all stations were part of short-term academic projects.

USR station
This is an experimental sugarcane field belonging to Power Plant Santa Rita (USR) located in the State of SP (Fig. 1c).This station measures (latitude 21°38'13" S, longitude 47°47'25" W), downward shortwave radiation (DSWR), albedo (α), downward longwave radiation (DLWR) and upward longwave radiation (ULWR) that were collected between February-2005 and January-2006.This place has an average altitude of 552 m and corresponds to sugarcane plants in the first cycle of regrowth (Oliveira et al., 2018;Silva et al., 2015).

Gridded datasets
Five different modern gridded datasets were used.The RB components were extracted directly from ERA5-Land, GLDAS and GLASS.Air temperature (Ta), DSWR, and incident solar radiation at the top of the atmosphere (Ro), used to compute DLWR with the SEBAL algorithm (Section 2.4), were extracted from the Xavier and CERES-SYN datasets.Table 1 summarizes the characteristics of each dataset.
ERA5-Land is based on running the land component from ERA5 reanalysis of the European Centre for Medium-Range Weather Forecasts (ECMWF) but without coupling to the atmospheric models (Cao et al., 2020).This product has a spatial resolution of 0.1°× 0.1°, with hourly temporal frequency, from 1981 to present (Pelosi et al., 2020).The radiation scheme performs calculations of the shortwave and longwave radiative fluxes using the predicted values of temperature, humidity, cloud, and monthly-mean climatologies for aerosols and the main trace gases (Muñoz-Sabater et al., 2021).
The Global Land Data Assimilation System (GLDAS) reanalysis is a project lead by the National Aeronautics and Space Administration (NASA).The RB components are available in a spatial resolution of 0.25°× 0.25°, temporal resolution of 3 hours, for the 2000-present period (Oliveira et al., 2016).The radiation fluxes are calculated as a function of atmospheric transmissivity and emissivity, which are determined by cloud type and amount, derived from NOAA satellites (Rodell et al., 2004).
The Global Land Surface Satellite (GLASS) products are produced from multiple satellite observations, exploring the use of multiple algorithms for the same product to improve accuracy and stability, optimizing the use of temporal signatures in remote sensing data and the existing satellite high-level products (Liang et al., 2013;Zhao et al., 2013).Albedo and DSWR products are available in frequency of 8 days and 1 day, respectively.Both products have spatial resolution of 0.05°× 0.05°.The GLASS albedo product is produced from MODIS data, based on two direct estimation algorithms from surface reflectance, top of atmosphere radiance, and a statisticsbased temporal filtering fusion algorithm (Liang et al., 2013).The DSWR product of GLASS is generated based on an improved look-up table method using both polarorbiting and geostationary satellite data, including MODIS, Meteosat Second Generation (MSG) SEVIRI, the Multi-functional Transport Satellite (MTSAT)-1R, and the Geostationary Operational Environmental Satellite (GOES) Imager (Zhang et al., 2019(Zhang et al., , 2014)).
Clouds and the Earth's Radiant Energy System (CERES) is a mission of NASA that provided the climate community a 20-yr record of observed top-of-the-atmosphere (TOA) fluxes (Doelling et al., 2016).The CERES synoptic (SYN) product incorporates derived fluxes from the geostationary satellites (GEOs) in 1°x 1°spatial resolution and daily temporal frequency.OBS-Brazil dataset  et al., 2016).For conventional stations, DSWR was estimated using the Ångström-Prescott equation, while at automatic stations, DSWR was directly measured using Eppley thermopile pyranometers.

Radiation Balance (RB)
The RB can be defined, by its four components, as the difference between incoming and outgoing energies at the Earth surface, expressed as: With the exception of albedo (dimensionless), all components were calculated in W/m 2 .The DSWR is the radiation received directly or indirectly from the sun by a horizontal plane on the Earth's surface, and was extracted from the ERA5-Land, GLDAS and GLASS datasets.Albedo is the fraction of DSWR reflected by the Earth's surface, supplied by the ERA5-Land, GLDAS and GLASS datasets.The ULWR refers to longwave radiation emitted by the Earth's surface towards the atmosphere, extracted from the ERA5-Land and GLDAS datasets.The DLWR is the longwave radiation emitted by the atmosphere towards the Earth's surface, calculated by the ERA5-Land and GLDAS datasets, and through an adaptation of the SEBAL algorithm.This methodology was proposed by Bastiaanssen et al.
(1998) to calculate the energy balance on surface, and the DLWR is obtained by: where σ is Stephen-Boltzmann constant (5.6697 x 10 -8 W/ m 2 .K 4 ), ɛ a represents the atmospheric emissivity, T a (K) was taken from Xavier dataset, and τ is the one way atmospheric transmissivity, which is calculated in this study as the relationship between solar radiation incident on the surface and at the top of the atmosphere, using the following equation: where DSWR and Ro were extracted from Xavier dataset and CERES-SYN, respectively.

Statistic validation
The performance of the datasets in the calculation of the RB components was determined by comparison reanalyses, satellite-based and blended data with USR, PDG and Marambaia observations.Four statistical indices were used: correlation coefficient (r), bias, root-mean-square error (RMSE) and mean relative error (MRE).The linear relationship between estimates and observations is explain by r.The bias is the tendency to overestimate or underestimate the error.The RMSE is the general error in the predictions in relation to the actual observed value.The MRE is a measure of forecast accuracy, expressed as a percentage.These statistical indices are widely used in validation studies of the RB components (Ferreira et al., 2020;Oliveira et al., 2016;Silva et al., 2015;Verma et al., 2016;Zeng et al., 2020).It is noteworthy that, although there are few stations in relation to the size of the territory, they represent different soil cover conditions, environmental and, in particular, distinct weather and cloud conditions in the study area.
To understand the variations of each evaluated products, the main variables that influence the RB components are analyzed.Previous studies show that DSWR, Albedo, DLWR and ULWR are strongly determined by cloud cover fraction (CF), enhanced vegetation index (EVI), Ta and land surface temperature (LST), respectively (Jiao et al., 2015;Oliveira et al., 2018;Wang et al., 2018;Wild, 2016;Zuluaga et al., 2021).Except for Ta (directly measured at all stations) the other variables were obtained from MODIS sensor products.

DSWR
The DSWR is the basic energy for biological, physical and chemical processes (Zhang et al., 2020), as well as being an increasingly attractive resource to meet growing energy demands through photovoltaic energy conversion (Wild et al., 2015).For validation of the DSWR, Table 2 shows the performance results of ERA5-Land, GLDAS and GLASS to estimate DSWR with respect to stations.Figure 2 shows the comparison between observations and the datasets based on scatterplots analyses.
The standard deviation (SD) indicates that the monthly variability of datasets ranges from 30% to 40%, compared to their annual mean values (Table 2).ERA5-Land shows a statistically significant underestimation of DSWR for all stations (Fig. 2a-c, Table 2).GLDAS and GLASS do not present statistically significant bias, except for GLDAS at the PDG station (bias = 10 W/m 2 , Table 2).Differences between GLASS and GLDAS are evident when analyzing the error values, which is smaller in GLASS (Table 2).The latter displays values of RMSE < 40 W/m 2 and MRE < 18% for all stations, while the lowest GLDAS values are RMSE = 43 W/m 2 at PDG and MRE = 21% at USR. ERA5-Land presents the largest errors at the Marambaia station (RMSE = 93 W/m 2 and MRE = 49%) and the smallest at the PDG station (RMSE = 55 W/m 2 and MRE = 25%), still much larger than those presented by GLASS.Verifying the ERA5-Land at the USR station, the r values (0.7-0.9) suggests a moderate-strong relationship between the datasets and observations (Table 2).Among all datasets, GLASS displays the highest concentration of scatter points on the 1: 1 line in the three stations (Fig. 2).
The comparison between gridded products and local station measurements can induce errors, because the pixel assumes unique values for spatially heterogeneous surface, which may not correspond to the observations surface type (Huang et al., 2016).Thus, spatial resolution plays an important role, as in the case of GLDAS, where the proper scatter (Fig. 2d-f), may be related to its grid (0.25°lat/lon), as argued by Oliveira et al. (2016) in the Amazon.However, ERA5-Land, despite having a more refined spatial resolution (0.1°lat/lon), presents larger scattering (Fig. 2a-c) and higher errors (Table 2) in comparison to other datasets.
Figure 3 presents the daily DSWR variability by month in the three stations.As expected, the maximum values of observed DSWR (362 W/m 2 at USR, 340 W/m 2 at PDG, and 371 W/m 2 at Marambaia), is observed for the austral summer.This season takes place between December and February, when the solar zenith angle over the Southern Hemisphere is the lowest and, therefore, more radiation is received at the top of the atmosphere.The minimum values of observed DSWR appear in September, in the beginning of spring (26 W/m 2 at USR and 25 W/m 2 at PDG), and in July, during winter (34 W/m 2 at Marambaia).All datasets are able to reproduce the DSWR seasonal variability.During summer, the highest DSWR values are displayed, as well as greater variability (longer box plots), associated with changes in rainfall and cloudiness, caused by South American Monsoon (SAM).This, begins at Equatorial Amazon in the spring, and spreads rapidly to the east and southeast of the country during summer, boosting the SACZ (Garcia and Kayano, 2015).In contrast, in winter, the lowest DSWR values and the least variability (shorter box plots) are shown, related to the dry season and the cloudless sky that mark the end of SAM (Garcia and Kayano, 2015).In this period the cloudiness of the southeastern region results from the presence of substantially cold frontal systems (Zandonadi et al., 2015).
Considering that Zuluaga et al. (2021) found that cloud cover is the main factor that contributes to the DSWR variations in southeastern Brazil, Fig. 4 shows the correlation between cloud fraction (CF), and the observed and estimated DSWR standard deviation.At the USR (Fig. 4a) and PDG (Fig. 4b) stations, values of r (> 0.8) indicate that, in general, datasets are able to simulate the strong influence that CF has on DSWR variations.In these two seasons, ERA5-Land presents the main lags in the spring-summer transition (October-December), when SACZ starts to act on the region.The Marambaia station (Fig. 4c) exhibits greater CF variability, probably related to constant advection from the ocean, making it more difficult for datasets simulations.Here, ERA5-Land and GLDAS show difficulty in simulating DSWR variability during spring and much of summer (December and January).GLASS displays correlation values closer to the observations, with some differences during autumn and winter, when CF is mainly related to frontal systems.
In short, the combination of geostationary and polar orbit satellites, high spatial resolution, and algorithms for different cloud conditions, makes GLASS the best option for DSWR studies in southeastern Brazil.In addition to the previous one, all datasets are restricted to continental areas, which should influence the large errors presented in pixels near the coast (Kara et al., 2007;Pelosi et al., 2020), as in the case of the Marambaia station.

Albedo
Albedo is a fundamental parameter of the RB; since it controls the energy budget through the regulation of the DSWR quantity reflected by the surface (He et al., 2018).Table 3 contains the statistical analyses of the datasets in estimating the surface albedo.Figure 5 shows the scatterplots between observed and estimated albedo from the datasets.
Observed data show a mean albedo of 0.189 but varying substantially during the studied period (SD = 0.016).These values are underestimated by the datasets with a mean bias of -0.054 (Table 3), and a concentration of scatter points below the 1: 1 line (Fig. 5a).Despite these results, GLDAS and ERA5-Land show positive statistically significant correlations.At the PDG station, biases (Table 3) show a reasonable performance for ERA5-Land (0.003) but statistically significant underestimation of albedo for GLDAS (-0.0048) and GLASS (-0.022).Moreover, these datasets do not follow the high variability of the observations as demonstrated by reduced standard deviation (SD = 0.018).ERA5-Land and GLDAS show negative correlations, but not statistically significant.The RMSE range from 0.020 to 0.051, and the MRE from 10.351% to 29.152% (Table 3).The low variability of the datasets is reflected in the line-shaped scatterplot in Fig. 5b.
In opposite to that has been found for the DSWR (Section 3.1), at the Marambaia station the best agreement is found for albedo estimates among the datasets (Table 3).These results are explained, in part, by the low variability of observations (SD = 0.009), which perhaps can be better treated by the reanalysis and satellite data.GLASS has the highest correlation (r = 0.695), while ERA5-Land exhibits a negative and statistically significant correlation (r = -0.407).The scatters in Fig. 4c show the reasonable fit between GLASS and observations, as shown by a bias close to 0 (Table 3).The GLASS delivers the smallest errors (RMSE = 0.006, MRE = 3.592%) and GLDAS and ERA5-Land outperformed GLASS errors by almost 5 times (Table 3).
Considering that the albedo depends on the characteristics of the surface (vegetation in this case), each station is discussed separately.For this, Fig. 6 shows the variability of albedo and the enhanced vegetation index (EVI) by month in the three stations.EVI is a frequently used remote-sensing vegetation phenological metric, which is optimized to resist atmospheric and soil background effects (Wang et al., 2017).
In February (regrowth phase), due to the absence of leaves, the higher reflectance of the soil determines and contributes to the highest (lower) albedo (EVI) value of the period (α = 0.22, EVI = 0.26) (Oliveira et al., 2018;Williamson et al., 2016).Between March and July, during the tillering phase, albedo decreases to approximately 0.19 related to the appearance and development of sprouts, leading to larger EVI up to 0.42.Between August and September, the albedo continues to decrease until 0.17 due to the total coverage of the soil by the foliage of the plants (EVI = 0.46) favoring absorption of solar radiation.Finally, between December and January, the maturation and senescence of the leaves appear (EVI = 0.41), and the albedo value returns to values close to 0.20 (Fig. 6a).These results are in line with those found by Oliveira et al. (2018) and Scarpare et al. (2016).
The PDG station presents the highest albedo values (α ≈ 0.17) during the spring when the renewal of the foliage generates greater vegetative vigor (EVI = 0.42).The lowest albedo values (α ≈ 0.15) correspond to autumn/ winter (Fig. 5b).According to Oliveira et al. (2018), the reduction in albedo is due to the fact that during the dry season (winter) DSWR penetrates the canopy more easily due to less foliage (EVI = 0.35), and the soil covered by dark plant litter absorbs more radiation.
According to Carvalho et al. (2015), the characteristic vegetation that surrounds the Marambaia station, presents its peak of leaf fall during the dry season (July-September) and beginning of the rainy season (October-November) with mean EVI value of 0.45.This, together  with the high content of organic matter in the soil, contributes to the reduction of albedo to 0.12 (Fig. 5c).However, between December and June, the vegetation has its greatest vegetative vigor (EVI ≈ 0.56) reaching an albedo of up to 0.15.
Turning to the GLDAS product is noticed that it is able to reproduced the albedo seasonality pace at USR and Marambaia stations, but with a clear underestimation (Fig. 6a, c).At the PDG station, albedo is also underestimated and remains practically constant throughout the year (Fig. 6b).These values may be associated with the fact that the GLDAS uses a global static land cover dataset, based on observations from the AVHRR in 2000 (Ro-    , 2004).ERA5-Land presents practically the same albedo cycle (with the similar values) at USR and PDG stations (Fig. 6a-b), which leads to the conclusion that despite its high resolution there is no differentiation between USR and PDG land cover.At Marambaia station an out-of-phase pattern of the albedo is shown with respect to observations (Fig. 6c).This is very likely related to caveats in ERA5-Land along coastline regions primarily reproducing sea-land mean albedo (Pelosi et al., 2020).GLASS matches the seasonal variations of observed albedo in the 3 stations (Fig. 6), but values are underestimated in USR and PDG (Fig. 6a-b).According to Liu et al. (2013), GLASS presents deficiencies related to the quantity (and quality) of data that is applied to training the regression algorithms used to calculate the albedo, mainly in tropical areas.
Additionally, two factors must be considered: 1) albedo observations are local, while GLDAS and GLASS use satellites, whose nominal resolution may be 1 km, but generally perform spatial averages to remove the effects of cloudiness.2) Satellite detection leads to a sensitivity limit in the decimals of the albedo values and even to systematic errors.These two facts help to understand that the albedos observed by GLDAS and GLASS are practically constant throughout the year.

DLWR
The DLWR, also known as thermal infrared energy, is considered a fundamental indicator of the effect of atmospheric greenhouse gases (water vapor, CO 2 , etc.) on the climate (Tang et al., 2021).However, DLWR observations are rarely available due to the cost of the instruments for their measurement (Kruk et al., 2010).Figure 7 shows the comparison of daily DLWR estimates and observed values.The statistical results of the comparisons are summarized in Table 4.
The observed values show that the DLWR variability (SD) at the Marambaia station is twice as high as at the USR station (Table 4).The r values of all datasets are statistically significant, however, at USR, the correlations are strong (r ≥ 0.7), while at Marambaia this is moderateweak (0.4 ≤ r ≤ 0.5).The complexity of estimating DLWR leads to different biases among datasets.ERA5-Land presents statistically significant bias, with overestimation at USR (bias = 14 W/m 2 ) and underestimation at Marambaia (bias = -4 W/m 2 ).GLDAS and SEBAL show completely different biases with values between -4 and 2 W/m 2 (without statistical significance) at USR, and underestimations between -45 and -26 W/m 2 (with statistical significance) at Marambaia (Table 4).The GLDAS, however shows the best fit with the observed DLWR at USR (Fig. 7b).It should be noted that at both stations (Table 4).
Figure 8 shows observed and estimated temporal variability of DLWR.Largest DLWR are noted in the USR (≈ 375 W/m 2 ) and Marambaia (≈ 420 W/m 2 ) stations du-ring spring/summer, in agreement to maximum cloudiness and precipitation across the Southeast region (Coelho et al., 2016).In opposite, the lowest DLWR values appear during the dry period (autumn/winter), with ≈ 320 W/m 2 at USR, and ≈ 398 W/m 2 at Marambaia (Fig. 8).This pattern is associated with the fact that the emissivity and atmospheric temperature (see Eq. ( 2)) present their maximum values during summer, and minimum values in winter (Ferreira et al., 2012).The influence of water vapor should not be disregarded.For instance, higher DLWR values at Marambaia station is associated with maritime advection of water vapor onto the continent in line with the sea breeze effect.This might increase the atmospheric humidity inducing higher DLWR (Brito and Oyama, 2014;Marques et al., 2010).
According to Wang et al. (2018), even on cloudy-sky days, Ta is a main controlling factor that influences the surface DLWR. Figure 9 shows the correlation between Ta and DLWR, measured at each station.Both places show high and statistically significant correlations (r > 0.8).At the USR station (Fig. 9a), observed and estimated DLWR fit nicely the pace of the Ta seasonal cycle.In Marambaia (Fig. 9b) disagreements between observation and estimation are exhibited during autumn and winter, probably due to limitation of ERA5-Land and blended data to reproduce the Ta daily variability.In this period, air temperatures may be affected by periodic stratocumulus clouds.During the spring/summer, DLWR from ERA5-Land matches the observations.The weak seasonal march of Ta in Marambaia is also reflected in the low variability of DLWR.In general, with all limitation, the datasets are able to capture the DLWR seasonality in both stations.However, caution should be taken if these data have to be used on high frequency, such as daily values.

ULWR
The ULWR is the component of the RB that mainly represents the thermal radiation capacity of the Earth's surface, dominating at night, in high latitudes and during most of the year in the Polar Regions (Jiao et al., 2015;Qin et al., 2020).Table 5 and Fig. 10 show the comparison between observed ULWR and estimates from two datasets.
The variability of ULWR over the measurement period at the two stations is shown in  depends on the DSWR amount -unlike the albedo which essentially depends on the type of ecosystem and the surface humidity conditions.In this sense, Fig. 12 shows the correlation between LST and ULWR.At the USR station (Fig. 12a); it is clear that the good performance of the datasets is related to the strong correlation (r > 0.94) that exists between the ULWR and LST cycles.In Marambaia (Fig. 12b), although the correlations are moderate (0.59 ≤ r ≤ 0.72), they are statistically significant.The main differences shown by the datasets are related to the underestimation of the effect of LST on ULWR, mainly in the case of GLDAS, and with some disagreements in the direction of the variations during the spring/summer especially.The optimal performance of datasets in the ULWR estimate, can be related to the fact that LST, in ERA5-Land and GLDAS, is one of the most accurate variables in relation to reality, due to the correction of bias in the data assimilation processes (Hersbach et al., 2020;Rodell et al., 2004).The slight advantage of ERA5-Land may be related to its high resolution (0.1°) compared to GLDAS (0.25°).

Conclusions
In this study, several datasets were evaluated on the southeast region of Brazil, which is a region with abrupt environmental changes caused by its economic growth and high population density.The components of the radia-tion balance from five datasets (obtained directly or indirectly) were compared with observations from three stations, located inside surfaces with different physical characteristics.For this, we use indicators that estimate the accuracy of the datasets (r, bias, RMSE and MRE).
The results showed that ERA5-Land offered the best performance in estimating albedo, DLWR and ULWR.In the case of DSWR, GLASS delivered the most accurate values in relation to the observations.Both datasets proved to be an adequate alternative to estimate the components of the radiation balance in southeastern Brazil.
For the DSWR, DLWR and ULWR components, the datasets presented the biggest errors in the Marambaia station, probably because these are products restricted to the land areas.These products can also be subject to parameterization errors in the radiative transfer models, such as limitations to capture variations generated by dominant meteorological systems (cloud changes) in the spring/ summer period.Albedo was the most problematic component due to the low seasonal variability shown by datasets in all stations.The worst albedo results were observed in the USR station, which corresponds to a sugarcane field.The datasets were unable to accompany the high variability of albedo due to the phenological cycles of the crop.
Analyzes conducted in this study highlight uncertainties that still represent posing a challenge for the research community in the development of products that compute the radiation balance components.Future studies  should explore the combination or merging of different datasets aiming at correcting errors in the coastal areas and the albedo estimates.Adequate calculations of the radiation balance components will lead to fundamental advances for the hydrological and ecological communities to improve their estimates of sensitive and latent heat, evapotranspiration, gross primary production, and surface climate projections at large.

Figure 1 -
Figure 1 -Study area characteristics and localization of stations.

Figure 2 -
Figure 2 -Comparison between DSWR estimations from datasets and observed in USR (a, d, g), PDG (b, e, h) and Marambaia (c, f, i) stations.ERA5-Land is red, GLDAS is green and GLASS is blue.

Figure 3 -
Figure 3 -Variability of daily DSWR by month in the a) USR, b) PDG and c) Marambaia stations from observations and datasets.Box plot include the interquartile range (25th-75th percentiles), median (horizontal line), mean (black circles), maximum and minimum values (black dots).

Figure 4 -
Figure 4 -Correlation (r) between monthly cloud fraction (CF) from MOD08_M3 and standard deviation (SD) of DSWR, in the a) USR, b) PDG and c) Marambaia stations.Significant r values at 95% level (pvalue ≤ 0.05) are accompanied by *.

Figure 5 -
Figure 5 -Comparison between albedo estimations from datasets and observed in a) USR, b) PDG and c) Marambaia stations.

Figure 6 -
Figure 6 -Variability of monthly EVI from MOD13Q1 and weekly albedo by month in the a) USR, b) PDG and c) Marambaia stations from observations and datasets.Error lines correspond to standard deviation.

Figure 7 -
Figure 7 -Comparison between DLWR estimations from datasets and observed in USR (a-c) and Marambaia (d-f) stations.

Figure 8 -
Figure 8 -Variability of daily DLWR by month in the a) USR and b) Marambaia stations from observations and datasets.Box plot include the interquartile range (25th-75th percentiles), median (horizontal line), mean (black circles), maximum and minimum values (black dots).

Figure 9 -
Figure 9 -Correlation (r) between observed monthly air temperature (Tair) from a) USR and b) Marambaia stations, and monthly DLWR.Significant r values at 95% level (p-value ≤ 0.05) are accompanied by *.
Figure 10 -Comparison between ULWR estimations from datasets and observed in USR (a,c) and Marambaia (b,e) stations.ERA5-Land is red and GLDAS is green.

Figure 11 -
Figure 11 -Variability of daily ULWR by month in the a) USR and b) Marambaia stations from observations and datasets.Box plot include the interquartile range (25th-75th percentiles), median (horizontal line), mean (black circles), maximum and minimum values (black dots).

Figure 12 -
Figure 12 -Correlation (r) between Land Surface Temperature (LST) from MOD11C3 and monthly mean ULWR in the a) USR and b) Marambaia stations.Significant r values at 95% level (p-value ≤ 0.05) are accompanied by *.

Table 1 -
Characteristics of gridded datasets used in this study.Ta and DSWR from 735 weather stations across Brazil.These were interpolated in a gridded of 0.25°x 0.25°for DSWR and 0.1°x 0.1°for Ta, in the 1980-2017 period (Xavier

Table 2 -
Statistic performance of datasets to estimate DSWR compared to the observations (Obs) from stations.Values in bold indicate significant values at 95% level (p-value ≤ 0.05).

Table 3 -
Statistic performance of datasets to estimate albedo (α) compared to observations (Obs) from stations.Values in bold indicate significant values at 95% level (p-value ≤ 0.05).
Tabela 4 -Statistic performance of datasets to estimate DLWR compared to station observations (Obs).Values in bold indicate significant values at 95% level (p-value ≤ 0.05).

Table 5 -
Statistic performance of datasets to estimate ULWR compared to station observations (Obs).Values in bold indicate significant values at 95% level (p-value ≤ 0.05).