Eddy-covariance CO 2 fluxes over Itaipu lake , southern Brazil

In this work, we present CO2, latent heat and sensible heat fluxes measured over the reservoir of the Itaipu Hydroelectric Power Plant (Paraná State, Brazil) during 2013. A tower was installed at a small island in the reservoir, where an Eddy Covariance system, with supplementary equipments and analysers, was deployed. The objective of this work was to determine the magnitude of CO2 fluxes and their variation throughout the year. CO2 flux displayed seasonality: in warm months there was a predominance of negative CO2 daytime fluxes and positive CO2 nighttime fluxes, which we attributed to photosynthesis/respiration in the reservoir; in cold months there was a predominance of negative CO2 fluxes in both periods, with CO2 air concentration apparently imposing the signal of the CO2 fluxes and the strong winds intensifying them. The range of 90% of the CO2 fluxes measured in this work (−102.68 to +151.72 μg m −2s−1 at the 30-min. time scale) is comparable to those observed in natural lakes and reservoirs around the world. On the average, the reservoir acted as a source of CO2, with an overall mean flux of +12.78 μg m −2s−1.


INTRODUCTION
In the last decades, continental water bodies have been found to be potentially important sources of greenhouse gases to the atmosphere. In particular, reservoirs of hydroelectric plants have been under intense scrutiny since the work of Rudd et al. (1993), which for the first time presented evidence that they could be important sources of CO 2 and CH 4 .
Currently, the overall carbon balance of lakes and reservoirs remains subject to considerable uncertainty, and their role as sources or sinks in the carbon cycle remains, to a large extent, unknown (Huotari et al., 2011;Mendonca et al., 2012). In several cases, a reservoirs and lakes have been reported to be large accumulators of carbon in their sediments (Tranvik et al., 2009;Mendonca et al., 2012;Ometto et al., 2013;Bernardo et al., 2017). As sources of atmospheric CO 2 and CH 4 , however, there is comparatively less uncertainty: most studies generally report net carbon dioxide and methane emissions into the atmosphere through the water surface of freshwater ecosystems (Rosa et al., 2003;Soumis et al., 2004;Kemenes et al., 2011;Huotari et al., 2011;Mammarella et al., 2015;Erkkilä et al., 2018) and, in the case of hydropower reservoirs, as a result of "degassing" as well (the process of ebullition of CH 4 and CO 2 as the water pressure drops suddenly downstream of the turbines) (Galy-Lacaux et al., 1997;Soumis et al., 2004).
Clearly, the subject of carbon balance of lakes and reservoirs is a complex one, and considerable efforts will be needed before a more definitive picture of their overall contribution to emission and sequestration of carbon, in its various forms, can be reached. In this work, we adopt as an assumption that CO 2 fluxes across water surfaces fluctuate over time in sequences of positive and negative values according to local drivers. With no a priori assumption about the reservoir role (if it is a source or a sink of CO 2 ), we will concentrate on the question of quantifying CO 2 exchanges across the water surface of a large subtropical reservoir by means of the eddy covariance (EC) method.
The EC technique is widely accepted for the measurement of greenhouse gas (GHG) fluxes at the ecosystem scale: it has been applied to all kinds of terrestrial ecosystems (Xu & Baldocchi, 2004;Lewicki et al., 2007;Suni et al., 2003;Hatala et al., 2012). For water bodies, however, its application is more difficult due, among other factors, to the requirement of a stable measuring platform, power supply and logistical difficulties. This partly explains the choice of the chamber method and the mass-transfer method (which requires measuring the partial pressure of CO 2 in water, usually by means of an analytical water-air equilibration approach, and in the air), which are relatively simple to apply (Kutzbach et al., 2007;Paranaíba et al., 2018), as the principle measuring techniques in water body studies of GHG emissions (Rosa et al., 2003;Zhu et al., 2012;Ometto et al., 2013;Abril et al., 2014;Pacheco et al., 2015). However, floating chambers have been criticized for causing modification of the flow at the water-air interface (Richey et al., 2002;Schubert et al., 2012). Another limitation is that point measurements in space and time of these methods need to be extrapolated for the whole water body, and are sometimes used to obtain estimates of long-term average GHG fluxes (Vesala et al., 2006;Podgrajsek et al., 2014;Paranaíba et al., 2018).
In Brazil, there has been a considerable effort to estimate GHG emissions from the reservoirs of hydroelectric plants (dos Santos et al., 2006;Kemenes et al., 2011;Ometto et al., 2013;Marcelino et al., 2015;Vale et al., 2017), partly due to their importance in the country's energy generation matrix: 60.3% of Brazil's electrical power are produced by hydroelectric plants (Empresa de Pesquisa Energética, 2018). The magnitudes of the net emissions of greenhouse gases to the atmosphere reported in those studies varied significantly: for example, Kemenes et al. (2011) estimated fluxes of 2.9 ton C/MWh from the Balbina hydroelectric reservoir. In comparison with a baseline emission per MWh from an equivalent coal powered thermoelectric plant (0.3 ton C/MWh), Balbina would emit almost ten times more GHG per MWh. On the other hand, Ometto et al. (2013) reported much lower emissions from 8 Brazilian hydropower reservoirs: their emissions varied from 0.01 to a maximum of 0.55 ton C/MWh, with only two reservoirs having emissions comparable to a thermal powerplant. All those studies resorted to the chamber method. The limitations of the chamber method were discussed above, and were present in those studies as well: temporal coverage was limited; seasonality had to be inferred from sparse measurements; and the measurement campaigns were limited by weather conditions (the method can't be used under high winds, for example ).
In contrast to chamber measurements, the EC method allows direct measurement of the turbulent flux on a continuous basis. The technique is non-intrusive and the measurement is representative of a region upwind of the measuring tower which is of the order of a few hundred meters, called the "footprint" (Schuepp et al., 1990).
In this work, we present 11 months of CO 2 flux measurements (with gaps: 149 days in total) at a large reservoir of a hydroelectric plant in Southern Brazil (Itaipu). The main objective is to quantify the fluxes of CO 2 across the water surface. Fluxes of sensible and latente heat were also measured and are briefly described as they are needed for the calculation of atmospheric stability and density corrections. We apply a footprint-based methodology to select the fluxes which come exclusively from the water surface in a reservoir where the water level fluctuates. The measurements allowed to determine the seasonality of CO 2 fluxes at the Itaipu reservoir, with the lake functioning mostly as a source of CO 2 to the atmosphere.

Site and measurements
Itaipu reservoir is located in the valley of the Paraná River, at the border between Brazil and Paraguay. The reservoir reached its normal operating level at the 220-m elevation (above mean sea level) for the first time in June 1984. At this level, the lake has a surface area of 1350 km 2 .
The drainage basin formed between the Itaipu dam and the previous dam located approximately 400 km upstream from it on the same river (Sérgio Motta Hydroelectric Power Plant) has an area of 147000 km 2 . This area presents three main types of land use: natural forests (∼ 12%), agriculture (∼ 48%) and pasture (∼ 40%) (Brasil, 2014). Most of the forested area is concentrated on the Western margin (Paraguayan), but a 30-m wide belt of reforestation exists along the Eastern (Brazilian) margin as well. Deciduous trees characterize the forest in this area. Crops are found on both sides of the lake.
A small micrometerological tower was set up on a flat, almost submerged, hill top (latitude −25 • 03ˈ25.72"; longitude −54 • 24ˈ33.67"; and altitude 218 m ASL). Note that the hill top is two meters below the normal operating level. At lower water levels, a small island appears. This provides very good conditions for EC measurements, with a fetch of 3000 m to the North, 1500 m to the West, 2000 m to the South and 500 m to the East (see Figure 1). In Figure 1 we also plotted an approximation of the reservoir's bed surface around the EC station, downloaded from Instituto de Terras Cartografia e Geociências do Paraná (2020).
At the EC station, a sonic anemometer (Campbell Scientific Instruments -CSI CSAT3), an infrared CO 2 and H 2 O analyzer (Licor LI-7500), and four thermocouples (CSI FW03) were installed to make high-frequency (20 Hz) turbulence measurements. The thermocouples were installed above, below and in the center of the sonic path, and in the center of the optical path of the infrared analyzer. The sensors were deployed on a small metal tower and placed at 3.76 m above the ground (Figure 2).
Low-frequency measurements (0.1 Hz) were also made of mean CO 2 concentration (Vaisälä GMP343; two at 1.77 m and two at 3.66 m); mean temperature and relative humidity (CSI CS500; at 2.85 m); mean atmospheric pressure (CSI CS100; at 1.73 m); mean downwelling solar radiation (Kipp & Zönen; at 2.67 m) and rainfall intensity (MRI; at 3.00 m). The pluviometer suffered frequent clogging by wasp nesting, and daily data from a nearby land station (Itacora station operated by Águas Paraná, located at latitude 25 • 06ˈ33.99" and longitude 54 • 23ˈ58.99") was used to detect rainfall periods for quality control of the LI-7500 measurements. To measure water temperature we installed two CSI L108 temperature sensors in a nautical buoy (Alcântara et al., 2013) moored 3 km northwest of the EC site (see Figure 1) at the surface (skin surface temperature) and 25 cm below the water level. Although it was planned to measure several variables at the buoy, including pH, a lightning strike interrupted its measurements soon after our micrometeorological station was set up (note that the water temperature measurements were independent of the rest of the sensors at the buoy, and proceeded without problems). Although there are no simultaneous measurements (other than water temperature) at the buoy and the micrometeorological station, we used the water pH measured at the buoy from July to November 2012 to help to interpret some of our CO 2 flux observations in 2013 (data available at SIMA (Sistema Integrado de Monitoramento Ambiental, 2020)) (see sections 3.4 and 4).
All the high-frequency raw turbulence measurements were made continuously and stored in a computer's disk running at the site, except for power interruptions, when the solar panels were not able for some reason to recharge the batteries, in 10-minute data files. The lowfrequency data were stored as 10-min. averages. We call the 10 min. measurement periods "blocks". Non-continuous measurements proceeded from December 2012 through April 2014. After July 2013, a robust power supply system was installed that increased the number of days with data collection considerably. The measurements presented in this work are from January to November 2013. Table 1 shows the beginning and end of each measurement period when the station was actively measuring, as well as the total number of blocks for those periods and the total number of blocks selected in a quality control process described in the next section.

Eddy covariance data processing and analysis
In a first round of data quality control, we eliminated spikes from the blocks by removing points lying outside the interval defined by five times the standard deviation (±5σ) around a 5-minute average from a linear autorecursive filter (the linear filtering is easier to implement, and computationally faster, than a centered running average) (Vickers and Mahrt, 1997). If the total number of spikes in each block was less than 50, the resulting gaps were filled with the mean from each run; otherwise, the quality control criterion was to eliminate it from further analysis. We also corrected the time lag between CSAT3 and LI-7500 sensor measurements, with the finewire thermocouple placed in the middle of the LI-7500 path. The CO 2 concentration data were shifted by an amount corresponding to the peak of the cross-correlation function between temperature and CO 2 fluctuations, for each run. The same procedure was done between temperature and vertical velocity. This methodology corrects the delay in the internal processing of the sensors and the loss of covariance generated by the spatial separation between CSAT3 and LI-7500 (Qi et al., 2015).
In addition, we grouped three consecutive 10-minute blocks to get the CO 2 fluxes from 30-minute blocks. In a procedure similar to that established by Vickers & Mahrt (1997), to select stationary 30-minute blocks, we calculated CO 2 fluxes whose CO 2 kinematic fluxes of the 30-minute block ( 30 F ) and its 10-minute sub-blocks ( 10 F ) were in the range of ( ) The above-mentioned quality control procedures selected the number of blocks from each of the measurement periods shown in Table 1, column 5.
A 2-D rotation (Finnigan et al., 2003) was applied to each 10-min. block that passed the quality control process to align the x axis with the mean wind direction and to set the mean vertical wind velocity to zero. The fluctuations were extracted after linear detrending (Moncrief et al., 2004) over each 10-min. block.
Before the beginning of the experiment, the zero and span of the LI-7500 were adjusted per the manufacturer's instructions. After this, the sensor was deployed in the micrometeorological station, where was kept during the whole campaign. The remoteness and access conditions of the site made field calibration difficult. Thus, we used a very stable slow-response GMP343 as a surrogate standard. The adopted procedure is described in full in Armani (2019) and Armani et al. (2020) and consists of correcting the LI-7500 CO 2 measurements with multiple regressions adjusted to the data from each of the periods in Table 1, but for daytime and nighttime periods separately. We used auxiliary mean (10 min.) meteorological observations of temperature (θ), water vapor concentration (ρ v ) and atmospheric pressure (p) to correct the LI-7500 CO 2 measurements (ρ c,l ) as follows: based on the assumption that GMP343 measurements converted to CO 2 mixing ratios are correct (r c,g ), we calculated the linear regression to obtain turbulent fluctuations of CO 2 mixing ratios with In Equation 1, the δa's are obtained from ' a a a δ = + , where a is any of the variables r c,g , ρ c,l, , θ, ρ v and p; ' a is the mean of the 10-minute blocks and a is the average of the a measurements during the whole period, but calculated separately for daytime and nighttime. Turbulent CO 2 fluctuations were grouped in 30-minute samples to obtain CO 2 fluxes with where d ρ is 30-min. mean dry air density and w′ is the turbulent fluctuation of vertical velocity. The procedure outlined above for the calculation of F c was carefully checked under the assumption of perfect calibration, where it was shown to produce identical results to the WPL correction applied separately to 10-min. data blocks.
Water vapor fluxes (E) were corrected with the WPL correction (Webb et al., 1980), and sensible heat fluxes were obtained from In Equations (4)-(5) p c is the specific heat of air, ρ is the mean density of air and v µ is the ratio of the molecular weights of dry air to water vapor. To obtain the latent heat fluxes (LE), we multiplied E by the latent heat of vaporization: L = 2462 kJ kg -1 . In (2)-(5), all variables obey a Reynolds decomposition, with overbars denoting mean quantities, and primes denoting turbulent fluctuations.
For some periods, the reservoir's water level was low enough to render above water a significant amount of land in the island over which the EC station was installed. Therefore, we Table 1. Periods during which the micrometeorological station was operating, its corresponding total of 10-minute blocks and the total number of blocks selected in a quality control (Q.C.) process. Times are in local time (LT) (local time = UTC -3).

5/14
obtained an estimate of the expected contribution of this land surface to the fluxes measured by the tower. To take into account the heterogeneous footprint we used the analytical 1-D model (along the wind direction) of Hsieh et al. (2000), (with a cross-wind gaussian distribution as proposed by Gryning et al. (1987)), where . k 0 4 = is von Kármán's constant, D and P are constant over different stability ranges (see Table 2), and u z is given by where m z is the height of the measurements above water (which varies over time), and 0 z is the momentum aerodynamic roughness length of the surface.
When the water surface level is above 219 m ASL, the island is essentially flooded, and the EC footprint is all over water. A corresponding . 0 z 0 0001 = m was used in these situations (Panofsky & Dutton, 1984). Whenever the water level was below 219 m ASL, an effective roughness that took into account the heterogeneous footprint of water and the land surface of the island was calculated iteratively using the method proposed by Hutjes et al. (2010),  (Panofsky & Dutton, 1984), with the index i, which selects water or land, running over each grid of the discretization of the footprint along the wind direction. The procedure clearly needs to be iterative, as the contribution of water and land to the footprint depends of 0ef z , but converges quickly.

Footprint analysis
From Figure 1, it is seen that the closest distance from the island to the reservoir's margin is approximately 500 m to East. Overland distances from the EC sensors to the water are 95, 140, 70 and 93 m to the North, South, East and West at the lowest 216-m water level. A composite shape, made up from 3 ellipses, that grows linearly with water level was adopted to represent the island's contour, and is depicted as a dashed line in Figure 3. In that figure, we show the footprint contours for the  In Figure 3, the lines show the 90%, 70% and 50% footprint contours, whereas the dashed line indicates the island's contour at the lowest water level of the range. Notice that this gives a conservative view of the contribution of the land surface.
Most of the measurements were made while the water level was above 219 m, totaling 58.61% of the blocks. 25.02% of the blocks were measured while the water level range was 218-219 m; 13.43% in the range of 217-218m; and only 2.94% bellow 217m. The footprint analysis shown in Figure 2 clearly indicates that the EC tower does not "see" fluxes from the mainland (the footprint does not reach the margins of the reservoir), but that there is considerable flux from the island itself for the ranges of 216-217 and of 217-218m (Figures 2c and 2d). Thus, we removed all the measurements made when the water level was below 218m. This eliminates a further 16.37% of the fluxes that passed all the quality control procedures described above. The number of remaining runs after the footprint analysis for each month is listed in Table 3.

Meteorological variables and water surface temperature
In Figure 4 we show the daily means of water surface temperature, air temperature, atmospheric pressure, specific humidity and CO 2 air concentration, as well as daily rainfall. Interruptions of plotted lines or data points indicate the periods when the sensor in question was not operating.
Daily mean air temperatures are in the range of 10-30 °C most of the time, and surface water temperature varies much less, as expected, in the range of 20-30 °C. Data are missing for most of the wintertime, but during springtime there are large temperature fluctuations in phase with opposite atmospheric pressure fluctuations associated with the passage of weather systems. Figure 4 also shows that air CO 2 concentration is higher in colder months than in warmer months, suggesting the effect of a lower biological activity related to photosynthesis both in the water and on the surrounding land. The effect of air density variation on CO 2 concentrations is also quite clear in this figure, since this concentration varies with air temperature, especially in the abrupt variations measured in August and September.
Wind roses for daytime and nighttime, for all the data available after quality control and footprint analysis are shown in Figure 5. They are not too different, with more westerly winds during the night. There is a very low frequency of winds coming from the East, which is the direction closest to the margin (c.f. Figure 1), therefore assuring that the measurements are highly representative of the water surface. The most frequent directions reflect the topographical effect of the reservoir's orientation.

Sensible and latent heat fluxes
The overwhelming majority of measurements were made in the range of −1 ≤ ζ ≤ +1, where ζ is Obukhov's stability variable, which is fairly typical of stability conditions found in the surface layer. Furthermore, most measurements were performed under unstable atmospheric conditions, totaling 58.5% of all measurements.
The hourly means of the H and LE fluxes over each period are shown in Figure 6a and b. Note that the total number of available runs is different for each month, which affects the accuracy of the averages shown. Still, Figure 6 provides a convenient way to summarize flux patterns. The most noteworthy features of H are its low values in comparison to those over land surfaces (typical of water surfaces), and that it is most often positive during daytime, with negative values occurring during nighttime and in the late afternoon in some months. On the average, LE is positive throughout the 24-hour period, with the highest values occurring in daytime around noon, sometimes with a second peak in the afternoon.

CO 2 fluxes: daily and seasonal variation
In Figures 7, 8 and 9 we show three 4-day periods of measurements in detail (hereafter named Sample Period 1, 2 and 3, respectively) reasonably representative of the CO 2 patterns observed throughout the experiment. Sample Period 1 extends over March 13-17 (in Period III); during this period, we observe diurnal CO 2 uptake and nocturnal CO 2 emission, suggesting that photosynthesis and respiration are driving CO 2 concentrations in water. Sample Period 2 extends over July 27-31 (in Period V), and represents the somewhat unexpected pattern of nighttime CO 2 uptake and daytime emission. Sample Period 3, during August 12-16, belongs to Period VI, and intense winds generated fluxes of greater magnitude. In Figures 7, 8 and 9 we show the observed interplay of several environmental variables: surface water and air temperature at 3.76 m, atmospheric pressure, solar radiation and CO 2 concentration at 3.66 m, sensible (H) and latent (LE) heat fluxes, wind speed and direction, and CO 2 fluxes.
As already mentioned above, in Figure 7, negative CO 2 fluxes are driven by photosynthesis. During this period water temperature is greater than air temperature, with larger sensible heat fluxes during daytime. The solar radiation intensity of the Sample Period 1 is higher compared to the other periods selected for this section: Figure 7 shows the solar radiation reaching 800 W m −2 , while in Figures 8 and 9 the maximum solar radiation is 600 W m −2 . Note in Figure 7 the strong daily variation in CO 2 concentration in the atmosphere, which is probably a consequence not only of the CO 2 fluxes from the water surface, but also of the land CO 2 fluxes around the reservoir.
Episodes of daytime CO 2 emission and nighttime CO 2 uptake by the reservoir were common in July (Figure 8). It is noted that the fluxes during this period are smaller (in absolute value) than the fluxes during the other two periods (all of F c , H and LE).
During Sample Period 3, Figure 9, strong winds are observed, which generated larger fluxes. Note the drop in air temperature by approximately 14 • C between 13th and 15th Aug and the increase in atmospheric pressure during this period, showing the passage of a climate system.

9/14
In theory, under ideal conditions, CO 2 fluxes depend on the CO 2 concentrations in air and water. Unfortunately, there were no measurements of CO 2 in the water during the experiment to assess the CO 2 water-air concentration difference. However, continuous water pH measurements performed from July to November 2012 at the buoy showed in Figure 1, indicate the existence of pH seasonality and a daily pH cycle.
In Figure 10 we selected two periods of 2012 with solar radiation patterns roughly analogous to those of Sample Periods 1 and 2: In the October 2012 14-20 period (analogous to Sample Period 1), the water pH displayed a marked daily cycle, ranging from 7.7 to 8.7, which corresponds to daytime CO 2 uptake and nighttime emission, with the pH increasing during the day and decreasing at night. In the July 30 -August 05 period (analogous to Sample Period 2), the water pH varied over a much smaller range (7.0-7.5) and there was no discernible daily cycle. This suggests that the CO 2 concentration in water was approximately constant during this period. Table 4 shows the corresponding CO 2 flux values as well as the values of the main flux drivers (solar radiation, wind speed and CO 2 air concentration) averaged over each measurement period. The CO 2 flux drivers clearly influenced the sign of the average CO 2 flux in periods II to IX, as shown in Table 4. In period V, for example, we observed the highest mean CO 2 concentration in the air, the second highest mean wind speed, and a negative CO 2 mean flux during both daytime and nighttime. In periods III, VII and VIII, when solar radiation was only lower than in periods II and IX, the mean daytime CO 2 fluxes were negative, probably due to photosynthesis. Despite the higher solar radiation of periods II and IX compared to other periods, the mean concentrations of CO 2 in the air were the lowest of all periods, favoring positive CO 2 fluxes.
The pattern of the observed CO 2 fluxes during 2013, in Figure 11, clearly shows their seasonality. The mean of all fluxes measured in daytime was −0.07 µg m −2 s −1 , and +25.62 µg m −2 s −1 in nighttime. The overall mean was +12.78 µg m −2 s −1 . The nighttime Table 4. Period averages of nighttime, daytime and 24h CO 2 fluxes, and of their main drivers (solar radiation, wind speed and CO 2 air concentration).   RBRH, Porto Alegre, v. 25, e43, 2020 10/14

Period
Eddy-covariance CO2 fluxes over Itaipu lake, southern Brazil fluxes of periods V and VI indicated CO 2 uptake in the reservoir and approximately zero fluxes, respectively. The daytime fluxes measured in periods IV, V and IX also showed approximately null average fluxes.

DISCUSSION
The observations of CO 2 flux as a function of wind direction are shown in Figure 12, where they are further classified according to daytime or nighttime. The good fetch conditions for most directions, already noted in Figure 3, are confirmed, with very few measurements coming from East, which is the direction closest to the margins. Also noteworthy is the specific pattern of negative (absorption) and positive (emission) values of the fluxes: for example, as already observed, there are many situations of negative fluxes during nighttime. This is very different from a land environment, where most nighttime fluxes are positive due to the absence of light for photosynthesis. This is strong evidence, also, that our measurements are not being affected by local advection from the margins, where both tall vegetation and agricultural fields can be found.
It is well known that the water pH responds to the CO 2 concentration in water: when CO 2 is absorbed in water, water pH decreases, and vice-versa (Potes et al., 2017). If the water pH in 2013 displayed the same pattern as observed 2012 (under similar radiation forcings) in the Itaipu reservoir, our hypothesis is that in the periods with higher incoming solar radiation (such as in Sample Period 1), photosynthesis removes CO 2 from the water during daytime and respiration replenishes it at night, resulting in the variation of pH shown in Figure 10a. Thus, the CO 2 water concentration contributed to the CO 2 concentration difference between water and air, producing larger (in absolute value) CO 2 fluxes. In periods of lower incoming solar radiation, such as in Sample Period 2, the water pH varies considerably less, and this suggests that the CO 2 water concentration varies less as well. This variation between the CO 2 concentration and pH in the water is a commonly reported pattern in the literature, as shown by Finlay et al. (2009) in freshwater bodies and Duarte et al. (2008) in saline lakes. Therefore, in Sample Period 2 apparently the CO 2 fluxes are mainly forced by the variation of CO 2 concentration in air. Observe in Figure 8 that CO 2 fluxes vary in opposite phase to the CO 2 concentration in the air, which justifies the negative nighttime CO 2 fluxes in Sample Period 2: the higher the CO 2 concentration in the air, the more negative the CO 2 fluxes are and vice-versa. Eugster et al. (2003) also found nighttime negative CO 2 fluxes, but they argued that these fluxes were not from the lake, because they were measured during extremely stable atmospheric conditions and CO 2 concentrations in the lake were greater than atmospheric concentrations. Differently from Eugster et al. (2003), during Sample Period 2, although most of the CO 2 negative fluxes were measured at low wind speeds, the measurements took place under unstable atmospheric conditions.
We also observed some periods with high wind speed intensifying the turbulent fluxes. According to Liu et al. (2016), synoptic events may increase the mixing of the water column by both convection and the mechanical mixing of water by wind. In fact, ultimately the scalar fluxes are intensified by wind speed, as shown for CO 2 fluxes by Macintyre et al. (2013) and for energy fluxes by Blanken et al. (2000). As discussed in McGillis et al. (2001), the gas transfer rate between water and air is higher under high wind conditions, probably due to the thinning of the diffusive layer of the water surface, as well as due to the increased turbulence in the water column. We must however emphasize that the foregoing discussion is qualitative, and that more research into the subject is needed, with simultaneous and continuous measurements of CO 2 flux and CO 2 concentration both in water and in the air.
With the exception of the winter periods (periods V and VI), the nighttime fluxes were all larger than the daytime fluxes (Table 4). Liu et al. (2016) also found higher nocturnal fluxes than daytime fluxes in a freshwater reservoir. They showed that CO 2 fluxes measured over a year with Eddy Covariance at the Ross Barnett Reservoir, located in Mississippi State -USA, were approximately 70% higher than those measured during the day.

11/14
The emission of CO 2 from continental waters is a consequence of CO 2 surface water saturation, generated by the biological respiration of organic carbon. Generally, the main source of organic carbon in reservoirs is allochthonous (Bernardo et al., 2017), but this depends on the age of the reservoir. As shown by Barros et al. (2011), newly implanted reservoirs emit more greenhouse gases due to the biodecomposition of organic matter from the flooded areas. Over time, the concentration of this organic matter decreases, driving down the emission of greenhouse gases as well. According to Teodoru et al. (2011), only in the first 15 years of the reservoir the main source of carbon is the flooded biomass. Therefore, since the Itaipu reservoir has been in existence since 1984, it is very likely that the greenhouse gases are generated mainly by allochthonous carbon.
During 2012, the Balcar (Brasil, 2014) project measured the diffusion fluxes of CO 2 from the Itaipu reservoir using the chamber method at 45 points distributed throughout the reservoir surface. These points were sampled four times in measurement campaigns that took place in January, May, August and October 2012. The CO 2 fluxes deemed representative of the reservoir ranged from approximately +3.47 µg m −2 s −1 to +16.20 µg m −2 s −1 (Brasil, 2014), which encompasses the average CO 2 fluxes obtained in this work: +12.78 µg m −2 s −1 . As already mentioned, an advantage of the eddy covariance method is that it allows continuous measurement of CO 2 fluxes. Thus, we identified that 90% of the measured CO 2 fluxes were in the range of −102.68 to +151.72 µg m −2 s −1 , which are comparable to those measured in natural lakes with eddy covariance. For example, Anderson et al. (1999) observed, for a natural lake in Minnesota, USA, a range of −7.24 a +78.19 µg m −2 s −1 ; the fluxes measured by Vesala et al. (2006) and Huotari et al. (2011) and Mammarella et al. (2015) in a natural lake of Finland were in the [+5.79, +11.58] µg m −2 s −1 , [−38.61, +144.8] µg m −2 s −1 and [−110, +220] µg m −2 s −1 ranges, respectively; the fluxes measured by Jonsson et al. (2008) in a natural lake in the north of Sweden varied from −0.77 a +1.54 µg m −2 s −1 ; and the fluxes measured by Eugster et al. (2003) in a natural lake in Alaska ranged from −12.16 and +36.2 µg m −2 s −1 .

CONCLUSIONS
In this work, we used the Eddy Covariance method to measure sensible heat fluxes, latent heat fluxes and carbon dioxide fluxes in the reservoir of the Itaipu Hydroelectric Power Plant during 2013.
Through a footprint analysis it was found that 84% of the fluxes measured at Itaipu came from the reservoir water surface. The other 16% of the fluxes were discarded as most of the fluxes came from the exposed island ground where the station was installed. With the footprint analysis we found that the reservoir margins do not interfere with the station flux measurements.
In general, the reservoir emitted more carbon dioxide into the atmosphere than it absorbed, and it was possible to observe three distinct patterns in CO 2 fluxes. The first was daytime CO 2 uptake and nighttime emission, which can be attributed to photosynthesis in the reservoir water and possibly by submerged vegetation. The second pattern was prevailing nighttime CO 2 uptake and daytime emission, which probably can be attributed to the daily cycle of CO 2 concentration in the air, accompanied by approximately constant water CO 2 concentration. The third pattern was wind speed-intensified fluxes. This pattern is clearly associated with the passage of weather systems by the region.
The CO 2 fluxes measured in this work varied within a range comparable to those measured in other lakes and reservoirs. Most of the fluxes were positive, with higher (in absolute values) values at night. The average of all CO 2 fluxes measured during daytime, nighttime and 24h were −0.07 µg m −2 s −1 , +25.62 µg m −2 s −1 and +12.78 µg m −2 s −1 , respectively.