SEASONALITY OF WATER AND HEAT FLUXES OVER A TROPICAL FOREST IN EASTERN AMAZONIA

. We used the eddy covariance technique from July 2000 to July 2001 to measure the fluxes of sensible heat, water vapor, and CO2 between an old-growth tropical forest in eastern Amazonia and the atmosphere. Precipitation varied seasonally, with a wet season from mid-December 2000 to July 2001 characterized by successive rainy days, wet soil, and, relative to the dry season, cooler temperatures, greater cloudiness, and reduced incoming solar and net radiation. Average evapotranspiration decreased from 3.96 + 0.65 mm/d during the dry season to 3.18 ? 0.76 mm/d during the wet season, in parallel with decreasing radiation and decreasing water vapor deficit. The average Bowen ratio was 0.17 1 0.10, indicating that most of the incoming radiation was used for evaporation. The Bowen ratio was relatively low during the early wet season (December-March), as a result of increased evaporative fraction and reduced sensible heat flux. The seasonal decline in Bowen ratio and increase in evaporative fraction coincided with an increase in ecosystem CO2 assimilation capacity, which we attribute to the growth of new leaves. The evaporative fraction did not decline as the dry season progressed, implying that the forest did not become drought stressed. The roots extracted water throughout the top 250 cm of soil, and water redistribution, possibly by hydraulic lift, partially recharged the shallow soil during dry season nights. The lack of drought stress during the dry season was likely a consequence of deep rooting, and possibly vertical water movement, which allowed the trees to maintain access to soil water year round.


INTRODUCTION
Tropical forests play a key role in the hydrology of Amazonia. About half of the moisture that falls in the Amazon basin originates from evapotranspiration off upwind forest, and the other half originates from evaporation off the tropical Atlantic Ocean (Nobre et al. 1991). The regional pattern of precipitation within Amazonia is linked to the large-scale pattern of landsurface evaporation. At the same time, the distribution of forest in the tropics is linked to rainfall, with closedcanopy evergreen forest occurring in areas with at least 2 m annual rainfall (Walter 1984). Amazonia is therefore best understood as a coupled system, where vegetation affects the patterns of precipitation (da Rocha et al. 1996a) and precipitation affects the distribution and activity of vegetation. One of the goals of the Large-scale Biosphere-Atmosphere Experiment in Amazonia (LBA) is an improved understanding of the components that determine this coupling, including the seasonal patterns of, and biological and physical controls on, forest evapotranspiration.
Early observations of forest evapotranspiration in Amazonia focused on the Manaus area, where Shuttleworth et al. (1984) and Fitzjarrald et al. (1988) used micrometeorological techniques during pioneering studies. Shuttleworth (1988) subsequently combined field measurements and model-based extrapolation to calculate the monthly variation of evaporation. Roberts et al. (1990) relied on methods from plant physiology to infer that primary forest is not very sensitive to seasonal soil moisture deficit. Nepstad et al. (1994) emphasized the role of deep roots in allowing evapotranspiration to continue during the dry season in large regions of Amazonia. Much of the early work on forest evaporation in Amazonia was done before year-round flux observation became feasible, and, consequently, was confined to relatively short intensive field campaigns. While the technical hurdles preventing yearround micrometeorological observation were largely resolved in the 1990s (Wofsy et al. 1993), recent work on forest-atmosphere exchange in the Amazon has focused on CO2 exchange, rather than energy exchange (Fan et al. 1990, Grace et al. 1995).
We established a micrometeorological field station in June 2000 to make continuous observations of the

Measurements
The flux measurements were made from a 67 m tall tower (Rohn 55G, Rohn, Peoria, Illinois, USA). The data acquisition computer and some of the instruments were operated in an air-conditioned hut 8 m south of the tower base. Power was provided by a diesel generator located 800 m south of the tower. The measurements discussed here were made from 1 July 2000 to 5 July 2001. Valid meteorological data were obtained during 91% of this period and valid surface fluxes were obtained during 85% of this period. One large data gap occurred in April 2001 when the sonic anemometer failed. The data acquisition system used five networked data loggers (CRlOx or CR23x; Campbell Scientific, Logan, Utah, USA), which delivered raw 4or 0.5-Hz data to a computer on site. The wind and temperature at 64 m altitude were measured at 4 Hz with a three-axis sonic anemometer (Campbell Scientific). The CO2 and H20 densities at 64 m were measured with two independent infrared gas analyzers (an open-path Li-Cor 7500 and a closed-path Li-Cor 6262 or 7000; Li-Cor, Lincoln, Nebraska, USA). The current analysis relied most heavily on the fluxes calculated with the open path gas analyzer, which were judged more reliable for the humidity flux based on spectral analyses. The turbulent fluxes were calculated as the 30-min covariances of vertical wind velocity and either temperature or H20.

The fluxes were rotated to the plane with no mean vertical wind (McMillen 1988).
Observations of the physical environment were recorded at 0.5 Hz, including precipitation at 64 m (TE525 rain gauge; Texas Electronics, Dallas, Texas, USA), incoming short-wave radiation at 64 m (CM6B pyranometer; Kipp & Zonen, Delft, The Netherlands), net radiation at 64 m (Q*7.1 ventilated net radiometer; REBS, Seattle, Washington, USA), soil moisture at 20 locations from 5-to 250-cm depth (Campbell Scientific CS615 water content reflectometers), and soil heat flux at 2 cm depth in five locations (REBS HFT3.1 heat flux plates).
Calibration curves for the soil moisture probes were developed using large blocks of soil collected at 5 cm and 25 cm depth. The calibrations were carried out by automated gravimetry in our laboratory at the University of Sdo Paulo. Probes were inserted into intact blocks of soil that were shipped from the site. The soil was initially saturated and then allowed to dry gradually for several days with gentle indoor ventilation. The wave-guide transit time (sensor output) and the mass of the soil block were recorded every 20 min until no significant variations were observed. The sample was then dried in an oven at 1080C to correct for the remaining water content. The calibration curve was 0 (m3/m3) = 2.745 -9.429t + 10.212t2 -3.297t3 (R2 = 0.97) for soil collected at 5 cm, and 0 (m3/m3) = 1.759 -6.830t + 8.114t2 -2.790t3 (R2 = 0.98) for soil collected at 25 cm, where t is the transit time in ms.

Data analysis
Energy budget closure provides a measure of the absolute accuracy of the measured fluxes of sensible heat flux (H) and latent energy (LE). If all measurements are accurate, the energy budget should close such that (Rn -G -Sbc -C -A) -(H + LE) where Rn is the measured net radiation, G is the measured soil heat flux, Sbc is the energy stored in biomass and canopy air beneath 64 m, C is the chemical energy for net CO2 exchange, and A is the energy advected by horizontal wind. Rn, G, H, and LE were measured directly. Sbc was calculated using the empirical relationships reported by Moore and Fisch (1986) for tropical forest in Manaus: Sc = ST + Sq = 16.7AT, + 28.0Aq, where S, and Sq are the sensible heat and latent heat storage in the canopy air column (W/m2), Sb is the energy stored in the biomass (W/m2), ATr is the hourly change in air temperature (K), Aq, is the hourly change in specific humidity (g water/kg air), and AT,. is the hourly change lagged one hour. Advection (A) is very S24 HUMBERTO R. DA ROCHA ET AL. Ecological Applications Special Issue difficult to measure, and is hypothesized to be insignificant over homogeneous, flat terrain. Chemical storage (C) is generally small, and was not considered. The Penman-Monteith equation describes the dependence of evaporation on the meteorological (air temperature and humidity, available energy, and wind speed) and biological (canopy stomatal conductance) conditions. In studies where the evaporation and meteorological conditions are independently measured, the Penman-Monteith equation can be inverted to calculate the canopy conductance to water vapor (gc), which is analogous to stomatal conductance on a ground-area basis. The canopy conductance, and the aerodynamic conductance to water vapor (gv), which describes the turbulent transport from 64 m to the outside of the leaf boundary layers, were calculated as S= re = raA(Rn -G)+ pc,8eg-(A + y) where Rn is the net radiation, G is the soil heat flux, LE is the flux of latent energy, p is the air density, c, is specific heat at constant pressure, 8e is the water vapor pressure deficit, A is the rate of change of saturated water vapor pressure with temperature, y = (pc, / eL), L is the latent heat of vaporization, p is the atmospheric pressure, and e = 0.622. The variables Tv and Tm in the right-hand side of Eq. 5 account for the effects of atmospheric stability and were calculated following Verma (1989). The ratio ln(zo/zv) was set to 2 following Garratt (1992) and Grace et al. (1995); zo is the aerodynamic roughness length (a measure of how rough the top of the canopy is and its effectiveness in retarding the flow of the overlying air) and zv is the roughness length for water vapor.

Climate and radiation
The local climate was hot and humid, with reduced rainfall from August to December (Fig. la). The total precipitation from 1 July 2000 to 1 July 2001 was 2200 mm, though this sum may underestimate the true precipitation as a result of observational gaps. About onethird of the annual precipitation fell from 15 July to 14 December, with the remainder falling from late December to July. In subsequent analyses we refer to the period 15 July to 14 December as the dry season and to the remaining period as the wet season. There was a tendency for increased precipitation from 1300 and 1600 hours local time, presumably as a result of increased convection (Fig. 2b). Intense rainfall occurred occasionally during the dry season. The wet season was marked by many consecutive days with moderate rainfall (Figs. la and 2c).  lb). Seasonal change in cloud cover was the main controller of incident radiation, with solar angle playing a secondary role. The transition from dry to wet season coincided with a marked reduction in incoming radiation.
The surface air temperature varied little over the year (Fig. ic), a pattern that is typical for. the region (Culf et al. 1996). The maximum daily temperature was 24-320C and the minimum was 20-250C (Figs. Ic and 2a). The dry season was 1-30C warmer on average than the wet season. A few periods with cooler than average temperatures were observed during the dry season, re-flecting the incursion of cold fronts from the south. The daily maximum and minimum temperatures and the daily temperature range were reduced during the three months centered on January (Fig. Ic), coincident with increased cloudiness. The water vapor content in the air was consistently 17-19 g water/kg air (Fig. Id). The daily average water vapor deficit decreased from 700 Pa in the dry season to 200 Pa in the wet season (Fig.  Id), coincident with the decline in daytime air temperature.

Soil moisture
Shallow (5 cm beneath the surface) soil moisture ranged from 0.47 m3 water/m3 soil in the wet season to 0.3 m3 water/m3 soil in the dry season (Fig. le). Deep (250 cm beneath the surface) soil moisture was comparatively constant year round, ranging from 0.46 m3 water/m3 soil in the wet season to 0.42 m3 water/ m3 soil late in the dry season. The shallow soil moisture varied markedly during the dry season in response to storms, while the deep soil moisture declined gradually during the dry season.
The soil moisture content late in the dry season increased with depth from 0.356 m3 water/m3 soil at 5 cm to 0.414 m3 water/m3 soil at 200 cm (Fig. 3) patterns of moisture withdrawal during extended periods without rain (Fig. 3). The daily withdrawal of water in early December declined with depth from 0.001 to 0.002 m3 water-m-3 soild-' in the upper 160 cm of soil to 0.0005 m3 water.m-3 soil-d-1 at 200 cm depth (Fig. 3). The pattern of water withdrawal with depth presumably reflects the distribution of root activity, with fewer active roots beneath 1.6 m depth. The daily total withdrawal of water in the upper 2 m of soil was 2.5 to 4.5 mm/d, which is in excellent agreement with the daily evaporation measured simultaneously by eddy covariance (Fig. 4a). The soil moisture content at 2 m decreased moderately during daytime and remained relatively constant at night. This diel pattern implies significant extraction of water by roots at 2 m depth (Fig. 3). The day-to-day extraction of deep water remained constant as the dry season progressed, and the soil at 2 m remained moist year round, implying that the trees maintained access to soil water throughout the dry season (Figs. le, 3, 4a).
The shallow soil water content during rain-free periods decreased during daytime and recovered partially during nighttime (Fig. 3), a pattern that we attribute to the nocturnal redistribution of water. Lopes (2001) reported a similar nocturnal recovery for both tropical forest and grassland growing on silt-clay podzols in southern Amazonia. The nocturnal redistribution of soil water can occur either by flow through plant roots, a process referred to as hydraulic lift (Caldwell et al. 1998), or flow through the bulk soil. Hydraulic lift occurs when root systems provide a low-resistance bridge between shallow dry soil and deep moist soil. Water flows through roots from moist, deep soil to shallow dry soil at night, resulting in an increase in shallow-soil water content. Hydraulic lift has been reported in semiarid shrublands, conifer and temperate forests, and several types of cropland. Nocturnal soil moisture redistribution can also occur by capillary flow through the bulk soil in response to strong vertical gradients in matric potential.
The nocturnal recharge integrated throughout the top 60 cm of soil was 0.3 mm/d (Fig. 4b). This rate of recharge is equivalent to 10% of the daily evapotranspiration (Fig. 4a). The nocturnal recharge may be important for forest function. The redistribution of water may help the trees avoid drought stress by increasing the efficiency with which deep roots extract water. Likewise, redistribution may increase microbial activity ini the shallow soil by improving the local moisture status.

Energy balance
Energy budget closure provides a measure of the absolute accuracy of the measured fluxes of sensible heat flux (H) and latent energy (LE) (Fig. 5). Energy storage associated with changes in the temperatures of trunks and stems and in the temperature and humidity of air within the canopy can be substantial in tall forest. While this storage approaches zero when integrated over 24 h, the daytime magnitude can significantly affect the amount of available energy (the net radiation after subtracting the energy stored). From 2 to 6 W/m2 of energy was stored in the biomass during daytime, and a comparable amount of energy was released from the biomass at night (Fig. 6). Similarly, 0-3 W/m2 was stored in the air column during daytime and lost at night. The seasonality of the canopy energy storage followed the variation of the daily air temperature

FIG. 6. (a) Biomass energy storage averaged over 24 h (line connecting filled squares for 5-d means) and averaged over the 12 daytime hours (line connecting open squares). (b) Canopy energy storage averaged over 24 h (line connecting filled squares) and averaged over the 12 daytime hours (line connecting open squares). The dry season is shaded. range, with a decrease in the wet season and an increase in the dry season (Figs. Ic, 6b).
The energy budget indicated that turbulent exchange underestimated available energy by 13% (Fig. 5). Our failure to completely close the energy budget is consistent with observations from many other sites. The discrepancy is unlikely to result from uncertainty in the soil heat flux, since the soil heat flux was only 2% of the daytime net radiation. Moreover, the fraction GIRn observed is typical for closed-canopy forests (Alvala et al. 1996). The analysis of the energy budget suggests that the turbulent fluxes reported here may underestimate the true fluxes by about 13%, or possibly more depending on the accuracy of the net radiation measurement. This possibility should be kept in mind when interpreting the results, including the annual integrated evaporation.

Seasonal patterns of water and heat fluxes
Ecosystem evaporation represents the sum of evaporation from soil surfaces, leaves during gas exchange, and plant surfaces following precipitation. The evaporation ranged from 1.5 to 6 mm/d, with an annual average of 3.45 + 0.81 mm/d. The annually integrated evaporation was -1300 mm, or 60% of the observed precipitation. Evaporation and sensible heat flux increased in the dry season and decreased in the wet season, coincident with changes in cloudiness and net radiation (Fig. 7). The observed evaporation is comparable to the rate of 3.5 mm/d reported by Shuttleworth (1988), which was based on a combination of observations and model runs for a forest near Manaus.
The observed evaporation is also comparable to the annual rate of 3.7-4.0 mm/d reported by da Rocha et al. (1996b), which was based on model runs for three forest sites in Amazonia (Grace et al. 1995, Gash et al. 1996, Sellers et al. 1996. The observed increase in evaporation in the dry season is similar to that predicted by Shuttleworth (1988) and da Rocha et al. (1996b).
Daily integrated net radiation was the main controller of day-to-day variation in latent and sensible heat flux, with high turbulent fluxes occurring on sunny days. The 24-h average net radiation was 70-180 W/m2, with an average of 140 W/m2 in the dry season and 113 W/m2 in the wet season ( Fig. 7a and Table 1). The inter-and intraseasonal patterns of latent and sensible heat flux were similar to the trends in radiation (Fig.  7a-c). Evaporation was relatively high (-4 mm/d; Table 1) and constant from day to day in the dry season (cv = 16%), and low (-3.2 mm/d) and variable in the wet season (cv = 25%). Likewise, net radiation was relatively constant from day to day in the dry season and variable in the wet season. The seasonal pattern of sensible heat flux was broadly similar to the seasonal pattern of radiation, with a minimum centered in January. The 24-h average heat flux was 21 W/m2 in the dry season and 16 W/m2 in the wet season. The variance in H was slightly larger in the wet season than the dry season, a pattern that we attribute to the variance in radiation (Table 1).
While daily integrated radiation was the main controller of daily evaporation, it did not explain all of the day-to-day variance. The evaporative fraction, calcu-  lated as the percentage ratio of latent heat flux to net radiation, varied seasonally, reaching a minimum of 65-75% from May to October and a maximum of 75-100% from December to March (Fig. 8a). Likewise, the Bowen ratio (sensible heat flux divided by latent heat flux, a measure of energy partitioning) varied seasonally, with a minimum coinciding with the maximum evaporative fraction (Fig. 8b). The increase in evaporative fraction from December to March cannot be explained based on a seasonal change in potential evaporation. The water vapor pressure deficit was relatively low during much of this period (Fig. 8c), as a result of cooler air temperatures (Fig. Ic). The direct effect of a decrease in water vapor pressure deficit is a decrease in evaporative fraction and an increase in Bowen ratio, which are opposite the trends observed.

Bowen ratio Season
The seasonal shift in energy partitioning may be partially related to the frequency of precipitation (see Fig.  2c). Frequent, moderate-to light-intensity storms increase the overall fraction of precipitation retained on plant surfaces. This intercepted precipitation subsequently evaporates rapidly, since there is no stomatal limitation, resulting in higher overall rates of evaporation. In contrast, heavy, infrequent storms increase the fraction of precipitation that infiltrates into soil. The seasonal shift in rainfall frequency (Figs. la, 2c) may partially explain the decrease in Bowen ratio and increase in evaporative fraction observed early in the rainy season (Fig. 8a, b). Rain fell very frequently during this period (Figs. la, 2c), a pattern that should result in higher rates of evaporation by increasing the fraction of interception loss.
The seasonal pattern of energy exchange was probably also controlled in part by changes in tree physiology. The decline in evaporative fraction in May preceded the onset of the dry season, and the increase in evaporative fraction in November preceded the end of the dry season (Figs. le, 8a), indicating that the seasonal pattern of evaporative fraction cannot be explained entirely by changes in meteorology. The canopy surface conductance (gc), which is analogous to stomatal conductance on a ground-area basis, provides a useful measure of the effect of physiology on energy exchange. The canopy surface conductance was com-paratively large from December until early April (Fig.  9), suggesting that the seasonal changes in evaporative fraction and Bowen ratio were caused in part by changes in leaf area index (LAI) or leaf physiology. The seasonal patterns of evaporative fraction, Bowen ratio, and canopy conductance are broadly similar to the seasonal pattern of daytime CO2 uptake described for the site by Goulden et al. (2004). Goulden et al. found that CO2 uptake at a given light intensity was greater from October to April than from May to September, a pattern they attributed to a seasonal increase in LAI. It is likely that these observations are related, and that the October-to-April increases in canopy photosynthesis, canopy conductance, and evaporative fraction, and decrease in Bowen ratio, are mechanistically linked through seasonal changes in LAI.

Diel patterns of water and heat fluxes
The diel patterns of latent and sensible heat flux (Fig.  10a, b) were closely tied to the intensity of sunlight. The sensible heat flux reached a maximum before noon local time, and was typically negative at night. The latent heat flux reached a maximum shortly after noon local time, and approached zero at night. The sensible heat flux peaked at <300 W/m2, with a dry season 24h average of 21 W/m2 and a wet season 24-h average of 18 W/m2 (Fig. 10b, Table 1). Latent heat flux peaked at 800 W/m2, with a wet season 24-h average of 92 W/m2 and a dry season 24-h average of 115 W/m2 (Fig.  10a, Table 1). The integrated daily Bowen ratio was 0.17 over the year, indicating that most of the incoming solar energy was used to evaporate water (Fig. 10c, Table 1).
The aerodynamic conductance (gav) varied diurnally, reaching a minimum at night and a maximum shortly before noon (Fig. 1 ib). The peak is strongly correlated with a similar peak in the horizontal wind speed at 64 m at the same hour (not shown). The aerodynamic conductance increased rapidly in the morning, when thermal instability caused by increasing solar energy eroded the previous night's stable boundary layer and caused rapid mixing above the forest. The diel pattern of gav varied seasonally, with a faster increase in the first hours of daytime in the dry season than the wet season. The increased morning gav during the dry season may result from an increase in daytime radiation, which should generate greater kinetic energy in the boundary layer. Alternatively, the seasonal trend in gav may be a result of a regional trend toward increased kinetic energy during the austral winter and spring (Peixoto and Oort 1994). The canopy conductance gc generally peaked before noon and declined in the afternoon (Fig. 1 la), a pattern similar to that observed at other Amazonian sites (Shuttleworth 1988, Grace et al. 1995). This diel pattern was broadly similar to that of gay, reaching a minimum at night and a maximum shortly before noon. The peak g, was -24 mm/s on average, which is a factor of --2.5 less than the peak gay, indicating substantial stomatal limitation to evaporation. The surface conductance de- clined in the afternoon, presumably as a result of stomatal closure (Fig. 11 c). Goulden et al. (2004) analyzed data from the same site and found that canopy CO2 uptake reached a maximum at around 1000 local time, and declined steadily in the afternoon. Goulden et al. attributed the afternoon decline to either stomatal closure caused by high evaporative demand, or an effect of high temperature on photosynthetic biochemistry, or an intrinsic circadian rhythm, or a combination of the three. The afternoon decline in g& in Fig. 11 corresponds well with the decline in CO2 uptake reported by Goulden et al. (2004). These two phenomenon are likely related, though we are unable to identify the specific causal mechanism.
The Bowen ratio increased shortly after sunrise and reached a maximum in the morning, before declining in the afternoon (Fig. 10c). The diurnal change in Bowen ratio resulted from a slight asynchrony between the energy fluxes, with the sensible heat flux peaking earlier in the day than the latent heat flux (Fig. 10a, b). The diel trend in Bowen ratio resulted from the diel patterns of air temperature (Fig. 2a) and vapor pressure deficit (Fig. 1lc). The observed afternoon decline of g, (Fig. 1 la)  ratio, provided that the vapor pressure deficit between the atmosphere and the leaf intercellular spaces remained constant. However, the afternoon decline in Bowen ratio (Fig. 10c) indicates that the afternoon increase in vapor pressure deficit (Fig. 1 1c) caused by increasing temperature (Fig. 2a) overwhelmed the effect of decreasing in g, (Fig. 1 la).