What drives the seasonality of photosynthesis across the Amazon basin? A cross-site analysis of eddy flux tower measurements from the Brasil flux network

We investigated the seasonal patterns of Amazonian forest photosynthetic activity, and the effects thereon of variations in climate and land-use, by integrating data from a network of ground-based eddy ﬂux towers in Brazil established as part of the ‘Large-Scale Biosphere Atmosphere Experiment in Ama- zonia’ project. We found that degree of water limitation, as indicated by the seasonality of the ratio of sensible to latent heat ﬂux (Bowen ratio) predicts seasonal patterns of photosynthesis. In equatorial Amazonian forests (5 ◦ N–5 ◦ S), water limitation is absent, and photosynthetic ﬂuxes (or gross ecosystem productivity, GEP ) exhibit high or increasing levels of photosynthetic activity as the dry season progresses, likely a consequence of allocation to growth of new leaves. In contrast, forests along the southern ﬂank of the Amazon, pastures converted from forest, and mixed forest-grass savanna, exhibit dry-season declines in GEP , consistent with increasing degrees of water limitation. Although previous work showed tropical ecosystem evapotranspiration ( ET ) is driven by incoming radiation, GEP obser- vations reported here surprisingly show no or negative relationships with photosynthetically active radiation ( PAR ). Instead, GEP ﬂuxes non-water limited equatorial forest sites peak in the dry season, in correlation with high dry season light levels. The higher photosynthetic capacity that follows persists into the wet season, driving high GEP that is out of phase with sunlight, explaining the negative observed relationship with sunlight. Overall, these patterns suggest that at sites where water is not limiting, light interacts with adaptive mechanisms to determine photosynthetic capacity indirectly through leaf ﬂush and litterfall seasonality. These mecha- nisms are poorly represented in ecosystem models, and represent an important challenge to efforts to predict tropical forest responses to climatic variations.


a b s t r a c t
We investigated the seasonal patterns of Amazonian forest photosynthetic activity, and the effects thereon of variations in climate and land-use, by integrating data from a network of ground-based eddy flux towers in Brazil established as part of the 'Large-Scale Biosphere Atmosphere Experiment in Amazonia' project. We found that degree of water limitation, as indicated by the seasonality of the ratio of sensible to latent heat flux (Bowen ratio) predicts seasonal patterns of photosynthesis. In equatorial Amazonian forests (5 • N-5 • S), water limitation is absent, and photosynthetic fluxes (or gross ecosystem productivity, GEP) exhibit high or increasing levels of photosynthetic activity as the dry season progresses, likely a consequence of allocation to growth of new leaves. In contrast, forests along the southern flank of the Amazon, pastures converted from forest, and mixed forest-grass savanna, exhibit dry-season declines in GEP, consistent with increasing degrees of water limitation. Although previous work showed tropical ecosystem evapotranspiration (ET) is driven by incoming radiation, GEP observations reported here surprisingly show no or negative relationships with photosynthetically active radiation (PAR). Instead, GEP fluxes largely followed the phenology of canopy photosynthetic capacity (Pc), with only deviations from this primary pattern driven by variations in PAR. Estimates of leaf flush at three

Introduction
The Amazon basin represents a major component of regional and global carbon and hydrological cycles. Globally significant variations in these cycles could be induced by climate change (Betts et al., 2004;Monteith, 1965), but predictions for the future of Amazon forests under climate change vary widely (Friedlingstein et al., 2006). Predictions of vegetation responses to long-term climate change are not easily tested, but common mechanisms also control response to short-term climatic variations, including seasonal variations. Thus, understanding seasonal and spatial variation of forest metabolism is an important basis for understanding ecological responses to climate generally. In contrast to temperate zones, where the seasonality of ecosystem metabolism is evident and straightforwardly dominated by the contrast between an active growing and a dormant season, the seasonality of photosynthetic activity of evergreen tropical forests is not so obvious.
The dominant seasonal rhythm in many Amazonian forests is that of rainfall, with a distinct dry season, when precipitation inputs fall below 100 mm month −1 , varying in length from 0 months (no dry season in the ever-wet northwestern Amazon) to 5 months in central eastern equatorial Amazonia, and in southwestern edge (Sombroek, 2001) (Fig. 1). The dry season brings a decrease in cumulus cloud cover and, in areas close to the equator with low variation in top-of-atmosphere solar energy, an increase in solar radiation at the surface. The dry season also brings increased smoke from biomass burning (Koren et al., 2004;Oliveira et al., 2007), shifting the composition of direct and diffuse radiation.
Many vegetation modeling studies have represented Amazon forests as water-limited, and so predicted dry season declines in transpiration and/or photosynthesis (Botta et al., 2002;Dickinson and Henderson-Sellers, 1988;Lee et al., 2005;Nobre et al., 1991;Tian et al., 1998). Early empirical studies near Manaus both supported (Malhi et al., , 1998 and opposed (Shuttleworth, 1988) the water-limitation view. However, an accumulating suite of evidence is beginning to clarify the picture. Eddy flux measurements begun as part of the Large-scale Biosphere-Atmosphere Experiment in Amazonia (LBA) clearly show that evapotranspiration at some central Amazon sites is not water limited but largely driven by the availability of net radiation Hutyra et al., 2007). Recently, integrated analysis across a network of towers confirms this behavior across much of the Amazon (Hasler and Avissar, 2007;Negrón Juárez et al., 2009;da Rocha et al., 2009;Fisher et al., 2009;Costa et al., 2010a,b). This evidence suggests that intact Amazon forests could be more resilient to variations in climate than most models have predicted. Deep tree roots (Nepstad et al., 1994) combined with hydraulic redistribution by roots  and increased availability of sunlight (Potter et al., 1998) have been suggested as explanations for these observations.
Remote sensing data from the moderate resolution imaging spectroradiometer (MODIS) suggest that intact forest photosynthesis increases as the dry season progresses across a large area of the central Amazon. The MODIS Leaf Area Index (LAI) product shows a widespread pattern of dry-season LAI increases that peak during the early to mid-dry season (Myneni et al., 2007), while the MODIS Enhanced Vegetation Index (EVI) shows equatorial Amazon forests continually "greening-up" throughout the dry season as availability of sunlight increases. By contrast, areas with extensive conversion of forests to pasture or agricultural management show the opposite "brown-down" trend (Huete et al., 2006). Remote sensing of regional-scale chlorophyll fluorescence (Fs) in the Amazon (by the new Japanese Greenhouse gases Observing SATellite, GOSAT, which may give a more direct index of photosynthesis than existing technologies (Frankenberg et al., 2011), suggests patterns broadly similar to those observed via EVI (Fs vs EVI, R 2 = 0.52) (Lee et al., 2013).
Tower-based estimates of photosynthetic flux derived from observations of net ecosystem exchange of carbon dioxide via eddy covariance method are perhaps the most reliable and direct measure of ecosystem scale photosynthesis to "ground truth" remote sensing indices and to test models. Individual tower studies have shown a range of results, including: (1) clear increases in dry season forest photosynthesis in the central eastern Amazon Hutyra et al., 2007;Saleska et al., 2003), (2) dry season decreases in the southwest Amazon , and (3) no detectable seasonality in the far eastern Amazon . These studies have pointed to a range of possible biophysical drivers for carbon assimilation, including length of dry season, cloud cover, access to deep water, and disturbance history, among others. Site-specific results suggest the need for a consistent, integrated analysis across sites and biomes of the seasonal patterns in eddy flux-derived ecosystem-scale photosynthesis to present a coherent picture of its spatial variability and possible drivers across the Amazon basin.
Our objective is to use eddy covariance (EC) data from a network of flux towers, installed as part of the 'Large-Scale Biosphere Atmosphere Experiment in Amazonia' (LBA) project , to address the question of what environmental and biological factors control the seasonality of photosynthesis in these tropical Amazonian ecosystem. We also sought to provide an integrated, consistently processed dataset, which could be used for comparison to model simulations of the LBA "Data-Model Intercomparison Project (LBA-DMIP). To this end, we conducted the first integrated analysis of basin-wide patterns of ecosystem photosynthesis across the Amazon. This effort is possible by the collaboration from a range of research groups who have contributed their observations to an integrated and unified database, known as Brasil flux network . The nine sites (Table 1) from this database are: four tropical forests and one converted pasture/agricultural site located along the main stem of the Amazon (equator to 3 • S), one tropical dry forest and an adjacent converted site located in southern Amazonia (∼10 • S), one seasonally flooded ecotone (∼10 • S), and one non-Amazonian savanna in southern Brazil.
First, we describe the seasonal patterns of ecosystem photosynthetic fluxes, and canopy photosynthetic capacity across different Amazonian ecosystems, latitudinal gradients, and disturbance history. Then, we test hypotheses about control of photosynthesis by analyzing relationships between environmental drivers and overall photosynthetic fluxes and capacity. We initially focus on two main contrasting ideas drawn from the literature: that photosynthesis is limited by water availability, and that photosynthesis is limited by  Goulden et al., 2004;Miller et al., 2004). CAX: Sa, L. (Museo Goeldi), da Rocha, H. (USP, Brazil) Souza Filho et al., 2005). K77: Fitzjarrald, D. (SUNY, Albany, USA) (Sakai et al., 2004). RJA: Manzi, A. (INPA, Brasil), Cardoso, F. (UFR, Brazil) von Randow et al., 2004 Rocha et al., 2002). a Andreae et al. (2002). b Nepstad et al. (2002). c Zanchi et al. (2009). d Fidelis and Godoy (2003). e Domingues et al. (2005). f Doughty and Goulden (2008 Oliveira et al. (2005). l Malhi et al. (2009). available energy. Finally, after finding that photosynthetic seasonality is not well-explained by environmental variables alone, we advance the hypothesis that it arises instead from the interaction between biologically determined leaf phenology and environmental drivers. We explore this third hypothesis with a simple model of canopy phenology that assumes photosynthetic capacity arises from the balance between loss of capacity to litterfall and increase of capacity from new leaf-flush.

Site descriptions
We compiled original flux data from nine LBA eddy covariance towers in the Brazil flux network ( Fig. 1 and Table 1). From east to west and north to south, these sites are: the Reserva Cuieiras near Manaus (K34 forest), the Tapajós National forest, near Santarém (K67 and K83 forests, and K77 pasture/agriculture), the Caxiuanã National forest near Belém (CAX forest), the Reserva Jarú (RJA forest) and Fazenda Nossa Senhora (FNS pasture), near Ji-Paraná-Rondônia, the Tocantins-Javaes site (JAV seasonally flooded ecotone) on Bananal Island, and the Reserva Pe-de-Gigante in Sao Paulo state (PDG savanna).
The equatorial forest sites are located near the main stem of the Amazon River, close to the equator (∼3 • S). The Manaus K34 forest tower is the westernmost of the equatorial Amazonian sites, located 60 km north of the city of Manaus (Araújo et al., 2002a). Santarém tropical moist forest towers (K67 and K83) are near the confluence of the Tapajós and Amazon rivers, midway between Manaus and the Atlantic coast. Santarém K77 was a forest area converted to pasture (∼1990) and later (November 2001) was burned and plowed, followed by planting of non-irrigated rice (February 2001), and other crops Sakai et al., 2004;Wick et al., 2005). Caxiuanã National Forest, (CAX) is a dense lowland forest located close to the Baia de Caxiuanã and the city of Belém (Pará) North-Eastern Brazil .
Southern Amazon sites include Reserva Jarú (RJA), a tropical wet and dry (Peel et al., 2007) forest located in the state of Rondônia (Brazil) latitude ∼ 10 • S, 100 km north of Ji-Paraná. Fazenda Nossa Senhora da Aparecida (FNS) is a cleared forest, pasture close to the RJA forest tower, located on the Igarapé Lourdes Indian Reservation, which is about 30 km northwest of the city of Ji-Paraná (Gu et al., 2001;von Randow et al., 2004). The Bananal Island site (JAV) is a seasonally flooded (forest-savanna) ecotone, located on the state of Tocantins close to the Javaes River . Finally, south of the Amazon basin proper, is Reserva Pe-de-Gigante, in Sao Paulo state (PDG), also known as Sao Paulo-cerrado, a tropical woodland savanna site (da Rocha et al., 2002).
An analysis by Rosolem et al. (2008) showed that climatic and weather patterns at each site during the periods of flux observation (Supplement Table 1 and Fig. 2), were generally representative of the long-term climatology of each site, in particular: (1) precipitation was statistically consistent with the climatology and (2) temperature showed a slight tendency (not statistically significant) toward higher values. Only temperature at the seasonally inundated site (JAV) was ∼1 • C higher than climatology.
Mean monthly Amazon basin precipitation ( Fig. 1) is 182 mm month −1 with a range from 75 to 304 mm month −1 , based on the Tropical Rainfall Measuring Mission (TRMM) data product from 1998-2010 (3B43-v7 derived by combining TRMM satellite data, GOES-PI satellite data, and a global network of gauge data) (NASA, 2010). The spatial and temporal patterns of Amazon rainfall are associated with the behavior of the Intertropical Convergence Zone (ITCZ) centered over the Basin December to February (Horel et al., 1989). At the southern Amazon sites, the start of the dry season falls in the range of March to May as the ITCZ moves toward the NW (Central America) where it remains ∼June to September.
The equatorial Amazon (5 • S-5 • N) seasonal dry period induces a peak in maximum incoming solar radiation. A defined dry season (<100 mm month −1 ) is absent in the NW corner of the basin, where we observe the highest annual precipitation values (mean monthly precipitation > 250 mm) (Fig. 1). Toward the south (>5 • S), the onset of the dry season (when cloud cover is low) is in May and extends until the end of Aug, out of phase with the period of highest top of the atmosphere radiation (Jul to Sep, depending on latitude). At these southern sites, the onset of the wet season thus coincides with the increase of top-of-atmosphere incoming radiation (TOA).

Net and gross ecosystem exchange (NEE and GEE) of CO 2 fluxes
Sites in the Brasil flux network used the eddy covariance (EC) method to measure the turbulent fluxes of CO 2 (F co 2 in mol CO 2 m −2 s −1 ), sensible (H in W m −2 ) and latent heat (LE in W m −2 ), and momentum ( in kg m −1 s −2 ). For a detailed site description, instrumentation type and basic methods for processing the raw eddy-flux high frequency data and aggregation to hourly intervals, refer to the site-specific references and information listed in Table 1 and Supplement Table 2, and previous work by Keller et al. (2004), da Rocha et al. (2009) and Saleska et al. (2009).
Starting with half-hourly data provided from each site's operators, we calculated net ecosystem exchange (NEE in mol CO 2 m −2 s −1 , with fluxes to the atmosphere defined as positive) as the sum of turbulent eddy flux above the forest (F co 2 ) and the time rate of change in CO 2 stored within the canopy (the "storage flux", S co 2 ) (Wofsy et al., 1993).
Some forest sites had instrument malfunctions that prevented the acquisition of storage flux (S co 2 ) measurements for long periods, and we corrected for missing storage in these cases by using one of three methods for estimating the storage flux, as described in Restrepo-Coupe et al. (in preparation), and summarized here in Table 1. It is well known that when significant CO 2 can accumulate in the canopy airspace (as in forests) the absence of storage fluxes prevents accurate characterization of diel patterns of CO 2 exchange (Wofsy et al., 1993;Iwata et al., 2005). Similarly, S co 2 is in some cases important when determining seasonal patterns of NEE and its components. This is because the magnitude of S co 2 is inversely correlated with the strength of turbulent mixing (e.g. as measured by friction velocity u*), and turbulence mixing varies seasonally. We believe that our correction for missing storage data is adequate for the purposes of comparing seasonal patterns across sites. However, several of the forest sites in this analysis suffered from long gaps in storage data, especially at Caxiuanã, and our findings for this site are therefore less robust.
Next, we quality-checked and aggregated NEE and environmental driver data to an hourly time-step. For quality control and assurance, we used the energy balance equation to help us identify outliers in CO 2 , H 2 O and sensible heat fluxes (F c , LE, and H, respectively). We treated data as unreliable and removed it from the dataset when energy flux residuals were greater than 3-standard deviations away from the linear regression (Rn = LE + H, where Rn is the net radiation in W m −2 ). Similarly, we filtered data for consistency between sonic anemometer and infrared gas analyzer measurements, on the one hand, and meteorological data (air temperature and water vapor concentration from HMP45 sensors, and wind speed from cup anemometers).
We removed isolated observations that were preceded and followed by a missing value. Any residual spikes from the hourly NEE data were identified, following the method proposed by Papale et al. (2006) and modified by Barr et al. (2009).
Because NEE can be erroneously low when turbulence is low, and because nighttime turbulence often varies seasonally, failure to correct for low turbulence could introduce a seasonal bias. We therefore followed the standard practice (Aubinet, 2008) of correcting nighttime hourly NEE values for periods of low turbulent mixing by filtering out periods when friction velocity (u* in m s −1 ) was below a threshold value (u * thresh ), and filling those times with estimates averaged from nearby valid measurements (those taken during periods that had more vigorous mixing). We filled each nighttime missing value from an average of all valid nighttime values within a 5-day window. If there were less than 36 valid hourly values within the window, the window was expanded until 36 valid points were encompassed (to a maximum window size of 30 days).
Site-specific u * thresh values were calculated by dividing the dataset into three-month intervals and for each interval creating a look up table (LUT) of nighttime NEE as a function of seven air temperature (T a ) bins of equal u* (m s −1 ) sample size (see Restrepo-Coupe et al., in preparation). The selected u * thresh was chosen as the lowest u* bin for which corresponding NEE (averaged across temperature bins) was at least 99% of the maximum NEE bin average (Barr et al., 2009;Papale et al., 2006).
At the K77 pasture/agricultural site near Santarém, we used NEE values calculated from a nocturnal accumulation method (Sakai et al., 2004) without applying u* correction per se, as calm nights were the norm at this ecosystem (pastures have lower turbulence compared to forests) and previous analysis showed that even with u* filtering, nighttime NEE was underestimated (Sakai et al., 2004). There are no CO 2 storage measurements for the two non-forest sites, FNS (pasture) and PDG (savanna); here we assumed NEE is equivalent to F co 2 .
We derived gross ecosystem exchange (GEE) from tower measurements of daytime NEE by subtracting estimates of ecosystem respiration (R eco ), which in turn we derived from the u * thresh filtered nighttime NEE Hutyra et al., 2007;Saleska et al., 2003). We did not use a temperature function to extrapolate nighttime NEE to daytime R eco , as is done at some higher latitude sites (Goulden et al., 1996), because we did not observe any withinmonth correlation between nighttime hourly NEE and nighttime T a at any of the sites. Assuming daytime R eco is the same as nighttime R eco may underestimate the magnitude of daytime R eco (and hence, of GEE) if R eco is temperature sensitive, but we believe that any such underestimation will have little effect on the seasonal pattern, which is the focus here. Hourly R eco is calculated as the average of valid nighttime NEE within a centered 5-day wide window, assuming at least 36 valid hours of nighttime NEE were available. If not, we expanded the window from 5 days until sufficient valid data were included (up to 31 days).
As defined above, GEE is a negative value (GEE = NEE + R eco , with NEE negative in the daytime, and R eco positive); we express ecosystem-scale photosynthesis, or Gross Ecosystem Productivity (GEP), as a positive value by equating it to negative GEE. GEP is equal to negative GEE assuming negligible leaf re-assimilation of dark respiration and insignificant CO 2 recirculation within the canopy air space, below the EC system (Stoy et al., 2006). Missing GEP data is filled using a look up table, calculated for a moving window of 11-21 or 31 days (with window size set to achieve a minimum of 16 h of valid data), of mean GEP compared to photosynthetic active radiation (PAR), divided into bins of width 200 mol m −2 s −1 ) and time of day (2-h wide bins). The hourly bins account for the higher sensitivity of photosynthetic activity to light in the morning compared to afternoon  and the moving window accounts for sensitivity to seasonally varying soil moisture.
We aggregated hourly GEP to 16-day time periods for purposes of presentation, assuming that at least 4 days were available with at least 21 h of observations each (Fig. 2). Gaps were not filled further. The ∼16-day window is a time scale representative of important ecological processes, in particular, leaf appearance to full expansion (Jurik, 1986) and corresponds to the MODIS vegetation index product interval.

Canopy photosynthetic capacity and Bowen ratios
Pc is an index of canopy photosynthetic capacity that is independent of varying light levels; we estimate it as the mean ecosystem productivity (GEP) for a fixed range of PAR (725 < PAR < 925 mol m −2 s −1 ), following Hutyra et al. (2007).
Pc is closely related to canopy-scale Light Use Efficiency (LUE), since dividing by the fixed PAR level for which it is defined (∼825 mol m −2 s −1 ) gives the average slope of the light response curve in units of LUE (e.g. mol CO 2 fixed per mol PAR). It is also significantly correlated (R∼0.47, p < 0.0005) with seasonal patterns in LAI (as measured optically by an LAI-2000 instrument) an independently derived index of photosynthetic infrastructure (Supplement Fig. 5).
Canopy-scale photosynthetic capacity arises from factors in addition to LAI (including variations in photosynthetic capacity at the leaf scale), but the correlation with LAI confirms that Pc is in fact capturing not just flux but underlying infrastructure. Interpretation of Pc strictly as an index of canopy-scale photosynthetic infrastructure that is independent of environmental conditions may still be confounded by the covarying seasonality of other environmental factors, in particular, vapor pressure deficit (VPD). VPD should be higher in the dry season, and higher VPD should induce stomatal closure that suppresses canopy photosynthesis at any given light level (Lasslop et al., 2010;Lloyd and Farquhar, 2008;Wehr et al., 2013). We therefore compared our results, where possible, to a version of Pc defined for different fixed VPD-bins to test the effect of removing VPD changes on the results (see Section 3.3).
In this analysis, we refer to normalized values for GEP and Pc, dividing by the maximum value within the annual cycle to give, GEP GEP max −1 and Pc Pc max −1 , as our focus here is on investigating differences in seasonal patterns across sites, not on the absolute differences in carbon balance, which may be less certain. For the 16-day intervals, GEP and Pc were calculated in units of gC m −2 d −1 . Evapotranspiration (ET) and photosynthetic activity (GEP) are considered tightly related processes and both fluxes will decrease during periods of water stress or low incoming radiation and ambient temperatures. As an indicator of water stress, we calculated the Bowen ratio of sensible (H) to latent heat flux (LE), similar to Yuan et al. (2007) and da .

Environmental driving variables
A suite of meteorological sensors measured environmental driving variables: photosynthetically active radiation (PAR), net radiation (Rn), precipitation (Prec), and air temperature (T a ) among others ) (Supplement Table 1 and Fig. 2). Soil moisture (Â in m 3 m −3 ) was measured using frequency-domain reflectometry (TDR) at K34 (Teixeira et al., 2003), K83  and at the control site of the Tapajos Throughfall Exclusion Experiment (TEE control ∼ 4.5 km from the K67 tower) (Brando et al., 2006). We used consistency of relationships between similar sensors to identify and remove faulty periods and obtain a more complete meteorological time series (e.g. Rn and SW down in W m −2 ).
In addition to the flux and meteorological measurements, we used satellite-derived data from the Tropical Rainfall Measuring Mission (TRMM) (NASA, 2010) to obtain a consistently filled precipitation time series across all sites. We think this is justified by the observed correlation between TRMM and tower data when tower data were available (Supplement Fig. 1 and Table 2). Dry season is defined following the conventional definition for the basin, when precipitation was less than 100 mm month −1 (Shuttleworth, 1988).
For the estimation of reliable PAR, we compared PAR to short wave radiation (SW down in W m −2 ) data at Brasil flux network sites where both were available. At Caxiuanã (CAX), Santarém K83 and K77, Manaus (K34), and Fazenda Nossa Senora (FNS), raw PAR data showed a decline over time, probably attributable to PAR sensor degradation, that we corrected by scaling to short wave radiation measurements, which showed no evidence of degradation. For these corrections, we derived correlation coefficients between calibrated PAR observations and short-wave incoming radiation measurements (rgs in W m −2 ), where PAR ∼ 2 × rgs (Papaioannou et al., 1993;Szeicz, 1974). PAR values at Pe-de-Gigante (PDG) and Bananal Island (JAV) were abnormally high. To make PAR (mol m −2 s −1 ) consistent to other sites we used our derived PARshortwave radiation relationship to scale PAR at PDG and JAV by multiplying the complete time-series by 0.79 and 0.86, respectively. PAR and Rn (W m −2 ) were composited to 16-day intervals by simple average across all daytime and nighttime hours, as daylength (a function of time of year and latitude) is a significant driver of the variation in radiation received. Information regarding light sensors and other tower instrumentation can be found in Supplement Table 2.
We calculated hourly values of top-of-atmosphere (TOA) radiation in W m −2 units, following the method of Goudriaan (1986), in which the solar constant (1370 J m −2 s −1 ), is modified according to orbital mechanics that can be determined by site latitude, day of year, and local time. We generated 16-day composites by simple averaging of calculated hourly values across all day and nighttime hours in each 16-day interval.

Leaf-flush model
In order to better understand the underlying processes that drive canopy scale photosynthetic assimilation, we implemented a simple model that assumes that tower-measured Pc (gC m −2 d −1 ), is determined by the balance between the construction of new photosynthetic capacity (via leaf flush) and the loss of photosynthetic capacity (via litterfall): where leaf flush and litterfall quantify the mass flow of leaf matter (gC m −2 ground d −1 ) into and out of the canopy, respectively. Leaf flush and litterfall produce and remove photosynthetic capacity of the canopy according to the amount of photosynthetic capacity added or removed per unit mass of leaf, given by A max × SLA, the product of average leaf level photosynthetic assimilation per unit area, A max (gC m −2 leaf d −1 ) and specific leaf area, SLA (m 2 leaf gC −1 ). For unit conversion, we accounted carbon for 49% of the biomass (Chambers et al., 2001). We implemented the leaf-flush model at both Santarém forest sites (K67 and K83) and at Manaus (K34), where data for litterfall (Figueira et al., 2008;Luizao and Schubart, 1987;Rice et al., 2004) and leaf-level gas exchange parameters (Carswell et al., 2000;Domingues et al., 2005) were available. SLA values were set to 0.0032 for K67 and K83 (Domingues et al., 2005) and to 0.0038 m 2 gC −1 for K34 (Carswell et al., 2000). In a first iteration, at both Tapajos forests, we assumed A max to be constant and comparable to mean leaf carbon assimilation at saturating light (A max ) at 80% height, using the equation A max = 0.042 × (relative height) + 5.34 (Domingues et al., 2005), A max = 8.66 gC m −2 d −1 assuming most of the photosynthesis will be happening at the top of the canopy 100-80% relative height. For Manaus K34, A max was set to 7.36 gC m −2 d −1 following results presented by Carswell et al. (2000).
This approach likely oversimplifies canopy dynamics by assuming that the photosynthetic capacity of new and falling leaves is the same. We could improve this method by using distinct estimates of the leaf photosynthetic capacity of old and new leaves and constructing an explicit age model for the leaf pool. In the absence of more complete data, we used a sensitivity analysis where we assume that A max decreases linearly starting 2 months before the peak in leaf-fall annual cycle, reaching 80%, 60% or 40% of mean A max . Thus assuming that most of the abscised leaves would show a lower A max and it will have an effect on the hole-ecosystem rate of assimilation.
We estimated dPc/dt by first smoothing the Pc time series using a LOESS (locally weighted scatter plot smoothing) regression, and then taking the derivative of a 2nd order polynomial fitted to 3 measurements (previous, target and next 16-days). Finally, we solved Eq. (1) for leaf flush at each 16-day time intervals.
This approach (Pc-flush) is similar to that employed by Doughty and Goulden (2008), who derived seasonal variations in leaf flush from variations in canopy leaf area index (dLAI/dt, estimated from light measurements above and below the canopy), except that we used the eddy flux derived quantity, Pc, instead of optically derived LAI. We compared these two approaches. For the LAI based approach (LAI-flush), we used LAI measurements (made with the LI-2000) on the control forest plot of the near-by drought experiment (Brando et al., 2010;Juárez et al., 2007), plugging these into a variation of Eq. (1) that replaced Pc/A max with LAI.

Uncertainty analysis on GEP and Pc
Errors or uncertainties in characterizing the mean seasonality of photosynthetic patterns arise from at least three sources: systematic measurement bias (e.g. from failing to correctly adjust for biases arising during periods of low turbulence, Goulden et al., 1996), random sampling error, and variation in seasonality due to interannual variability in climate or other driver variables. We account for the latter two sources of uncertainty by calculating the standard deviation in measured flux in each 16-day period across the multiple years of each site's observations. However, this approach will be unlikely to account for uncertainty due to systematic measurement bias (because the same bias can persist across multiple years).
We estimated the uncertainty of measurement bias as that associated with choosing the appropriate u * thresh for filtering out periods with unrepresentatively low fluxes. We used the bootstrap method, drawing samples (with replacement) from each of the cells of the quarterly look-up table described in Section 2.2 for choosing u * thresh (u* − T a − NEE), to create 1000 bootstrap sample sets of NEE versus u*, with each sample preserving the nighttime air temperature distribution. Each bootstrap sample gives an estimate of u * thresh . The resulting 4000 mean values (1000 samples and 4 quarters per year) are fitted to a normal distribution and the mean and 95% confidence interval of this distribution defines the uncertainty range for u * thresh , as reported in Table 1. This confidence interval in u * thresh , when propagated to results derived from NEE (including for GEE), gives an associated uncertainty range, or error, on the derived flux that arises from uncertainty in our ability to correct for the bias caused by low turbulence.
The uncertainty in correcting for bias error was added in quadrature, using a sum-of-squares approach, to the uncertainty in interannual variability (described above) to give a total uncertainty on the seasonality (Err 2 tot = Err 2 bias +Err 2 interannual ). This total uncertainty is plotted as a shaded area for GEP and Pc time series in Fig. 3.

Dataset availability
All driver and flux data, as well as relevant processing codes (in Matlab) needed to reproduce this analysis are publicly available for download at http://eebweb.arizona.edu/faculty/saleska/data.htm and at http://www.climatemodeling.org/lba-mip/. . Light gray boxes indicate dry season conditions (precipitation lower than 100 mm month −1 ). All lines represent the mean seasonal cycles, an average over all data years available for each site and the gray areas around the mean are the standard deviations when more than one year of data is available.

Environmental conditions across Brasil flux network tower sites
Environmental and climatological conditions at the Brasil fluxsites (Supplement Table 1), show differences in seasonal patterns across the basin. Annual precipitation values ( -2008 were highest at the CAX forest site (2463 mm) and lowest at the PDG savanna site (1494 mm). Precipitation at Santarém and Bananal Island forest sites (K67, K83, and JAV) is approximately 1800 mm year −1 and compares to rainfall at the Manaus forest (K34) and the Rondônia forest and pasture sites (RJA, FNS), both ∼2000 mm year −1 . Dry season precipitation values, in contrast to annual values, were lower at the southern site (RJA) than at the equatorial sites. In general, dry-season length at our measurement sites is approximately five months, with CAX and K34 having the shortest dry seasons, 3 and 2 months, respectively.
The seasonality of sunlight and incoming energy at the surface arises from interacting effects of cloud cover (which reflects or absorbs radiation before reaching the surface, and peaks with precipitation), and the seasonality of incoming energy at the top-of-the atmosphere (which depends on latitude and the seasonal tilt of the earth's axis with respect to the sun). Surface PAR levels at equatorial Amazon sites show a seasonal cycle with a minimum in the wet season (January-March), and a maximum in the dry-season (∼July to December), when reduced cloud cover increases atmospheric transmittance just as top-of-atmosphere radiation reached its second maximum with the September equinox (Fig. 3). Despite greater seasonality in top-of-atmosphere radiation at southwestern sites (RJA, FNS) 10 degrees south of the equator, there was less seasonality in surface radiation. This is a consequence of the period of minimum cloud-cover in the dry season (May-August), which is out of phase with maximum top-of-atmosphere radiation (October-February). Since dry season minimum cloud cover (maximum atmospheric transmittance) nearly coincides with minimum top-of-atmosphere radiation (June solstice), there is no obvious dry season peak in the observed surface solar radiation (Fig. 3). Moving further south to the savanna site at 22 • S, the top-of-atmosphere radiation variability increases further, dominating the cloud-cover effect, so there is a strong minimum of surface PAR in the dry season, even though transmittance is maximized due to low cloud cover (April-October).

Seasonality of photosynthetic activity, GEP
Three distinct seasonal patterns in GEP were observed: First, across the equatorial Amazon forest sites (K34, K67, K83, and CAX), GEP increases, or is maintained at high levels, as the dry season progresses (Figs. 3 and 4b and e). The equatorial site converted to agricultural use (K77) showed distinctly different patterns from the nearby equatorial forests, with large dry-season reductions in GEP, presumably a consequence of the loss of deep roots that can access deep soil water. Coinciding with the GEP decline at K77 is a sharp increase in Bowen ratio (sensible:latent heat flux ratio, Fig. 4c and f and Supplement Fig. 2), indicating significant water limitation relative to the nearby forests, which show little change in Bowen ratio.
In the second pattern, at southwestern Amazon sites (forest at RJA and pasture at FNS), GEP declines throughout the dry season. At these sites, the dry season tends to be more extreme than at equatorial forests and, notably, there is no distinct peak in solar radiation (Fig. 3).
An exception to the southern site pattern is the Amazon ecotone site on Bananal Island (JAV) where increasing dry-season GEP peaks simultaneously with PAR and a late dry-season decline in ET and increase in Bowen ratio, which remains high for the rest of the wet season ( Fig. 4c and f). However, this site is seasonally flooded during the wet season, and may not experience as extreme soil dryness as the other southern sites. The seasonal wet season flooding (in February) is associated with declining GEP. This could be associated with anoxia, or it could be due to difficulty in interpreting eddy flux measurements during inundation. Periods of flooding are inconsistent with the standard eddy flux assumption that biosphere-atmosphere exchange captures all fluxes associated with biological activity, because significant carbon export may occur via flowing water.
The third pattern is the site in the savanna (or cerrado) biome (Fig. 3, savanna site PDG). Here, the trend from high or increasing dry season photosynthetic activity (in equatorial forests) to moderate dry season photosynthetic decline (in southern forests) reaches its extreme in the transition from forest to cerrado (Fig. 3). In comparison to forest vegetation, savanna photosynthetic metabolism exhibits a much larger range, experiencing an average reduction of 80% during the driest part of the dry season, when incoming sunlight also reaches its minimum.

Seasonality of photosynthetic capacity, Pc
Photosynthetic capacity (Pc) (Fig 3, second row), calculated from GEP within a narrow fixed range of moderately high light levels, removes the effect of differences in incoming PAR on productivity, giving an index of underlying photosynthetic infrastructure of the canopy Hutyra et al., 2007). Pc is closely related to canopy-scale Light Use Efficiency and is significantly correlated with seasonal LAI (Supplement Fig. 5).
The patterns of Pc are broadly similar to those seen for GEP: At equatorial Amazon sites, Pc either increases over the course of the dry season (CAX, K67, K83) or remains approximately constant (K34). Pc at these sites, like GEP, also remains high well into the wet season, before declining to reach a minimum at the end of the wet season ( Fig. 4a and d).
In contrast, at southern Amazon forest sites (RJA, JAV), and especially at the savanna site (PDG or at sites converted from forest (K77, FNS), Pc decreases as the dry season progresses and rapidly recovers after the onset of the rainfall.
We tested whether these observations of Pc seasonality could be an artifact of seasonal variations in VPD by plotting the seasonality of Pc defined for fixed VPD bins -at low (0-1 kPa), medium (1-2 kPa), and high (2-3 kPa) VPD levels -and comparing the results to our standard Pc definition across all VPD bins (Supplement Fig. 6). Fixed VPD bin Pc seasonalities were noisier and missing for some time-periods (because datasets constrained by both fixed light and VPD bins were smaller), but they showed patterns similar to those in Fig. 4: stable or increasing Pc at non-water limited sites (K34, K67, K83, and CAX), and Pc declines at the other (water-limited) sites (Supplement Fig. 6). We therefore conclude that the observed seasonal patterns of Pc are robust to the effect of variations in VPD.

Controls on seasonality of GEP and Pc
Across biomes, climates, and land uses, we observed two distinct dry season patterns (increases versus decreases in photosynthetic indices as the dry season progressed) that were largely predictable from an index of water limitation, the ratio of sensible to latent heat flux (i.e. Bowen ratio). In general, sites with access to water reserves sufficient to enable latent heat fluxes to maintain the same proportion to sensible heat fluxes in the dry season as in the wet season (Fig. 4f) also sustain or increase measures of photosynthetic capacity (Fig. 4d) or flux (Fig. 4e) in the dry season. By contrast, sites exhibiting significant increases in Bowen ratio during the dry season -whether due to land use change which removed deep roots (K77 agricultural site), natural limitation on access to deep water (RJA forest site), or absence of water during intense dry seasons (PDG savanna site) -show declines in photosynthetic capacity or flux during the same period.
However, despite access to water resources (as indicated by Bowen ratios) and high levels of dry season PAR, equatorial Amazon sites where GEP and Pc increased during the dry season showed little evidence that seasonal variation in photosynthetic activity or capacity was directly driven or limited by available energy (Fig. 4). Indeed, instead of positive correlation with PAR, average daytime GEP decreased with radiation at almost all sites across the basin, including those that exhibit dry season increases in photosynthesis ( Fig. 5 top panels).
We note that this result from regression analysis is also inferable from the broad patterns observed in Sections 3.2 and 3.3, in which photosynthetic activity and capacity at equatorial sites increase throughout the dry season (when sunlight is high) but also remain high in the first part of the wet season (when sunlight is declining). These indices of photosynthesis reach their minimum at the end of the wet season, even as available sunlight is increasing. This is in distinct contrast to controls on water fluxes: as noted in previous studies (Hasler and Avissar, 2007;da Rocha et al., 2009), LE, is well-correlated with net radiation (Rn) across most sites, even accounting for reduced sensitivity at southern forest sites (Fig. 5,  second row).
Note that consistent with standard practice, we investigated PAR as a driver of GEP and net radiation as a driver of LE. This practice arises because photosynthetic activity is proportional to the number of photons absorbed in the 400-700 nm spectral range, and not to their energy (Alados and Alados-Arboledas, 1999), whereas evaporation is driven by total energy flux (W m −2 ) across the solar spectrum (as captured in Rn). However, it is important to note that these results are robust regardless of which measure of incoming energy is used).
Despite the weak or negative association between GEP and radiation on average (Fig. 5), variation in PAR well-explained the deviation of GEP from Pc, when GEP was expressed as the fraction of photosynthetic capacity utilized: GEP Pc −1 (the fraction of photosynthetic capacity, Pc, utilized in GEP) is consistently positively correlated with PAR at all tower sites (Fig. 6a) (GEP = k PAR Pc, with k ≈ 0.8-1.2 m 2 s mmol −1 for forest sites).

Leaf-flush model and analysis
GEP seasonality may be explained primarily by Pc seasonality, with differences between GEP and Pc explainable by light, but environmental factors do not explain Pc seasonality. To explore the seasonal patterns of allocation and the mechanisms that drive Pc we examined results from a simple leaf-flush model (Eq. (1)) at sites with sufficient data (Santarém K67 and K83, and Manaus K34). All central Amazon forests showed a peak in leaf flush during the middle of the dry season (Fig. 7). This result follows from the constraint of sustained or increasing canopy photosynthetic capacity (Pc) in the face of increased losses of leaf matter due to higher dry season litterfall (Rice et al., 2004).
Comparisons between a leaf-flush model driven by Pc (Pcflush) and a LAI based approach (LAI-flush), showed that the two approaches produced consistent results at the K67 forest where we were able to compare them -there is a lag between LAI-flush and Pc-flush, which we expected, since leaves start to physically appear before their photosynthetic capacity fully develops. It takes ∼10-30 days for those new leaves to reach maximum A max values as reported for tropical evergreen species by Sobrado (1994). We can maximize the coefficient of determination of the relationship between the LAI-flush and Pc-flush (excluding the litterfall term), by lagging LAI-flush going from 0.13 (non-lagged) to 0.40 with a slope of 1.00 (32-day lag) (Supplement Fig. 3). These leaf-flush model results were also consistent with observations of Coussarea racemosa A. Rich (common sub-canopy tree) foliage and flower production, showing flushing-flowering coinciding with peak irradiance during the onset of the dry season (Brando et al., 2006). These comparisons indicate that leaf-flush seasonality as seen by our simplified model based in changes in ecosystem photosynthetic capacity can be used as a proxy at other sites to represent the time when the new leaves start to contribute to the CO 2 uptake.
The results from a sensitivity analysis for three leaf-flush models where A max linearly decreased 2 months prior to the peak in litterfall showed a more clearly dry season peak in comparison to the leaf-flush model where we kept A max constant throughout the year (Supplement Fig. 4). Thus, as decreasing ecosystem A max before abscission should be counterbalance by a higher number of more productive leaves in order to maintain the observed increasing levels of Pc-a balance between quality vs. quantity of leaves.
Seasonality of leaf productivity (from the leaf-flush model) was anti-correlated with the seasonality of wood productivity during tree growth obtained from dendrometry measurements at K67 Rice et al., 2004;Saleska et al., 2003), K83  and K34 (Pereira da . Correlation coefficients in the Tapajos (K67: R = −0.76, K83 R = −0.64, p = 0.000) indicate distinct differences in seasonal carbon allocation: whereas leaf production peaks in the dry season, wood production peaks in the wet season, reaching maximal values between mid-February and the end of April when soil moisture is high. This relationship was non-significant at the K34 forest near Manaus (R = 0.1, p = 0.71).
Multivariate regression models showed that leaf flush was statistically associated with sunlight, 0-40 cm volumetric soil water content (Â 0-40 cm ), and their interaction, in a range of models (Table 2 and Supplement Table 3). Akaike's Information Criterion (AIC) selected the best models (  between PAR (mmol m −2 s −1 ) and soil moisture Â 0-40 cm (e.g. at K67 Â 0-40 cm = −0.98 × PAR + 0.66, R 2 of 0.21), the magnitude of coefficients should be interpreted cautiously. However, sunlight was statistically significantly positively associated with leaf flush in all models at all sites (Supplement Table 3), suggesting that this association is a robust result.

Discussion
Our goal is to better understand the seasonal patterns and drivers of Amazonian photosynthesis and productivity, and to provide a consistently processed database to test models of Amazon ecosystem function. Here we discuss three main questions Table 2 Linear regressions obtained by a nonlinear mixed-effects regression model for leaf-flush (Eq. (1)) versus combinations of 16-day average PAR (mmol m −2 d −1 , measured at each site separately) and soil volumetric water content, (m 3 m −3 , measured at K83), for Santarém forest sites (K67 and K83) and Manaus (K34).  (3) What is the role of intrinsic canopy phenology in understanding photosynthetic seasonality?

What is the seasonality of photosynthesis across the Amazon Basin?
We divided GEP seasonality results into three broad patterns. The first pattern -sustained or increasing GEP throughout the dry season at equatorial forest sites -confirms previous individual site carbon-cycle studies near Santarem (Saleska et al., 2003;Goulden et al., 2004;Hutyra et al., 2007), and is also broadly consistent with large scale remote sensing observations of moderate dry season "green up" that used the MODIS Enhanced Vegetation Index (EVI) as a proxy for photosynthesis (Huete et al., 2006). These results, when combined with recent analyses showing that the dependence of forest evapotranspiration on the energy available to drive water vapor fluxes did not differ between dry and wet seasons (Costa et al., 2010b;da Rocha et al., 2004;Fisher et al., 2009;Hasler and Avissar, 2007;Hutyra et al., 2007); provide strong evidence against the idea, represented in many earlier earth-system models, that water-limitation constrains photosynthesis in equatorial Amazonian forests.
Instead, sustained or increasing dry-season GEP might be considered consistent with light limitation Huete et al., 2006), as these sites experience dry season peaks in irradiance, due to reduced dry-season cloud cover (see further discussion below in the section on mechanisms). This interpretation is broadly consistent with the recent generation of ecosystem modeling studies, designed to represent dry-season increases in photosynthetic metabolism (Baker et al., 2008;Kim et al., 2012).
The second pattern, observed at the southern Amazon forest site (RJA at 10 • S) and the savanna biome (PDG at 20 • S), is modest (RJA) or sharp (PDG) dry season declines in GEP. This pattern may be evidence of water stress at these sites: the southern Amazon forest (RJA) has shallower soil than the equatorial forest sites (Table 1, Kruijt et al., 2004), and may not have access to deep soil water resources, while the savanna (PDG) has climatology typical of water limited savanna biomes (with the lowest annual rainfall and driest dry season of the eddy flux sites). However, we cannot rule out simultaneous co-limitation by light, since, in contrast to the central Amazon sites, they do not experience dry season increases in light.
The contrast in photosynthesis patterns between equatorial forests and the southern forest is consistent with the analysis of water fluxes in Costa et al. (2010a), which showed that at the equatorial Amazon forests, net radiation controls ET, while at the southern forest (RJA) the relationship between ET and Rn was statistically less significant (Fig. 5).
A recent study based on newly available observations of chlorophyll fluorescence from the GOSAT satellite (Lee et al., 2013) concluded that regional-scale photosynthesis exhibits dry season declines in Amazon regions with moderate dry season length. This appears at odds with observations reported here for the equatorial Amazon, but consistent with those of the southern Amazon. Since Lee et al.'s (2013) findings applied only to large areas that encompassed both our equatorial and southern region towers, we cannot yet judge whether GOSAT observations are consistent with those from eddy towers.
The third pattern is associated with land-use change. The converted sites (K77, FNS) showed distinctly different seasonality from the nearby forests, with relative reductions in dry season productivity, particularly at the equatorial K77 site, where the forest pattern of dry season increases in GEP was completely reversed. In the southern Amazon, where the RJA forest site showed some dry season decline in photosynthesis, conversion at FNS enhanced this behavior. These dry-season reductions in GEP are presumably a consequence of the loss of deep roots that can access deep soil water. Simultaneous with lowered photosynthesis, dry season evapotranspiration (ET) is also reduced at the equatorial K77 site, apparently constrained by the low precipitation . This is in contrast to the nearby forest sites (K83 and K67) where ET substantially exceeds rainfall during the dry season.
Declines in photosynthesis at savanna and converted sites are likely caused by water limitation, as indicated by seasonal cycle of Bowen ratios (Fig. 4c and f). The ratio shows that relative to the approximately constant sensible to latent heat flux ratio in forests (Supplement Fig. 2), conversion or shift in biome from forest to savanna causes energy fluxes to the atmosphere to be carried much more by sensible heat, as water supply to drive the latent heat flux is sharply reduced.
In sum, differences in seasonal GEP patterns appear closely associated with difference in climatology, biome type, and land use (Fig. 3). Along a North-South climate gradient, differences in GEP patterns correspond to declines in precipitation and/or shifts in the seasonality of irradiance. Near the equator irradiance peaks in the relatively cloud-free dry season; in the southern Amazon there is no seasonality (dry-season minimum in top-of-atmosphere irradiance nullifies the effect of reduced cloud cover); and at the high-latitude southern savanna, strong seasonality gives a dry-season minimum in irradiance (opposite in phase to the equatorial sites).
Next, we discuss specific tests for whether annual cycle in environmental variables is associated with photosynthesis patterns.

Do environmental drivers explain observed patterns in photosynthetic seasonality?
How are equatorial Amazon forests able to sustain high levels of photosynthetic capacity, even throughout dry seasons that many of the early model studies suggested should lead to water-limited reductions in photosynthesis?
Observations of exceptionally deep tree roots (up to 18 m) (Hodnett et al., 1996;Nepstad et al., 1994), of hydraulic redistribution by roots Oliveira et al., 2005), and of seasonal shifts in the water supply for ET from shallow (3 m) to deep (>7 m) soil layers between wet and dry seasons (Bruno et al., 2006) suggest mechanisms by which water limitation may be circumvented. These mechanism, could allow tropical forests to maintain access to non-transient stores of deep soil water, even when the surface becomes sufficiently dry to limit biological activity. Such mechanisms were missing from early models, as many were built around more studied temperate and agricultural systems, and which typically represent forests as having 2-3 m rooting depths.
If the deep-soil water hypothesis explains the maintenance of dry season photosynthesis, this predicts that removing access to deep soil water should reverse dry season GEP patterns. This is precisely what is observed with conversion of equatorial forest (e.g. K77, Fig. 4), which induces the effect of the forest-savanna transition by converting forest from a dry season "green-up" to a "brown-down" pattern, presumably due to the removal of deep roots, which allow for access to stored water during dry periods.
Thus, this study confirms earlier suggestions that many Amazon forests have mechanisms that allow them to escape water limitation. However, a new result of this study is the general finding that, across both sites that are water limited and those that are not, photosynthesis also shows no simple relation to sunlight (Fig. 5). At all Amazonian sites bi-weekly GEP shows a weak or declining relation with increasing light, contradicting assumptions of many models that days with high light are consistently more productive than days with low light. Only at the savanna site (PDG) was there a (weakly) positive relation between 16-day average GEP and PAR (Fig. 5 far right panel).
These patterns imply that combined canopy-scale evaporation and transpiration (i.e. total latent energy flux, tightly correlated with available energy) are decoupled from photosynthesis (which is uncorrelated or negatively correlated with available energy) (Fig. 5). If plant transpiration is coupled to photosynthesis, as is usually assumed, this suggests that the contribution of non-plant regulated evaporation must vary in tandem with photosynthetically coupled transpiration to maintain whole-ecosystem water flux that remains correlated with incoming energy.
We considered whether the lack of correlation between GEP and sunlight, and the divergence of this observation with LE response to incoming energy, was a consequence of photosynthetic saturation with respect to PAR, at least in the dry season. However, the available evidence suggests it is not. In particular, because sunlight still drives fluctuations in GEP relative to photosynthetic capacity (Fig. 6), it is unlikely that GEP is saturated. A more likely explanation, is simply that the underlying canopy photosynthetic capacity, a dominant control on photosynthetic flux, follows a seasonality determined by endogenous biological seasonal cycles of litterfall and leaf flush, and these are only indirectly related to sunlight (see Section 4.3, below).
In sum, canopy photosynthesis (in contrast to ET) cannot be modeled as a simple increasing function of available energy or sunlight. Consequently, observed patterns of photosynthesis and potential driving variables are not only inconsistent with ecosystem models that assume water limitation in these forests (Lee et al., 2005;Nobre et al., 1991;Werth and Avissar, 2002), but also with non-water limited light use efficiency models that are based on the assumption of constant photosynthetic capacity interacting with varying light levels (Xiao et al., 2005).

What is the role of intrinsic canopy phenology in understanding photosynthetic seasonality?
Taken together, the analysis presented here suggests that there is significant seasonal variation in standing canopy photosynthetic capacity (Pc, second row, Fig. 3), even in these evergreen forests. In other words, we find that canopy phenology, not environmental variability, is the dominant control on the seasonality of tropical forest photosynthesis, across all studied regions. The dominant seasonality of photosynthetic flux (GEP) simply follows that of the seasonality of the canopy photosynthetic capacity. GEP deviations from the primary pattern set by Pc are explained by variations in sunlight availability, which effectively regulates the fraction of photosynthetic capacity utilized (GEP/Pc) (Fig. 6). This suggests that if canopy photosynthetic capacity were constant (as many ecosystem models assume for tropical evergreen forests), then the seasonality of GEP at non water-limited sites would indeed be well-explained by the seasonality of sunlight.
For example, a recent study of Amazon forest photosynthetic seasonality using the new method of satellite-observed fluorescence (Lee et al., 2013) assumed that LAI was constant, leading them to conclude that "GPP changes are more related to environmental conditions" than to changes in canopy photosynthetic infrastructure, a conclusion not supported by the tower-derived observations reported here.
As noted by Doughty and Goulden (2008) based on observations at the K83 site alone, this seasonal variation in canopy photosynthetic capacity poses a challenge for modeling efforts. Though many ecosystem models allow Pc to be specified exogenously, e.g. as varying LAI (BIOME-BGC (Thornton et al., 2002), LPJ (Sitch et al., 2003), CASA (Potter, 2003), Orchidee (Krinner et al., 2005) and other Land Surface Models), few are able to prognostically simulate variation in photosynthetic infrastructure in evergreen tropical forests (for a simple recent approach to this problem, however, see Kim et al., 2012).
This raises the question: what controls the seasonal variation in canopy photosynthetic capacity? and in particular, why is it not well correlated with sunlight? Here, we used a simple the leaf flush model to characterize the underlying process that give rise to canopy photosynthetic capacity, and investigate their relation to environmental drivers. The leaf flush model embodies the constraint that changes in canopy photosynthetic capacity must arise from the difference between the creation of new capacity (leaf flush) and the loss of existing capacity (leaf abscission and litterfall). Leaf flush and litterfall likely arise from evolutionary strategies that are mostly beyond the scope of this study, but we can hypothesize here about which strategies might be adaptive in which climates. Litterfall at the three equatorial Amazon sites peaks in the dry season (Rice et al., 2004;Figueira et al., 2008). Though the seasonal cycle of leaf abscission has been interpreted as implying that LAI and productivity should decline in the dry season, our observations of simultaneously increasing dry-season photosynthetic capacity at these sites requires a dry-season peak in leaf flush sufficient to more than compensate for the peak in litterfall (Fig. 7). This dry season peak in leaf flush, together with multiple regression analysis (Table 2 and Supplement Table 3) showing consistent positive correlation, across sites and regression models, between leaf flush and light, suggests that leaf growth is light limited.
High dry-season light levels drive new leaf production in the dry season, which in turn drives increases in Pc and GEP that should level off, or maximize, only when leaf production declines to zero, at the end of the dry season; this expected late dry-season peak in GEP, lagging that of leaf flush, is especially visible at the Tapajos sites (Fig. 7). The leaf flush model thus helps explain features of the observed seasonal patterns of photosynthesis and light. If leaf growth (not photosynthetic capacity directly) is light limited, then the higher photosynthetic capacity that follows from high dry season leaf production should lag leaf growth, and persist into the wet season, also driving high GEP during the dry-wet season transition. This pushes GEP out of phase with sunlight, explaining the negative observed relationship between GEP and incoming radiation.
We hypothesize that seasonality in tropical forest Pc arises when limiting resources for leaf growth are seasonal. This hypothesis is supported by observations in this study and in the literature. For example, if the growth of leaves is light-limited, production of new leaves should occur during the season of maximal irradiance (a pattern typically seen in studies of tropical vegetation phenology, e.g. Borchert, 1980;Wright and Van Schaik, 1994), and equatorial climates with high dry-season light levels should benefit vegetation that grows leaves in the dry season (as seen in Fig. 7).
This benefit should also control allocation patterns within individuals, such that during high-irradiance periods, allocation to leaf production should be prioritized over allocation to growth of other organs. Just such a seasonal cycle in allocation patterns can be seen between leaf flush and wood growth (Fig. 7) in some of our equatorial sites, where production of leaves (which grow in the dry season) is out-of-phase with that of wood (which grows in the wet season). This pattern could be altered when other resource constraints become more pressing than that of light. For example, in seasonally inundated floodplain tropical forests, where anoxia limits respiratory requirements of wood growth during wet seasons, wood growth phenology should also be shifted into the dry season, as has been observed in tree-ring records from seasonally flooded Amazon forests (Dezzeo et al., 2003).
By contrast, in climates with no strong seasonality in irradiance (as at the southern forest site, RJA), light-based growth strategies would not be advantageous, and seasonal patterns would not favor dry season leaf production. There is only one southern upland Amazonian forest site (RJA), and we do not have long-term litterfall data from this site, so our ability to test this hypothesis for southern sites is limited. However, the dry season declines in Pc observed here would be consistent with absence of dry season leaf growth following from no increase in dry season irradiance at this site.
Stronger conclusions about whether southern forests are truly water limited (e.g. due to relatively shallower soils as in von , or simply experience little or no resource benefit from elevated dry season metabolism, would require more observations in this region of the Amazon.

Conclusions
Integrated study across Amazonian eddy flux towers provide new insights into the drivers on the seasonality of photosynthetic metabolism in tropical systems, the different factors which influence metabolism in different biomes and climate zones, and the ways in which human changes to the landscape alter ecosystem functions along with the changes in structure.
First, although southerly Amazon forests, strongly seasonal savanna, and sites converted from forest to agricultural uses exhibit photosynthetic and energy flux patterns consistent with waterlimitation, other sites do not. In particular, high or increasing levels of photosynthetic activity observed during the dry season in equatorial forests, together with no water limitation observed in the seasonality of the ratio of sensible to latent heat flux (the Bowen ratio), provide strong evidence against the common idea that water-limitation constrains photosynthesis in these forests.
Second, even though our analysis suggests that the seasonality of photosynthesis in equatorial Amazonian forests is not water limited, nor is it limited by sunlight in a simple way. Rather, photosynthetic fluxes largely follow the biologically determined phenology of canopy photosynthetic capacity, which itself is driven by patterns of leaf flush and litterfall. Light can affect canopy photosynthetic capacity indirectly, and with time lags, by driving leaf growth.
Third, equatorial climates benefit vegetation that can grow leaves in the dry season, when surface solar radiation peaks. By contrast, southerly forests may not allocate to new leaves to increase their photosynthetic activity and capacity in the dry season, either because water limitation prevents them, or because lack of seasonality in surface radiation implies that no benefits accrue from such strategies.
This analysis of the seasonality of photosynthetic metabolism provides a window onto the mechanisms controlling photosynthetic metabolism of Amazon vegetation, showing that responses to seasonal climate variation are complex, and pose challenges to models seeking to represent Amazon forest responses to climatic variation.