In Amazon forests, the relative contributions of climate,
phenology, and disturbance to net ecosystem exchange of carbon (NEE) are not
well understood. To partition influences across various timescales, we use a
statistical model to represent eddy-covariance-derived NEE in an evergreen
eastern Amazon forest as a constant response to changing meteorology and
phenology throughout a decade. Our best fit model represented hourly NEE
variations as changes due to sunlight, while seasonal variations arose from
phenology influencing photosynthesis and from rainfall influencing ecosystem
respiration, where phenology was asynchronous with dry-season onset. We
compared annual model residuals with biometric forest surveys to estimate
impacts of drought disturbance. We found that our simple model represented
hourly and monthly variations in NEE well (
The Amazon's tropical forests are pivotal to global climate, exchanging large, globally important quantities of energy and matter, including atmospheric carbon (Betts et al., 2004). Amazon forests contain 10 %–20 % of Earth's biomass carbon (Houghton et al., 2001). Increased emissions of the forest's carbon can therefore accelerate climate change, and attention is now focused on how vulnerable this large reservoir of carbon will be to a potentially drier future climate (de Almeida Castanho et al., 2016; Farrior et al., 2015; Duffy et al., 2015; Longo et al., 2018; McDowell et al., 2018). Characterizing the response of present-day Amazon rain forest carbon balance to climate and drought disturbance is a necessary step to improving predictions of future vulnerability.
Eddy covariance
Partitioning the exogenous and endogenous influences upon eddy covariance NEE
is possible using statistical modeling (Barford et al., 2001; Yadav et al.,
2010; Wu et al., 2017). To partition influences upon NEE in a 20-year eddy
flux record in a temperate New England forest, Urbanski et al. (2007) used a
statistical modeling approach: by representing hourly NEE merely as response
to exogenous meteorology and annually integrating their results, they
concluded that meteorology did not explain the accelerated uptake seen in
annually integrated NEE. They hypothesized that residual uptake was due to
long-term forest regrowth and succession, a hypothesis that was corroborated
by biometric measurements of increasing canopy foliage and accelerating
mid-successional tree biomass accrual. This novel partitioning framework for
NEE has not previously been applied to any tropical forest, in part because
long-term eddy covariance coverage of tropical forests is lacking
(Zscheischler et al., 2017). A simple statistical framework may allow
tropical forest
On seasonal timescales, tropical evergreen forests undergo endogenous changes in GEP via the phenology of leaf flush and abscission (Doughty and Goulden, 2008; Restrepo-Coupe et al., 2013; Wu et al., 2016). The seasonal dependency of productivity has motivated the development of rooting depth and phenology sub-models in dynamic global vegetation models (DVGMs) (Verbeeck et al., 2011; De Weirdt et al., 2012; Kim et al., 2012). These sub-models have led to complexity in the modeled mechanisms controlling the GEP seasonal cycle without necessarily improving accuracy. It is necessary to quantify the magnitude and timing of phenology's effect on the GEP seasonal cycle after accounting for the integrated hourly response to sunlight.
On interannual to decadal timescales, endogenous changes in forest NEE can arise from disturbance and recovery (Nelson et al., 1994; Moorcroft et al., 2001; Chambers et al., 2013; Espírito-Santo et al., 2014; Anderegg et al., 2015). The km67 eddy flux site in the Tapajós National Forest (TNF) presents a unique opportunity to study the potential legacy of disturbance caused by drought. This eastern Brazilian Amazon forest lies on the dry end of the rainfall spectrum for tropical evergreen forests (Saleska et al., 2003; Hutyra et al., 2005). A severe El Niño drought in 1997–1998 was followed by disturbance, evidenced by a large and heavily respiring coarse woody debris (CWD) pool in 2001. Subsequent NEE measurements showed a 4-year transition from being a net carbon source in 2002 to nearly carbon-neutral in 2004 and 2005 (Hutyra et al., 2007). The observed disequilibrium state led researchers to the hypothesis that RE was high but dissipating and that the forest will continue to transition into equilibrium, becoming a sink throughout the decade (Pyle et al., 2008). Conversely, this hypothesis implies that any new disturbance should drive the forest back into disequilibrium, becoming a source again. We test these predictions using meteorological records; forest inventories of aboveground biomass (AGB) and CWD; and an additional 3.5 years of eddy flux data, resumed after a 2.5-year interruption, collected since prior studies.
In this study, we test hypotheses related to controls of NEE on multiple timescales at an eastern Amazon rain forest. Specifically, we seek to answer the following questions: (1) what were the effects of exogenous meteorology upon NEE across hourly to yearly timescales? (2) What is the seasonal effect of canopy phenology upon NEE? Is phenology synchronized with wet/dry seasonality? (3) Major basin-wide droughts occurred in 1998 before eddy flux measurements began, and they were reported again in 2005 and 2010 (Zeng et al., 2008; Philips et al., 2009; Lewis et al., 2011; Doughty et al., 2015) during the span of measurements. Did any of these basin-wide droughts affect the TNF in particular? What was the impact of drought upon interannual variability and the decadal trend in NEE? Furthermore, which NEE component, GEP or RE, was perturbed most by drought? Overall, we statistically partitioned the multiple influences on NEE across timescales from hours to an entire decade of eddy flux and forest inventory measurements.
The Tapajós National Forest (TNF) is located to the southeast of the
convergence of the Tapajós and Amazon rivers in Pará, Brazil. The
forest site is on the dry end of the spectrum of evergreen tropical forests,
receiving 1918 mm of annual rainfall and experiencing a 5-month-long dry
season from July 15 to December 14, defined by average monthly precipitation
of less than 100 mm (Hutyra et al., 2007). Temperature and humidity average
25
Hourly fluxes of NEE were calculated using the sum of hourly turbulent eddy
fluxes plus the hourly change in height-weighted average
Nighttime NEE measurements were filtered for low turbulence. We used a
turbulence threshold filter of
We used established gap-filling models to obtain annual NEE totals. Gross
ecosystem productivity (GEP) was gap-filled using a hyperbolic fit curve
between GEP and photosynthetically active radiation (PAR) (Waring et al.,
1995). For RE, we adapted the method by Hutyra et al. (2007), who calculated
missing, filtered, and daytime hours using 50
Meteorological variables measured at km67 included PAR, temperature, and specific humidity. Downward drifts in PAR
data due to a degrading sensor were corrected by de-trending a time series of
midday PAR observations in the top 95th percentile of each month (Longo,
2014). This threshold included substantial information about the sunniest
hours, throughout which intensity should remain constant between years for
any given month. We scaled the radiation time series using the proportion
between the fitted trend and the initial fitted value. Simultaneous total
incoming shortwave-radiation measurements allowed us to partially fill
missing periods of PAR data using a relationship derived from linear
regression in simultaneously measured hours (
Rainfall measurements were greatly underestimated at this site because of a faulty tipping bucket rain gauge. We discarded site-based data and calculated a distance-weighted synthetic hourly rainfall time series from a network of nearby meteorological stations, with locations ranging from 10 to 110 km away from km67. More information on the meteorological network is available in Fitzjarrald et al. (2008). Detailed information about the subsequent calculations of the synthetic precipitation data set and PAR drift correction are available in Longo (2014).
Additionally, the Brazil National Institute of Meteorology (INMET) has a
station at Belterra, located 25 km away from km67, with daily precipitation
totals dating back to 1971, which were used to corroborate the seasonal and
long-term trends at km67. Correlation between these two monthly data sets for
the years 2001–2012 was
To assess how disturbance coincided with changes in NEE, we conducted surveys
of coarse woody debris (CWD). These surveys capture the magnitude and
dynamics of the respiring pool of dead tree biomass. Transect subplots were
surveyed in 2001 for pieces greater than 10 cm in diameter (Rice et al.,
2004). Bootstrapped confidence intervals were quantified by resampling
subplot totals (
Because CWD surveys were conducted infrequently, we inferred mortality from
aboveground biometry surveys in 1999, 2001, 2005, 2008, 2009, 2010, and 2011.
Trees larger than 10 cm diameter at breast height were surveyed and were
converted to biomass using non-species-specific equations (Chambers et al.,
2001a) based on sampling previously established protocols for this site (Rice
et al., 2004; Pyle et al., 2008). Mortality biomass was inferred by tallying
biomass of dead trees that were alive in the prior survey. Sometimes, trees
were missed by the census surveyors before they could be confirmed dead or
were found again. In 2012 we assigned missing trees that were not later found
alive an equal probability of dying in all surveyed years in which they had
been missing (Longo, 2014). We used tree mortality to model CWD over time
using a simple box model with a first-order rate equation:
Our low-parameter empirical model represents the mean response of NEE to
hourly and seasonal changes in exogenous meteorology and seasonal changes in
phenology throughout the decade. We used our model to diagnose interannual
nonstationarity in model residuals, which correspond to endogenous ecosystem
changes in photosynthesis and respiration rates between years, give or take
random measurement error and unaccounted for model terms. We fit the model to
the entire 7.5-year interrupted eddy covariance record of raw,
Atmospheric moisture and diffuse radiation, in addition to radiation, are
also known to affect photosynthesis at tropical sites on short timescales
(Kiew et al., 2018), by affecting stomatal closure and hence controlling the
degree to which photosynthetic uptake saturates at high PAR. We tested a
higher-parameter model based on a light and moisture model representing
exogenous changes to LUE from Wu et al. (2017) to examine whether these
meteorological variables added explanatory power to our model at monthly and
longer timescales. This model adjusts LUE by multiplying terms that account
for effects of vapor pressure deficit (VPD:
This forest site has coincident deficits in rainfall and ecosystem RE during
the dry season (Saleska et al., 2003; Goulden et al., 2004) due to
desiccation of dead wood, leaf litter, and other substrates for heterotrophic
respiration (Hutyra et al., 2008). To depict this reduced dry-season RE, we
set dry-season
We tested three different seasonal timings for the phenology factor variable:
(1)
After subtracting hourly NEE
We partitioned both NEE
Annual sums of NEE in kg ha
NEE has a large diurnal cycle relative to its mean seasonal cycle, with a
mean diel range of 25.05
We examined our distance-weighted interpolated estimate of km67 rainfall for trends and droughts. Our precipitation estimate was consistent with previous estimates of precipitation for this site and region, with a minimum of 1595 mm in 2005 and maximum of 2137 mm in 2011 (Saleska et al., 2003; Nepstad et al., 2007). While 2005 annual precipitation was a minimum, no previous groundwater deficits in carbon exchange, latent heat flux, or sensible heat fluxes were observed during this year (Hutyra et al, 2007). Our measurements did not indicate that any drought occurred during or immediately preceding period 2 of NEE measurements. In fact, period 2 annual rainfall totals increased on average by 20 % relative to period 1. The dry season in 2009 was longer than average, lasting 6 months (Fig. 2a). Mean annual radiation was expectedly anti-correlated with annual rainfall. Accordingly, period 2 experienced 4 % less mean annual PAR than period 1.
Measurements of total CWD (black squares with 95 % bootstrapped
CI error bars) and subsets of CWD
Our synthetic decade-long rainfall record corresponded closely with the nearby INMET Belterra measurements, although INMET Belterra had on average 220 mm of rainfall more per year, likely due to differences in circulation and convection between the km67 forest and Belterra pasture land surface (Fitzjarrald et al., 2008). Annual rainfall totals throughout the decade of eddy flux measurements of 2002–2011 lay well within the historical variability of annual rainfall since 1972, which experienced a range of 974 to 3057 mm of annual precipitation (Fig. 2b). The second- and third-lowest annual precipitation totals (1391 and 1218 mm, respectively) occurred during 1997–1998, during a major El Niño event, which persisted from June of 1997 to June of 1998 (Ross et al., 1998) and corresponded with a 9-month-long dry season, the longest in the historical record.
We examined measurements of CWD over time to assess whether a disturbance
might have impacted the period 2 carbon balance. Compared to CWD stocks in
2001 of 48.6 (
Example time series of NEE
Model parameter values (95 % confidence intervals in
parentheses) and
A box model of CWD (Eq. 2) allowed us to estimate the transient behavior of
the CWD pool throughout years in which it was not directly measured (Fig. 3).
The CWD mortality input rates
Assuming that the large initial CWD pool arose from a past disturbance,
hypothetically following the 1997–1998 El Niño drought, we ran the CWD
box model (Eq. 2) backward in time to estimate the magnitude of such a
disturbance. We assumed that the disturbance occurred in 1998 because 1999
and 2000 were not characterized by below-average rainfall. Severe drought
events have been accompanied by increased mortality and canopy turnover rates
in intact Amazon forests (Leitold et al., 2018). Because the CWD measurement
was made in July of 2001, we calculated the box model CWD value to the end of
the El Niño drought in June 1998 using the same respiration rate,
Optimized parameter values for our model are included in Table 1. Our model predicted 81 % of the variance in observed hourly NEE and captured 94 % of the amplitude of the diurnal cycle. The only hourly independent variable in the model was PAR; hourly NEE in our model was therefore predominantly driven by changes in sunlight. Modeled hourly variability frequently captured the difference in magnitude in NEE between high- and low-uptake events (example time series shown in Fig. 4).
Mean daily data–model residuals averaged over all 7.5 years:
In our best-fitting model parameterization, phenology was asynchronous with
the dry season (Table 2). Over the mean seasonal cycle, removing this
seasonal phenology parameterization resulted in positive residual NEE from
15 June to 14 September, hence overpredicting uptake during this time
(Fig. 5a). Our final model, however, simplistically corrects for this
positive anomaly, adjusting NEE by 16 % (Fig. 5b; Table 2). Although this
seasonal transition appears to be more gradual over the season, our
simplistic, low-parameter phenology representation was chosen for parsimony.
While the seasonal timing of respiration,
Mean seasonal cycle of NEE
Our model predicted monthly mean NEE well (
Part of the remaining seasonal variability was explained by random
measurement error: 95 % bootstrap confidence intervals representing
hourly measurement errors in monthly mean NEE had an average range of
1.07
A higher-parameter model with VPD and diffuse radiation from Wu et al. (2017)
explained additional variance in hourly NEE but not in monthly NEE (Table S1
in the Supplement). The BIC score for this model (
Hourly changes in PAR and seasonal changes in precipitation were integrated annually to determine yearly sums of modeled NEE. Therefore, interannual variability was controlled by precipitation and sunlight. Phenology did not vary interannually; therefore it did not affect interannual variability in modeled NEE.
We disaggregated the meteorological influence on NEE, represented by our
model (Eq. 2), from long-term changes in forests' ecological efficiency by
examining the annually integrated hourly model residuals. In 2002, there was
a total of 1.2 MgC ha
On average, period 2 saw a 20 % increase in annual precipitation relative to period 1. Abbreviated dry-season lengths and lack of radiation from increased cloudiness in period 2 resulted in less modeled net uptake relative to period 1.
We partitioned observed and modeled NEE into RE and GEE. Interannual
variations in RE were accurately represented as changes in wet- and dry-season
length (Fig. S1 in the Supplement). The range in annual residual RE is
therefore small compared to that of annual residual GEE (Fig. 7b). In 2002,
mean model GEE had 0.85
Hourly changes in NEE were due predominantly to changes in sunlight (Fig. 4). Phenology only played a small role in modeled hourly variability, improving the fit of our model by only 1 % relative to a model that only used meteorology and lacked a phenology parameterization (Table 2).
Seasonal changes, on the other hand, were due to a combination of sunlight,
rainfall inputs, and phenology (Fig. 6). The model parameterization contained
a seasonal decrease in respiration (
VPD and diffuse radiation do not explain significant additional variance in NEE relative to our model (Eq. 2) at monthly timescales (Tables 1, S1). The relative importance of phenology at monthly timescales, compared to that of VPD and diffuse radiation, is consistent with other findings regarding GEP at our research site: moving from finer to coarser temporal resolution, the influence of exogenous meteorology becomes outweighed by that of exogenous ecosystem changes such as those in phenology (Wu et al., 2017).
Seasonal changes in LUE are well explained by canopy leaf age and demography both at this site and at a comparatively wetter forest site in Manaus, showing good agreement with a model informed by camera and trap-based observations of leaf flushing and shedding (Wu et al., 2016). Our single midyear parameter simplistically upshifts the trough in a more continuous seasonal oscillation between low and high LUE (Fig. 5) because we lacked independent variables explaining the seasonal oscillation.
The seasonally asynchronous nature of phenology-mediated LUE establishes a middle ground in debates over whether the eastern Amazon canopy is enhanced or “greens up” during the dry season (Huete et al., 2006; Myneni et al, 2007; Samanta et al., 2012; Morton et al., 2014; Bi et al., 2015; Guan et al., 2015; Saleska et al., 2016). Changes to the canopy's LUE do indeed occur, but not synchronously with the dry season at our site (Fig. 5). Evidence from previous studies at the TNF suggests that changes in phenological LUE result from carbon allocation shifting from stem allocation to the turnover and production of new leaves (Goulden et al., 2004), supporting the prevailing hypothesis that tropical trees have been selected to coordinate new leaf production ahead of dry-season peaks of irradiance (Wright and van Schaik, 1994). The GEP seasonal cycles at additional evergreen Amazon forest sites are not well described by sunlight alone (Restrepo-Coupe et al., 2013). Averaging over seasonal windows is therefore likely to miss a potential inter-seasonal depletion and enhancement of canopy LUE if additional regions of evergreen Amazon forest similarly exhibit seasonally asynchronous phenology.
Interannual variation in phenology is represented mechanistically in phenology and LUE sub-models, which have been optimized using km67 eddy flux data but nonetheless fail to reproduce the observed midyear GEP decrease at this site. Kim et al. (2012) present a light-triggered phenology scheme, which assumes higher modeled leaf turnover rates and higher maximum leaf photosynthesis during the dry season, and hence produced higher dry-season GEP. Their model produced leaf-flushing rates that lagged behind observations and contradicted observations that light-controlled GEP decreases midyear at km67 (Fig. 5). Another phenology scheme has been developed by De Weirdt et al. (2012), which attributes excess leaf allocation to the turnover of new, more efficient leaves but nevertheless overpredicted midyear GEP at km67 relative to their prior model. Wu et al. (2016), on the other hand, successfully represent the GEP seasonal cycle using their model of leaf age and demography but relied on observations of canopy leaf fluxes. Their model, however, does not provide a meteorologically triggered mechanism for seasonal leaf shedding and flushing. Therefore, until such a trigger can be identified, models that mechanistically represent phenology are primed to make erroneous predictions about the interannual and long-term consequences of changing seasonal lengths for the Amazon carbon balance.
Annual totals of measured NEE exhibited an unpredicted trend: despite previous hypotheses that the years after period 1 would continue to trend downward towards more uptake (Hutyra et al., 2007; Pyle et al., 2008), the ecosystem returned to a moderate carbon source in all 3 years of period 2 (Fig. 1). We examined whether the reversal of the period 1 trend throughout period 2 could be explained by exogenous changes in climate or an endogenous biophysical change. We developed the model selection framework to partition these two sources of variability.
Our model represented NEE well across a variety of timescales (Figs. 4, 5,
7). On yearly timescales, interannual differences in NEE
The overall magnitude of the carbon source/sink, however, was highly
sensitive to the choice of
We examined the possibility that a systematically high bias in 2002 PAR could
result in an overprediction of 2002 GEP and erroneously cause a positive
2002 residual. We found that PAR was appropriately drift-corrected by
corroboration with
Additional meteorological variables such as the VPD and diffuse radiation did not appear to explain residual NEE in 2002. A model including these variables did not explain the positive NEE and GEE anomaly in 2002 (Fig. S3). The annual means of both VPD and CI in 2002 lay within their decadal range, making high VPD or low diffuse radiation an unlikely explanation for low photosynthetic uptake. These meteorological factors did not appear to significantly impact interannual changes in NEE, consistent with previous findings regarding GEP at this site (Wu et al., 2017).
We cannot rule out that the 2002 source may be a measurement artifact, caused for example by disturbance following tower construction. We note, however, that tower construction was completed almost a year before the measurements we used, with preliminary data collection occurring during 2001 (Saleska et al., 2003). We examine the possibility that 1998 drought-based disturbance impacted forest GEP through 2002 in Sect. 4.2.2.
Our multiple records of meteorology adjacent to our research site (Fig. 2),
which we used to inform our simple model of NEE, can also shed light on the
larger discussion of recent droughts in the Amazon. Previous reports of
21st-century droughts in this region are inconsistent. For the 2010 Amazon
drought, Lewis et al. (2011) show that water deficits were minimal in
the eastern Amazon region, consistent with our findings. However, Doughty et
al. (2015) report ubiquitous detrimental effects of the 2010 drought
basin-wide, including a
Our model overpredicted photosynthetic uptake in 2002 but predicted RE well
(Figs. 7b, S1), suggesting that a drought disturbance in 1998 persistently
affected forest GEP, not RE, through 2002. These findings contradict a
previously established hypothesis that legacy effects of a prior drought
disturbance increased RE in 2002 via increased CWD respiration
(
CWD measurements from the km67 site suggest that there was major disturbance
before measurements of
Identifying the cause of the reduced 2002 GEP is beyond the scope of this
statistical modeling study. It is possible that the 1997–1998 El Niño
drought not only killed entire trees but also damaged living trees through
hydraulic failure and partial limb death, affecting canopy photosynthesis for
subsequent years. An analysis of over 1000 temperate forest census sites
suggests that recovery of live tree biomass accumulation may be delayed by up
to 4 years after drought (Anderegg et al., 2015). Following the 2005 and
2010 western droughts, findings from forest inventories (Brienen et al.,
2015) and remote sensing (Saatchi et al., 2013) suggested that legacy
effects from tropical forest droughts can also persist for 4 years or
more. Drought cavitation due to xylem embolisms reduces hydraulic
conductivity, leading to whole-tree mortality (Choat et al., 2012), initiating
a classic disturbance-recovery scenario in which felled trees generate canopy
gaps for early successional seedlings and saplings to immediately capitalize
on newly available light, causing
The decade-long record of eddy flux at km67 in the TNF demonstrated
unpredicted trends in 7.5 years of measured NEE. Our simple, low-parameter
empirical model could represent interannual differences in NEE as integrated
continuous responses to changes in meteorology, with the exception of the first
year, suggesting that increased moisture and decreased sunlight, not an
interim disturbance, were responsible for the elevated period 2 carbon
source. Although overall magnitude of the carbon source/sink was highly
sensitive to the specific choice of
Our model represented a seasonal midyear decline in GEP. Our representation of phenology follows set calendar dates and cannot distinguish between various hypotheses concerning the environmental trigger for seasonal leaf shedding and flushing. DVGMs and other numerical simulation ecosystem models that represent phenology as a response to light-triggered leaf flushing or root water constraints do not tend to represent the seasonal cycle of GEP accurately and are therefore in danger of overpredicting the future response of photosynthesis to longer dry seasons resulting from climate change.
Our finding that reduced photosynthesis, not increased respiration, contributed to the high NEE source in 2002 modifies the previous hypothesis that the 1997–1998 El Niño drought disturbance affected NEE via respiration. Our findings support a corollary hypothesis that partial drought-induced damage to still-living trees can impact whole-ecosystem photosynthesis adversely for multiple years, which is consistent with findings from regional- and global-scale forest biometric studies (Anderegg et al., 2015; Brienen et al., 2015). In order to understand how drought disturbance uniquely impacts forest recovery, observational studies and plot-based manipulation experiments are needed in conjunction with models. Such future research is needed to determine the return times for droughts at which persistent forest biomass loss and collapse may occur.
The eddy flux data used in this study are available online
via the Lawrence Berkeley National Laboratory (LBNL) AmeriFlux network
database at
The supplement related to this article is available online at:
MNH and SCW designed the study. MNH performed the statistical analysis. ML and JW conducted post-processing of meteorology data. MNS, LFA, and PBC conducted forest biometric surveys and provided data. RT, RdS, and DRF provided auxiliary meteorological data. NRC, LRH, BD, JWM, KTW, SRS, and SWC conducted measurements and eddy covariance data pre-processing. All authors contributed to writing the manuscript.
The authors declare that they have no conflict of interest.
This work was supported by funding from a National Science Foundation PIRE fellowship (OISE 0730305) and a US Department of Energy grant (DE-SC0008311). Edited by: Paul Stoy Reviewed by: two anonymous referees