Amazon Forests: Unexpected Seasonal Fluxes and Disturbance-Induced Losses

The net ecosystem exchange of carbon dioxide was measured by eddy covariance methods for 3 years in two old-growth forest sites near Santare´m, Brazil. Carbon was lost in the wet season and gained in the dry season, which was opposite to the seasonal cycles of both tree growth and model predictions. The 3-year average carbon loss was 1.3 (conﬁdence interval: 0.0 to 2.0) megagrams of carbon per hectare per year. Biometric observations conﬁrmed the net loss but imply that it is a transient effect of recent disturbance superimposed on long-term balance. Given that episodic disturbances are characteristic of old-growth forests, it is likely that carbon sequestration is lower than has been inferred from recent eddy covariance studies at undisturbed sites. The terrestrial biosphere currently sequesters 20 to 30% of global anthropogenic CO 2 emis-sions ( 1 , 2 ). Amazoˆnia has been suggested to be a major contributor to observed interannual variations in this sink ( 3 ). The underly-ing causes are unclear. One hypothesis suggests that C uptake in old-growth Amazonian forests may be stimulated by rising atmospheric CO 2 ( 4 – 7 ). Also, water availability limits tree growth, potentially reducing C uptake in dry periods ( 4 , 8 ). Thus, in El Nin˜o years, when much of Amazoˆnia experiences low rainfall ( 9 ), old-growth forests could release CO 2 ( 10 ), a notable problem if El Nin˜o frequency increases in future climates


Carbon in Amazon Forests: Unexpected Seasonal Fluxes and Disturbance-Induced Losses
The net ecosystem exchange of carbon dioxide was measured by eddy covariance methods for 3 years in two old-growth forest sites near Santarém, Brazil. Carbon was lost in the wet season and gained in the dry season, which was opposite to the seasonal cycles of both tree growth and model predictions. The 3-year average carbon loss was 1.3 (confidence interval: 0.0 to 2.0) megagrams of carbon per hectare per year. Biometric observations confirmed the net loss but imply that it is a transient effect of recent disturbance superimposed on long-term balance. Given that episodic disturbances are characteristic of oldgrowth forests, it is likely that carbon sequestration is lower than has been inferred from recent eddy covariance studies at undisturbed sites.
The terrestrial biosphere currently sequesters 20 to 30% of global anthropogenic CO 2 emissions (1,2). Amazônia has been suggested to be a major contributor to observed interan-nual variations in this sink (3). The underlying causes are unclear. One hypothesis suggests that C uptake in old-growth Amazonian forests may be stimulated by rising atmospheric CO 2 (4-7). Also, water availability limits tree growth, potentially reducing C uptake in dry periods (4,8). Thus, in El Niño years, when much of Amazônia experiences low rainfall (9), old-growth forests could release CO 2 (10), a notable problem if El Niño frequency increases in future climates (11). Evidence for uptake of 0.1 to 0.5 Mg C ha Ϫ1 year Ϫ1 of CO 2 has been inferred from long-term tree data on forest plots (12,13), which is consistent with hypothetical CO 2 fertilization (4) but remains controversial (14,15). A surprisingly large uptake (1.0 to 5.9 Mg C ha Ϫ1 year Ϫ1 ) (16)(17)(18)(19) was inferred from several eddy covariance studies in Amazonian forests, much greater than the accumulation rates pre-dicted for CO 2 enrichment or observed in trees (12) and soils (20). If it were representative of Amazônia (16,21,22), the associated sink in undisturbed forests would be 0.5 to 3.0 Pg C year Ϫ1 , comparable to global terrestrial C uptake and inconsistent with atmospheric inversion studies (3). It is possible that loss processes [deforestation (23) and evasion of CO 2 from rivers (24)] are larger than previously thought, or flux sites could have selectively avoided locations affected by disturbance, missing associated C losses (25,26). Alternatively, the flux studies may have overestimated uptake.
We measured C exchange for two oldgrowth Amazonian forests in the Tapajós National Forest near km 67 (02°51ЈS, 54°58ЈW) (27) and km 83 (03°03ЈS, 54°56ЈW) (28-30) of the Santarém-Cuiabá highway, just south of Santarém, Pará, Brazil, using eddy covariance methods (31). The site receives 1920 mm year Ϫ1 of rainfall and has a 7-month wet season (months with Ͼ100 mm of precipitation), representing the 25th to 30th percentile of each metric among Amazon forests ( fig. S2). Moist tropical forests like the Tapajós, which are drier than the more extensive wet forests, may be models for the future Amazon (which is predicted to become drier with climate change) (10,11).
We simultaneously tracked changes in the forest structure and monitored C that was stored in live wood (resulting from tree recruitment, growth, and mortality) and dead wood (resulting from tree mortality inputs and respiration losses) by means of biometric observations at km 67 (32). Long-term rates of C storage were determined by comparing inventories from 1984 and 2000 at km 83 (33). Thus, we independently checked C balances that had been inferred from flux data and studied the processes mediating C uptake and loss.
Eddy covariance data from the two sites together cover 1 July 2000 to 1 August 2003 nearly continuously. Hourly data for net ecosystem exchange (NEE) (34) were obtained for 83% of 31,000 hours, including above-canopy fluxes and CO 2 storage within the canopy. Gaps were caused by calibrations, intense precipitation events, maintenance, and equipment faults. The two sites are well-suited for eddy covariance measurements, have similar weather (Fig.  1A), and exhibit a similar dependence of NEE on sunlight (Fig. 1B). NEE was underestimated during calm nighttime conditions at both sites (Fig. 1C); we applied corrections using procedures developed for mid-latitude forests and adapted, with rigorous testing, for Tapajós (35). Cumulative values for NEE were indistinguishable after applying this correction ( fig. S7).
Carbon was taken up in the dry season and released in the wet season ( Fig. 2A), opposite to seasonal variations for non-El Niño years in two ecosystem models (36, 37) ( Fig. 2A) and inverse to the seasonality observed for tree growth (Fig. 2C). Seasonal precipitation patterns were very similar in models and observa-tions ( Fig. 2D), indicating that the discrepancy is likely mechanistic rather than reflective of differences in environmental forcing.
Seasonal variations in NEE reflect the influence of moisture and sunlight on photosynthesis and respiration. Photosynthesis [gross ecosystem production (GEP) (38)] responded weakly to seasonal changes in precipitation, but respiration responded strongly, with the peak wet-season month (March) 40% higher than the minimum dryseason month (November) (Fig. 2B) (39). The models, however, predicted that GEP was water-limited and thus more sensitive to precipitation than respiration (36, 37 ). Trees in the Tapajó s access deep soil water during dry periods and show little evidence of stress (29,40); rates for wood increment decline in the dry season, but remain substantial and notably increase just before rains return (Fig. 2C), suggesting an adaptive mechanism rather than a response to seasonal forcing (30). In contrast, respiration in soils, litter, and coarse woody debris (CWD) is concentrated near the surface and consequently inhibited by desiccation during the dry season ( Fig. 2C) (30).
Seasonal variation of NEE was stronger in Tapajós (700 kg C ha Ϫ1 month Ϫ1 minimum to maximum) ( Fig. 2A) than in forests near Manaus to the west (17, 41) (ϳ150 kg C ha Ϫ1 month Ϫ1 minimum to maximum) (41), and it had the opposite phase. Little seasonality was observed at Caxiuanã to the east (19). These differences derive from the longer dry seasons at Tapajós (42), which desiccate detrital material and reduce respiration, a pattern also observed in temperate forests (43) and mesquite woodlands (44) during dry periods.
The seasonality of C exchange has two important implications. First, the same model constructs (such as tree rooting depth and surface-litter layer thickness) affect seasonal and interannual variations of C exchange. The models predict an annual net C loss during El Niño droughts (4,8), but if a drought increases tree mortality (45,46) while inhibiting respiration, C losses are more likely to occur after the rains return and the wood decomposition accelerates.
Second, seasonal variations are important in global inversion studies. In the wet season, deep convection transports surface air to the free troposphere, diluting the effect of ecosystem fluxes on the near-surface CO 2 observations used in global inversions (47) and masking any net release of CO 2 similar to that observed here. In contrast, dry season fluxes are manifest in nearsurface observations used in global inversions. Thus, if the higher amplitude reversed phase of seasonal C exchange in drier forests such as the Tapajós dominates seasonal variation for Amazônia, inversions incorporating biosphere models that predict the wrong seasonal phase will overestimate Amazonian C uptake.
We derived annual C balances for Tapajós forests by integrating NEE data over the observation period. Cumulative losses to the Uncertainties arise from two sources: corrections for nighttime underestimation of flux under conditions of low turbulence (represented by the range in parentheses) (35) and conventional statistical uncertainty [repre-sented as Ϯ95% confidence interval (CI)]. Nighttime corrections added 3.8 to 5.8 Mg C ha Ϫ1 year Ϫ1 at km 83, and 2.0 to 3.8 Mg C ha Ϫ1 year Ϫ1 at km 67, with bounds based on a conservative analysis of the dependence of apparent respiration rates on turbulence statistics (35). The greater prevalence of windy nights at km 67 resulted in smaller corrections for that site. Aggregating the different kinds of error and the errors across site years gave a mean loss of 1.3 (combined confidence interval: 0.0 to 2.3) Mg C ha Ϫ1 year Ϫ1 (Fig. 3B).
Biometric observations at km 67 over an overlapping 2-year period also indicated a net loss of C (2.0 Ϯ 1.6 Mg C ha Ϫ1 year Ϫ1 ) (table S1) from above-and belowground pools of live and dead wood, with substantial contributions from respiration by notably large stocks of CWD. Generous allowance for C allocation to roots and to soil, within constraints from direct measurements of belowground root mass near the km 67 tower (48) and empirical studies of tropical soil C turnover (20), would not change the observation of net C loss (table S1 and Fig. 3B). Thus, biometric data support the carbon budget derived from eddy covariance data. This is the first eddy covariance study in old-growth Amazonian rain forests that fails to indicate C sequestration. Had we not corrected for evident underestimation of nighttime fluxes, as has been done at most eddy flux sites worldwide (49), our measurements would indicate large uptake, comparable to previous Amazonian studies in which this correction was not applied (18,19). Our algorithm for correcting nighttime fluxes (35) is supported by biometric data, convergence of eddy flux results for the two towers using independent corrections, and entirely independent estimates of nighttime fluxes at km 67 using radon as a tracer (50). Notable uncertainties are associated with the nighttime flux correction, reflecting imperfect understanding of the details of canopy transport (51)(52)(53), but the available evidence (49) strongly suggests that uncorrected fluxes will overestimate uptake at most sites.
The respiration losses from the large stock of dead wood at km 67 exceeded the uptake of C by live biomass, despite high growth rates (54), reflecting disequilibrium of the vegetation assemblage (table S1A). Notably, the C was taken up mostly by small trees (table S1A), which also increased in stem density (4.8% recruitment into the Ͼ10-cm size class) (55,56). But no net long-term change in aboveground biomass was observed over the course of 16 years at km 83 (table S1B). We infer that the forest at km 67 is responding to a recent episode of high mortality, possibly triggered by drought associated with the strong El Niño events of the   (table S1). Potential bias errors (boxes), and 95% CI due to sampling uncertainty (error bars) are shown.
1990s (57). Dead trees were deposited on the forest floor, simultaneously opening canopy gaps that drive the current growth and recruitment of small trees.
The observations suggest that present C losses at km 67 are part of a cycle of disturbance and recovery superimposed on longterm balance and that substantial transient fluxes lag mortality events. Disturbance and recovery cycles, characterized by high C loss in the decade after disturbance followed by many decades of moderate uptake, may be inferred from the extensive ecological literature on disturbance dynamics (58) and have been predicted for old-growth Amazonian forests (25); the current study empirically quantifies the associated transient fluxes.
Episodic natural disturbances complicate estimates of regional C balance. If disturbance occurs on spatial scales larger than flux tower footprints or ecological plots, but smaller than landscape scales, most areas will be "recently undisturbed" and will take up C. But relatively rare "recently disturbed" areas will release C at rates sufficient to offset this uptake. Because study sites tend to avoid recent disturbance, regional C sequestration tends to be overestimated (59). Evidently, the disturbance history of study sites and the distribution of disturbance across the landscape (25) must be known to assess largescale influences on the C budget, such as response to CO 2 fertilization.
Our results focus attention on the importance of respiration and disturbance for understanding the present and future C balance of Amazonian forests. The unexpected seasonality of C exchange, dominated by moisture effects on respiration, calls into question hypothesized C losses during El Niño-Southern Oscillation droughts and highlights the importance of using accurate seasonal fluxes for inverse atmospheric model studies. The observation of disturbance-induced loss implies that large-scale C balance depends critically on large-scale disturbance dynamics. Consequently, if study sites avoid recent disturbance, extrapolations to the whole basin will consistently overestimate C sequestration by Amazonian forests. 31. Materials and methods are available as supporting material on Science Online. 32. For the km 67 biometry study (27,49) , we surveyed 20 ha of forest near the eddy-flux tower in July 1999 and resurveyed the same area in July 2001 (obtaining estimates of recruitment, growth, and mortality), including stocks of CWD. CWD respiration was calculated by applying density-specific respiration rates measured in the Amazon rainforest near Manaus to measured decay class-specific CWD mass at km 67 (27). Sampling uncertainty in all estimates was quantified using bootstrap analyses, with 95% CIs reported. 33. The km 83 long-term biometry study (28,49) compared a 1984 inventory of large trees (diameter at breast height Ͼ 55 cm) in 48 ha around the eddy-flux tower location with an inventory of the same area in 2000 to analyze the change in large tree biomass. The measured small tree:large tree biomass ratio was used to estimate changes in smaller trees. Errors were quantified using a sensitivity analysis that bracketed a highly conservative range of plausible uncertainty. Flux contributions resulting from a change in CWD stocks over the period, which were not measured in the km 83 study, are likely much less important than in the km 67 study because of the long time between surveys (about twice the turnover time of CWD). 34. The NEE is the biotic flux from the forest, calculated as the sum of the eddy flux at the top of the canopy and the storage flux (the change in CO 2 stored in the canopy below the sensor). 35. We corrected for evident underestimation of nighttime NEE (Fig. 1C and figs. S4 and S5) by filtering out measurements taken during periods of weak mixing, indicated when u* is below a given threshold (U* thresh ), and replacing them with an estimate based on nearby NEE obtained during well-mixed conditions (when u* Ͼ U* thresh ). Analysis (49) (figs. S4 and S5) indicated that a best-estimate U* thresh of ϳ0.22 m sec Ϫ1 selects nighttime NEE measurements representative of total ecosystem respiration (these are 40 and 20% of total nighttime data at km 67 and km 83, respectively). We estimated a conservative uncertainty associated with this nighttime correction by calculating a low estimate for C balance using U* thresh ϭ 0.17 m s Ϫ1 and a high estimate using U* thresh ϭ 0.30 m s Ϫ1 . We believe that the probability of the true balance falling outside of this range is very small, but it is not rigorously quantifiable (49