5.1. Site Differences in Dynamics and Structure
 The gross fluxes of growth and CWD respiration were larger in the TNF than the BDFFP plots, indicating higher rates of turnover and a more dynamic forest in the Tapajós. More carbon was taken up through growth of live trees in the TNF, particularly in the smallest size classes (Figure 4), despite the slightly lower overall stem counts (Table 2). Malhi et al.  and Baker et al.  suggested that on a continental scale, forests in the western Amazon are more dynamic than central and eastern Amazon forests, where mean wood density is inversely correlated with forest dynamics [Malhi et al., 2004]. We compare sites on a smaller scale (within the central to eastern Amazon) and with very similar mean wood density (0.66 g•cm−3 for TNF and 0.69 g•cm−3 for BDFFP). The site differences in dynamism we found on this smaller scale may still fit in with the continental trends described by Malhi et al. , particularly where some of the dynamism observed in the TNF may reflect past disturbance.
 The sites also showed stand structural differences in the distribution of live trees consistent with the observed differences in growth and stand dynamics. The BDFFP site had higher stem densities and greater biomass in small and middle size classes. We have observed that small trees in the TNF accumulated carbon more rapidly than at BDFFP: Trees in the 20 – 35 cm size class added an average of 0.047 ± 0.0057 MgC per MgC biomass, while the same sized trees in the BDFFP plots added 0.019 ± 0.0012 MgC per MgC biomass. These slower growth rates could lead to differences in age structure, where trees of the same size are older in Manaus: Vieira et al.  found that tree ages average ∼380 years in Manaus, but only average ∼200–280 years in the TNF. This suggests that the greater dynamism at the TNF may be a long-established and ongoing phenomenon.
 Although the two sites show comparable overall mortality in terms of carbon lost from the pool of live biomass (Table 3), the total proportion of carbon in standing biomass lost through mortality is about 25% higher in the TNF. This site difference is reflected in the higher annualized mortality rates in the TNF (1.9% of stems in the TNF, 1.6% of stems in BDFFP), that suggest higher turnover rates at TNF. Notably, the TNF showed higher stem mortality in the smallest size classes on both an absolute and per stem basis (Figure 7). The higher stem mortality in larger size classes in the BDFFP (Figure 7) leads to the equal estimates of overall carbon flux in mortality due to the weight of large trees in carbon estimates. In the TNF, higher mortality in the smallest trees may partially account for lower live tree stem counts and lower biomass in the small and middle size classes at the TNF. Higher mortality in the smallest size classes could be an underlying cause for the higher growth rates observed in the smallest size classes of TNF trees, and ultimately a driver of the higher overall turnover rates in the TNF, though because mortality is not are well constrained as other quantities, this possibility remains speculative.
 The higher mortality in small trees (<35 cm DBH) in the TNF could reflect the dominance of Coussarea racemosa A.Rich.ex.DC., a small-stature understory tree, which makes up 24% of trees <30 cm DBH at km 67, compared to <0.2% of trees <30 cm DBH in the BDFFP. C.racemosa tends to reach maximum size of ∼30 cm DBH, and showed a 2.8% annual mortality from 2001 to 2005, 40% higher than the km 67 site average of ∼2%a−1. Understory trees typically have higher mortality rates than canopy trees [Nascimento et al., 2005], but the prevalence of C.racemosa in particular might bias stem turnover for the smallest size classes at TNF.
 The forest in Manaus has shown evidence for seasonal water limitation in forest carbon flux [Malhi et al., 2002; Araújo et al., 2002], while the TNF has shown no signs of similar dry-season C uptake or evapotranspiration reductions due to water stress [Hutyra et al., 2007; Bruno et al., 2006] despite a longer dry season and lower annual rainfall at TNF. Continued dry season C uptake has been attributed to deep roots accessing deep water during dry season and thus maintaining a green and functioning canopy [Nepstad et al., 1994]. Canopy carbon uptake at TNF increases late in the dry season [Hutyra et al., 2007; Goulden et al., 2004], possibly as a result of greater light availability and patterns in forest phenology [Huete et al., 2006; Saleska at al., 2007]. The combination of adequate water availability and a long, sunny dry season could be a factor contributing to the greater gross uptake in live biomass observed for the TNF, possibly driving higher mortality and turnover due to increased plant competition.
 Differences between TNF and BDFFP in size class structure appear to reflect differences in age structure, which in turn could result from differences in growth rates. It could be argued that the longer dry season at the TNF leads to greater carbon uptake in live trees, in turn leading to the stand structural differences observed in this study. Though the causality cannot be unambiguously attributed, due to the short period of study, the consistency among greater gross fluxes, greater carbon uptake in small trees, greater small tree mortality, and differences in stand and age distributions clearly indicate that the TNF is currently a more dynamic forest than the BDFFP.
5.2. Sampling Limitations and Spatial Pattern
 Our analysis of subplots showed the sampling of some demographic components of carbon flux was not quite adequate. Specifically, mortality and recruitment were not well sampled for either site (Figure 8, BDFFP: coefficient of variation for recruitment: 22%, for mortality: 21%; TNF: coefficient of variation for recruitment: 26%, for mortality: 24%) Much of the uncertainty around these demographic parameters in the TNF could have resulted from the subsampling of trees 10–35 cm DBH, the only size classes where recruitment occurs and important sizes classes for mortality. Notably, the much larger sampling area for small trees in the BDFFP (more than double, with 20 ha versus ∼8.5 ha in the TNF) reduced the coefficient of variation by only a few percentage points (3–4%), suggesting the inherent difficulty in adequately sampling these dynamic demographic quantities. The episodic nature of mortality and the influence of individual large tree deaths on individual subplots likely complicate such estimates.
 Though mortality and recruitment fluxes may not be as well constrained as other fluxes in this study, they exert less influence on the overall estimates of net flux of carbon in forest biomass. With mortality, the flux of biomass from the pool of live biomass to CWD (mortality) does not factor into the net flux (growth + recruitment – CWD respiration). With recruitment, the flux of carbon into the pool of live biomass due to recruitment is ∼10% of the total flux of carbon into biomass for both sites, with the remaining ∼90% due to growth of live trees, which is well sampled.
 Another potential problem in the subplots analysis for the TNF is that the locations of individual 1 ha subplots in the TNF were constrained by the configuration of randomly located transects, unlike the stratified random placement of the individual BDFFP plots. The spatial constraints on TNF subplots could lead to spatial auto-correlation; however, our analysis of distance and similarity showed no signs of spatial trends in the measured forest quantities: plots located in close proximity were not more similar. Moreover, if TNF subplots were auto-correlated, they would be more likely to show lower variability: the reverse of what we found.
 Subplots were also used for spatial analysis which showed no pattern of variation at the scale of 1–50 km: Plots located in close proximity showed the same range of similarities (mean squared difference) as plots separated by larger distances. While there may be fine scale spatial variability in carbon dynamics (e.g., gap dynamics), these processes do not appear to influence variability at the landscape scale; i.e., there are not some areas more prone to tree falls or other sources of fine scale variability at the landscape scale (∼1 to ∼50 km). This consistency within both sites suggests that a localized patch of sampling (of adequate size) can provide a reasonable estimate of the larger matrix of forest on the 50 km scale; i.e., the research plots at km 67 do characterize biomass for terra firme forest in the TNF in general, and an eddy flux tower “foot print” may be considered representative of the surrounding forest. It is important to note, however, that by design, plots in this study did not sample across obvious topological or edaphic gradients, where landscape scale spatial variation might be apparent, and thus an adequate local sample may not reflect the surrounding forest where there are notable shifts in forest type or conditions.
 The differences between the two sites, TNF and BDFFP were consistent, possibly due in part to low internal variability within each site discussed above. This spatial consistency indicates that the factors regulating the site differences act across the landscape scale and over sufficiently long time periods to shape both stand structural qualities and dynamics. Such forest-wide causes of site differences may include climatological, edaphic, and/or meteorological conditions. In the BDFFP area, the poor quality of weathered, acidic soils, or the dissected terrain might limit tree growth rates and favor denser, slower growing species in the forest outside of Manaus [Nascimento et al., 2005; Vieira et al., 2005]. While the larger gross fluxes in live biomass growth and CWD decomposition in the TNF may reflect influences of soils, climate regime, topology, they are most clearly linked to disturbance and recovery. The strong signal of disturbance recovery and disequilibrium detected in the TNF defines the carbon dynamics observed there and cannot be separated from other, long-term influences on carbon balance.
5.3. Disequilibria, Disturbance, and the Importance of CWD
 At both sites, CWD plays a pivotal role in ecosystem carbon dynamics and is a key indicator of underlying causes of site differences. CWD stocks at both the TNF and BDFFP led to a significant shift in the observed net carbon balance: without consideration of the CWD pools the data show moderate (∼0.8 MgC•ha−1•a−1) uptake in the TNF and small uptake (∼0.3 MgC•ha−1•a−1) in the BDFFP plots (Figure 6, left hand bars only), leading to the conclusion that both sites are net storing carbon in biomass [cf. Vieira et al., 2004; Baker et al., 2004]. This result would seem to support the conclusion that sequestration of carbon is widespread in Amazonian forests [cf. Baker et al., 2004; Phillips et al., 2002]. But inclusion of the CWD carbon flux turned the TNF into a significant net source (−1.25 ± 0.45 MgC•ha−1•a−1) and changed the carbon balance from a small sink to carbon neutral (0.18 ± 0.29 MgC•ha−1•a−1) for the BDFFP. Moreover, with CWD included, land-based measurements more closely matched the tower-based NEE measurements in TNF, which indicated a small source transitioning to neutral between 2001 and 2005 [Hutyra et al., 2008, Table 3].
 Site comparisons of CWD could be complicated by the differences in wood density and decay classifications. In the case of this study, the decay classes used in the BDFFP study did not precisely match those used in the TNF study, but they were consistent. For example, the density values of the “sound” decay class used in the BDFFP sites (0.69 Mg•m−3), falls between the density values used in the TNF for decay classes 1 and 2 (0.60, and 0.70 Mg•m−3, respectively). Site-specific densities applied for each site may be the most appropriate choice where CWD wood density varies spatially [Chao et al., 2008]. Though the compatibility of the size class and densities applied at each site cannot be tested with the data available, the efficacy of the CWD methods for each site is supported by independent estimates. The BDFFP estimate (∼16 MgC•ha−1) concurs with that of Summers  (∼15 MgC•ha−1) who directly estimated CWD in 3, 1 ha plots only 20 km south of the BDFFP area. Likewise, the TNF CWD estimate (∼40 MgC•ha−1) is similar to the estimate provided by Palace et al.  (∼41 MgC•ha−1) for trees ≥ 10 cm DBH in unlogged forest near km 83 in the TNF.
 Diagnosis of the factors controlling CWD cannot, unfortunately, be inferred just from measurements of the CWD pool. The CWD distribution between standing and downed trees, which might be expected to provide an indication of disturbance mechanisms (blow-downs versus standing dead), was variable at both sites over a similar range (10–30%), likely a reflection of the high variability of both mortality and disturbance in space, time and intensity. Rice et al.  found standing to be 18% (8.6 MgC·ha−1) of CWD; Clark et al.  found 12% (3.1 MgC·ha−1); Palace et al.  report standing CWD ranged from 12 to 17% of total, depending on site and treatment (some sites had experienced reduced impact logging). This consistent variability across sites indicates that the proportion of standing CWD is not a good indicator of site differences or causes of mortality.
 The mortality rates measured over the study period were also comparable in terms of carbon lost from the pool of live biomass, though the TNF mortality was proportionally higher by stem count (1.9% in the TNF, 1.6% in BDFFP). The TNF mortality was not high enough, however, to account for the much larger CWD pool at the TNF. The very large stock of CWD distributed throughout the TNF must have resulted from excessive (higher than current observations) mortality occurring at our measurement sites in the TNF before the start of the study. Strong drought conditions were measured in the region of the TNF during the 1997–98 ENSO event, and Rice et al.  proposed that this drought could have produced a mortality pulse leading to the imbalance in CWD observed at km 67 in 1999–2001. In the BDFFP plots, Williamson et al.  documented increased mortality rates during the 1997–1998 ENSO, though the CWD pools in the BDFFP plots do not show excess CWD or a residual imbalance during the time period of this study. However, the magnitude of the mortality increase reported for BDFFP plots (from 1.12% − 1.91%) was considerably smaller than other reported ENSO-related mortality increases (2–3% [Condit et al., 1995]) [Leighton and Wirawan, 1986]. See review in Clark [2004b, Table 1]. While there are no tree mortality data for the TNF during the time period of the ENSO, Rice et al.  estimated that mortality of roughly 5% of the standing biomass, persisting for 5 years, would be required to yield the excess CWD found in the TNF in 2001.
 Another possible source for excessive CWD could include nonfatal limb and branch falls which contribute to the CWD, but would not be accounted for with mortality-derived estimates of CWD inputs [Rice et al., 2004; Chambers et al., 2001a]. Palace et al.  found that using mortality rates to estimate coarse wood production underestimates CWD production by 30–50%, based on direct measurements of coarse wood inputs in the TNF. In this study, CWD values were limited to ≥10 cm, a minimum size that may exclude many limb falls. Palace et al.  reported a coarse wood production value for large pieces (≥10 cm) of 4.7 Mg•ha−1•a−1, slightly lower than our mortality derived estimate of 5.2 Mg•ha−1•a−1, suggesting we have not underestimated the inputs into our pool of CWD.
 Our linearized box model, based on Figure 2, further illuminates disequilibria of the biomass pools at TNF. Figure 2 provides a quantitative visualization of the mass balances of the major stocks of organic matter at the two sites. At TNF, only the smallest size class was approximately in steady state, with total inputs (recruitment + growth in class = 1.27 MgC•ha−1•a−1) approximately balancing outputs (export + mortality = 1.19 MgC•ha−1•a−1; net flux 0.08 ± 0.08 MgC•ha−1•a−1; see Table 4). The second size class experienced strong growth, but exported more biomass to the next larger class than it retained. The two largest size classes grew at significant rates, and the CWD pool declined rapidly (∼6%•a−1). In contrast, at BDFFP, the live biomass stocks were each close to being in balance (within error), and the CWD pool may have been accumulating at a slow rate. The apparent imbalances in each size class indicate that the structure of the forest at the TNF was shifting significantly at the time of measurement, while the BDFFP was much closer to a steady condition. The differences shown for the net fluxes of the major pools belie the similarity that might have been expected given the comparable totals of biomass.
Table 4. Values for Carbon Pool and Fluxes by Size Class Used to Build Box Model for the TNF and the BDFFP Sitesa
|Class||Mass (MgC•ha−1)||Growth Within Size Class (MgC•ha−1•a−1)||Growth Into Size Class (MgC•ha−1•a−1)||Mortality (MgC•ha−1•a−1)||Respiration Rate (a−1)||Net Change (MgC•ha−1•a−1)|
|B1 10–20 cm DBH||21.03 ± 0.35||0.79 ± 0.038||0.48 ± 0.021 (recruitment)||0.45 ± 0.047||-||0.08 ± 0.081|
|B2 20–35 cm DBH||30.03 ± 0.70||0.84 ± 0.079||0.74 ± 0.048||0.62 ± 0.09||-||−0.33 ± 0.16|
|B335–60 cm DBH||42.67 ± 0.53||0.81 ± 0.045||1.29 ± 0.10||0.53 ± 0.063||-||0.74 ± 0.14|
|B4 >60 cm DBH||60.80 ± 1.16||0.40 ± 0.040||0.83 ± 0.066||1.01 ± 0.12||-||0.22 ± 0.15|
|CWD||43.9 ± 5.06||-||-||-||0.123 ± 0.0001||−2.8 ± 0.64|
|B1 10–20 cm DBH||24.82 ± 0.30||0.52 ± 0.018||0.25 ± 0.008 (recruitment)||0.33 ± 0.016||-||−0.16 ± 0.21|
|B2 20–35 cm DBH||48.55 ± 0.71||0.81 ± 0.041||0.60 ± 0.12||0.70 ± 0.040||-||−0.17 ± 0.33|
|B3 35–60 cm DBH||65.84 ± 1.18||0.74 ± 0.060||0.88 ± 0.24||1.00 ± 0.068||-||0.28 ± 0.29|
|B4 >60 cm DBH||27.35±1.15||0.19±0.040||0.32±0.19||0.54±0.077||-||−0.03±0.21|
|CWD||16.2 ± 1.15||-||-||-||0.123 ± 0.0001||0.58 ± 0.13|
 If we imagine that the system evolves forward and backward in time with fixed transition and growth frequencies, the box model can also be used to place a bound on the date for a mortality pulse by computing the date back in time where the modeled CWD pool would become “excessive” or not biologically possible. If the pulse of excess mortality had occurred as long ago as 1992, the initial CWD pool would have to have been 90.3 MgC·ha−1, equivalent to ∼38% total woody biomass (Figure 9). A CWD pool of this magnitude would imply an input pulse of ∼60 MgC·ha−1, or 25% of standing biomass, representing a major, visible die-off of the forest. If the CWD increase happened in 1998, the CWD pool would have been 54 MgC·ha−1 (∼26% total woody biomass), implying an input of > 20 MgC·ha−1 provided by excess mortality. This input would correspond to ∼10% mortality or about 7 times the rate observed during the nondrought years of our observations, a mortality level comparable to the total mortality observed in the 3-year dry-down experiment of Nepstad et al.  in the TNF.
Figure 9. Modeled trajectories of carbon in different size classes of live trees and CWD for the TNF. Results from box model show short-term reduction of the CWD pool followed by long-term accrual of carbon in forest biomass (inset). Symbols mark the baseline values from 2003 used to construct the model.
Download figure to PowerPoint
 The large pool of CWD remained unbalanced in the TNF through the 6 years of this study, though the magnitude of the C imbalance for at km 67 diminished in the 2001 to 2005 interval, relative to the 1999 to 2001 interval (Figure 6a), confirmed by eddy flux data that showed NEE approaching zero. Increased mortality (2.87 MgC•ha−1•a−1) at km 67 from 2001 to 2005 offset the reduction of the CWD pool via CWD respiration (4.84 MgC•ha−1•a−1), while large-scale transects T3 and T6 showed lower mortality (2.4 and 2.1 MgC•ha−1•a−1, respectively), and edged closer to overall carbon balance, with error bars spanning zero in some cases (Figure 6a).
 How much longer could the imbalance in the TNF last? Using a simple model of CWD respiration at km 67 [Hutyra et al., 2008] and assuming tree growth and recruitment rate constants remain constant, we estimate that the CWD pool could be reduced to the point where CWD respiration would be matched by growth and recruitment of live biomass sometime in 2011 (Figure 9), perhaps a bit earlier  if mortality were lower than observed at km 67 from 2001 to 2005. The continued CWD imbalance in this study suggests that carbon release following a disturbance could last as long as 10–15 years.
 Note that near equilibration of the CWD pool in 2007–2011 indicated in Figure 9 does not imply that the TNF will reach a steady state in that time. The simple model has eigenvalues corresponding to time scales λ = 8.1, 28, 52, 77, and 100 years. The decay rate of CWD fixes the shortest time scale but adjustments among the size classes take much longer (Figure 9, inset): tree boles decay faster than they are constructed. Moreover, the model assumes continuation of current fluxes, which could change with changing environmental conditions or disturbance.
5.4. Implications and Conclusions
 The measurements reported here of fluxes between major stocks of biomass, and complementary eddy flux data provide a strong foundation for quantitatively defining the factors regulating carbon stocks and fluxes in these Amazonian forests. We found significant differences between two sites in the central and eastern Amazon, with TNF showing notably larger gross fluxes in live and dead biomass than the BDFFP plots. Major carbon pools were close to steady state in the BDFFP plots, but significantly out of equilibrium at TNF. Disequilibrium was found on multiple levels in the TNF, with a large CWD imbalance and significant shifting in live tree size class structure. The stand structural changes represent a legacy that, based on observed imbalances and comparison with BDFFP data, will tend to persist for over a decade.
 Legacies of probable disturbance were captured and quantified in the measurements at TNF and our analysis shows the TNF responding rapidly to apparent disturbance by releasing carbon to the atmosphere. Our box model predicts that this release will be short (∼10 years) and followed by an extended period of uptake and adjustments of forest structure, in the absence of a major disturbance. Our data quantify, at the landscape scale, the phenomena of disturbance—recovery widely discussed in the absence of such data hitherto [e.g., Körner, 2004; Clark, 2004a; Moorcroft et al., 2001].
 This analysis of two sites showing disparate carbon dynamics does not allow estimation of the net carbon balance of the entire Amazon Basin. While our two sites are internally consistent in carbon balance, the way they fit into the basin as a whole remains unknown. Baker et al.  suggested some sites in the Western Amazon are close to equilibrium, with low CWD and no signs of recent disturbance. Most likely, the sites in this study represent points in a continuum of disturbance-recovery cycles, where the TNF carbon dynamics presented here exemplify carbon dynamics of more recently disturbed sites. The long-term balance of the Amazon Basin will be determined by the abundance and longevity of sites within this disturbance-recovery continuum: carbon accumulation at a majority of sites may be more than compensated for by episodic release of CO2 during and shortly after major disturbances. Indeed, two recent studies have reported a majority of forest inventory plots showing accumulation of live biomass along with infrequent plots showing excessive biomass loss [Feeley et al., 2007; Chave et al., 2008]. To assess trends in Basin-wide carbon balance, we require further observations (remote sensing, eddy covariance, biometric, etc.) that can accurately detect, over large areas for long times, the ensemble of mortality events that involve single trees or small groups of trees.