For more accurate projections of both the global carbon (C) cycle and the
changing climate, a critical current need is to improve the representation of
tropical forests in Earth system models. Tropical forests exchange more C,
energy, and water with the atmosphere than any other class of land
ecosystems. Further, tropical-forest
The near-future research effort should be on development of a set of widely
acceptable benchmarks that can be used to objectively, effectively, and
reliably evaluate fundamental properties of land models to improve their
prediction performance skills. (Luo et al., 2012).
Improved modeling of tropical-forest carbon (C) cycling is urgently needed
for projecting future climate and for guiding global policy concerning
greenhouse gases. Tropical forests are major players in the global
In addition to the ongoing effects of deforestation and fires, climate
change is likely to magnify the biome's large role in global C cycling.
Tropical forests are being rapidly moved into new climate territory (Wright
et al., 2009). One Earth system model (ESM) has projected that, during the
next 25 years, up to 70
Divergent projections (colored lines) of the changes in
tropical net ecosystem production through this century from seven of the
CMIP5 climate models. The key identifies the models. Dashed lines: models
that include coupled carbon–nitrogen (C–N) biogeochemistry; solid lines:
models lacking explicit nutrient cycling. The ensemble mean is indicated by
the heavy black line, and gray shading indicates the range of 1 standard
deviation (1
Projecting the future integrated effects of climatic and atmospheric change
on tropical forest
To improve current global
A comparison of CASA and CN model outputs to estimates derived by combining the limited field data with estimates of unmeasured components (from Randerson et al., 2009, with permission from © 2009 Blackwell Publishing Ltd.).
Field observations from tropical forests can help develop and validate
models in multiple ways. First, for each C cycle model, the prescribed and
diagnostic ecosystem metrics for the biome should be comparable to the
relevant field data. For instance, do the modeled leaf area index (LAI),
aboveground live biomass, and aboveground wood production fall within the
95
A fundamental consideration for model–data interactions is comparing “apples
to apples.” The field studies to date in tropical forests have addressed
only some of the forest attributes and processes involved in
The other side of the apples-to-apples issue is that, for data–model
comparisons, many C cycle models may require development to include or
output those specific ecosystem attributes that have been field-quantified
in tropical forests (e.g., aboveground wood production, leaf litterfall).
Similarly, the land surface models may need to be restructured to better
represent properties for which only part of the system state can actually be
observed (e.g., predicting surface-soil organic
Two further aspects will determine the usefulness of data–model comparisons. One is the need for the field researchers to clearly communicate the underlying methods and their limitations. The other is that the modelers carefully evaluate field-based observations and take into account their limitations for use in model–data exercises.
As discussed above, some reported observations of C cycle attributes are
based partly on direct measurements and partly on extrapolation. An example
would be total fine-root production as estimated by extrapolating
surface-soil measurements to the unstudied deeper soil layers (e.g., Doughty
et al., 2014). Similarly, the tower-based eddy covariance technique measures
forest-level net ecosystem exchange (NEE) of
Field measurements can be comparable to the predictions of global NPP
models (and could be eventually used for parameterizing them) only when they
are collected by a systematic stratified design, and are therefore
representative of the given region. (Simova and Storch, 2016) … extrapolations and predictions of forest properties based on sparsely
and/or non-randomly distributed field plots are no longer acceptable for
understanding tropical forests in regional or global carbon cycles. (Marvin
et al., 2014) A single plot corresponds to one sample of the forest, and it is unlikely
to represent the whole landscape-scale environmental variability. (Chave et
al., 2004)
Most land surface models attempt to predict landscape-scale fluxes and
pools. Field studies should therefore provide distributed measurements that
span the within-landscape variability. When a forest is instead sampled in
only one or two small (
For typical land surface models, which operate on a scale of 0.5
Two classes of models contrast with the ESMs in explicitly representing the small-scale within-landscape heterogeneity caused by the patchwork of disturbance and recovery phases observed in the real world. Demographic models such as the Ecosystem Demography model (Moorcroft et al., 2001; Medvigy et al., 2009; Fisher et al., 2015) are designed in part to capture the variation between recently disturbed and old-growth forests. Similarly, individual-based models such as TFS and LPJ-GUESS (Fyllas et al., 2014; Pappas et al., 2015) explicitly represent the within-landscape spatial heterogeneity. With those models the smaller-scale observations, such as those from individual hectares, can be usefully compared directly to the model output.
Key outputs from the global models concern the long-term trends in C cycle
attributes in each biome due to both climate change and increasing
atmospheric [
Anomalies of pantropical mean temperature (black) and the ENSO multivariate index (gray) compared to the period of 1960–1990. (from Malhi and Wright, 2004, by permission of the Royal Society).
When a long data series does exist for a given C cycle attribute or process,
climatic and/or [
Effect of length of data series on the correlation of tree growth with minimum temperatures at La Selva, Costa Rica. Data labels: year 1 of each segment of the series (from Clark and Clark, 2011, with permission from the Association for Tropical Biology and Conservation).
For model–data fusion, benchmark field data should be accompanied by several
classes of supporting information. Geographic coordinates of the study site
are required for spatially explicit model tests. Site elevation (meters above sea
level) locates the finding along the lowland-montane continuum of tropical
forests. Given the likelihood of interannual and directional changes in
forest
Ideally, model runs should be set up for individual test bed sites to best allow consideration of site-specific circumstances. Where these types of model–data fusion are planned, a much larger set of auxiliary data, including high-resolution local meteorological data, soil physical properties (texture, depth), and vegetation properties relevant to the question being posed, is potentially useful.
Using the criteria above (direct field measurements, landscape-scale
sampling, sufficiently long data series), we have extracted, from the
literature, examples of robust ecosystem-level field observations of
Table 1 provides a capsule summary of our findings, which are detailed in the following sections. As illustrated in the table, C cycle attributes vary across space and/or time. Model predictions are typically for a single state in a given place and time. Increasingly, however, model predictions are made across a range of parameters (Zaehle et al., 2005; Fischer et al., 2011), initial conditions (Lombardozzi et al., 2014), driving data (Fox et al., 2009; Viskari et al., 2015), and structural variations (Fisher et al., 2015; Medlyn et al., 2015), resulting in ranges of predictions that can be compared against observations, which themselves are known to have errors. Therefore, it is not strictly necessary that observational benchmarks have very low confidence ranges, but it is necessary to document that range of observations and the natural variability that the observations span.
Summary of the characteristics of field observations of ecosystem C cycling in lowland old-growth tropical forests, from the example data presented in this paper (in the tables or text, or footnotes here). n.d.: no benchmark field observations yet identified from this biome. Attribute abbreviations are defined in the text.
LAI observations in lowland old-growth tropical forests; ht: height.
Landscape-scale estimates of aboveground biomass in lowland old-growth tropical forests.
Estimates are based on diameters of all live stems in 9–72
Landscape-scale estimates of coarse woody debris in lowland old-growth tropical forests.
Standing dead:
Field observations for this often prognostic model parameter are
method-dependent and typically underestimate (see Table 2). Forest-level
LAI can be assessed in the field directly, if laboriously, through
replicated leaf harvests from the canopy top to the forest floor. To date,
however, only one study (Clark et al., 2008) has directly assessed it this
way in a tropical forest (LS site, Table 2). Harvested LAI at the
55 4.6
The total ecosystem
All field observations of live aboveground biomass in tropical (and nontropical) forests are indirect, un-validated estimates for just the larger stems (EAB, estimated aboveground biomass). For multiple reasons (see below), it remains unclear how the existing EAB values for this biome can best serve the models.
To derive EAB, all live stems in a stand above some diameter limit (usually
10
For testing models against field observations of tropical-forest biomass
(see Cleveland et al., 2015), a separate important issue is the
within-forest spatial heterogeneity of EAB. For example, within a 10
Many models, particularly those that simulate forest demographics, use allometric equations to relate stem diameter to biomass. They also typically use estimated production of woody biomass to calculate diameter increments. In such cases, comparisons of both biomass and diameter increment for the same forest are therefore only sensible if the same allometric scaling is used. Again, detailed knowledge both of the data products (including EAB) and of model structures is critical.
Current ILAMB benchmarks for tropical regions include maps of aboveground
biomass across the biome based on remote sensing products (e.g., Saatchi et
al., 2011; Baccini et al., 2012). Large divergences between these maps
(Mitchard et al., 2014) highlight the unresolved uncertainties due to
method issues for both the remotely sensed data and the field observations
(e.g., un-validated allometries, landscape-scale samples vs. a single
1
Estimates of tropical-forest CWD span a wide range and are method-dependent (see Table 4). The different methods in current use can produce significantly different estimates for the same site and time (e.g., the two 2005 estimates for JH-CLAY, Table 4). The spatial heterogeneity of standing and fallen CWD within tropical forests calls for landscape-scale sampling. CWD stocks are also likely to significantly change through time due to the temporal variation in tree mortality in this biome (see below).
Estimates of fine-root stocks based on multiple hectares within each lowland old-growth tropical forest.
Dead roots:
Highly replicated, landscape-scale field
observations of this
As illustrated in Table 5, the methods used to quantify fine roots vary in multiple ways, including the maximum diameter of evaluated roots, the depth of soil cores, and whether or not dead roots are included. These method variations make cross-site comparisons and model benchmarking difficult.
A separate critical issue affects observations of fine-root stocks in all
forest types, boreal to tropical: fine-root sampling in forests is usually
restricted to the surface soils. No study has quantified fine roots all the
way down the soil column in any tropical forest (see Table 5). The soils
underlying these forests are often many meters deep. Nepstad et al. (1994)
found live roots down to at least ca. 18
There are as yet no stand-level observations of coarse roots in any forest type. In tropical forests, the field sampling for these spatially variable organs has been confined to harvesting the root systems of selected individual trees (e.g., Niiyama et al., 2010) or to sampling coarse roots in pits or trenches away from trees, thus missing their tap roots and other large roots (e.g., Castellanos et al., 1991; Veldkamp et al., 2003). A recent survey of the available harvest data (Waring and Powers, 2017) found that root : shoot ratios for individual trees from old-growth tropical forests averaged ca. 0.65, indicating the importance of this biomass component. Notably, this ratio strongly contrasts with the 0.21 multiplier commonly used to extrapolate tropical-forest coarse-root biomass from estimated aboveground live biomass (e.g., Malhi et al., 2009; Girardin et al., 2010; Quinto-Mosquera and Moreno, 2017).
SOC estimates based on sampling to
SOC is strongly underestimated
in all forest types (boreal to tropical) because it is rarely if ever
quantified to depth (Jobbagy and Jackson, 2000). The limited tropical data
in hand for subsurface SOC indicate that total SOC can dominate the
The incompletely quantified SOC is a particularly critical data gap for
tropical forests. There is accumulating evidence that the huge
A second issue in tropical forests is that SOC shows marked spatial
variation on all scales: from one square meter to the next (Powers, 2006)
and across the major edaphic changes (topography, soil types; see Richter
and Babbar, 1991) within a forest. An example of this within-forest
heterogeneity is the significant difference in cumulative SOC content
between two major soil types at the LS site (Table 6). Distributed
and replicated sampling is therefore required to quantify this important
The eddy flux method has been criticized for uncertainty in its nighttime
measurements. This is especially obvious in tropical areas, where nighttime
turbulence is not well developed. Nevertheless, … Convincing results can be
obtained from daytime eddy flux measurements… (Tan et al., 2013) It is clear that the choice whether or not to filter and replace nighttime
[Amazon forest eddy flux] data represents the single major uncertainty in
the whole estimation process. The choice can turn a very large carbon sink
into a moderate one or even into a small source. (Araújo et al., 2002)
For NEE at longer time steps (days to years), however, estimates based on
the eddy flux technique in tropical forests do not provide reference-level
field benchmarks for the models. Multiple issues for this technique in these
forests create large uncertainties about the magnitude and even the sign of
such estimates. The prevalence of still air conditions at night (e.g.,
70–80
… there is no way of directly measuring the photosynthesis or daytime
respiration of a whole ecosystem of interacting organisms; instead, these
fluxes are generally inferred from measurements of net ecosystem–atmosphere
Although GPP estimates have been produced by tropical-forest eddy covariance
studies, the sole
The biometric components of total NPP in
tropical forests (Mg
Alternatively, bottom-up biometric approaches have been used to estimate GPP
for some tropical-forest sites (e.g., Doughty et al., 2014; Malhi et al.,
2015). These studies, carried out in a single 1
Similarly, existing eddy flux
estimates for whole-forest respiration in this biome remain questionable due
to multiple issues: (1) the uncertainty of the NEE estimate from which
Benchmark-level field observations of these two fractions of
No benchmark field observations are available for total NPP. As is the case in all other forest types (Clark et al., 2001a), the field studies in tropical forests have been restricted to a subset of NPP components (Table 7). Those that remain unquantified could sum to a substantial fraction of total NPP (see also Clark et al., 2001a, b; Litton and Giardina, 2008; Cleveland et al., 2015). For the models, the sum of the NPP components assessed in the field provides a lower bound for total NPP.
Two NPP constituents missing from the field studies (Litton and
Giardina, 2008) and from most models (Fatichi et al., 2014) so far are the amounts
of new fixed
Opportunities for data–model fusion will be maximized by developing the
C cycle models to explicitly specify those NPP components that have been
assessed in the field. As recently reported by Negrón-Juárez et al. (2015),
only three of the 10 ESMs in CMIP5 report “leaf NPP”, “wood NPP”, and “root NPP”. The different
production components are functionally distinct. In a landscape-scale field
study at the Costa Rican LS site, the several field-quantified NPP
components varied independently through 12
Landscape-scale estimates of the components of fine litterfall
(leaf, reproductive, twig) in lowland old-growth tropical forests. Grd. traps:
In tropical forests, biometric aboveground
NPP is typically dominated by short-lived tissues (Clark et al., 2001b).
These are assayed as shed “fine litterfall” collected in litter traps (Table 8).
Fine litterfall varies spatially within each tropical forest. When
assessed in 18 0.5
In field studies of
biometric NPP (termed NPP*; Clark et al., 2001a), leaf litterfall over a
given study interval is typically taken as a surrogate for leaf production
over that interval. Stand-level leaf production itself has not been
quantified in the field in tropical forests. In most tropical forests, leaf
litterfall is the largest contributor to aboveground NPP* (Clark et al.,
2013, and included references). It can be a misleading surrogate for leaf
production in terms of both mass and timing. One method issue is the
difficulty of quantifying the very large fallen leaves in tropical forests
(e.g., 3
One potential approach for models would be to explicitly include the processes of herbivory and decomposition losses that occur between leaf production and leaf shedding, therefore facilitating a direct comparison. In lieu of this, model–data comparisons should take into account the low bias of leaf litterfall observations. In cases in which leaf litterfall is conflated with leaf production for the purposes of determining allocation to the leaf fraction, the resulting allocation underestimate might lead to underestimation of LAI.
A separate issue is that the seasonal timing of leaf production can differ
from that of leaf litterfall, as found by Reich et al. (2004) in a
Venezuelan tropical forest (in most species studied, although there was some
degree of correlation). In many tropical forests, leaf litterfall typically
peaks at the time of the yearly maximum soil dry-down (Wagner et al., 2016);
this timing can be distinct from that of actual leaf production. Such a
timing disjunct will complicate attempts to evaluate the seasonality of
tropical-forest NPP and
Estimates of twig
litterfall should be treated as a lower bound for twig production. In
tropical forests, twig litterfall (Table 8) is likely to strongly
underestimate actual production due to substantial mass loss before
collection. In a New Guinea rain forest, when Edwards (1977) compared
canopy-collected live twigs with a diameter
The
biometric surrogate for reproductive production, reproductive litterfall
(Table 8), is likely to undervalue production by at least 50
For multiple reasons, this NPP component merits attention for the models.
Many land surface models do not specifically include the carbon allocation
to reproduction; this omission implies corresponding overestimates of stocks
of other carbon pools (e.g., roots, stems, leaves). Demographic models, in
contrast, typically do specify reproductive allocation, which is needed to
drive forest recruitment (Moorcroft et al., 2001). Secondly, reproductive
tissues are nutrient-rich (e.g., in nitrogen, phosphorus, and cations) and
thus likely play a significant role in the cycling of those nutrients.
Reproductive status could influence nutrient resorption and thus
reallocation of carbon (Tully et al., 2013). A third issue is that this
production component could be responding to climatic and/or [
Landscape-scale estimates of aboveground wood production (EABI,
As for aboveground
woody biomass (above), field estimates of aboveground wood production, also
termed EABI (estimated aboveground biomass increment), are unverified and
highly uncertain. This production component is based on measurements at two
successive censuses of the diameters of all live stems in the study plot
that exceed an arbitrary diameter limit (usually 10
As for estimates of aboveground biomass, because EABI depends on an unverified allometric relationship between stem diameter and stem biomass, all values of this metric involve unquantifiable uncertainty. When different allometries are applied to the same set of diameter data, different estimates of EABI can be produced (e.g., duplicate estimates at site TAP-KM67; Table 9). Determining which if any of such estimates are reasonable would require follow-up on-site verification of the underlying allometry (Clark and Kellner, 2012).
Estimates of fine-root production (
Given the heterogeneity of biomass dynamics within a tropical forest,
data–model fusion exercises and site-level model testing call for
landscape-scale field data for EABI. Individual-based or demographic models (e.g., ED, Moorcroft et al., 2001)
that address the small-scale spatial heterogeneity within a forest
landscape are the exceptions to this. In spite of this metric's unquantifiable uncertainty, when
estimated on the landscape scale and in the same way over a long series of
successive periods, repeated annual estimates can provide valuable guidance
for the models with respect to both long-term trends in this productivity
component and its climatic and [
Field estimates of fine-root
production at the landscape level in tropical forests provide a useful lower
bound for this NPP component. Due to the method challenges, fine-root
production has not been well quantified in any forest type, boreal to
tropical. In the tropical-forest biome, because of the notorious variation
in fine-root stocks on all spatial scales (Espeleta and Clark, 2007; Powers
et al., 2005), robust assessment of fine-root production for a given forest
would require highly replicated and distributed sampling. Unfortunately,
this production component has only rarely been assessed in multiple hectares
of a tropical forest (Table 10). A second critical limitation is that the
field measurements to date in this biome have been confined to the surface
soil (0 to
Variable methods for assessing fine-root production (different soil depths
and root sizes, inclusion or exclusion of dead roots; Table 10) also make
cross-site comparisons difficult. The usual approach in tropical forests,
in-growth cores, is likely to strongly underestimate production due to lags
before root in-growth and the likelihood of roots dying and decomposing
before soil cores are retrieved; in a temperate pine forest, production
estimates based on in-growth cores averaged 54
Estimated mortality-driven biomass loss (
… [in a steady-state landscape] about 98.0 to 99.7 … a more comprehensive sampling scheme that includes large-area data
(e.g., large plots and remote sensing) and robustly characterizes
disturbance size distribution is required to understand tropical forest
dynamics and its impact on carbon balance. (Di Vittorio et al., 2014)
An observational finding important for the C cycle models is the strong temporal variation in tropical-forest tree mortality. Mortality spikes have been observed in both neotropical and Asian tropical forests in extreme climatic events such as the strong El Niños of 1982/83 and 1997/98 and the 2005 Amazon drought (Clark, 2004; Williamson et al., 2001; van Nieuwstadt and Sheil, 2005; Phillips et al., 2009).
Some models specify stochastic dynamics of tree death (Fyllas et al., 2014; Smith et al., 2014). Many models attempt to simulate the responses of tree mortality to changes in vegetation stress (McDowell et al., 2013; Powell et al., 2013) but more aggregated models typically use a simple turnover parameter (Galbraith et al., 2013, reviewed by McDowell et al., 2013). Introducing more robust mortality benchmarks based on combining structured ground data with satellite observations (e.g., Kellner and Hubbell, 2017) and also explicitly linking large mortality losses to extremes of climatic stressors (e.g., Phillips et al., 2009) should help modelers move towards a more process-based representation of tropical-forest mortality.
Climatic and [
Local meteorological records for lowland old-growth tropical forests (one example site).
Qa/Qc:
Site codes and descriptors for the field sites in the benchmark data tables. MAP: mean annual precipitation; MAT: mean annual temperature.
A valuable class of benchmarks for the C cycle models will be
landscape-scale field observations of the decadal changes in and
climatic and
To illustrate this type of response benchmarks Table 12 lists the
significant relationships revealed by a 12-year landscape-scale study of
annual biometric aboveground NPP (ANPP*) in a Costa Rican forest (Clark et
al., 2013). Through that period, one of the four biometric ANPP* components,
EABI, showed highly significant negative impacts from two climatic stressors
and a small positive response to increasing [
Sparse and intermittent climatic monitoring in all tropical regions makes the interpolated global gridded climatic datasets unreliable for this biome (see Deblauwe et al., 2016). In addition, sub-daily meteorological records are critically needed for driving C cycle models. High-quality climatic records from tropical-forest field sites would be particularly important resources for model–data fusion exercises and merit inclusion among the benchmark field observations of the ILAMB effort.
For a catalogue of such local climatic records, key accompanying information
should include whether the data are from a ground-level met station or from
above-canopy sensors, and whether the records have been screened, corrected
to maintain internal consistency, and gap-filled. At the example site in
Table 13, multiple adjustments to the records were required after
the manual instruments were relocated and then augmented with an automated
system (see Clark and Clark, 2011). The calculation (
A community-consensus catalogue of the benchmark-level field observations
directly relevant to
Data catalogues need to be “living” resources, constantly updated as new information comes in and as ecological insights and methods develop in each biome. For the ongoing updating, a web-based, moderated system would seem to be the strongest approach. With such a system, field researchers worldwide could actively participate, continuously offering new field observations for consideration and also correcting or augmenting current entries. Proposed updates, however, should be prescreened by a team of volunteer researchers and modelers with the relevant expertise.
We have identified here examples of reference-level field observations from lowland old-growth tropical forests. Now what is clearly needed is a much broadened discussion among the wider tropical research community, both to refine the benchmark criteria for these forests and to contribute observations on a continual basis going forward. A similar parallel effort is also greatly needed to identify data benchmarks for the highly distinct C cycling processes taking place in degraded and successional tropical forests, which may account for half or more of the forest area across the tropics (Chazdon, 2014). Yet a different set of benchmarks would be needed to characterize C cycling in tropical montane forests, an ecologically distinct class of tropical forests.
Our effort here provides a starting point for addressing the modeling
community's need for reference-level field observations from the
tropical-forest biome. As is evident from our review, the field data for our
target forests are woefully sparse, and the uncertainties around the major
No datasets were used in this article.
All authors collaborated in the writing of the paper.
The authors declare that they have no conflict of interest.
This work was made possible by the support of the US Geological Survey John Wesley Powell Center for Analysis and Synthesis. The support from the Powell Center included funding the participation of Shinichi Asao. Deborah A. Clark was supported by US National Science Foundation LTREB grants DEB-1147367 and DEB-1357112. Tana E. Wood was supported by US Department of Energy, Terrestrial Ecosystem Sciences grant DE-SC0011806 and by the USDA Forest Service International Institute for Tropical Forestry in collaboration with the University of Puerto Rico. Xiaojuan Yang was supported by the Next-Generation Ecosystem Experiments-Tropics and the Biogeochemistry–Climate Feedbacks Scientific Focus Area (BGC Feedbacks SFA) of the US Department of Energy, Office of Science, Office of Biological and Environmental Research. Peter B. Reich was supported by the US Department of Energy, Office of Science (DE-SC0012677). Sasha Reed was financially supported by US Department of Energy, Terrestrial Ecosystem Sciences grant DE-SC0011806 and by the US Geological Survey Ecosystems Mission Area. Any use of trade, firm, or product names is for descriptive purposes only and does not imply endorsement by the US Government. David B. Clark constructively commented on the paper. Edited by: Anja Rammig Reviewed by: two anonymous referees