Published online 11 February 2004

THE ROYAL
SOCIETY

Error propagation and scaling for tropical forest

biomass estimates

Jerome Chave^*, Richard Condit^, Salomon Aguilar^, Andres Hernandez^,

Suzanne Lao^ and Rolando Perez^
^Laboratoire Evolution et Diversité Biologique UMR 5174 CNRS/UPS, bâtiment 4R3, 118 route de Narbonne F-31062
Toulouse, France
^Center for Tropical Forest Science, Smithsonian Tropical Research Institute, Unit 0948, APO AA 34002-0948, USA

The above-ground biomass (AGB) of tropical forests is a crucial variable for ecologists, biogeochemists,
foresters and policymakers. Tree inventories are an efficient way of assessing forest carbon stocks and
emissions to the atmosphere during deforestation. To make correct inferences about long-term changes in
biomass stocks, it is essential to know the uncertainty associated with AGB estimates, yet this uncertainty is
rarely evaluated carefully. Here, we quantify four types of uncertainty that could lead to statistical error
in AGB estimates: (i) error due to tree measurement; (ii) error due to the choice of an allometric model
relating AGB to other tree dimensions; (iii) sampling uncertainty, related to the size of the study plot;
(iv) representativeness of a network of small plots across a vast forest landscape. In previous studies, these
sources of error were reported but rarely integrated into a consistent framework. We estimate all four
terms in a 50 hectare (ha, where 1 ha = lO** m^) plot on Barro Colorado Island, Panama, and in a network
of 1 ha plots scattered across central Panama. We find that the most important source of error is currently
related to the choice of the allometric model. More work should be devoted to improving the predictive
power of allometric models for biomass.
Keywords: above-ground biomass; allometric equation; error propagation; sampling; tropical forest

1. INTRODUCTION response of tropical forests to global change (Clark 2002;

Chave et al. 2003; Baker et al. 2004b). The biomass stocks
Permanent sampling plots have long been used in ecologi-
of tropical forests remain poorly resolved at the regional
cal studies for assessing how much biomass is held in eco-
scale (Fearnside 1996; Houghton et al. 2001). Indeed, it
systems (Olson et al. 1983; Fearnside 1996; Brown 2002).
is difficult to integrate site-specific and heterogeneously
Tree AGB is strongly correlated with trunk diameter
collected data to draw regional-scale conclusions about
(Brown & Lugo 1992; Brown 1997; Clark et al. 2001a),
tree densities, turnover rates or biomass stocks of tropical
and it is therefore possible to use forest inventory data to
forests, and it thus seems precarious to extrapolate such
estimate the stocks and changes in AGB in those inven-
local data to larger scales without assessing how represen-
tories. Recently, plot data have been influential in creating
tative these data are. Dynamic global vegetation models
new hypotheses on the dynamic coupling between tropical
of the new generation have made remarkable progress
forests and the atmosphere (Phillips & Gentry 1994; Phil-
towards integrating ecological processes across scales for
lips et al. 1998; Malhi & Grace 2000). It has been sug-
tropical forests (Hurtt et al. 1998; Bugmann & Solomon
gested that the tropical forest biome could be responding
2000; Foley et al. 2000; Cramer et al. 2001, 2004;
en masse to global change, leading to shifts in species
Moorecroft et al. 2001), but these models need to be cali-
composition and to an overall increase of the turnover rate
brated.
(Phillips & Gentry 1994; Phillips et al. 1998; 2002a,è).
Figure 1 depicts one strategy for converting forest plot
This renewed interest in tropical forest inventories has also
data into regional-scale AGB estimates (Brown et al. 1989;
motivated a new literature on methodological aspects that
Brown 1997; Houghton et al. 2001). Each tree in a plot
have greatly improved our confidence in biomass data esti-
is measured, tagged and identified (Clark 2002; Phillips
mated from plots (Shell 1995; MacDicken 1997; Condit
et al. 2002a); an allometric equation is used to relate its
1998; Higuchi et al. 1998; Chave et al. 2001; Clark et al.
diameter to an AGB estimate (Brown 1997). The plot-
2001a,b; Clark 2002; Phillips et al. 2002a; Brown 2002).
level estimate is then summed over all the trees to obtain
Nevertheless, difficulties in assessing data quality in for-
a stand-level AGB estimate. For carbon sequestration
est inventories lead to continuing debate on the functional
issues, the quality of this estimate depends on the plot size.
In addition, the landscape-scale environmental variability
should be integrated by replicating the measurement in
" Author for correspondence (chave@cict.fr). other plots of the same forest (Clark & Clark 2000; Keller
One contribution of 17 to a Theme Issue 'Tropical forests and global et al. 2001; Nascimento & Laurance 2002). These steps
atmospheric change'. integrate a variety of techniques that all contain some

Phil. Trans. R. Soc. Land. B (2004) 359, 409-420 409 © 2004 The Royal Society
DOI 10.1098/rstb.2003.1425
410 J. Chave and others Error propagation for biomass estimates

error
propagation

quality of the data

diameter of one tree
allometric
quality of the model
model
V
AGB of one tree
sum over
size of the sampled area
trees
V
AGB of one plot
average over
representativity of
plots
V the plots
AGB of the forest

Figure 1. The error propagation for estimating the AGB of a tropical forest from permanent sampling plots.

uncertainty, yet there is no consistent methodology for In addition, there is a second source of error due to the con-
propagating uncertainty across scales (but see Ketterings struction of the allometry: the model's parameters are usually
et al. 2001). Because errors due to these problems add up, estimated using a regression on the log-transformed variables:
each needs to be quantified carefully and independently.
ln(AGB)=ln(/(D,/i,p)) + e (2.1)
In the present contribution, we assess the different sources
of error associated with AGE estimates from forest inven- The residuals ^ represent the departure from a perfect allometry,
tories, and present calculated examples for a moist tropical and are normally distributed. The s.d. a of these residuals rep-
forest of central Panama. resents the uncertainty in the AGE estimation due to the
allometry itself. Easkerville (1972) noticed the following bias:
the expectation of AGE is <AGE> =/(D,H,p)<exp(ö>, and it is
2. METHODS
incorrect to state that {exp(Q) is equal to exp({^)) = exp(O) = 1.
(a) Uncertainty on tree level AGB estimate Indeed, the random variable exp (O is log-normally distributed
The first potential source of error is the tree measurement and the expectation of exp (^) is {exp(Q) = exp(a^/2). This last
process. Stems, of diameter 10 cm or greater, at 130 cm above term is often called the correction factor, CF (Brown et al. 1989;
the ground or above any trunk deformity, are tagged, located Ter-Mikaelian & Korzuldiin 1997; Hughes et al. 1999). An
and their diameter D is measured. As any ecological variable, unbiased estimate of AGE is, therefore
D is likely to be measured with some imprecision, and special
(AGB> = CFx/(D,H,p). (2.2)
conventions should be applied (Shell 1995; Condit 1998; Phil-
lips et al. 2002a). In particular, special attention should be paid The uncertainty on the estimate of AGE associated with the
to irregular-shaped trunks. allometric model is measured by the s.d. o^ = -v/CF^ • 1 X
We denote the standard error associated with the diameter (AGB). Thus, for a given allometric model, one can estimate
measurement as a¡^. It is expected to be an increasing function both the expected AGE held in a tree and the s.d. of this esti-
of D. When a height measurement H is also taken, the related mate.
error is denoted af¡. These two error terms covary as D and H In the following, we assume that the measurement and the
are positively correlated in most woody plants. Finally, a wood allometric uncertainties are independent sources of variability.
specific gravity value p (oven-dry weight over green volume; The overall uncertainty on the AGE estimation of a single tree
Chudnoff (1984)) can be associated with each tree, either by therefore is CTX + c«.
direct measurement (e.g. from tree cores) or using databases
that provide the mean p for the species to which the tree belongs (b) Allometric model selection error
(Brown 1997; J. Chave, T. Baker and H. C. Muller-Landau, Ideally, as is the case for temperate trees (Ter-Mikaelian &
unpublished results). The corresponding error a^ could be due Korzukhin 1997), each species should have its own biomass
to a misidentification of the tree, or to a variation in p within equation, based on a large sample size. This programme is
or among conspecific trees (Carvalho et al. 1995; Wiemann & unrealistic for tropical forests. Tropical forest allometric models
Williamson 2002). used for AGE estimation suffer from three important short-
Errors in trunk diameter, height or density measurement all comings: (i) they are constructed from limited samples; (ii) they
result in error in estimating the AGE, which is usually taken are sometimes applied beyond their valid diameter range; (iii)
from an allometric model of the form AGB =f(D,H,p). This they rarely take into account available information on wood spe-
error is propagated to the AGE estimate through the allometric cific gravity.
model by expanding the model function / in Taylor series. In Many of the published models are based on harvest experi-
Appendix A, we compute the measurement s.d. for the AGB ments performed in a single forest and based on, typically, less
estimate, a^^. than 50 harvested trees. We suggest that the number of trees

PhU. Trans. R. Soc. Lond. B (2004)

 Error propagation for biomass estimates J. Chave and others 411

used to calibrate allometric models is a major source of variation _0.80

of the AGB estimate when different models are selected. We B 0.70 J
used several published allometric models to assess this problem o
3 0.60
of model selection. We selected six published allometric models
reconstructed directly from the original datasets (trees 10 cm or •| 0.50
cd
more in diameter) with 39-187 sampled trees (Lescure et al. ^0.40
o
1983; Yamakura et al. 1986a,è; Overman et al. 1994; Joyce in
^ 0.30
Brown 1997; Araújo et al. 1999; Chambers et al. 2001). AU of u
these models were used to estimate AGB stocks outside the ^0.20
biogeographic zone where they were constructed, either alone, i 0.10
or included in pan-tropical allometric models. Further, they all
used similar sampling strategies (but the dataset of Joyce cited in 50 100 150 200 250
Brown 1997 remains poorly documented). To test for variation midpoint value of diameter class (cm)
among allometric models' predictions, we use a pan-tropical
equation that relates the AGB (in kilograms) to the trunk diam- Figure 2. Wood specific gravity for different diameter classes
eter (in cm) and the wood specific gravity (in grams per cubic in the BCI 50 ha plot. The decrease in wood density was
centimetre) deduced from a compilation of 634 trees of diameter significant across size classes (p < 0.001, r^ = 0.97).
10 cm or greater (J. Chave (and 11 others), unpublished
results) :
specific gravity. If, in this dataset, large trees have lighter wood
/(D,p) =^exp(-3.742 + 3.4501n(D) (2.3) than small trees, then this factor would induce a systematic bias.
We used a compilation of tree harvest datasets and confirmed
- 0.1481n(D)^).
that there was no such bias (results not shown; J. Chave (and
We independently assessed the model selection effect by using 11 others), unpublished results).
a rarefaction technique on the pan-tropical dataset. A set of 10-
400 trees was drawn at random from the sample of 634 trees, (c) Minimal single plot size
and it was used to construct a regression model from which we Tree plots are usually 0.2-100 ha in size (Houghton et al.
estimated the AGB of the 50 ha BCI plot. We replicated this 2001). Methodologies related to the establishment of plots have
procedure 1000 times, and computed the variance among our been covered in the literature (Shell 1995; Condit 1998). More
estimates. This enabled us to assess the true error associated recently, the limitations of this methodology in relation to the
with using a model constructed from an imperfect knowledge of estimation of AGB stocks and changes have also been covered
the allometric relationships in tropical trees. (Clark et al. 2001a,b; Clark 2002; Phillips et al. 2002a). For
Second, many published allometric models lack reliable data assessing the uncertainty on the stand-level AGB estimate in one
for the largest diameter classes, so they cannot be used to esti- plot, one should consider how well the census has been perfor-
mate the AGB held in large trees (Clark & Clark 2000; Hough- med, and how large the census was. We refer to this as the
ton et al. 2001). We therefore assessed the potential error caused within-plot sampling uncertainty.
by extrapolating the models beyond their range of applicability. Methodological sources of error include the incorrect esti-
This uncertainty is of the order estimated AGB for the largest mation of the plot area, trees missed, measured twice or dead
trees, and more importantly, it does not average out for large trees counted as alive. In addition, tree-level errors average out
sample sizes, because it reflects our limited knowledge of the in large plots, and for this reason too it is advisable to establish
model itself for the largest diameter classes. We attempted to large permanent sampling plots. This aspect has been largely
minimize this uncertainty as much as possible. We used the pan- overlooked in the literature before the 1990s. Klinge & Rodrig-
tropical equation to correct for the AGB of extrapolated trees ues (1973) wrongly concluded that one 0.2 ha plot was enough
for the other published models, as we have the best confidence to estimate tree AGB with good confidence, by assuming that
in this equation for the largest trees. the AGB was normally distributed among 10 m X 10m sub-
Finally, few of these models use information on wood specific plots. Since rare large trees contribute a large fraction of the
gravity, known to vary over a fivefold range in tropical tree spec- overall AGB, the distribution of AGB across 10 m X 10m sub-
ies. Including wood specific gravity as a predictive variable plots is far from normal (Chave et al. 2003).
improves the quality of the AGB estimate. For instance, in the Lianas, multi-stemmed trees and re-sprouting trees are often
BCI 50-ha plot, large trees tend to have a much lower wood not recorded during the censuses. This leads to an underestim-
density than small trees (see figure 2). We corrected the pub- ation of the stand-level AGB. Correction factors have been com-
lished allometric models by including a dependence on wood puted in studies where this information is available, and the
specific gravity of the form uncertainty on these correction factors contributes to the stand-
level error. Similarly, BGB is usually estimated from other stud-
<AGB) : CFx^/(D), (2.4) ies' averages (Malhi & Grace 2000; Houghton et al. 2001), more
Pav
rarely from diameter-BGB allometries (Ovington & Olson
where the AGB estimate is a linear function of tree-level wood 1970). In this study, we have no data on lianas or on BGB, and
specific gravity divided by the average specific gravity for the plot we cannot therefore consider errors in those areas.
where the allometric equation was constructed. The parameter In the Panama data, trees less than 10 cm in diameter (and
PJ,^, is the ratio of the total oven-dry weight of the trees used to 1 cm or more) were measured in the whole 50 ha BCI plot and
construct the equation, over their fresh volume. Multiplying in sub-plots of the Marena plot network of plots (see § 3). For
directly the model by a factor p/p¡,^ as above is only valid if the trees less than 10 cm in diameter, we used a single equation
trees used to construct the original allometry all have the same modified from the model devised by Hughes et al. (1999) for a

PhU. Trans. R. Soc. Lond. B (2004)

412 J. Chave and others Error propagation for biomass estimates

moist tropical forest of south Mexico (Los Tuxtlas). The (1993). To estimate rates of error, we performed a dou-
model was ble-blind re-measurement of 1715 trees in 1995 and 2000
(Condit 1998) and fitted the discrepancies with a sum of
/(D,p) =-^exp(-1.9703 + 2.11661n(D)). (2.5) two normal distributions. The first describes small errors
Pav
and has an s.d. {SD{) proportional to the trunk diameter;
(d) Landscape-scale representativity the second has a fixed larger s.d. (SD2). The 1715 errors
A single plot corresponds to one sample of the forest, and it were best fit with SDi = 0.0062 xD + 0.0904,
SD2 = 4.64 (all units in centimetres), with 5% of the trees
is unlikely to represent the whole landscape-scale environmental
subject to the larger error. For example, the diameter of
variability. Among the possible biases, there is a tendency for
researchers to select attractive forests (the 'majestic effect' of a 30 cm tree has a typical error of 0.27 cm (95%
Shell 1995) or to avoid disturbed forests. The landscape-level probability) or of 4.63 cm (5% probability). The uncer-
tainty associated with the height estimate is due to the
AGB estimate should be assessed by establishing a network of
inherent measurement problem of tree height. Tree
plots randomly distributed over the landscape, to assess the
variability of forest types. heights were taken for over 1000 trees of diameter 10 cm
or greater in 80 different species. Based on this dataset,
we assume that the error in height is ca. 10% of the esti-
3. MATERIAL AND STUDY SITES mated value.
We quantified the uncertainty associated with the esti- We assume an s.d. of 10% of the mean wood specific
mation of AGB of a single tree, assuming that the allo- gravity for all species. This figure is based on 50 neotrop-
metric method is unbiased, for the forest of the Panama ical tree species for which more than six different estimates
Canal Zone, central Panama. We have already provided were available from a total of 43 literature sources (see
estimates of the AGB held in the 50 ha permanent sam- Appendix C). A detailed report on this dataset is beyond
pling plot on BCI. We used diameter measurements for the scope of the present publication and will be the topic
over 200 000 trees of diameter 1 cm or greater, combined of a forthcoming publication (J. Chave, T. Baker and
with tree heights modelled from diameter-height H. C. Muller-Landau, unpublished results). For species
regressions that had been developed for 80 common tree missing wood specific gravity estimates, we used a mean
species. We used literature data on wood specific gravity of 0.58 g cm^^ and the same error of 10%.
for 123 species occurring in the BCI plot (Chave et al. The measurement error on the AGB can be deduced
2003). Here, we use the BCI plot and reassess various firom the equations provided in Appendix A for the pan-
sources of error in the previously published AGB estimate. tropical model used for trees of diameter 10 cm or greater.
In the present contribution, we use only the most recent We find o-M = 0.165(AGB) and (7A = 0.313<AGB). Hence,
census of the BCI plot, conducted during the year 2000. the uncertainty on the AGB estimation of a single tree of
We addressed the issue of within-plot sampling using diameter 10 cm or greater is 47% of the estimated AGB,
sub-plots within the BCI plot. We also investigated the partitioned into 31% due to the allometric model and 16%
landscape-scale sampling problem, by using a network of due to the measurement uncertainty. However, this error
45 plots distributed across the watershed of the Panama averages out at the stand level. For the model used for
Canal, henceforth called the Marena plots. These plots trees of diameter less than 10 cm (equation (2.5)), the
were originally set up to study the variation of floristic uncertainties are CTM = 0.234(AGB) and CTA = 0.547<AGB),
composition in forests across the north-south climatic and this model predicts that 7.66 Mg ha^' are in trees less
gradient of this region (Pyke et al. 2001), spatial turnover than 10 cm in diameter.
in diversity (Condit et al. 2002) and differential forest
response to drought (Condit et al. 2004). Each plot is 1 ha (b) Allometric model selection error
in size and has all trees of diameter 10 cm or greater The different allometric models estimated AGB from
tagged, mapped and identified to species or morphospe- 215 to 461 Mgha"' with a mean of 347 Mgha"' and an
cies, except for 154 trees out of 22 955 (0.7%) that remain s.d. of 77 Mg ha^', before correcting for variation in wood
unidentified. A total of 775 species or morphospecies were specific gravity (table 1).
identified. In 35 of the plots, trees of diameter 1 cm or Using the rarefaction method, we estimated the s.d. dif-
greater were censused in central sub-plots of 40 m x 40 m ferently because of the choice of the allometric model
(0.16 ha), whereas all trees of diameter 1 cm or greater (figure 3). For a sample size of 300 trees, we found a mean
were censused in another 10 1 ha plots. Finally, trees of AGB estimate of 263 Mgha"' with an s.d. of 3.1%) of the
diameter 10 cm or greater were inventoried in nine mean, but this figure increased to over 10% of the mean
additional smaller plots (0.32 ha each). This represented for samples of 50 trees or less.
a total sampling effort of ca. 49 ha. These plots spanned Equations that included wood specific gravity predicted
a variety of environmental types and successional ages 218-334 Mgha^' in trees of diameter 10 cm or greater
(Appendix B; see also Pyke et al. (2001); note that some (mean: 284Mgha"', s.d.: 37Mgha"' or 13% of the
errors in table 1 of Pyke et al. (2001) have been corrected mean).
in Appendix B). Next, we assessed the uncertainty due to the AGB esti-
mation in large trees for the same eight allometric models.
The extrapolated AGB represented 7-30% of the total
4. RESULTS
AGB, depending on the model. We used the pan-tropical
(a) Uncertainty on tree-level AGB estimate equation to correct for the AGB of extrapolated trees. In
The uncertainty associated with the diameter measure- the 50 ha plot, 46 trees are larger than 148 cm in diameter,
ment in the BCI forest was discussed in Condit et al. the larger diameter in the sample used to construct the

PhU. Trans. R. Soc. Lond. B (2004)

 Error propagation for biomass estimates J. Chave and others 413

Table 1. AGB estimates (in megagrams per hectare) for the BCI 50 ha forest based on eight different equations involving diameter,
developed for different forests.
(The AGB estimate for trees of diameter 10 cm or greater varied significantly among equations, even when the problem of using
an equation beyond its acceptable range was corrected. We provide the reference for the original data, the region of this study,
the number of trees of diameter 10 cm or greater, the maximal diameter. The column headed 'order' refers to the order of the
polynomial regression of In(AGB) versus ln(Z)) (chosen to minimize the variance in the residuals. Chambers et al. (2001)), and
column CF provides the correction factor for this regression, max dbh, maximum stem diameter at breast height.)

correction
number max uncorrected correction large trees
reference location of trees dbh Pav order CF AGB p/p^v + p/p^v

Araújo et al. (1999) Para, Brazil 127 138 0.68=' 2 1.070 375 307 315
Chambers et al. (2001) Manaus, Brazil 161 120 0.691^ 2 1.065 330 266 278
Overman et al. (1994) Colombia 51 98.2 0.62= 1 1.046 351 292 274
Yamakura et al. (1986a,è) Kalimantan, 38 130.5 0.7= 1 1.074 461 334 310
Indonesia
A. Joyce, in Brown (1997) Costa Rica 92 116 0.52'ä 1 1.028 215 218 220
Lescure et al. (1983) French Guiana 187 118 0.66 1 1.080 428 322 288
Chave et al. (in prep.) pan-tropical 634 148 0.6 1 1.092 324 278 268
Chave et al. (in prep.) pan-tropical 634 148 0.6 2 1.091 293 260 260

" From species-level information, assuming a fresh wood specific gravity of 1.05.
*' From a neighbouring forest, value published by Fearnside (1997).
= Estimated from information in Yamakura et al. (1986a,è) and in Suzuki (1999).
'^ Low-quality estimate from the value of the dominant species at La Selva (Pentaclethra macroloba).
= Estimated from combining allometric equations published in Overman et al. (1994).

35 1 Table 2. Uncertainty in the BCI 50 ha plot, and normality of

the AGB estimates in sub-plots of various sizes.
30 (The normality test was performed on the skewness and on the
kurtosis. Asterisks show the significance level of the normality
c 25 test CP < 0.10; ""p < 0.05). For sub-plots of size 50 m X 50 m
and greater, the distribution was normal (2000 census; cf.
20 Chave et al. (2003)).)

:: 15
sub-plot size number of
(m) plots s.d. skewness kurtosis
ä 10

10x10 5000 385.08 5.42 47.05

20x20 1250 187.64 2.57 10.62
25x25 800 149.44 1.88 5.34
o 50 100 150 200 250 300 350 400 20x50 500 119.52 1.54 3.47
sample size in the allometry 50x50 200 77.55 0.57* -0.24"
100x100 50 42.01 -0.06" -0.12"
Figure 3. Coefficient of variation (CV= lOOxs.e.m. over
mean AGB) as a function of the sample size used to
construct the allometric equation. Each point corresponds to
the CV constructed from 1000 independent selections out of (c) Minimal single plot size
the pan-tropical dataset of 632 trees. For example, with We used the BCI 50 ha plot to evaluate the stand-level
allometric equations constructed from 50 trees, the s.e.m.
sampling uncertainty, under the assumption that the allo-
was 10% of the mean AGB.
metric model is perfect. The tree-level uncertainties aver-
age out at the stand scale. For example, in a typical plot
pan-tropical equation. The largest tree is a Hura crepitens of one-quarter of a hectare, the error on the AGB estimate
of 246.8 cm in diameter. Approximately 17 Mgha^' was is 10%) of the mean. In Chave et al. (2003), we assessed
held in trees greater than 148 cm in diameter, which cor- the uncertainty on our AGB estimate based on our limited
responds to the largest tree used to construct the pan-trop- sampling of the forest. We showed that the AGB held in
ical model. The BCI forest has a high density of very large the sub-plots of a 50 ha plot is not autocorrelated, even for
trees, and this problem is unlikely to be as important in very small sub-plots: two neighbouring sub-plots of size
other forests. When corrected for very large trees, the pre- lOmXlOmto lOOmxlOOm are not significantly more
dicted AGB estimate for the BCI forest ranged between similar in their AGB stock than two randomly chosen
220 and 315Mgha"' (mean: 277Mgha"', s.d.: plots. We also developed a test of normality for the data.
30 Mg ha^' or 11% of the mean). The 'best estimate' equ- Table 2 gives the first moments of AGB distributions,
ation predicted 260 Mg ha^' for trees of diameter 10 cm together with tests of normality, for sub-plots of varying
or greater, very close to the value reported in Chave et size. This shows that the size of one-quarter of a hectare
al. (2003). is the minimal size such that the normality criterion is

PhU. Trans. R. Soc. Lond. B (2004)

414 J. Chave and others Error propagation for biomass estimates

satisfied in this forest, in agreement with Clark & Clark in some datasets. However, we did consider the error
(2000, section 3.5). Although this figure might vary terms that are unavoidable in ecological studies: for
slightly with the stem density in the plot, it can be taken example, imprecision on the measurements, and on the
as a reasonable guideline. In Phillips et al. (1998), only estimate of wood specific gravity. These are obvious
two of the 68 plots (Queensland, Australia) were less than sources of error, yet we contend that they are not the larg-
one-quarter of a hectare in size, so this constraint in the est ones.
error assessment is not too stringent. Analysing the structure of the existing allometric
regressions, we found an intrinsic source of error not due
(d) Landscape-scale representativity to the size of the census plot, but to the sample available
Several environmental factors, e.g. edaphic and topo- to construct the allometric model itself (i.e. harvested
graphic constraints or climatic gradients, might bias the trees). For the 50 ha plot, an error of greater than 20% on
extrapolation of AGB estimates to the landscape scale. the AGB estimate was due to the choice of the allometric
This is a serious problem if the plot is located in a forest equation. We then corrected these equations by including
patch that is not representative of the surrounding forest. wood specific gravity and showed that AGB varied signifi-
Recent studies suggest that a total sampling size of ca. cantly across diameter classes. This reduced the error to
5 ha, or 20 plots of one-quarter of a hectare allows a land- ca. 13% of the mean. Because none of these equations
scape-scale estimation of the AGB with an error of ± 10% was designed to estimate the AGB of trees beyond a lim-
within 95% confidence (Clark & Clark 2000; Keller et al. ited range, we used a pan-tropical regression model (i.e.
2001). On BCI, we used the AGB estimates from the 200 based on the largest sample size and with the broadest
one-quarter of a hectare sub-plots to assess a similar meas- diameter range) to estimate the AGB of the largest trees.
ure of uncertainty on AGB, and we found that it was 7% This also led to a significant reduction of the error, to ca.
of the mean (Chave et al. 2003). However, this estimate 10% of the mean AGB. Finally, we included wood specific
cannot be thought of as a landscape-scale one, as it results gravity in the allometric equation, which resulted in a sig-
from a single plot measurement. nificant reduction the across-model variation.
The landscape-scale variation of the AGB estimate for This intermodel comparison was supplemented by a
the BCI forest was therefore assessed using the Marena rarefaction study of a pan-tropical dataset of weighted
dataset, a network of plots scattered across the Panama trees. The quality of an AGB estimate should depend on
Canal watershed (Appendix B). The mean BA in trees of the size of the dataset used to construct the allometric
diameter 1 cm or greater for the Marena plots was model. The dataset was a compilation of various literature
31.3 ± 5.5 m^ ha^' (s.d. computed across plots), similar to studies, and the data might be heterogeneous. However,
the within-plot BCI value (30.6 ± 5.1 m^ ha^O- One plot we suspect that the trend is a general one. This implies
was an outlier (plot 39: 56.2 vn? ha^'), owing to the pres- that allometric biomass models based on regional or pan-
ence of a massive Ficus (222 cm diameter) in a small plot. tropical compilations should be preferred to site-specific
We estimated a landscape-scale AGB estimate of models based on small sample sizes.
245 ± 57 Mgha^' (see plot-level results in Appendix B). The uncertainty resulting from the use of small plots
This high variance reflects the variability of environmental (type 3 error) was also considered. We reinforce previous
conditions, and of variations in forest disturbance history. results advocating the use of plots at least 0.25 ha in size
We then performed an ANOVA to assess whether the (Laurance et al 1999; Clark & Clark 2000; Keller et al.
environmental factors accounted for part of this varia- 2001). For the uncertainty related to the representativity
bility. First, we ran a linear model (SAS 8.02, SAS Insti- of a network of such plots in a landscape (type 4 error),
tute Inc.) including annual rainfall, length of the dry we confirmed that an area of ca. 5 ha is necessary to esti-
season, plot age, and geology of the substrate, and exclud- mate the landscape-scale AGB to within 10% of the mean.
ing the outlying plot number 39. The geology factor was As a synthesis of our results, we present a summary in
not significant (w = 49, p = 0.16), but the three other fac- table 3. For a total sampled area of 5 ha, our study reveals
tors were significant {p < 0.002). Plot age alone explained that the cumulated uncertainty on the estimate is ca. 20%
only 14% of the variance, probably because of the broad of the mean, with only a small fraction due to measure-
and somewhat ambiguous definition of this parameter in ment error (assuming unbiased measurement), 10% due
our study. We re-ran the model with the climate data only, to the allometric error and 10% due to the sampling error.
and found that both parameters were significantly corre- With larger plots, one can reduce the sampling uncertainty
lated with the AGB estimate (« = 54, /> < 0.001 for both but not the allometric uncertainty. We stress that such
variables) and explained 41% of the variance. The corre- conclusions may vary, depending on the forest under
lation between rainfall and AGB was positive, whereas study.
that between length of the dry season and AGB was nega-
tive. (b) Comparison ivith other landscape-scale studies
Several studies report results on AGB estimates for for-
ests at the landscape scale. We compare our results to
5. DISCUSSION
three similar studies: one at Los Tuxtlas Biological
(a) Relative importance of the sources of error Station, southern Mexico (Hughes et al. 1999, 2000), one
We have assumed that the AGB stock of a forest is esti- at La Selva Biological Station, Costa Rica (Clark & Clark
mated from sampling plots that have been correctly set up 2000), and one in the Manaus region, central Brazilian
and measured. We did not examine the importance of Amazon (Nascimento & Laurance 2002).
biases such as the measurement of trees at breast height The study done in Mexico used a nested sampling
when the stem is buttressed, although those can be present design to estimate the AGB in four plots ca. 0.79 ha in

Phil. Trans. R. Soc. Lond. B (2004)

 Error propagation for biomass estimates J. Chave and others 415

Table 3. Summary of the sources of error in the AGB estimation of a tropical forest.
(Type 1 error refers to the error made in the estimation of the AGB held in a single tree; this error averages out in plots. Type
2 error is that caused by the choice of the allometric model. Types 3 and 4 are two types of sampling error, which can be
minimized by large-sized, multi-plot, censuses. The reported values are examples for the forests of the Panama Canal watershed.)

s.e.m.
(percentage of
error type the mean) type of data

1. tree level error trees > 10 cm diameter 48 BCI plot•pan-tropical allometric model
trees < 10 cm diameter 78
2. allometric model before p correction 22 BCI plot•eight allometric models
after p correction 13
after large tree correction 11 BCI plot•pan-tropical allometric model
3. within-plot uncertainty 0.1 ha plot 16
0.25 ha plot 10
1 ha plot 5
4. among-plot uncertainty 11 Marena plots•pan-tropical allometric model

total 50 1 ha plots, after p and large 24

tree corrections

size (Hughes et al. 2000; table 5). A mean AGB of summarized in Malhi et al. (2002) is an important step for-
403 ± 50 Mg ha^' was found for this forest. However, they ward in this direction. Using the data of this project, Baker
report a very high density of large trees (as many as 23 et al. (2004a) have reported results for 59 forest plots across
trees greater than 70 cm in diameter per hectare), almost Amazonia, for a total sampled area of ca. 80 ha. They inves-
twice as high as the values commonly found in neotropical tigated three regions: northwestern Amazonia, southwest-
rainforests. Thus, it is possible that diameter measure- em Amazonia, and central and eastern Amazonia, and
ments were not taken above buttresses, which would found significant difference among these regions: using the
greatly overestimate the true AGB (Clark 2002). equation of Chambers et al (2001) they predicted 288, 258
The second study took place in the La Selva forest, and 347Mgha"' in these regions, respectively. Using
characterized by the overdominance of one tree species another equation, however, they consistently found AGB
(Pentaclethra macroloba (Willd.) Kuntze, Fabaceae), which figures 20% lower, confirming the crucial importance of the
constitutes over 35% of the AGB estimate (Clark & Clark choice of the allometric model.
2000). Clark & Clark (2000) used three sampling designs
to assess the landscape-scale variability (three 4-4.4 ha (c) Recommendations
plots, 18 0.5 ha plots and 1170 0.01 ha plots). They report Using allometric models to convert tree diameter data
a low AGB estimate (160.5-186.1 Mg ha"') and an into stand-level AGB estimates often leads to methodolog-
among-site sampling error of 4.2-8.4 Mg ha"', based on ical errors, and we have therefore quantified those errors.
trees of diameter 10 cm or greater. Their sampling error Plots where very many large trees are recorded (e.g. more
is consistent with that found in the present study, and than 15 trees greater than 70 cm in diameter per hectare)
measurement error was minimized as much as possible. should be double-checked. Only large enough stands
Though they did not account for the allometric error, and should be included in the analysis (greater than 0.25 ha).
their allometric equation may significantly underestimate These factors have fortunately been taken into account in
tree AGB, it is possible that the La Selva forest indeed the most recent AGB estimation protocol at the scale of
holds less AGB than the forests of central Panama. a regional forest network (Baker et al. 2004è). Moreover,
The third study took place in the forests of the central allometric equations constructed from very small sample
Amazon, in the BDFF project north of Manaus. This for- sizes and from trees spanning a small diameter range
est is characterized by the rarity of very large trees and the should be avoided, and only equations based on at least
abundance of hardwood species. Nascimento & Laurance 100 weighted trees should be used. Pan-tropical allometric
(2002) reported an AGB estimate of 325 ±31 Mgha"' models are, for the moment, the best available ones. The
(w = 20 1 ha plots). The AGB held in small trees (less than AGB of large trees should be carefully estimated,
10 cm in diameter) represented 21 Mgha"'. This study especially if their diameter exceeds the range for which the
used an allometry comparable to ours, and suggests that use of the allometric equation is valid. In this case, only
the central Amazonian forests hold, on average, ca. 20% a 'best guess' estimate can be produced. Wood specific
more AGB than the forests of central Panama, although gravity should be included in the allometric equation
the density of large trees is much lower. This is a clear wherever possible. AGB held in life forms other than trees
illustration of the importance of including wood specific of diameter 10 cm or greater should also be estimated (in
gravity in pan-tropical allometric models. particular trees less than 10 cm in diameter, lianas, and
In general, the lack of standardization to estimate tropical bamboos, when present). This may represent as much as
forest AGB results in great difficulty in comparing the pub- 10% of the total AGB stock. The landscape-scale varia-
lished values, and we hope that collaborative efforts will bility and issues of spatial autocorrelation of the data
help resolve this problem. The collaborative project should be carefully investigated.

PhU. Trans. R. Soc. Lond. B (2004)

416 J. Chave and others Error propagation for biomass estimates

The authors thank two anonymous reviewers for comments AGB associated with the measurement is measured by
that helped improve the quality of the paper. The US Depart- the variance
ment of Defense Legacy Program supported the census of the
Marena plots at Fort Sherman and Cocoli, both on former US rg/3ln(/)y
military land. The BCI plot has been supported by the AGB^ oAdlniD) HAd\n{H) J^\d\n(p))
National Science Foundation, the Smithsonian Scholarly Stud-
ies Program, the John D. and Catherine T. MacArthur Foun- (7^^1n(/)3ln(/)
dation, the WWF, the Earthwatch Center for Field Studies, + 2
^HDd\niH)d\niD) '
the Géraldine R. Dodge Foundation, and the Alton Jones
Foundation. The authors thank the many field workers who 3ln(/)/3ln(D) is the partial derivative of ln(/) with respect
tagged and measured over a third of a million trees, and the to ln(D), and a^^ = (ôDôH) represents the covariance
Smithsonian Tropical Research Institute for logistical and fin- between D and H. If, for example, the chosen model does
ancial support. This is a scientific contribution from the Center not depend on the total tree height, then d^f = 0, and
for Tropical Forest Science.
height does not contribute to the measurement uncer-
tainty. Most allometric models are of the form
f(D,H,p) = aD-H^p'.
APPENDIX A
For this class of models, the measurement error reads
2 \l/2
The measurement uncertainty on diameter, height and (7M = (AGB>U^ + ^ 84
P
2aß
DH
wood density can be directly added to the model uncer-
tainty: assuming the expected data D^, H^, p^ deviate by For example, if one assumes a 5% uncertainty on the
small amounts ôD, ôH, ôp normally distributed with zero measurement of diameter, 10% uncertainty on both
mean and with standard deviations ao, •H, cTp. This is height and wood density, a correlation coefficient of 0.8
propagated to the AGB estimate through the allometric between diameter and height, and a simple model
model by expanding the model function / in the Taylor f(D,H,p) = aD^Hp, the measurement uncertainty is
series. Moreover, we assume that f(D, H, p) = 21.6% of the AGB. For the model f(D,H>p) = aD", with
Îi(D)f2(H)f^(p). The uncertainty on the estimate of a close to 2.5, it is 12.5% of the AGB.

APPENDIX B

Summary information for the 54 Marena plots, and for the BCI plot. Coordinates: x,y (UTM), elevation (m). Plot
structure: size, number of trees, BA, fraction of BA in trees less than 10 cm in diameter. AGB: total AGB (megagrams
per hectare), fraction of AGB in trees less than 10 cm in diameter. Age type: 1, young secondary forest; 2, old secondary
forest; 3, old-growth forest. Geology: see legend in Pyke et al. (2001), table 2 (except symbol Tue: Marine rocks. Late
Eocene, sandstone and siltstone). Climate: annual rainfall, and length of the dry season (interpolated from weather
stations). UTM, universal transverse mercator coordinates.

number
eleva- of trees basal AGE geo- dry
plot tion size s= 10 area % BA (Mg % AGE age logy rainfall season
number UTMx UTMy (m) (ha) cm (m^ha-') < 10 cm ha-') < 10 cm Pav category type (mmyr" ')(days)

BCI 625 755 1 Oil 569 120 50 21 205 30.6 11.4 269 2.85 0.54 3 Tb 2530 135
1 614 857 1 031 786 20 400 34.2 12.2 257 4.08 0.45 1 Tct 2993 122
2 613 985 1 030 725 100 409 30.3 12.4 271 4.09 0.56 3 Tc 3072 123
3 614 674 1 023 802 180 366 37.9 5.9 297 1.8 0.43 2 Tc 3007 126
4 615 019 1 023 548 180 450 34.1 10.7 229 4.09 0.4 2 Tc 3000 127
5 637 158 1 012 428 40 364 28.4 11.1 243 3.33 0.51 1 Tgo 2414 136
6 637 984 1 012 395 30 480 22.6 16 179 5.66 0.58 2 Tgo 2394 137
7 638 144 1 012 886 60 381 26.9 21 219 6.86 0.56 2 Tgo 2438 136
8 637 732 1 013 699 50 560 29.9 11 229 3.38 0.52 3 pT 2456 136
9 638 365 1 013 754 410 503 32.1 11 265 3.57 0.52 3 pT 2889 136
10 625 402 1 Oil 039 90 403 27.2 16.5 223 5.19 0.55 3 Tcm 2529 135
11 623 291 1 Oil 065 60 449 33.5 9.3 254 3.31 0.46 3 Tcm 2516 135
12 628 587 1 014 891 10 521 24.3 17.7 181 6.33 0.57 1 Tbo 2497 134
13 629 529 1 015 836 55 647 27.9 9.3 194 3.37 0.48 1 Tcm 2576 133
14 625 125 1 012 545 60 381 22.5 12.9 195 4.44 0.58 3 Tcm 2535 134
15 637 861 1 012 976 70 457 26.9 18.8 181 6.99 0.47 2 Tgo 2455 136
16 641 464 1 Oil 328 160 467 28.7 11 235 4.01 0.52 3 pT 2502 138
17 641 108 1 Oil 888 120 464 31.8 10.4 231 3.36 0.45 3 pT 2471 138
18 622 785 1 010 903 58 431 29.3 11.3 215 3.99 0.46 1 Tcm 2511 135
(Continued)

Phil. Trans. R. Soc. Lond. B (2004)

 Error propagation for biomass estimates J. Chave and others 417

number
eleva- of trees basal AGB geo- dry
plot tion size s=10 area % BA (Mg % AGB age logy rainfall season
number UTMx UTMy (m) (ha) cm (m^ha-') < 10 cm ha-') < 10 cm Pav category type (mmyr" OCdays)

19 634 683 1 017 102 160 520 26.1 17.4 199 6.09 0.56 2 pT 2688 134
20 635 754 1 016 123 160 539 26.4 23.3 175 9.13 0.51 2 pT 2658 134
21 643 560 1 010 755 110 405 35.2 8.2 226 3.66 0.36 2 Tgo 2411 139
22 643 599 1 Oil 461 180 508 32.1 8.6 209 4.03 0.38 2 Tb 2514 138
23 645 805 1 008 575 30 590 29.4 14.5 184 7.1 0.44 1 Tic 2248 140
24 645 416 1 008 797 50 568 35.3 13.3 217 6.69 0.39 1 Tic 2280 140
25 632 003 1 003 751 110 600 27.3 24.7 178 9.52 0.56 1 pT 2334 140
26 633 322 1 003 529 50 490 26.8 22.2 218 7.38 0.61 1 pT 2252 140
27 648 907 1 004 027 180 395 34.6 5 258 1.68 0.42 2 Tl 2305 143
28 649 196 1 004 697 160 410 31.1 6.6 226 2.72 0.42 2 Tl 2294 143
29 649 678 993 573 100 357 36.3 6.4 265 2.22 0.4 2 Tb 1969 149
30 649 221 994 670 180 306 28 12.2 205 5 0.43 2 Tb 2096 148
31 637 474 1 034 700 343 498 30.3 15.1 254 4.7 0.58 3 pT 3292 125
32 639 832 1 034 475 340 537 26.7 n.a. 240 n.a. 0.54 3 pT 3293 126
33 647 620 1 038 364 600 0.32 222 38.4 n.a. 329 n.a. 0.51 3 pT 3615 125
34 660 393 1 041 453 210 0.32 174 39.2 n.a. 362 n.a. 0.51 3 pT 3107 126
35 656 577 1 045 987 830 0.32 188 36.2 n.a. 332 n.a. 0.54 3 pT 4002 123
36 661 790 1 039 037 200 0.32 256 34 n.a. 318 n.a. 0.61 3 pT 3029 127
37 688 165 1 030 609 600 0.32 260 33.8 n.a. 276 n.a. 0.55 3 pT 3134 137
38 600 714 962 862 810 0.32 277 37.8 n.a. 314 n.a. 0.53 2 pT 2517 155
39 601 167 966 019 660 0.32 204 56.2 n.a. 464 n.a. 0.49 2 pT 2401 154
40 670 204 1 026 675 160 0.32 135 33.3 n.a. 324 n.a. 0.52 3 pT 2623 135
41 674 032 1 027 111 280 0.32 223 35.7 n.a. 370 n.a. 0.61 3 pT 2743 136
Cl 651 916 993 636 50 281 37.3 5.1 232 2.08 0.32 2 pT 1888 149
C2 651 916 993 736 50 255 33.1 5.6 233 2.06 0.37 2 pT 1890 149
C3 651 916 993 836 50 249 29.2 5.9 219 1.99 0.4 2 pT 1892 149
C4 652 016 993 636 50 294 35.8 4.3 221 1.77 0.32 2 Tb 1887 149
Gl 625 935 1 006 017 55 424 21.4 20.9 169 7.56 0.6 2 Tb 2374 138
G2 625 975 1 006 497 60 390 22.5 15.3 183 5.31 0.54 2 Tb 2389 137
PI 620 342 1 008 821 80 441 26.4 10.2 202 3.3 0.49 Tue* 2531 135
P2 622 483 1 008 890 40 473 29.5 9.8 226 3.4 0.49 Tcm 2468 136
SO 612 610 1 026 067 140 464 30.4 9.7 246 2.93 0.51 Tc 3026 125
SI 612 710 1 026 067 140 531 30.4 11.2 266 3.15 0.58 Tc 3026 125
S2 612 710 1 026 167 140 500 29.5 14.2 263 4.09 0.61 Tc 3028 125
S3 612 710 1 026 267 140 516 32 10.5 277 3.06 0.55 Tc 3030 125
S4 612 710 1 026 367 140 849 31.2 14.7 187 5.92 0.49 2 Tc 3032 125
SH 612 647 1 026 097 140 0.96 503 32.4 11.9 292 3.46 0.58 1 Tc 3030 125

APPENDIX C

Oven-dry wood specific gravity for selected Amazonian tree species. Here, we only report the species for which more
than six different bibliographical sources were available (Détienne et al. 1982; Chudnoff 1984; for other partial lists, see
Chave et al. 2003; Baker et al. 2004a). All of these species have either a pan-neotropical or a pan-Amazonian distribution.

specific
gravity number of s.e.m. relative
family species name (g cm^^) estimates (g cm^^) error (%)

Anacardiaceae Anacardium excelsum (Bertero & Balb. ex Kunth) 0.39 0.04 11.07
Skeels
Anacardiaceae Astronium graveolens Jacq. 0.86 7 0.17 19.68
Anacardiaceae Spondias mombin L. 0.37 13 0.04 10.78
Araliaceae Schefflera morototoni (Aubl.) Maguire, Steyerm. & 0.43 14 0.07 15.83
Frodin (Continued)

Phil. Trans. R. Soc. Lond. B (2004)

418 J. Chave and others Error propagation for biomass estimates

specific
gravity numbiler of s.e.m. relative
family species name (s cm
(g cm^^")
^) estim
estimates (g cm^^) error (%)

Bignoniaceae Jacaranda copaia (Aubl.) D. Don 0.35 18 0.04 10.87

Bignoniaceae Tabebuia serratifolia (Vahl) G. Nicholson 0.92 9 0.04 4.49
Bombacaceae Ceiba pentandra (1^.) Ga.&ctn.. 0.31 14 0.10 30.97
Boraginaceae Cordia alliodora (Ruiz & Pav.) Oken 0.50 0.11 22.25
Burseraceae Tetragastris panamensis (Engl.) Kuntze 0.73 0.03 4.62
Celastraceae Goupia glabra Aubl. 0.72 15 0.05 6.39
Clusiaceae Calophyllum brasiliense Cambess. 0.55 9 0.06 10.46
Clusiaceae Symphonia globulifera L. f. 0.60 15 0.06 9.51
Combretaceae Terminalia amazonia (J. F. Gmel.) Exell 0.68 13 0.06 8.65
Euphorbiaceae Hieronyma alchomeoides Allemao 0.63 10 0.06 8.83
Euphorbiaceae Hura crepitans L. 0.38 10 0.04 9.77
Fabaceae Cedrelinga cateniformis (Ducke) Ducke 0.49 12 0.08 15.38
Fabaceae Dialium guianense (Aubl.) Sandwith 0.87 10 0.14 16.07
Fabaceae Dinizia excelsa Ducke 0.89 7 0.05 5.90
Fabaceae Dipteryx odorata (Aubl.) Willd. 0.93 13 0.04 4.38
Fabaceae Enterolobium schomburgkii (Benth.) Benth. 0.71 15 0.11 15.66
Fabaceae Hymenaea courbaril L. 0.76 12 0.06 8.15
Fabaceae Inga alba (Sw.) Willd. 0.61 7 0.05 8.00
Fabaceae Parkia péndula (Willd.) Benth. ex Walp. 0.54 10 0.15 28.02
Fabaceae Pentaclethra macroloba (Willd.) Kuntze 0.54 8 0.09 16.67
Fabaceae Pseudopiptadenia suaveolens (Miq.) J. W. Grimes 0.69 9 0.11 16.02
Flacourtiaceae Laetia procera (Poepp.) Eichler 0.64 11 0.06 9.02
Lauraceae Mezilaurus itauba (Meisn.) Taub, ex Mez 0.73 9 0.04 5.15
Lauraceae Sextonia rubra (Mez) van der Werff 0.55 11 0.07 12.46
Lecythidaceae Bertholletia excelsa Bonpl. 0.63 0.05 8.12
Lecythidaceae Couratari guianensis Aubl. 0.53 0.04 7.84
Lecythidaceae Eschweilera coriácea (DC.) S. A. Mori 0.83 9 0.06 7.23
Lecythidaceae Lecythis zabucajo Aubl. 0.86 7 0.04 4.35
Meliaceae Carapa guianensis Aubl. 0.53 13 0.07 13.14
Meliaceae Cedrela odorata L. 0.42 0.04 9.99
Meliaceae Guarea guidonia (L.) Sleumer 0.62 0.09 14.30
Moraceae Bagassa guianensis Aubl. 0.69 7 0.05 7.19
Moraceae Brosimum guianense (Aubl.) Huber 0.89 7 0.14 15.58
Moraceae Brosimum parinarioides Ducke ssp. parinarioides 0.60 7 0.13 21.69
Moraceae Brosimum rubescens Taub. 0.88 10 0.08 9.53
Moraceae Brosimum utile (Kunth) Pittier ssp. ovatifolium 0.49 10 0.08 16.78
(Ducke) ce. Berg
Moraceae Clarisia racemosa Ruiz & Pav. 0.58 17 0.07 11.97
Moraceae Madura tinctoria (L.) D. Don ex Steud. 0.78 0.09 11.59
Myristicaceae Virola sebifera Aubl. 0.45 0.06 12.92
Olacaceae Minquania guianensis Aubl. 0.75 0.08 10.89
Sapotaceae Manilkara bidentata (A. DC.) A. Chev. 0.87 9 0.04 4.70
Sapotaceae Manilkara huberi (Ducke) A. Chev. 0.92 9 0.04 3.86
Simaroubaceae Simarouba amara Aubl. 0.38 19 0.04 10.65
Tiliaceae Apeiba petoumo Aubl. 0.31 9 0.06 19.90
Vochysiaceae Erisma uncinatum Warm. 0.51 12 0.05 8.98

REFERENCES Baker, T. R. (and 17 others) 2004è Increasing biomass in

Amazonian forest plots. Phil. Trans. R. Soc. Lond. B 359,
353-365. (DOI 10.1098/rstb.2003.1422.)
Araújo, T. M., Higuchi, N. & Carvalho Jr, J. A. 1999 Com- Baskerville, G. L. 1972 Use of logarithmic regression in the
parison formulae for biomass content determination in a estimation of plant biomass. Can. J. Forest. 2, 49-53.
tropical rain forest site in the state of Para, Brazil. Forest Ecol. Brown, S. 1997 Estimating biomass and biomass change of tropical
Mngmt 111, 43-52. forests. Rome: PAO.
Baker, T. R. (and 17 others) 2004a Variation in wood density Brown, S. 2002 Measuring, monitoring, and verification of
determines spatial patterns in Amazonian forest biomass. carbon benefits for forest-based projects. Phil. Trans. R. Soc.
Global Change Biol. 20, 51-72. Lond. A 360, 1669-1683. (DOI 10.1098/rsta.2002.1026.)

Phil. Trans. R. Soc. Lond. B (2004)

 Error propagation for biomass estimates J. Chave and others 419

Brown, S. & Lugo, A. E. 1992 The storage and production of Feamside, P. M. 1997 Wood density for estimating forest
organic matter in tropical forests and their role in the global biomass in Brazilian Amazonia. Forest Ecol. Mngmt 90, 59-
carbon cycle. Biotropica 14, 161-187. 87.
Brown, S., Gillespie, A. & Lugo, A. 1989 Biomass estimation Foley, J. A., Levis, S., Costa, M. H., Cramer, W. & Pollard, D.
methods for tropical forests with applications to forest inven- 2000 Incorporating dynamic vegetation cover within global
tory data. Forest Sei. 35, 881-902. climate models. Ecol. Applic. 10, 1620-1632.
Bugmann, H. K. M. & Solomon, A. M. 2000 Explaining forest Higuchi, N., Santos, J., Ribeiro, R. J., Minette, L. & Biot, Y.
composition and biomass across multiple biogeographical 1998 Biomassa da parte aerea da vegetaçâo da ñoresta
regions. Ecol. Applic. 10, 95-114. tropical úmida de terra-firme da Amazonia brasileira. Acta
Carvalho Jr, J. R., Santos, J. M., Santos, J. C, Leitao, M. M. & Amazon. 28, 153-166.
Higuchi, N. 1995 A tropical forest clearing experiment by Houghton, R. A., Lawrence, K. T., Hackler, J. L. & Brown, S.
biomass burning in the Manaus region. Atmos. Environ. 29, 2001 The spatial distribution of forest biomass in the Brazil-
2301-2309. ian Amazon: a comparison of estimates. Global Change Biol.
Chambers, J. Q., dos Santos, J., Ribeiro, R. J. & Higuchi, N. 7, 731-746.
2001 Tree damage, allometric relationships, and above- Hughes, R. F., Kauffman, J. B. & Jaramillo, V.J. 1999
ground net primary production in central Amazon forest. Biomass, carbon, and nutrient dynamics of secondary forests
Forest Ecol. Mngmt 152, 73-84. in a humid tropical region of Mexico. Ecology 80, 1892-
Chave, J., Riera, B. & Dubois, M. A. 2001 Estimation of 1907.
biomass in a neotropical forest of French Guiana: spatial and Hughes, R. F., Kauffman, J. B. & Jaramillo, V. J. 2000 Eco-
temporal variability. J. Trop. Ecol. 17, 79-96. system-scale impacts of deforestation and land use in a
Chave, J., Condit, R., Lao, S., Caspersen, J. P., Foster, R. B. & humid tropical region of Mexico. Ecol. Applic. 10, 515-527.
Hubbell, S. P. 2003 Spatial and temporal variation in Hurtt, G. C, Moorcroft, P. R., Pacala, S. W. & Levin, S. A.
biomass of a tropical forest: results from a large census plot 1998 Terrestrial models and global change: challenges for
in Panama. J. Ecol. 91, 240-252. the future. Global Change Biol. 4, 581-590.
Chudnoff, M. 1984 Tropical timbers of the world. Washington, Keller, M., Palace, M. & Hurtt, G. 2001 Biomass estimation
DC: USDA Forest Service. in the Tapajos National Forest, Brazil; examination of sam-
Clark, D. A. 2002 Are tropical forests an important carbon pling and allometric uncertainties. Forest Ecol. Mngmt 154,
sink? Reanalysis of the long-term plot data. Ecol. Applic. 12, 371-382.
3-7. Ketterings, Q. M., Coe, R., van Noordwijk, M., Ambagau,
Clark, D. B. & Clark, D. A. 2000 Landscape-scale variation in Y. & Palm, C. A. 2001 Reducing uncertainty in the use of
forest structure and biomass in a tropical rain forest. Forest allometric biomass equations for predicting above-ground
Ecol. Mngmt 137, 185-198. tree biomass in mixed secondary forests. Forest Ecol. Mngmt
Clark, D. A., Brown, S., Kicklighter, D., Chambers, J. Q., 146, 199-209.
Thomlinson, J. R. & Ni, J. 2001a Measuring net primary Klinge, H. & Rodrigues, W. A. 1973 Biomass estimation in a
production in forests: concepts and field methods. Ecol. central Amazonian rain forest. Acta Cien. Venezoalana 24,
Applic. 11, 356-370. 225-237.
Clark, D. A., Brown, S., Kicklighter, D., Chambers, J. Q., Laurance, W. F., Fearnside, P. M., Laurance, S. G., Delamon-
Thomlinson, J. R., Ni, J. & Holland, E. A. 200le Net pri- ica. P., Lovejoy, T. E., Rankin-de Merona, J. M., Chambers,
mary production in tropical forests: an evaluation and syn- J. Q. & Gascon, C. 1999 Relationship between soils and
thesis of existing field data. Ecol. Applic. 11, 371-384. Amazon forest biomass: a landscape-scale study. Forest Ecol.
Condit, R. 1998 Tropical forest census plots. Berlin: Springer. Mngmt 118, 127-138.
Condit, R., Hubbell, S. P. & Foster, R. B. 1993 Identifying Lescure, J.-P., Puig, H., Riera, B., Leclerc, D., Beekman, A. &
fast-growing native trees from the neotropics using data from Bénéteau, A. 1983 La phytomasse épigée d'une forêt dense
a large, permanent census plot. Forest Ecol. Mngmt 62, en Guyane française. Acta Oecol./Oecol. Gen. 4, 237-251.
123-143. MacDicken, K. G. 1997 A guide to monitoring carbon storage in
Condit, R. (and 12 others) 2002 Beta-diversity in tropical for- forestry and agroforestry projects. Forest Carbon Monitoring
est trees. Science 295, 666-669. Program, Winrock International Institute for Agricultural
Condit, R., Aguilar, S., Hernandez, A., Perez, R., Lao, S., Development.
Angehr, G., Hubbell, S. P. & Foster, R. B. 2004 Tropical Malhi, Y. & Grace, J. 2000 Tropical forests and atmospheric
forest dynamics across a rainfall gradient and the impact of carbon dioxide. Trends Ecol. Evol. 15, 332-337.
an El Niño dry season. J. Trop. Ecol. (In the press.) Malhi, Y. (and 27 others) 2002 An international network to
Cramer, W. (and 16 others) 2001 Global response of terrestrial understand the biomass and dynamics of Amazonian forests
ecosystem structure and function to CO2 and climate (RAINFOR). J. Veg Sei. 13, 439-450.
change: results from six dynamic global vegetation models. Moorecroft, P., Hurtt, G. & Pacala, S. W. 2001 A method for
Global Change Biol. 7, 357-373. scaling vegetation dynamics: the ecosystem demography
Cramer, W., Bondeau, A., Schaphoff, S., Lucht, W., Smith, model. Ecol. Monogr. 71, 557-586.
B. & Sitch, S. 2004 Tropical forests and the global carbon Nascimento, F. E. M. & Laurance, W. F. 2002 Total above-
cycle: impacts of atmospheric carbon dioxide, climate ground biomass in central Amazonian rainforests: a land-
change and rate of deforestation. Phil. Trans. R. Soc. Lond. scape-scale study. Forest Ecol. Mngmt 168, 311-321.
B359, 331-343. (DOI 10.1098/rstb.2003.1428.) Olson, J. S., Watts, J. A. & Allison, L. J. 1983 Carbon in live
Détienne, P., Jacquet, P. & Mariaux, A. 1982 Manuel d'Ident- vegetation of major world ecosystems. Environmental Sciences
ification des Bois Tropicaux, Tome 3, Guyane Française. Nog- Division publication no. 1997. Oak Ridge: Oak Ridge
ent-sur-Marne, France: Centre Technique Forestier National Laboratory.
Trópica. Overman, J. P. M., Witte, H. J. L. & Saldarriaga, J. G. 1994
Fearnside, P. M. 1996 Amazonian deforestation and global Evaluation of regression models for above-ground biomass
warming: carbon stocks in vegetation replacing Brazil's determination in Amazon rainforest. J. Trop. Ecol. 10,
Amazon forest. Forest Ecol. Mngmt 80, 21-34. 207-218.

Phil. Trans. R. Soc. Lond. B (2004)

420 J. Chave and others Error propagation for biomass estimates

Ovington, J. D. & Olson, J. S. 1970 Biomass and chemical con- Suzuki, E. 1999 Diversity in specific gravity and water content
tent of El Verde lower montane rain forest plants. In A trop- of wood among Bornean tropical rainforest trees. Ecol. Res.
ical rain forest: a study of irradiation and ecology at El Verde, 14, 211-224.
Puerto Rico, vol. TID 24270 (ed. H. T. Odum & R. F. Ter-Mikaelian, M. T. & Korzukhin, M. D. 1997 Biomass
Pigeon), pp. 53-77. Springfield, VA: Clearinghouse for Fed- equations for sixty-five North American tree species. Forest
eral Scientific Technical Information. Ecol. Mngmt 97, 1-24.
Phillips, O. L. & Gentry, A. H. 1994 Increasing turnover Wiemann, M. C. & Williamson, G. B. 2002 Geographic vari-
through time in tropical forests. Science 263, 954-958. ation in wood specific gravity: effects of latitude, tempera-
Phillips, O. L. (and 10 others) 1998 Changes in the carbon ture, and precipitation. Wood Fiber Sei. 34, 96-107.
balance of tropical forest: evidence from long-term plots. Yamakura, T., Hagihara, A., Sukardjo, S. & Ogawa, H. 1986a
Science 282, 439-442. Aboveground biomass of tropical rain forest stands in
Phillips, O. L. (and 13 others) 2002a Changes in the biomass Indonesian Borneo. Vegetatio 68, 71-82.
of tropical forests: evaluating potential biases. Ecol. Applic. Yamakura, T., Hagihara, A., Sukardjo, S. & Ogawa, H. 1986è
12, 576-587. Tree size in a mature dipterocarp forest in Sebulu, East Kali-
Phillips, O. L. (and 17 others) 2002è Increasing dominance of mantan, Indonesia. Southeast Asian Stud. 23, 452-478.
large lianas in Amazonian forests. Nature 418, 770-774.
Pyke, C. R., Condit, R., Aguilar, S. & Lao, S. 2001 Floristic GLOSSARY
composition across a climatic gradient in a neotropical low-
land forest. J. Veg. Sei. 12, 553-566. AGB: above-ground biomass
Shell, D. 1995 A critique of permanent plot methods and BA: basal area
analysis with examples from Budongo forest, Uganda. Forest BCI: Barro Colorado Island
Ecol. Mngmt 11, 11-34. BGB: below-ground biomass

Phil. Trans. R. Soc. Lond. B (2004)