, 1452 (2008);
et al.Drew Purves,
Predictive Models of Forest Dynamics
www.sciencemag.org (this information is current as of May 19, 2009 ):
The following resources related to this article are available online at
version of this article at:
including high-resolution figures, can be found in the online
Updated information and services,
, 3 of which can be accessed for free:
cites 20 articles
related to this article
A list of selected additional articles on the Science Web sites
2 articles hosted by HighWire Press; see:
This article has been
This article appears in the following
in whole or in part can be found at:
permission to reproduce
of this article or about obtaining
Information about obtaining
registered trademark of AAAS.
is a Science 2008 by the American Association for the Advancement of Science; all rights reserved. The title
Copyright American Association for the Advancement of Science, 1200 New York Avenue NW, Washington, DC 20005.
(print ISSN 0036-8075; online ISSN 1095-9203) is published weekly, except the last week in December, by theScience
on May 19, 2009
10. J. C. Ritchie, G. M. MacDonald, J. Biogeogr. 13, 527
11. J. S. McLachlan, J. S. Clark, P. S. Manos, Ecology 86,
12. M. B. Davis, R. G. Shaw, J. R. Etterson, Ecology 86, 1704
13. R. J. Petit, A. Hampe, Annu. Rev. Ecol. Evol. Syst. 37, 187
14. C. Parmesan, Annu. Rev. Ecol. Evol. Syst. 37, 637
15. R. E. Latham, R. E. Ricklefs, Oikos 67, 325 (1993).
16. J. C. Svenning, Ecol. Lett. 6, 646 (2003).
17. K. J. Willis, A. Kleczkowski, M. New, R. J. Whittaker,
Ecol. Lett. 10, 673 (2007).
18. S. T. Jackson, C. Y. Weng, Proc. Natl. Acad. Sci. U.S.A.
96, 13847 (1999).
19. R. J. Petit, R. Bialozyt, P.Garnier-Géré, A. Hampe,
For. Ecol. Manage. 197, 117 (2004).
20. R. T. Pennington, C. W. Dick, Philos. Trans. R. Soc. Lond.
B Biol. Sci. 359, 1611 (2004).
21. C. W. Dick, E. Bermingham, M. R. Lemes, R. Gribel,
Mol. Ecol. 16, 3039 (2007).
22. B. C. Carstens, C. L. Richards, Evol. Int. J. Org. Evol. 61,
23. J. W. Williams, S. T. Jackson, Front. Ecol. Environ 5, 475
24. We thank M. Chipman and B. Clegg for constructive
comments. Support for this research came from the
Franco-American Fulbright Commission (F.S.H.), NSF
grants (F.S.H. and C.W.D.), the European Union
EVOLTREE Network of Excellence (R.J.P.), the Aquitaine
Regional Government of France (F.S.H. and R.J.P.), and
the Center for Tropical Forest Sciences (C.W.D.). We
thank R. Cheddadi for providing Fig. 1.
Predictive Models of Forest Dynamics
Drew Purves1and Stephen Pacala2
Dynamic global vegetation models (DGVMs) have shown that forest dynamics could dramatically alter
the response of the global climate system to increased atmospheric carbon dioxide over the next
century. But there is little agreement between different DGVMs, making forest dynamics one of the
greatest sources of uncertainty in predicting future climate. DGVM predictions could be strengthened by
integrating the ecological realities of biodiversity and height-structured competition for light, facilitated
by recent advances in the mathematics of forest modeling, ecological understanding of diverse
forest communities, and the availability of forest inventory data.
as much carbon asiscurrently inthe atmosphere,
and forest ecosystems harbor two-thirds of ter-
restrial biodiversity (3). The challenge of predic-
tive forest modeling is to forecast how this
collection of trees will develop in the future, in
response to the many perturbations to which it is
pollution, nitrogen deposition, the loss of polli-
nating and seed-dispersing animals, and the ef-
fects of increased atmospheric CO2, both direct
(the job of a leaf is to convert CO2into plant
material) and indirect (altered climate).
The most exciting recent advance in forest
modeling has been the appearance of dynamic
global vegetation models (DGVMs), which
simulate the distribution, physiology, and bio-
geochemistry of forests and other vegetation at
global scales, under present, historic, or simu-
lated future climates (4). DGVMs have shown
thatfuture changes in global forest carbon storage
system to anthropogenic CO2emissions over the
next century (5). However, because DGVMs
were developed recently, with limited informa-
tion, their predictions are currently highly un-
certain (Fig. 1), making vegetation dynamics one
of the largest sources of uncertainty in Earth sys-
tem models. Reducing this uncertainty requires
here are approximately a trillion canopy
trees on Earth (1) from around 100,000
species (2). The trees store approximately
work on several fronts. For example, physiolog-
data (6), and we need better models of distur-
bances, including fire (7) and land-use change
(8). But more fundamental improvements could
be achieved by incorporating the ecological real-
ities of biodiversity and competition for light. A
recent explosion in forest inventory data might
make this possible.
The only reason to build a trunk—to become
a tree—is to overtop your neighbors and capture
light before they do. This game-theoretic com-
petition for resources is responsible for the enor-
mous amounts of carbon stored in living trees
and in undecomposed organic matter and fossil
fuels, most of which began as wood. Foresters
and forest ecologists have developed individual-
based, height-structured models that can ac-
curately predict productivity (9) and species
composition (10). At every turn, these have re-
vealed nonlinearities in forest dynamics caused
by competition for light. For example, increased
growth leads to increased overtopping, which
loss; with the functions at each stage being non-
linear. In contrast, current DGVMs reduce whole
forested regions to the total biomass in compart-
ments (such as leaves, roots, and trunks), with
simple phenomenological rules for how the car-
bon generated from photosynthesis is allocated
competition among species [or at least among
plant functional types (PFTs)], which needs to be
represented to predict biome boundaries, fol-
lows rules with weak empirical support that
differ among models (11).
Therefore, DGVMs could be substantially
improved by basing them on the height-structured
models developed by foresters and forest ecolo-
gists. But because these models are individual-
based, this would require simulating every tree on
Earth, which would be immensely computation-
ally demanding. A more efficient approach would
be to derive so-called macroscopic equations to
scale correctly from the parameters governing in-
dividual trees to the dynamics of forested regions,
in the same way that the Navier Stokes equations
scale correctly from molecular motion to fluid dy-
namics. Recent progress implies that macroscopic
equations will soon form the basis of DGVMs.
Moorcroft et al. (12) introduced a demographic
method to scale up individual-based forest mod-
and to scale from stands to forests (13).
ulations of individual-based forest models, but
here arises the problem of biodiversity. The (ap-
proximately) 100,000 tree species vary hugely in
properties that drive the carbon cycle, such as
growth, mortality, decomposition of dead wood,
and their dependency on climate. Because of a
reduce biodiversity to a small number of PFTs,
within which all parameters are constant. The
PFTs represent simple morphological and bio-
sus needleaf or tropical versus temperate. But
these aggregations are unlikely to be optimal for
because the among-species differences within
birch) along with subtropical oaks; and evergreen
needleleaf contains cold-adapted spruces and firs
and heat-adapted pines. Even within a forest com-
posed of a single PFT, species parameters typically
mix of species, and hence parameters, found at a
given location is strongly correlated with climate
(15), with obvious implications for modeling the
fail to account for the fact that deleterious effects
canbemitigatedby increasesin thosespeciesbest
1Computational Ecology and Environmental Science Group,
Microsoft Research, Cambridge, UK.2Department of Ecology
and Evolutionary Biology, Princeton University, Princeton,
13 JUNE 2008 VOL 320
Forests in Flux
on May 19, 2009
of extant trees and select for warm-adapted species,
stantial increase in the amount of
However, adding biodiversity
and height-structured competition
into DGVMs would increase the
complexity of models that are al-
ready severely underconstrained.
DGVMs contain large numbers of
parameters, which are hand-
selected from literature values in
predictions to sparse observations
of ecosystem fluxes (such as pro-
ters are beginning to be objectively
estimated with measurements from
long-term dynamics of individuals,
ily, these dynamics are recorded in
more available recently. Forest in-
ventories consist of sample plots
within which trees are measured
The measurements are low-tech:
alive or dead. But the sample sizes
are large, running into millions of
trees in some cases (18, 19). The
few published biogeochemical analyses of forest
inventory data have yielded results with major im-
plications for our understanding of the global car-
bon cycle (18–20).
ments in forest inventories could simply be
summed to provide long-term average carbon
dynamics to compare with DGVM predictions.
But this approach discards most of the infor-
mation in the data. In contrast, if DGVMs were
based around models of individual trees, the
individual growth and mortality records could be
used to directly estimate key tree-level parame-
ters; although few if any inventories contain suf-
ficient information to estimate all parameters,
light, belowground carbon, nutrients, and seed
dispersal. In the low-diversity boreal and temper-
ate zones, the abundance of inventory data might
be sufficient to estimate parameters for every
dominant tree species. In addition to improving
predictions for the carbon cycle, this might allow
realistic predictions for particular species; for
example,climate-induced shifts in species ranges,
which to date have been predicted using only
correlative methods (21).
In high-diversity forests, species-specific pa-
must result in a loss of biological information, evi-
dence suggests that, with the correct aggregation,
this loss could be minimal. This is because wher-
ever parameters have been estimated for different
tree species, they have been found to be subject
to life history tradeoffs: strategic axes appearing
as among-species correlations in parameters (22).
different forest communities, such as the shade-
tolerance spectrum from fast-growing, short-lived
pioneers to slow-growing, long-lived species (22).
These tradeoffs imply that most of the effects of
biodiversity would be retained in models that re-
duced the state of a forest to the distribution of
individual trees along tradeoff axes, regardless of
taxonomic identity. Such models could capture the
effects of biodiversity on select aspects of forest
function (such as carbon dynamics), either by de-
axes or by treating the distribution of species as a
continuum. Either approach would require fewer
correspond closely to the discrete and continuous
lumping techniques used to model heterogeneous
systems of chemical reactions (23).
All of the above add up to a
proven individual-based, height-
plemented at global scales. We are
beginning to understand the trade-
off structure of forest communities
for the first time, we have millions
of observations of individual trees
and parameters of global models.
er properly, the result could be a
istic, better-constrained DGVMs.
A benchmark of success for this
endeavor might be that forest dy-
namics are no longer one of the
References and Notes
1. There are about 3 billion hectares of
forest at 1000 canopy trees per hectare.
2. S. Oldfield, C. Lusty, A. MacKinven,
The World List of Threatened Trees (World
Conservation Press, Cambridge, 1998).
3. Millennium Ecosystem Assessment,
Ecosystems and Human Well-Being:
Biodiversity Synthesis (World Resources
Institute, Washington, DC, 2005).
4. P. R. Moorcroft, Proc. R. Soc. London
Ser. B 270, 1215 (2003).
5. P. Friedlingstein et al., J. Clim. 19, 3337 (2006).
6. M. R. Raupach et al., Global Change Biol. 11, 378 (2005).
7. W. J. Bond, F. I. Woodward, G. F. Midgley, New Phytol.
165, 525 (2005).
8. G. C. Hurtt et al., Global Change Biol. 12, 1208 (2006).
9. J. Landsberg, Can. J. Forest Res. 33, 385 (2003).
10. S. W. Pacala et al., Ecol. Monogr. 66, 1 (1996).
11. W. Cramer et al., Global Change Biol. 7, 357 (2001).
12. P. R. Moorcroft, G. C. Hurtt, S. W. Pacala, Ecol. Monogr.
71, 557 (2001).
13. T. Kohyama, Ecol. Res. 20, 305 (2005).
14. R. Condit et al., Science 313, 98 (2006).
15. T. R. Baker et al., Global Change Biol. 10, 545 (2004).
16. D. Tilman, P. B. Reich, J. M. H. Knops, Nature 441, 629
17. C. Loehle, J. Biogeogr. 25, 735 (1998).
18. J. P. Caspersen et al., Science 290, 1148 (2000).
19. T. R. Baker et al., Philos. Trans. R. Soc. London Ser. B
359, 353 (2004).
20. H. C. Keeling, O. L. Phillips, Global Ecol. Biogeogr. 16,
21. A. J. Davis, J. L. Jenkinson, B. Lawton, B. Shorrocks,
S. Wood, Nature 391, 783 (1998).
22. B. Gilbert, S. J. Wright, H. C. Muller-Landau, K. Kitajima,
A. Hernandez, Ecology 87, 1281 (2006).
24. We thank A. Brändle, D. Coomes, S. Emmott, P. Friedlingstein,
A. Heil, M. Smith, and R. Williams for useful discussion.
Land uptake (GtC/yr)
1850 1900 195020002050 2100
Fig. 1. DGVMs have shown that the terrestrial biosphere could be crucial in
determining the future of Earth’s climate. But this figure [from (5)] shows how
divergent the predictions of DGVMs currently are. For comparison, current anthro-
pogenic CO2emissions are 7.6 ± 0.6 Gt of carbon/year. True DGVMs, with a
carbon cycle but a fixed distribution of PFTs. Some of the variation in Fig. 1 results
from different climate models, but a large spread was also seen when different
DGVMs were run uncoupled from global climate models under a common, fixed
coupled to the Lund-Potsdam-Jena DGVM; UVic-2.7, University of Victoria Earth
system climate model, version 2.7; UMD, University of Maryland coupled carbon-
climate model; HadCM3LC, Hadley Centre coupled climate-carbon cycle model.
VOL 320 13 JUNE 2008
on May 19, 2009