1 | INTRODUCTION
In many environments, episodic disturbance events result in the
loss of some or all of the aboveground biomass of a plant. A clas-
sic example of a high- disturbance environment is provided by the
tropical savanna biome, where fire and megafaunal assemblages
of browsers such as elephant and giraffe frequently topkill or se-
verely prune trees (Hoffmann et al. 2012; Holdo, 2006a; Pellew,
1983). In savannas, many tree species readily resprout from root-
stocks following topkill, and repeated cycles of topkill and re-
sprouting may keep individual trees in a “disturbance trap” (Grady
& Hoffmann, 2012; Trapnell, 1959). Given that whether topkill
occurs or not (especially in response to fire) is strongly related to
pre- disturbance tree size (Higgins et al. 2000; Holdo, 2006a), indi-
viduals that can grow fast enough between disturbance events to
avoid topkill have a higher likelihood of escaping this cycle (Bond
Received: 16 April 2020
Revised: 4 December 2020
Accepted: 10 January 2021
Interspecific variation in post- disturbance growth responses of
a savanna tree community and its implications for escaping the
Julienne E. NeSmith1 | Wayne Twine2 | Ricardo M. Holdo1,2
© 2021 The Association for Tropical Biology and Conservation
1Odum School of Ecology, University of
Georgia, Athens, G A, USA
2School of Animal Plant and
Environmental Scie nces, Unive rsity of
the Witwatersrand, Johannesburg, South
Ricardo M. Holdo, Odum School of
Ecology, University of Georgia, At hens,
GA 306 02, USA.
Email: rholdo@uga .edu
Andrew W. Mellon Foundation
Associate Editor: Je nnifer Power s
Handling Editor: Anand Osuri
Vegetation states in savannas are highly sensitive to tree growth rates, which deter-
mine whether individual trees can “escape” periodic disturbances. Resprouting trees
have lopsided shoot:root ratios and are often multi- stemmed, and these variables can
modify post- disturbance growth rates and therefore the probability of escape. To
date, few studies have systematically examined the implications of interspecific varia-
tion in these factors for escape. We conducted a two- year field experiment across 16
tree species in a South African lowveld savanna to quantify growth metrics following
topkill. We examined the dependence of growth on pre- disturbance stem size and
the relationship between growth rate and the tendency of trees to produce a few
large vs. many small resprouts following disturbance. We found that resprout growth
was strongly influenced by pre- disturbance size, but the strength of this relationship
did not vary across species. In contrast, our results showed that fast- growing species
tended to allocate resources toward a few dominant stems, while slow- growing spe-
cies allocated new biomass towards many smaller stems. Tree species that produced
a few large stems also tended to produce individual stems that were tall and thin,
further suggesting that the “few large vs. many small” axis is linked to intrinsic species
attributes. These findings have implications for understanding how interspecific vari-
ation in savanna tree communities may influence their ability to escape disturbance
escape, fire trap, multi- stemmed, persistence, savanna vegetation dynamics
NESMIT H ET al.
& Midgley, 2001; Midgley & Bond, 20 01). It has been shown that
the likelihood of “escape” is highly sensitive to tree growth rate
following disturbance (Higgins et al. 20 00; Wakeling et al. 2011).
Des pite the ove rriding importance of g rowth, however, the ex tent
to which post- disturbance grow th rates var y across species— and
the drivers of this variation— has not been studied systematically
across entire tree communities.
Following topkill, tree shoot:root ratios are altered dramatically
(Holdo, 20 06a; Rutherford, 1981). A topkilled tree has access to an
established root infrastructure, which allows for ready access to soil
resources and stored reserves (Bond et al. 2003; Holdo, 2006b). The
larger the size of the tree prior to topkill, the greater the potential ac-
cess to resources to drive resprouting and stem growth. It has in fact
been well established that growth following resprouting is positively
correlate d with pre- disturbance stem size (Grady & Hoffmann, 2012;
Holdo, 2006b; Schafer & Just, 2014) and that this relationship can
hold vegetation in a persistent, stable fire trap (Grady & Hoffman,
2012). It is unclear, however, how post- disturbance growth rates
and their dependence on pre- disturbance size vary across species
and the consequences of this interspecific variation for escape. For
example, if some species can leverage existing resources more ef-
fectively than others to resprout rapidly following disturbance, they
may be less prone to disturbance traps.
In addition to pre- disturbance size, growth rates following re-
sprouting have the potential to be affected by the number of re-
sprouts produced. Resprouting trees typically produce multiple
new stems, either from root tissue or from surviving shoots (Bond
& Midgley, 2001; Holdo, 2006b). In savannas, many multi- stemmed
trees undergo a thinning process that, in the absence of distur-
bance, culminates in a disturbance- resistant single- stemmed tree
(Bond & Midgley, 2001). The initial resprouts can rapidly develop
new leaves and therefore maximize photosynthesis to ameliorate
the respiratory costs of the surviving root system, but they may also
compete with each other and limit the rapid growth of a single re-
sprout (Chidumayo & Frost, 1996; Childes, 1984). The production of
a few large resprouts vs. many small ones may thus pose a conflict
for a tree recovering from periodic disturbances. A tree developing
fewer but more robust resprouts (a strategy that could be adopted
by inherently fast- growing species) may be better suited to reach
a disturbance- resistant size than a tree apportioning to multiple
smaller resprouts, particularly in fire- prone ecosystems where top-
kill is strongly size- dependent (Higgins et al. 2000). This might be
considered an “escape” strategy. In contrast, a tree with an inher-
ently slow growth rate may invest in many small stems, allowing it
to restore belowground reserves between disturbance events (a
“tolerance” strategy). Although interspecific differences in relative
investment toward a few large vs. many small resprouts have been
noted (Kaschula et al. 2005), to the best of our knowledge, this large
vs. small resprouting “shape” strategy has not previously been quan-
tified for a large number of species. Doing so will allow us to identify
general patterns and allocation strategies in tree communities.
We conducted a field experiment to quantify interspecific
variation in post- disturbance growth of a lowveld savanna tree
communit y in South Africa. We used this experiment to address
our first two objectives: first, to characterize interspecific variation
in post- disturbance growth rate, the time needed to recover pre-
disturbance size, and the relationship between pre- disturbance size
and post- disturbance growth; and second, to investigate whether
the propensity to allocate resources to a few large vs. many small
stems is related to interspecific variation in intrinsic growth rate.
We then used a simulation study based on the experimental data to
address our third objec tive: to quantify the effects of interspecific
variation in post- disturbance growth on escape probabilities under
different fire frequencies.
2 | MATERIALS AND METHODS
2.1 | Study system
We conducted the study between January 2016 and April 2019 at
the Wits Rural Facility (WRF) in Limpopo Province, South Africa.
WRF is characterized by low veld savanna on highly weathered
sandy soils. The area has a mean annual precipitation of 679 mm
(s.d. = 211 mm), mostly concentrated between November and April.
The tree layer is dominated by species in the families Combretaceae
(such as Terminalia sericea and Combretum collinum) and Fabaceae
(Dichrostachys cinerea). The grass layer is dominated by C4 peren-
nial grasses such as Hyperthelia dissoluta and Panicum maximum.
The study site contains impala (Aepyceros melampus), greater kudu
(Tragelaphus strepsiceros), and a few giraffe (Girafa camelopardalis).
The lowveld savanna ecosystem is strongly affec ted by fire, with
typical fire return intervals in nearby Kruger National Park being on
the order of 4.5 years (van Wilgen et al. 2003). The return time of
fires at WRF is between 5 and 10 years, depending on fuel load. Due
to the relatively small size of the propert y and extensive infrastruc-
ture that needs to be protected, managed fires are intentionally lit
under conditions that lead to relatively low fire intensit y.
2.2 | Data collection
In January 2016, we identified candidate trees at 14 focal sites across
WRF. To qualify as an experimental site, a location had to contain at
least five tree species. Potential sites were identified by driving along
existing roads within WRF until locations meeting minimum criteria
were encountered. Clustering our trees into sites allowed us to sep-
arate fixed species effects from varying local environmental effects
without confounding the two. Within a given site, the maximum dis-
tance between any two trees was ~200 m. Across the study area as a
whole, this distance was ~2.6 km. Overall, we mapped with GPS and
measured the diameter of all stems (above the basal swelling) for 493
individual candidate trees of 18 common species. We restricted this
candidate pool to trees with a basal diameter in the range 1– 5 cm (a
basal area range of 0.8– 19.6 cm2). We then used a stratified randomi-
zation procedure to select a final pool of 320 individuals of 16 species
NESMIT H ET al.
(after excluding two specially protected species) distributed across 12
sites. We note that trees exceeding our target size range are common
in the study area, and woody cover appears to be increasing over time
(W. Twine, pers. obs.). The causes of this apparent increase have not
been studied systematically, but could include infrequent, low- intensity
burns and the lack of browsing pressure. We chose this study system
so that we could quantify resprouting in the absence of unwanted dis-
turbances, with the goal of then making inferences for systems beyond
WRF with a range of disturbance frequencies. Accordingly, we focused
our study on trees <5 cm in basal diameter, since this size class is the
most vulnerable to fire (Holdo, 2005), which is a dominant source of
disturbance and habitat heterogeneity in savanna ecosystems (Bond,
2008; van Wilgen et al. 2003).
We measured the maximum height of all trees in our final study
sample. In February 2016, we cut all trees at a height of 25 cm above
ground level, including secondary stems if present. We surrounded
all cut stumps with cages (1 m tall × 0.3 m wide) constructed of
rebar poles and chicken wire to minimize effects of herbivory on
subsequent regrowth. Between Feb- Apr 2018, 24– 26 months post-
treatment, we removed all c ages and measured the diameter of
every resprouted stem, whether originating from the cut stump or
from root suckers. Root suckers were far less common than stump
resprouts and tended to be close to the cut stump and easy to trace.
Of the 320 original trees, we were able to locate 319 trees across 12
sites. We noted that the amount of c anopy cover varied substantially
across sites and individual trees. To account for possible confound-
ing effects of tree neighborhoods and/or light regime, we returned
to the site in Feb- Apr 2019 and took hemispherical photographs (tri-
pod and Nikon D90 camera with 4.5 mm f/2.8 EX DC HSM Nikon F
lens) directly over each tree to estimate the canopy gap fraction. We
took all photographs 1.2 m above ground level with the camera ori-
ented nor th while the sun was at a low angle. To estimate the canopy
gap fraction, we batch- processed hemispherical photographs using
the Hemispherical 2.0 macro plugin (Beckschäfer, 2015) for ImageJ
(Schneider et al. 2012). We calculated mean canopy gap fraction for
each tree. Prior to analysis, we categorized trees on the basis of po-
tential edge effects. We excluded trees that were within 4 m of a
road, termite mound, or power line clearing. Our final dataset con-
sisted of 223 individuals of 16 species across 12 sites (Table S1). The
vast majority of these trees survived cutting and our analysis there-
fore focuses on patterns of post- disturbance growth across species,
rather than survival.
We used the stem allometry data from our initial stem mea-
surements (stem diameter and height) to quantify the tendency of
trees to produce tall thin stems vs. short thick ones. In the absence
of functional trait data (e.g., wood density) for our experimental
trees, we used this variable as a species- level attribute that might
capture a gradient of strategies ranging from “escape” (tall thin
stems) to “resistance” (shor t thick stems). We used ordinary least-
squares (OLS) regressions of log height vs. log diameter to predict
the height (H10 hereafter; Table 1) of a standard stem of 10 mm
diameter. We only estimated H10 for species that exhibited a pos-
itive relationship between stem diameter and height, i.e., cases for
which there was enough power to reliably estimate H10. Our rea-
soning was based on the fact that a positive relationship between
height and diameter should be expected, but was undetectable
(or even negative) in some cases as a result of low sample sizes
or narrow ranges of stem size. We used a conservative threshold
(p < 0.1) to exclude three species lacking a positive diameter– height
relationship: Albizia harveyi, Euclea divinorum, and Gymnosporia sen-
egalensis. We only included single- stemmed trees in the regressions
(N = 175), given that a single height was measured per tree, and in
multi- stemmed trees, this height may not have necessarily corre-
sponded to the largest stem.
2.3 | Data analysis
We conducted all data analyses and simulations in R v. 3.3.1 (R
Development Core Team, 2011). We used a candidate model selec-
tion approach based on AIC (the Akaike Information Criterion) for
regression problems with multiple independent variables (Burnham
& Anderson, 1998), and repor t F- ratio statistics and P- values for bi-
variate tests of association.
We first calculated the summed cross- sectional (basal) area of
every tree from the 2016 (original stem plus any secondary basal
stems) and 2018 (all resprouts) data. We used 2016 tree basal area
(Bpre) and number of stems (Npre) as independent variables cap-
turing original (pre- disturbance) size and architecture (see Table 1
for variable definitions). We used 2018 aggregate basal area (Bpost),
the number of resprouts (Npost), and the basal stem diameter of the
largest resprout (Dpost) as dependent variables. To quantify the ex-
tent to which individual trees and species tend to produce one large
stem vs. many small stems following topkill, we calculated the ratio
Spost = Npost/Dpost as an additional dependent “shape” variable.
TABLE 1 Variable abbreviations used in the study, grouped by
category: initial (pre- disturbance) metrics, resprouting metrics,
stem allometry, and simulation variables
Symbol Category Description
Bpre Initial Pre- disturbance basal area (cm2)
Npre Pre- disturbance number of stems
Bpost Resprouting Post- disturbance (resprouted) basal
Npost Post- disturbance number of stems
Dpost Stem diameter of larges t resprout (cm)
Spost Resprout shape =Npost / Dpost (cm−1 )
H10 Allometry Predicted stem height assuming
standardized 10 mm basal diameter
FRI Simulation Fire return interval (y)
Ptopkill Topkill probability
D Stem diameter of largest resprout at
time t (cm)
B Resprouted basal area at time t (cm2)
NESMIT H ET al.
We examined the relationship between the dependent vari-
ables Bpost, Npost, Dpost, and Spost and the fixed effects Bpre,
Npre, and species using linear mixed models implemented with
lme in the R package nlme (Pinheiro & Bates, 2000), treating site
as a random effect. For each dependent variable, we fit five candi-
date models combining main effects and the interaction between
Bpre and species (Table 2), and compared the fit of the various
models using AIC (Burnham & Anderson, 1998). We tested some
additional models with other interaction effects but they failed
to converge. We log- transformed all of the continuous variables
to meet model distributional assumptions. To test for effects of
tree neighborhoods on growth, we conducted a separate analy-
sis because we lacked gap fraction data for some of our trees.
We added a fixed effect of log- transformed gap fraction to the
best overall Bpost mixed model (which included additive effects
of Bpre and species) and refit the model to the subset of trees
with gap fraction data to test the significance of the tree neigh-
As an alternative growth metric for interspecific comparisons,
we estimated recovery times (i.e., the time needed for a cut tree to
regain its original basal area) by assuming simple growth scenarios
following hypothetical cutting of a range of initial stem basal areas
across species. Our results showed clear relationships between re-
growth in the first two years and initial basal area, but we lacked
information on the pace of regrowth over time. To obtain this in-
formation, growth measurements over multiple time points are re-
quired. We used an unpublished dataset obtained from a resprouting
experiment conducted on three tree species in the nearby Kruger
National Park (KNP) between 2013 and 2016 to determine how
regrowth changes over time (R. Holdo, unpubl. data). We use time
series data from the one species common to the two study sites
(Terminalia sericea) obtained from 24 trees across eight locations in
KNP. The trees were cut in March 2013 and enclosed in cages to
preclude herbivory, and all stems were measured on five different
dates between initial cutting and March 2016. We converted stem
diameters into basal area, and used OLS regressions to fit linear and
quadratic (in terms of days since cutting) models to each tree. We
then tested for significant non- zero values of the coefficients using
OLS, treating individual trees as units of observation. We found no
support for nonlinear growth patterns (i.e., neither accelerating nor
decelerating over time) in KNP Terminalia and use this finding to sup-
port an assumption of a constant growth rate in WRF trees.
We estimated mean time to recovery by dividing a set of hy-
pothetical initial values of Bpre (ranging from 5 to 50 cm2) by the
predicted annual basal area increment for each species (e.g., an indi-
vidual starting with a pre- disturbance size 10 cm2 growing at a rate
of 2 cm2 y−1 would be expected to recover its pre- disturbance basal
area in 5 y). Based on our KNP results suppor ting an assumption
of constant growth across years, we first divided Bpost values by
2 to obtain estimates of annual (as opposed to t wo- year) basal area
increment for our 223 trees. We used a simple linear model (with
no random effects) to obtain coefficients describing the relation-
ship between Bpre and annual basal area increment across species.
We then used these coefficients to predict basal annual increment
for our range of hypothetical Bpre values. In order to incorporate
error in model coefficients into our time to recover estimates, we
used a boot strapping procedure. We generated 1000 samples (with
replacement) of 223 trees each, refitted the linear model to each
sample, and recalculated time to recover values across species
and hypothetical Bpre values. We then estimated means and 95%
confidence intervals for time to recover y across the bootstrapped
Finally, we tested for relationships between standardized values
(adjusted for species size differences and site effects) of Bpost and
Spost and H10 using OLS. To obtain these standardized values, we
calculated species- specific least- squares means values for Bpost
and Spost using the lsmeans R package (Lenth, 2016). The lsmeans
function did not converge for the best overall Spost model, which
included effects of Bpre, Npre, and species (Table 2). We therefore
simplified the analysis to include only individuals that were single-
stemmed prior to cutting (i.e., Npre = 1) and calculated Spost and
Bpost least- squares means values for the model with Bpre and spe-
cies as mixed effect s (Table 2). T his reduced our sam ple size by ~20%.
To account for effects of phylogenetic autocorrelation, we repeated
our regressions of Bpost and Spost vs. H10 using phylogenetic gen-
eralized least- squares (PGLS) regressions with the nlme package,
using the Pagel's lambda correlation structure, which assumes a
Brownian motion evolutionary model. We used a published phylog-
eny of South African woody species for nearby Kruger National Park
(Yessoufou et al. 2013), which included all of our tree species.
Model fixed effects
log(Bpost) log(Dpost) log(Npost) log(Spost)
log(Bpre) + species +
498.5 254.5 329. 2 462. 8
log(Bpre) × species 50 7. 5 2 57. 7 362.3 482.4
log(Bpre) + species 502.5 253 .1 353.9 471.3
log(Bpre) 53 7. 2 288.8 37 7.9 5 07.1
Intercept only 603 .9 315.0 418.1 508.2
Note: The overall best model (lowes t AIC) in each c ase is highlighted in bold.
TABLE 2 Model AIC (Akaike
Information Criterion) values for five
candidate linear mixed models relating
four dependent variables (Bpost, Dpost,
Npost, and Spost) to pre- disturbance
basal area (Bpre), pre- disturbance number
of stems (Npre), and species for trees in
Wits Rural Facility (WRF), South Africa
NESMIT H ET al.
2.4 | Simulation study
To more broadly evaluate how species differences and pre-
disturbance size might affect vegetation dynamics, we simulated
scenarios involving a range of fire frequencies and initial basal areas.
Although cutting does not necessarily capture all of the features of
fire disturbance, our cutting treatment was intended to simulate
topkill, which is a frequent outcome of fire in this system. Other dis-
turbances such as browsing often leave behind a significant portion
of the aboveground biomass. For each of three species (Combretum
collinum, Terminalia sericea, and Dichrostachys cinerea: fast, interme-
diate, slow- growing species, respectively), we simulated 100 years
of growth and disturbance for 1000 individuals of each species.
We started all individuals at the same initial basal area Bpre, which
ranged from 2 to 20 cm2. Assuming an initial topkill event at time 0,
we then allowed each individual to grow annually to size B using the
same functions of Bpre and species that we used to estimate time
to recovery. For each initial value of Bpre, we simulated six fire sce-
narios: mean fire return intervals (FRI) of 1, 2, 3, 5, 10, and 20 years.
Each year, following growth, we used a Bernoulli distribution with
probability 1/FRI to stochastically generate fire events, which then
either occurred or did not for 1000 individuals. To calculate the re-
sulting topkill probability in years with fire (there was no topkill in
fire- free years), we used a published fire mortalit y function from
Holdo (2005), where logit(Ptopkill) = 4.5– 3.4 * log(D), and D equals
stem diameter (in cm). Here, we assumed for simplicity that D was
equivalent to Dpost at any point in time, and that if the largest stem
in a tree died, then all smaller stems died as well. To estimate D in
any given year from B, we used species- specific allometric relation-
ships linking log(Dpost) to log(Bpost), which had adjusted R2 values
of 0.80, 0.85, and 0.53 for C. collinum, T. sericea, and D. cinerea, re-
spectively. Whenever topkill occurred, B reverted back to Bpre. If
at any point in a 100- year growth trajectory B surpassed the origi-
nal basal area Bpre for a given tree, we increased its value of Bpre
to equal the current value of B. At the end of a 100- year run, we
calculated the proportion of trees of each species that had reached
“escape size” for each combination of Bpre and FRI. We assumed
that trees largely escape the fire trap at a value of D = 5 cm, based on
the strong threshold response of topkill to fire around this diameter
observed in Holdo (20 05). We conducted 1000 runs of each simula-
3 | RESULTS
3.1 | Post- disturbance growth rate
Among the 223 trees included in the final analysis (following ex-
clusion of trees due to edge effects), the t wo- year survival rate
was 97.8%. This includes both single- stemmed (which comprised
78.5% of the total) and multi- stemmed trees. For Bpost, the best
fit to the data was provided by a model with additive effects of
Bpre and species (Table 2), with Bpost showing a strong positive
dependence on Bpre (Table S2, Figure 1). The same model pro-
vided the best fit to the data for Dpost (Table 2 and Table S2).
Unlike Bpost and Dpost, Npost and Spost were positively related
to th e num ber of stems pri or to cutt ing ( Table 2 and Table S2) after
controlling for pre- cut basal area Bpre, suggesting that tree size
and number of stems exert independent effects on the number of
resprouts. There was no support for an interaction between Bpre
and species for any of the dependent variables (based on AIC dif-
ferences between models with and without the interaction term;
Table 2), indicating that the slope of the relationship between Bpre
and the dependent variables did not vary across species. In the
case of Bpost and Dpost, the site random effect ter ms for the best
overall model had a standard deviation that was around half as
large as the residual standard deviations, but the site effects were
negligible fo r the best overa ll mode l in the case of Npo st and Spos t
The canopy gap fraction ranged between 9.7% and 100% (mean
=50.2%, median =43.3%, s.d. = 24.5%). Despite this large variation,
a mixed model regression showed no effec t of gap fraction on Bpost
while controlling for effects of Bpre and species (F1 ,185 = 2.06, n.s.).
The Bpost least- squares means values suggested substantial vari-
ation in resprout growth rate after adjusting for initial size differ-
ences and site effects (Figure 2a). Our time- to- recovery estimates
suggested broad variation across species (Figure S1): faster- growing
species, such as C. collinum, are predicted to recover their pre-
cutting basal area in ~8 y for a 5- cm stem (Table S3), whereas slower-
growing species such as Dichrostachys cinerea, S. burkei, and Euclea
divinorum can take almost five times as long.
FIGURE 1 Relationship between post- cutting (resprouted)
basal area Bpost and initial (pre- cutting) basal area Bpre across 16
tree species in Wits Rural Facility (WRF), South Africa. Bpost was
measured two years after cutting.
−2 −1 0123
Log Bpre (cm
Log Bpost (cm2)
NESMIT H ET al.
3.2 | Post- disturbance growth shape
The clear effect of species identit y on growth rate and shape fol-
lowing resprouting suggests that species differed not only in terms
of growth rate post- disturbance, but also in terms of relative alloca-
tion to a few large vs. multiple small resprouts (Figure 2). Spost was
negatively associated with Bpost across species (F1,15 = 5.53, p < 0.05;
Figure 3a). The relationship held when phylogeny was incorporated
into the regression (PGLS: t = 2.38, p < 0.05), suggesting that the re-
sult was not driven simply by evolutionary correlations between the
two variables. We also found a negative relationship between Spost
and H10 (F1,11 = 6.29, p < 0.05), indicating that tree species that tend
to produce short, stubby stems also tend to produce many small re-
sprouts following disturbance (Figure 3b). This relationship held when
phylogeny was taken into account (PGLS: t = 2.56, p < 0.05). However,
we found no relationship between Bpost and H10 (F1,11 = 0.13, n.s.)
FIGURE 2 Site- and size- independent least- squares means values (± 1 SE) for (a) post- cut ting basal area Bpost and (b) resprouting shape
Spost across tree species in WRF. We define shape as the number of resprouted stems Npost divided by the diameter of the largest stem
Dpost. The least- squares means control for site effects.
Bpost lsmean (log cm
Spost lsmean (log cm
FIGURE 3 Relationships between resprouting shape Spost least- squares means values (± 1 SE) and (a) post- cutting basal area Bpost least-
squares means values (± 1 SE) and (b) projected height of a 10- mm stem H10 across tree species in WRF. The least- squares means control for
−1.0 −0.5 0.00.5 1.0
Bpost lsmean (log cm
Spost lsmean (log cm
0.40.6 0.81.0 1.2
Spost lsmean (log cm
NESMIT H ET al.
3.3 | Simulations and escape probability
Our simulation results suggested a strong nonlinear relationship be-
tween pre- disturbance stem size and the probability of escape from
a disturbance trap (Figure 4). The shape of this relationship varied
with fire return interval, and there were marked dif ferences among
the three species (Figure 4). The shape of the relationship between
the probability of escape and pre- disturbance size also strongly de-
pends on fire frequency: at low fire frequency, escape probability
increases somewhat linearly with pre- disturbance size, but as fire
frequency increases, the relationship shifts toward a threshold re-
sponse, that is, escape probability increases rapidly as a function of
Bpre (Figure 4). This highlights the potential import ance of windows
of opportunity for resource acquisition at some point in the lifetime
of trees to ensure a reasonable chance of escaping disturbance traps
at high- disturbance frequencies.
4 | DISCUSSION
Our first aim was to characterize variation in post- disturbance
growth rate and time to recovery across species, and to test for
interspecific variation in the relationship between pre- disturbance
size and post- disturbance growth. We show that there is consider-
able variation in growth following resprouting: our estimates pro-
ject an almost five- fold difference in the expected time needed to
recover pre- disturbance size (starting from a pre- disturbance basal
area of 20 cm2) bet ween the slowest and fastest grower. Supporting
previous work (Bonfil et al. 2004; Grady & Hoffmann, 2012; Schafer
& Just, 2014), we found a strong dependence of post- disturbance
growth on pre- disturbance size (even two years after the original
disturbance event), but there was no evidence for variation in the
magnitude of this dependence across species. Fast- growing spe-
cies might be expected to exhibit a stronger relationship with pre-
disturbance size than slow- growing species, for example, assuming
that fast- growing species are more likely to invest in growth and
escape from future disturbances rather than in storage and the abil-
ity to tolerate disturbance. This finding simplifies efforts to model
community- level responses to repeated disturbance because spe-
cies differences can be treated as simple, additive intercept ef-
fects that can be overlaid onto pre- disturbance size effects. Our
mixed models also suggested that growth rate may be influenced
by unmeasured site- level variables (e.g., soil moisture or nutrient
concentrations), but that stem number and resprout shape are less
influenced by site- level factors and therefore perhaps more tightly
coupled to individual- level variables such as the number of buds
available for producing new resprouts.
Our second aim examined whether the propensity of individual
trees to allocate resources to a few large vs. many small stems is
related to interspecific variation in intrinsic growth rate. We found
clear interspecific differences in our resprouting shape metric,
suggesting that some species (e.g., Combretum collinum) tend to
allocate resources to one or a few large stems, whereas others
(e.g., Senegalia burkei and Dichrostachys cinerea) tend to produce
many small stems (Figure 2b). We found that this shape metric
was related to growth rate, with slower- growing species allocating
more resources to producing more stems rather than investing in
vertical growth. Slow- growing species like D. cinerea and S. burkei
have a relatively lower probability of escaping from a given dis-
turbance regime than faster- growing species such as C. collinum
or Peltophorum africanum (Figures 2a and 4), and invest in the pro-
duction of many small rather than one or a few large stems that
FIGURE 4 Simulated probability of escaping fire in Combretum collinum (Comcol), Terminalia sericea (Terser), and Dichrostachys cinerea
(Diccin) as a function of initial basal area Bpre and mean fire return inter val (FRI). The simulated “escape” size corresponds to a basal area
of 20 cm2 for the largest stem. Values shown represent median escape probabilities across 1000 model runs. The inter- quartile ranges are
omitted in panel c to facilitate visualization.
● ● ●
Proportion > 20 cm2
● ● ●
● ● ● ● ●●
● ● ●
NESMIT H ET al.
have a high probability of sur viving the next disturbance event.
One possible interpretation for this finding is that it may reflect a
strategic difference between fast- and slow- growing species that
have evolved in fire- prone ecosystems. Fast- growing species, hav-
ing a high likelihood of escaping fire, may invest in vertical growth,
whereas slow- growing species, which are intrinsically less likely
to escape, invest in the production of many small stems that can
help accelerate the recovery of stored reserves. This is because
allometric models show that stem mass increases more rapidly
as a function of diameter than leaf mass (Tredennick et al. 2013),
so small- diameter stems have relatively more leaf area than large
ones and therefore more photosynthetic capacity per shoot mass.
The tradeoff would be that a “many small” strategy exposes an
individual to repeated topkill from fire (a tolerance strategy).
We also recognize the possibility that the relationship bet ween
resprouting shape and growth rate could simply reflect the fact that
fast- growing species are farther along a self- thinning trajec tory
than slow- growing ones at the time of harvest. In other words, all
species could conceivably produce similar numbers of resprouts
immediately following topkill, but after a fixed period of time, this
number will have been reduced to a greater extent in fast- growing
than slow- growing species. A counter- argument to this possibilit y
is the finding that resprouting shape is related not only to whole-
tree growth rate, but also to individual stem allometry (i.e., height in
relation to diameter) (Figure 3b), which should be unrelated to the
self- thinning rate. This suggests that resprouting shape is related to
intrinsic functional differences among species. Trees with relatively
stocky stems (e.g., D. cinerea) tend to produce many small resprouts,
whereas trees with more slender stems (e.g., C. collinum) tend to
produce a few tall resprouts. This potential tradeoff is analogous to
the one between lateral and vertical branching patterns identified
in arid vs. savanna ecosystems, respectively, by Archibald and Bond
(2003). A key difference is that the Archibald and Bond (2003) trade-
off occurs within a single species, whereas the resprouting one is
manifested across species.
Our third aim was to investigate how dif ferent hypothetical
disturbance frequencies might affect escape probabilities across
species. Our simulations suggested that additive interspecific differ-
ences in post- disturbance growth alone lead to strong nonlinearities
in the escape probability of fast- vs. slow- growing species as a func-
tion of fire frequency. For example, our model predictions show that
under a 5- year FRI, a 10- cm2 C. collinum stem (~ 3.5 cm in diameter)
would have a 93% probability of eventual escape, T. sericea would
have a 64% escape probability, while for D. cinerea, the probability
approaches zero (Figure 4). By contrast, a 14 cm2 stem exposed to a
20- year FRI would have a far more similar escape probability (>98%)
in all three species. In other words, the effects of pre- disturbance
size tend to dilute the effects of interspecific differences as pre-
disturbance stem size increases, although this dilution effect is
only really manifested in the largest stems. The key implication of
these results is that, across a realistic range of fire return inter vals,
we might expect a great deal of interspecific variation in the likeli-
hood of escape, and this variation might be expected to impact the
community composition of the large tree community. Increasing
fire frequency might be expected to act as a selective filter, trap-
ping slow- growing species into small size classes, and progressively
limiting their recruitment into the canopy, which could become less
diverse than the understory. Just as only the fastest- growing indi-
viduals within a population might escape fire (Wakeling et al. 2011),
the same principle could apply to the fastest- growing species within
a communit y. A possible manifestation of this is suggested by the
structure of savanna tree communities in Serengeti, where there is
a strong mismatch between the species composition of the under-
story and overstory communities (Anderson et al. 2015). Notably,
D. cinerea often dom inat es the understory in Serengeti, but is almos t
completely absent in the overstory (Anderson et al. 2015).
We note that there was far more uncertaint y (i.e., larger con-
fidence intervals) in the simulations for the slower- growing than
the faster- growing species, particularly as pre- disturbance size in-
creased. Part of the reason for this uncertainty derived from our
treatment of error in the growth terms, which was overly conser-
vative. In our model, we generated a growth value for a given tree
in a given year from a distribution of possible growth rates, but this
growth rate was independent from year to year. In reality, faster-
growing trees will tend to grow consistently fast year after year, and
vice versa. In other words, in reality there is probably less growth
variation than we included in our model. The importance of this vari-
ation was demonstrated by Wakeling et al. (2011), who showed that
accounting for the distribution of growth rates across individuals (as
opposed to a species mean) can lead to rapid escape for a few indi-
viduals, with the remainder stuck in the fire trap. We are unable to
explore this facet of inter- individual variation in growth because we
lack repeated measures of growth to show that persistent individual
variation exist s, but we acknowledge that it likely occurs. Our sim-
ulations also ignore any effects of disturbance frequency on stored
carbohydrate depletion. As a result, our model might overestimate
resprout growth rates under high- disturbance frequencies under
which progressive depletion of starch reserves might ultimately slow
growth. We note that it has been shown in these savannas that top-
killed individuals, while experiencing significant reductions in stored
starch reserves compared to control individuals in the short term,
are able to replenish these reserves within a single growing season
(Schutz et al. 2009). Separately, repeated harvests in our system
within a single growing season show that resprout growth only di-
minishes under disturbance rates that would not occur in nature (W.
Twine, unpubl. data).
Overall, our findings have potentially important implications
for under standing the tree species composition of savanna ecosys-
tems under contrasting disturbance regimes. There are widespread
differences among species in the rate of basal area accumulation
between disturbance events. Furthermore, the way in which this
basal area is allocated compounds these differences, with fast-
growing species investing in a few taller stems and slow- growing
species tending to produce many small ones. These differences
could ultimately lead to divergent communities in tall and short
vegetation layers (e.g., the compositional decoupling described in
NESMIT H ET al.
Anderson et al. (2015)), with consequences for species diversity
and function. In par ticular, our results suggest some areas for fur-
ther research. First, more effort should be devoted to exploring
the existence of “resprouting strategies” across species, and in
particular whether a tradeoff occurs between investing in vertical
escape vs. investing in tolerance and allocation to storage. Studies
that quantify patterns of stem production in relation to carbon
allocation to storage as well as interspecific differences in survival
associated with storage patterns could provide a robust test of the
escape/tolerance hypothesis in fire- prone environments. Second,
our work suggests that further insights into observed patterns of
“decoupling” between understory and overstory tree communities
in savannas might be obtained by examining these patterns in re-
lation to fire history and species- specific functional traits related
Keala Cummings collected the initial data for the study. We thank
Cradock Mthabini for assisting with field work, and Cameron Watt
and Wits Rural Facility (University of the Witwatersrand) for facili-
tating access to the study site. We also thank t wo reviewers for their
suggested improvements to this manuscript. SANParks and Skukuza
Scientific Services granted access to Kruger NP. Funding was pro-
vided by the University of Georgia and the Mellon Foundation.
DATA AVAIL ABILI TY STATEMENT
Data available from the Dryad Digital Repositor y ht tps://doi.
org/10.5061/dryad.6hdr7 sr01 (NeSmith et al. 2021).
Julienne E. NeSmith https://orcid.org/0000-0002-6953-0152
Wayne Twine https://orcid.org/0000-0002-4163-198X
Ricardo M. Holdo https://orcid.org/0000-0002-2484-0587
Anderson, T. M., Morrison, T., Rugemalila, D., & Holdo, R. (2015).
Compositional decoupling of savanna canopy and understory
tree communities in Serengeti. Journal of Vegetation Science, 26,
3 8 5 – 3 9 4 .
Archibald, S., & Bond, W. J. (2003). Growing t all vs growing wide: Tree
architecture and allometry of Acacia karroo in forest, savanna, and
arid environments. Oikos, 102, 3– 14.
Beckschäfer, P. (2015). Hemispherical_2.0 – Batch processing hemi-
spherical and canopy photographs with Image J – User Manual.
Georg- August- Universität, Göttingen, Germany.
Bond, W. J. (2008). What limits trees in C4 grasslands and savannas?
Annual Review of Ecology, Evolution and Systematic s, 39, 641– 659.
Bond, W. J., & Midgley, J. J. (2001). Ecology of sprouting in woody plants:
The persistence niche. Trends in Ecology & Evol ution, 16, 45– 51.
Bond, W., & Midgley, J. J. (20 03). The evolutionary ecology of sprout-
ing in woody plants. International Journal of Plant Sciences, 164 ,
Bonfil, C., Cortés, P., Espelta, J. M ., & Retana, J. (2004). The role of dis-
turbance in the co- existence of the evergreen Quercus ilex and
the deciduous Quercus cerrioides. Journal of Vegetation Science, 15,
4 2 3 – 4 3 0 .
Burnham, K. P., & Anderson, D. R. (1998). Model selection and multi-
model inference: A practical information- theoretic approach, 2nd ed.
Chidumayo, E. N., & Frost, P. G. H. (1996). Population biology of miombo
trees. In B. M. Campbell (Ed.), The Miombo in transition: Woodlands
and welfare in Africa (pp. 59– 71). Centre for International Forestry
Childes, S. L. (1984). The population dynamics of some woody species
in the Kalahari sand vegetation of Hwange National Park. PhD
Disser tation. University of the Witwatersrand, Johannesburg.
Grady, J. M., & Hoffmann , W. A. (2012). Caught in a fire trap: Recurring
fire creates stable size equilibria in woody resprouters. Ecology, 93,
Higgins , S . I., Bond, W. J., & Trollope, W. S. W. (20 00). Fire, resprout-
ing and variability: A recipe for gr ass- tree coexistence in savanna.
Journal of Ecology, 88, 213– 2 29.
Hoffmann, W. A ., Geiger, E. L., Gotsch, S. G ., Rossatto, D. R ., Silva, L.
C. R., Lau, O. L., Haridasan, M., & Franco, A. C. (2012). Ecological
thresholds at the savanna- forest boundary: How plant traits , re-
sources and fire govern the distribution of tropical biomes. Ecology
Letter s, 15, 759– 768 .
Holdo, R. M. (2005). Stem mortalit y following fire in Kalahari sand veg-
etation: Effects of frost, prior damage, and tree neighbourhoods.
Plant Ecology, 108, 77– 86.
Holdo, R . M. (200 6a). Elephant herbivor y, frost damage, and topkill
in Kalahari sand woodland savanna trees. Journal of Vegetation
Science, 17, 50 9– 518 .
Holdo, R . M. (2006b). Tree growth in an African woodland savanna af-
fected by disturbance. Journal of Vegetation Science, 17, 369– 378.
Kas chula, S. A., Twine, W. C., & Scholes , M. C. (2 005) . The ef fect of catena
position and stump characteristics on the coppice response of three
savannah fuelwood species. Environmental Conservation, 32, 76– 84.
Lenth, R . V. 2016. Least- squares means: The R Package lsmeans. Journal
of Statistical Software 69, 33.
Midgley, J. J., & Bond, W. J. (2001). A synthesis of the demography of
African acacias. Journal of Tropical Ecology, 17, 871– 886.
NeSmith, J. E., Twine, W., & Holdo, R. M. (2021). Data from: Interspecific
variation in post- disturbance growth responses of a savanna tree
community and its implications for escaping the fire trap. Dryad,
Dataset. https://doi.org/10.5061/dryad.6hdr7 sr01.
Pellew, R. A . P. (1983). The impacts of elephant, giraffe and f ire upon
the Acacia- Tortilis woodlands of the Serengeti. African Journal of
Ecology, 21, 41– 74.
Pinheiro, J. C., & Bates, M. (2000). Mixed- effects models in S and S- PLUS.
R Development Core Team (2011). R: A language and env ironment for sta-
tistical computing. R Foundation for Statistical Computing.
Ruther ford, M. C. (1981). Survival, regeneration and leaf biomass
changes in woody plant s following spring burns in Burkea africana-
Ochna pulchra savanna. Bothalia, 13, 5 31– 5 52 .
Schafer, J. L., & Jus t, M. G. (2014). Size dependency of post- dis turbance
recover y of multi- stem med resprou ting trees. P LoS One, 9, e1056 00.
Schneider, C. A., Rasband, W. S., & Eliceiri, K. W. (2012). NIH Image to
ImageJ: 25 years of image analysis. Nature Methods, 9, 671– 675.
Schutz, A. E. N., Bond, W. J., & Cramer, M. D. J. O. (20 09). Juggling car-
bon: Allocation pat terns of a dominant tree in a fire- prone savanna.
Oecologia, 160, 235.
Trapnell, C. G. (1959). Ecological result s of woodland burning experi-
ments in Northern Rhodesia. Journal of Ecology, 47, 1 2 9– 168 .
Tredennick, A. T., Bentley, L. P., & Hanan, N. P. (2013). Allometric conver-
gence in Savanna trees and implications for the use of plant scaling
models in variable ecosystems. PLoS One, 8, e58241.
van Wilgen, B. W., Trollope, W. S. W., Biggs, H. C., Potgieter, A. L. F., &
Brocket t, B. H. (2003). Fire as a driver of ecosystem variability. In J.
T. Du Toit, K. H. Rogers, & H. C. Biggs (Eds.), The Kruger experience:
NESMIT H ET al.
Ecology and management of Savanna heterogeneity (pp. 149– 170).
Wakeling, J. L., Staver, A. C., & Bond, W. J. (2011). Simply the best: The
transition of savanna s aplings to trees. Oikos, 120, 1448– 1451.
Yessoufou, K., Davies, T. J., Maurin, O., Kuzmina, M., Schaefer, H., van
der Bank, M., & Savolainen, V. (2013). Large herbivores favour spe-
cies diversity but have mixed impacts on phylogenetic community
struc ture in an African savanna ecosystem . Journal of Ecology, 101,
Additional supporting information may be found online in the
Supporting Information section.
How to cite this article: NeSmith JE, Twine W, Holdo RM.
Interspecific variation in post- disturbance growth responses
of a savanna tree community and its implications for
escaping the fire trap. Biotropica. 2021;00:1–10. h t t ps : //d o i .