Ecological Modelling 457 (2021) 109681

Contents lists available at ScienceDirect

Ecological Modelling
journal homepage: www.elsevier.com/locate/ecolmodel

Modeling the persistence of plant populations in fragmented ecosystems

Maria C.A. Leite a , Rebecca Sauchuk b , Folashade B. Agusto c , Orou G. Gaoue d,e,f ,
Benito Chen-Charpentier g ,∗
a Department of Mathematics and Statistics, University of South Florida St. Petersburg, St. Petersburg, FL 3370, USA
b Department of Medical Sciences, University of South Florida Mosani College of Medicine, FL 33612, USA
c
Department of Ecology and Evolutionary Biology, University of Kansas, Lawrence, KS 66045, USA
d
Department of Ecology and Evolutionary Biology, University of Tennessee Knoxville, Knoxville, TN 37996, USA
e
Department of Geography, Environmental Management and Energy Studies, University of Johannesburg, APK Campus, Johannesburg, South Africa
f
Faculty of Agronomy, University of Parakou, Parakou, Benin
g
Department of Mathematics, University of Texas at Arlington, Arlington, TX 76019, USA

ARTICLE INFO ABSTRACT

Keywords: Ecosystem fragmentation is one of the main threats to species persistence via habitat reduction and isolation
Forest fragmentation which often lead to species extinctions. A question that has long been of interest is the minimum habitat
ODE models size that can sustain viable populations in fragmented landscape. Despite numerous empirical and theoretical
Linear matrix models
efforts on this topic, most studies fail to address this central question, and our mechanistic understanding of
Population dynamics
and capacity to predict the effects and outcomes associated with fragmentation stressor is still illusive. We
Amazonian herb
Minimum fragment size
develop an ordinary differential equation (ODE) based framework that incorporates the effect of the patch
area on the net population growth rate for a plant species in fragmented ecosystem via a general net growth
function. We investigate the minimum patch area needed to sustain a given plant species. We use data from the
Amazonian herb Heliconia acuminata to test our model. Furthermore, we compare the performance ODE model
and a linear matrix model to predict the observed data. We provide a general formula for a threshold value
for the fragment area, below which a plant population is not viable. For Heliconia acuminata, our ODE-based
model predicts a value for the minimum fragment area of ≈ 0.7 ℎ𝑎, which reflects the observed data and is
smaller than the value obtained using the matrix projection model. Our findings suggest that the Heliconia’s
mortality rate responds more negatively to fragmentation. Furthermore, we found that the ODE-based model
can serve as an alternative to the linear demographic model.

1. Introduction lasting effect on population dynamics as well as community assembly

and associated dynamics (Bruna, 2003; Gagnon et al., 2011; Kolb et al.,
The rate of species extinction and ecosystems degradation is increas- 2010; Zambrano and Salguero-Gómez, 2014; Bruna and Kress, 2002;
ing worldwide due to increasing chronic anthropogenic disturbances Fahrig, 2017; Matesanz et al., 2017; Rybicki et al., 2020; Lucas, 2020).
and of human population growth. Forest fragmentation and associated Most of these studies used linear matrix models (Bolker et al., 2009;
habitat loss and isolation are important causes of known species ex- Caswell, 2001; Cushing, 1998; Morris and Doak, 2003) based on field
tinctions and one of the main threats to species persistence (Andreazzi
data to project population dynamics response to habitat reduction.
et al., 2012; Broadhurst et al., 2008; Bruna, 2003; Brudvig et al., 2015;
Quoting (Bruna, 2003), ‘‘Matrix-based demographic models have
Flaspohler et al., 2010; Gagnon et al., 2011; Jacquemyn et al., 2012;
Kolb et al., 2010; Laurance et al., 2011; Tang et al., 2011; Vaughn become increasingly common in ecology because of their application
et al., 2014; Zambrano and Salguero-Gómez, 2014). Habitat patchiness, from life-history evolution to conservation and there is a broad body
including fragmentation, has been an important concept in ecology for of literature describing their construction, interpretation, assumptions,
over 80 years. For example, early work on the importance of habitat and limitations’’. Linear matrix models can consider the dependence of
heterogeneity by Gause (1934) suggests that persistence of predator– the birth or death rate on the population density but studies including
prey systems depends on the availability of separate refuges for prey. such dependence are not found often in the literature (Silva Matos
Also, there is empirical evidence that reduction in habitat size has long et al., 1999). However, several works demonstrate the existence of such

∗ Corresponding author.
E-mail address: bmchen@uta.edu (B. Chen-Charpentier).

https://doi.org/10.1016/j.ecolmodel.2021.109681
Received 17 March 2021; Received in revised form 20 July 2021; Accepted 22 July 2021
Available online 4 August 2021
0304-3800/© 2021 Elsevier B.V. All rights reserved.
M.C.A. Leite et al. Ecological Modelling 457 (2021) 109681

density dependent processes as well as its role in determining plant

population persistence in fragmented landscapes (Broadhurst et al.,
2008; Jacquemyn et al., 2012; Bruna and Kress, 2002; Jacquemyn
et al., 2003; Pachepsky and Levine, 2011; Lienert, 2004). Furthermore,
fragmentation studies that used these linear matrix models failed to
predict the system responses beyond observed degree of fragmentation
and often neglect the role of path isolation. Consequently, predicting
how fragmentation stressor impacts the plant population viability is
still illusive (Andreazzi et al., 2012; Rybicki et al., 2020; Lucas, 2020;
Didham et al., 2012).
Mathematical modeling with data integration is an important tool
to obtain insights to many aspects of ecosystem sustainability and
management. There exists a substantial body of theoretical studies
focusing on modeling the spread of one population in fragmented Fig. 1. Conceptual diagram of the mathematical model.

ecosystems (Brudvig et al., 2015; Cantrell et al., 2011; Cantrell and

Cosner, 2003; Ferrari et al., 2014; Hastings et al., 2004; Musgrave and
Lutscher, 2014; Murray, 2002; Neubert and Parker, 2004; Williamson, These data sets were collected in Central Amazonia, in the scope of the
1989). These range from ordinary differential equations based models Brazil’s Biological Dynamics of Forest Fragments Project — BDFFP. This
of meta-populations in patches to network models to models based project is one of the largest and longest experimental data collection
on reaction diffusion equations to structural equation models (SEM). endeavors to study the ecological impact of habitat fragmentation (Lau-
These theoretical models aid to understand the influence of forest rance et al., 2011). The experimental sites consist of fragments of size 1,
fragmentation beyond the limits of observed fragmentation examples. 10 hectares, and a continuous forest located in Brazil. The demography
This is critical given that replicated cases of forest fragmentation are and population dynamics of Heliconia acuminata in these patches has
often scarce, and in some cases experimentally manipulated (Laurance been previously studied (Bruna, 2003) using a linear matrix model.
et al., 2011) due to lack of study examples. However, most of these In this paper, we first introduce in Section 2 the mathematical model
theoretical studies often do not provide practical answers to the key and its underlying assumptions. We further describe the data sets and
question of the minimum or optimal fragment size that is needed to present the net population growth rate as well as death rate functions
protect the populations in fragmented ecosystem. and their parameterization in Section 3. In Section 4, we characterize
To further our mechanistic understanding of theoretical and prac- the dynamics of the model and the biological implications of the find-
tical conservation implications of forest fragmentation on species ex- ings. Specifically, we provide the formula to compute the area threshold
tinction, it is critical to develop a modeling framework that addresses and the specific value for the Amazonia herb in the fragmented system.
these issues. We propose a model that uses the flexibility of existing Although, the goal of the work by Bruna (2003) was not to predict
theoretical models on fragmentation by the practical use of linear short term changes in abundance, we have used it to compare the result
matrix models to illustrate how to investigate the minimum fragment obtained from the matrix project model with the results of our ODE
size necessary to maintain viable population. The model consists of a model. In Section 5 we investigate the ability of the ODE model to
single ordinary differential equation (ODE) and explicitly incorporates predict the observed population density of Heliconia acuminata in the
a density dependency on the net growth rate (recruitment rate minus year 2000 in the Amazonia system studied in Bruna (2003). Fig. 1
death rate) of the species population. The multiple mechanisms by shows the conceptual diagram summarizing the development of the
which fragmentation affects the recruitment rate is a complex topic to model.
disentangle. It has been largely investigated by a couple of researchers
(see Bruna, 2003; Brudvig et al., 2015; Jacquemyn et al., 2012; Lienert, 2. Mathematical model for fragmented ecosystem
2004; Bruna et al., 2004; Cramer et al., 2007, references there in) but it
remains elusive. Although it is critical to understand such mechanisms, In this section we develop a mathematical model that will integrat-
we do not address it in this paper. Instead we assume that plant ing experimental data from sites with distinct areas. The model is then
growth is dependent on population density (or biomass density) and used to investigate how fragmentation affects the dynamics of Heliconia
is modeled by a logistic type function. This particular form was chosen acuminata population. Let 𝑏(𝑡) represent the plant density (or density
because a logistic growth rate is one of the most common and simplest of biomass), express in number of plants per area (or units of mass
assumption in population models (see Allen, 2007.). per area), in a fragment of a given area and with carrying capacity
Furthermore, we assume that population net growth rate is affected 𝐾. The temporal dynamics of the population density is modeled by the
by the reduction of patch area through a general net growth function. following ODE with a logistic growth:
This assumption reflects the knowledge in the field that patch size ( ) ( )
𝑑𝑏 𝑏 𝑏
= 𝛽(𝐴)𝑏 1 − = [𝑟(𝐴) − 𝜇(𝐴)] 𝑏 1 − (1)
affects some vital rates via edge effects. How these interact to influence 𝑑𝑡 𝐾 𝐾
the asymptotic growth rate of a population is poorly understood. For where 𝛽(𝐴) = 𝑟(𝐴) − 𝜇(𝐴) is the net growth rate of the population,
instance, Bruna and Oli (Bruna and Oli, 2005) found at least one year 𝑟(𝐴) is the recruitment rate and 𝜇(𝐴) is the natural death rate of the
where the asymptotic growth rate of the Heliconia population was very population.
similar in the fragmented patches and the continuous forest. With model (1) we make the following assumption:
The model is used to broadly address the question: What is the min-
imum patch area needed to sustain any plant population? To provide A.1 The net growth rate 𝛽(𝐴) is assumed to depend on fragment size.
a practical example, we specifically use the model to determine the The specific form of the function 𝛽(𝐴) and its fitting details are
minimum patch size necessary to sustain the populations of Heliconia given in Section 3.
acuminata (Heliconiaceae) in a fragmented landscape. This Amazo- A.2 In Bruna (2003) the size of the continuous forest was not spec-
nian understory herb is about 30–50 cm tall with a maximum size ified. In this paper, we arbitrarily assume that a patch of 100
of about 1.6 m and native to the South American countries namely hectares behaves almost like a continuous forest.
Brazil, Bolivia, Colombia, French Guiana, Guyana, Peru, Suriname, A.3 It is reasonable to expect that there may be large variation in the
and Venezuela. We parameterized our ODE model by fitting the net carrying capacity of each site. However, due to lack of detailed
population growth rate function using the data sets in Bruna (2003). data we assume that the carrying capacity of each site is the

2
M.C.A. Leite et al. Ecological Modelling 457 (2021) 109681

carrying capacity of the site with the largest area. That is, we denoted by (𝑏𝑜𝑖 , 𝑡𝑖 ). A summary of the data used in this paper and a brief
assume that the carrying capacity is equal for all sites. Its value description on how it was adapted from the data published in Bruna
is taken to be the average of the observed value of the density (2003) is presented in Appendix A.1.
for the continuous forest in the year 1999 given in Bruna (2003). Next, we estimate the parameters for the function 𝛽(𝐴) (in Sec-
A.4 We assume that the patches are completely isolate, that is, there tion 3.2) and 𝜇(𝐴) (in Section 3.3). To accomplish this task we use
are no dispersal in or out of patches. This is a critical aspect function cftool in MATLAB – 9.5.0.1049112 (R2018b) to fit the
of this species’ biology as shown in Bruna (2003). In particular, functions to the data just described: (𝜆𝑝𝑖 , 𝑡𝑖 ) and (𝜆𝑜𝑖 , 𝑡𝑖 ). The best-
the study shows that, even though patches are not completely fit curve is assumed to be that which minimizes the sum of squared
isolated, the interactions between the species in the same patch residuals. Since the functions are nonlinear, we use the standard error
are much stronger than those between different patches. of the regression to assess the goodness-of-fit.
A.5 There are no seed bank. That is, there are no seeds that are stored
in the soil that survives for the next life cycles. This assumption
3.2. Dependence of 𝛽 on 𝐴, 𝛽(𝐴)
is consistent with the fact that seeds of tropical plants move
quickly into a ‘‘seedling bank’’ to escape intense pressure from
To establish the function 𝛽(𝐴) that will be parameterized with the
seed predators and pathogens (Bruna, 2002).
data, we assume that it satisfies the following properties: (i) 𝛽(𝐴) is a
A.6 There is no significant vital-death correlation. This means that
continuous and differentiable function on [0, ∞); (ii) when the area of
the probability of dying is not strongly correlated with the re-
the site is close to zero the growth rate approaches −∞. This models
production rate. So it can be assumed that the reproduction and
the fact that when the area of the patch is very small the biomass or
death rates are independent of each other. This assumption is
reasonable for Heliconia acuminata based on discussions in Bruna population density decreases very fast. (iii) When the area is very large
and Kress (2002). the growth rate levels out at a certain value. Two possible functions
A.7 The death rate 𝜇 is assumed to depend on the area. The partic- that satisfy conditions (i)–(iii) are given in Eqs. (2) and (3):
ular form of the function 𝜇(𝐴) and the details of its parameter- 𝐴 − 𝐴𝑐
𝛽(𝐴) = ln(𝜆(𝐴)) = 𝛽𝐹 . (2)
ization is given in Section 3. 𝐴
A.8 The growth of the population density is logistic. and
( ) ( )
𝑐1 𝐴 𝑐1 ∕𝑐2 𝐴
3. Model parameterization 𝛽(𝐴) = ln(𝜆(𝐴)) = ln = ln , (3)
1 + 𝑐2 𝐴 1∕𝑐2 + 𝐴

3.1. Data description Note that the available data contains the values of 𝜆, however, we need
to estimate 𝛽. The relation between 𝜆 and 𝛽 is
The data we use consists of three years of census data from 13
𝛽(𝐴) = ln (𝜆(𝐴)) . (4)
populations of the Amazonian understory herb Heliconia acuminata in
fragments with three distinct areas at Biological Dynamics of Forest Eq. (4) explains why the logarithm appears on both equation (2) and
Fragment Project, located 80 km north of Manaus, Brazil (see Bruna, (3). The parameter 𝐴𝑐 ≥ 0 in Eq. (2) is the critical value of the area for
2003, for details on methods and results). The data was collected in which the sign of the net growth rate changes from a negative value to
four sites with area of 1 ℎ𝑎, three sites with area of 10 ℎ𝑎, and six sites a positive value. We call this point the switching point. This is the point
in the continuous forest, which in this paper we consider to have an at which the plant density switches from increasing to decreasing or
area of 100 ℎ𝑎, (see Appendix A.1). vice-versa and we will show this fact in Section 4.
For each of the sites, Bruna (2003) provided the projected (or Next we proceed with the estimation of the parameters in the
asymptotic) population growth rate values for 1998–1999 and 1999– functions. We divide the fitting into two cases. In one case, we use
2000 transition years. These values are the dominant eigenvalue of the values of the projected 𝜆 for the years 1998–1999 and 1999–2000
each corresponding transition matrix. We called them the projected 𝜆 across the plots of same size and the corresponding areas, that is, we
values and denote them by 𝜆𝑝 . Furthermore, we denote by 𝜆𝑝𝑖 the differ- use the data set (𝜆𝑝𝑖 , 𝐴𝑖 ). In the other case, we utilize the values of
ent values of the projected 𝜆 values corresponding to the two transition the observed 𝜆 for the years 1998–1999 and 1999–2000 across the
years, 1998–1999 and 1999–2000, and fragments with three distinct
plots of same size and the corresponding areas. This is the data set
areas. The transition matrices corresponding to years 1998–1999 and
(𝜆𝑜𝑖 , 𝐴𝑖 ). The values of the data sets are given in Tables A.1 and A.2,
1999–2000 are also given in Bruna (2003).
Appendix A.1. For each data set we parametrize both functions given
Additionally, Bruna (2003) gave the values of the observed asymp-
in equations (2) and (3). There are two parameters to be estimated in
totic population growth rate for each transition year 1998–1999 and
each formula, 𝛽𝐹 , 𝐴𝑐 in Eq. (2) and 𝑐1 , 𝑐2 in Eq. (3). For each data set
1999–2000 for each site with distinct areas. These values were com-
and each of the functions we follow two approaches. In one approach,
puted in Bruna (2003) by dividing the observed plant density at year 𝑡
we fix one parameter and estimate the other. In the other approach, we
by the observed plant density at year 𝑡 − 1. In this paper we call them
estimate both parameters. When fixing one parameter we choose to fix
observed 𝜆 and denote them by 𝜆𝑜𝑖 . Thus, we have two sets of data ( )
𝛽𝐹 = ln 𝑐1 ∕𝑐2 = ln 𝜆𝐹 because it can be estimated from the available
points (𝜆, fragment area): a data set consisting of projected values of 𝜆
data. It is estimated by averaging over all values of the asymptotic
denoted by (𝜆𝑝𝑖 , 𝐴𝑖 ) and a data set consisting of observed 𝜆 denoted by
growth rate (𝜆) of the continuous forest (sites with area equal to 100 ℎ𝑎)
(𝜆𝑜𝑖 , 𝐴𝑖 ).
presented in Appendix A.1. Then compute the natural logarithm of
The plant mortality rate for each fragment was computed from
the survival probability in the transition matrices and is listed in the resulting value. The values for the estimate of 𝛽𝐹 using the two
Appendix A.1. data sets (𝜆𝑝𝑖 , 𝐴𝑖 ) and (𝜆𝑜𝑖 , 𝐴𝑖 ) are given in Table A.1, Appendix A.1,
The values of the population density available correspond to the where they are denoted by 𝛽𝐹 𝑝 and 𝛽𝐹 𝑜 , respectively. In particular,
year 1999. The methodology for computing the values of the density when performing the fitting with the data set (𝜆𝑝𝑖 , 𝐴𝑖 ) and fixing one
in 1998 and 2000 at each site is given in Appendix A.1. For each year parameter, we set
1998, 1999, and 2000, we obtained two values of plant density. The ( )
ln 𝑐1 ∕𝑐2 = 𝛽𝐹 = ln 𝜆𝐹 ≈ 0.0214. (5)
first set of values are derived from the projected 𝜆𝑝 , which we denote
by (𝑏𝑝𝑖 , 𝑡𝑖 ) where 𝑖 = 1, 2, 3 corresponds to years 𝑡1 = 1998, 𝑡2 = 1999, and Then, we estimate the parameter 𝐴𝑐 by parametrizing the function
𝑡3 = 2000. The second set of values are derived from the observed 𝜆𝑜 , in Eq. (2) and estimate 𝑐2 by fitting the function in Eq. (3). When

3
M.C.A. Leite et al. Ecological Modelling 457 (2021) 109681

Fig. 2. Fitting results for 𝛽(𝐴) functions with the area expressed in ℎ𝑎 and 𝛽 in year−1 . (a) using projected data (𝜆𝑝𝑖 , 𝐴𝑖 ) and Eq. (2). The curve depicted with black dashed line
is obtained when both parameters, 𝐴𝑐 and 𝛽𝐹 , are estimated with Matlab cftool. The solid blue line is the fitted curve with 𝛽𝐹 = 0.0214 fixed and the value of 𝐴𝑐 estimated via
cftool. (b) using projected data (𝜆𝑝𝑖 , 𝐴𝑖 ) and Eq. (3). The curve represented with solid blue line is the fitted curve with 𝛽𝐹 = ln(𝑐1 ∕𝑐2 ) = 0.0214 and the value of 𝑐2 estimated via
cftool. The black dashed line curve is the fitted curve when both 𝑐1 and 𝑐2 are estimated with cftool package. (c) and (d) graphs similar to (a) and (b), respectively but using
observed data (𝜆𝑜𝑖 , 𝐴𝑖 ) instead of projected data.

Table 1 significant difference in the goodness-of-fit associated to the fitting of

Parameters values and the corresponding Fit Standard Error (RMSE ) when fitting the two distinct function forms given in Eqs. (2) and (3). This feature
function 𝛽(𝐴). When the value of 𝛽𝐹 is not estimated, either using the function
presented in Eq. (2) or equation (3), the values of 𝛽𝐹 projected and 𝛽𝐹 observed are
is also visible in Fig. 2.
taken to be 0.0214 and 0.1066, respectively (see Eqs. (5) and (6).). From Table 1, column 6, we see that the value of 𝛽𝐹 estimated
Function Par. Unit With 𝜆𝑝 projected with 𝜆𝑜 observed using Matlab cftool is 0.1032 when the function in Eq. (2) is adopted
and it is ln(28.39∕25.60) = 0.1034 when the fitting is done with the
Value 𝑅𝑀𝑆𝐸 Value 𝑅𝑀𝑆𝐸
function in Eq. (3). These values differ by less than 3.2% from the
𝐴𝑐 ha 1.753 0.02178 0.6547 0.06573
value estimated directly from the data (𝜆𝑜𝑖 , 𝐴𝑖 ), which is 𝛽𝐹 = 0.1066
(2) 𝛽𝐹 fixed .
(see Eq. (6)). On the other hand, when the fitting is performed with
𝐴𝑐 ha 2.406 0.02062 0.3706 0.06496
the data set (𝜆𝑝𝑖 , 𝐴𝑖 ) (column 4, Table 1) 𝛽𝐹 is estimated to be equal
𝛽𝐹 year−1 0.01088 0.1032
0.01088 and ln(38.02∕37.61) = 0.01084 when the fitting is based on the
𝑐2 ha−1 28.5 0.02184 13.84 0.06573
(3) 𝑐1 fixed .
function (2) and (3), respectively. However, the value of 𝛽𝐹 estimated
from the data (𝜆𝑝𝑖 , 𝐴𝑖 ) is 0.0214 (see Eq. (5)), which is ≈ 50% higher
𝑐1 (year×ha)−1 38.02 0.02061 28.39 0.06496
𝑐2 ha−1 37.61 25.60
than the values obtained by fitting with Matlab cftool.
Based on these results, we select the function in Eq. (2) to model the
dependence of 𝛽 on the area of the fragment. This function is fed into
model (1) leading to two models: one model parameterized by (𝜆𝑝𝑖 , 𝐴𝑖 ),
using the data values (𝜆𝑜𝑖 , 𝐴𝑖 ) we proceed similarly but equation (5) is and the other parameterized by (𝜆𝑜𝑖 , 𝐴𝑖 ). In both models, 𝐴𝑐 is replaced
replaced by by the corresponding value estimated via Matlab cftool and given in
( ) Table 1. Similarly, to the parameter 𝛽𝐹 is assigned the corresponding
ln 𝑐1 ∕𝑐2 = 𝛽𝐹 = ln 𝜆𝐹 ≈ 0.1066 (6)
value estimated directly from the data and given in Eqs. (5) and (6).
The results of the both fittings are summarized in Table 1 and in
Fig. 2. 3.3. Dependence of 𝜇 on 𝐴, 𝜇(𝐴)
Results summary: We found a significant correlation between frag-
ment area and growth rate 𝛽. Thus, validating our assumption A.1, the Observe that the fitting of the mortality function 𝜇(𝐴) is not
growth rate is dependent on the area (A), that is 𝛽 = 𝛽(𝐴). Furthermore, necessary for parametrizing model (1). However, we present the pa-
when using the data points (𝜆𝑝𝑖 , 𝐴𝑖 ), the Fit Standard Error (RMSE) is rameterization of 𝜇(𝐴) to estimate the minimum fragment area at which
approximately the same (either ≈ 0.021 or ≈ 0.022) across the four the plant mortality equals its natural mortality rate, which is assumed
fittings (see column 5 in Table 1). Similarly, when using the data points to be the average mortality rate observed in the sites with area equal
(𝜆𝑜𝑖 , 𝐴𝑖 ), it follows that the RMSE across the four fittings is either 0.065 100 ℎ𝑎 (continuous forest). This quantity is important as it helps in
or 0.066 (see column 7 in Table 1). This result suggests that there is no understanding how fragmentation affects plant vital rates.

4
M.C.A. Leite et al. Ecological Modelling 457 (2021) 109681

𝐴−𝐴
Table 2 The condition 𝛽(𝐴) = 𝛽𝐹 𝐴 𝑐 = 0 holds true when 𝐴 = 𝐴𝑐 . At this
The estimated value of 𝑑1 and corresponding Fit Standard Error (RMSE)
value of area 𝛽(𝐴) = 𝑟(𝐴) − 𝜇(𝐴) = 0. That is, the recruitment rate 𝑟 and
obtained when fitting function 𝜇(𝐴) in Eq. (7).
the dead rate 𝜇 are equal. Thus, the initial density of biomass does not
Par. Value Unit 𝑅𝑀𝑆𝐸
grow or decay. The system will stabilize at the initial value of the plant
𝑑1 2.488 (ha)−1 0.1022
density.
We next consider the case 𝛽(𝐴) ≠ 0 and study the local stability of
the equilibrium states 𝑏 = 0 and 𝑏 = 𝐾. We first compute the Jacobian
matrix of the system, which is given by:
( )
2𝑏
𝐽𝐴𝑚 = 𝛽(𝐴) 1 − (10)
𝐾
𝐴−𝐴
At equilibrium, 𝑏 = 0, 𝐽𝐴𝑚 = 𝛽(𝐴) = 𝛽𝐹 𝐴 𝑐 . Since 𝛽𝐹 > 0 it follows
that 𝐽𝐴𝑚 is positive if 𝐴 > 𝐴𝑐 and negative if 𝐴 < 𝐴𝑐 . Therefore, 𝑏 = 0
is a locally asymptotical stable (l.a.s) equilibrium if 𝐴 < 𝐴𝑐 while it is
unstable if condition 𝐴 > 𝐴𝑐 holds true.
When considering the equilibrium state 𝑏 = 𝐾, the Jacobian 𝐽𝐴𝑚 =
−𝛽(𝐴) and the equilibrium state 𝑏 = 𝐾 is unstable if 𝐴 > 𝐴𝑐 and it is
l.a.s if 𝐴 < 𝐴𝑐 . The results can then be summarized in the following
proposition:

Proposition 1. Consider the model of forest fragmentation given by (8),

where the parameters have nonnegative values. If 𝐴 = 𝐴𝑐 , the plant density
stabilizes at its initial value. If 𝐴 < 𝐴𝑐 then the equilibrium 𝑏 = 0 is l.a.s
Fig. 3. Fitted curve for 𝜇(𝐴) function given in Eq. (7) represented by a solid red line. while the equilibrium 𝑏 = 𝐾 is unstable. On the other hand, if 𝐴 > 𝐴𝑐 then
The data is depicted with red circles The area is expressed in ha and 𝜇 in year−1 .
the equilibrium 𝑏 = 0 is unstable and the 𝑏 = 𝐾 is l.a.s.

Biological implications. If the fragment size is below the switching

We assume that the function 𝜇(𝐴) satisfies the following properties: point 𝐴𝑐 , the population growth rate 𝛽(𝐴) becomes negative indicating
(i) 𝜇(𝐴) is continuous and differentiable; (ii) 0 ≤ 𝜇 ≤ 1; (iii) when decreasing population size towards the equilibrium state 𝑏 = 0, which
the site area is close to zero the death rate approaches 1. This models implies that the persistence of the herb, Heliconia acuminata, is in jeop-
the fact that when the area of the patch is very small the death rate is ardy. If the fragment size is above the critical value 𝐴𝑐 then, Heliconia
very high. Furthermore, we assume that when the area is very large the acuminata population level stabilizes at the carrying capacity of the
mortality rate levels out at a certain value, which is the natural death system 𝐾. From the fitting we can see that the average value of the
rate of the plant if there is no fragmentation in the system (continuous critical patch area 𝐴𝑐 is approximately 1.75 ℎ𝑎, based on the projected
forest). One possible function that satisfies these conditions is data (𝜆𝑝𝑖 , 𝐴𝑖𝑖 ). And the fragment critical area 𝐴𝑐 is estimated to have
an average value of approximately 0.655 when the observed data set
𝜇(𝐴) = (1 − 𝜇𝐹 ) exp(−𝑑1 𝐴) + 𝜇𝐹 , (7) is used to fit the function. Thus, the switching point 𝐴𝑐 , estimated
where 𝜇𝐹 is the natural plant mortality when the site is not fragmented with observed data (𝜆𝑜𝑖 , 𝐴𝑖 ), is lower (about 63%) than the switching
(continuous forest) and 𝑑1 is the system decay constant. The value of 𝜇𝐹 point estimated with the data (𝜆𝑝𝑖 , 𝐴𝑖 ). Therefore, our model predictions
is the average of the mortality data values corresponding to sites with based on the observed data indicate that the plant population has a
area equal to 100 ℎ𝑎, which is the continuous forest. Those values are higher resilience to fragmentation than the prediction based on the
provided in Appendix A.1, Table A.4 and their average is 𝜇𝐹 = 0.0336. projected data. Recall that the values of 𝜆𝑝𝑖 in the project data (𝜆𝑝𝑖 , 𝐴𝑖 ),
In the same table, the values of 𝜇 for each plot with area of 1 ℎ𝑎, 10 ℎ𝑎 are derived from the linear matrix model. These results are illustrated
can be found. Eq. (7) together with the data is used to estimate the in Fig. 4. When the fragment size is precisely equal to the threshold
parameter 𝑑1 . The parameterization is carried out, as when fitting 𝛽(𝐴), value, population size that was initially in the patch does not change,
via MATLAB cftool and the results are given in Table 2 as well as in which indicates that the plant can persist in the system whenever the
Fig. 3. initial population size is different from zero.
Results summary: The value of system’s decay constant is estimated to The results summarized in Proposition 1 are illustrated in Fig. 4.
be 𝑑1 = 2.488 per year. Based on this value, straightforward calculations This figure depicts numerical simulations of the model (1) param-
show that the mortality rate in a fragment reaches the natural death eterized with the projected data set (𝜆𝑝𝑖 , 𝐴𝑖 ) as well as numerical
rate when the area is equal or greater than ≈ 4.1 ℎ𝑎. At 𝐴 ≈ 4.1 ℎ𝑎 the simulations of the model parameterized with the data (𝜆𝑜𝑖 , 𝐴𝑖 ).
mortality in the site differ 0.1% from 𝜇𝐹 .
5. Model predictability capabilities
4. Effect of fragment size on herb population density
After parametrizing the model (1) with the two data sets (projected
As discussed in Section 3.2, we adopt the function in Eq. (2) to data set (𝜆𝑝𝑖 , 𝐴𝑖 ) and observed data (𝜆𝑜𝑖 , 𝐴𝑖 )), we test their predictability
describe the relation between net growth rate 𝛽 and area. That is, against the experimental data consisting of plant population density
𝐴−𝐴
𝛽(𝐴) = 𝛽𝐹 𝐴 𝑐 . Thus, the model given in Eq. (1) reduces to (𝑏 corresponding to years 𝑡1 = 1998, 𝑡2 = 1999 and 𝑡3 = 2000, that
( ) 𝐴 − 𝐴𝑐 ( ) is, against the data set (𝑏𝑜𝑖 , 𝑡𝑖 ) given in Appendix A.1, Table A.2. For
𝑑𝑏 𝑏 𝑏
= 𝛽(𝐴)𝑏 1 − = 𝛽𝐹 𝑏 1− (8) a patch of a given area, the initial condition used to run the models,
𝑑𝑡 𝐾 𝐴 𝐾
the one parameterized by (𝜆𝑝𝑖 , 𝐴𝑖 ) and the one parameterized by the
with 𝛽(𝐴) = 𝑟(𝐴) − 𝜇(𝐴).
The equilibrium states of the system are obtained by solving the observed data (𝜆𝑜𝑖 , 𝐴𝑖 ), is the average value of the observed population
equation resulting from setting the right-hand side of (8) equal to zero. density in year 1998. More precisely, we used the initial value of
Simple calculations shows that, for fixed value of area 𝐴, the solutions 194, 342, 388 when the area was fixed at 1 ℎ𝑎, 10 ℎ𝑎, and 100 ℎ𝑎,
are: respectively. The results are shown in Fig. 5.
𝐴 − 𝐴𝑐 Additionally, fixing a patch area, we investigate how well the
𝛽(𝐴) = 𝛽𝐹 = 0, 𝑏 = 0, 𝑏 = 𝐾. (9) model parameterized by the set (𝜆𝑜𝑖 , 𝐴𝑖 ) performs to reproduced the
𝐴

5
M.C.A. Leite et al. Ecological Modelling 457 (2021) 109681

Fig. 4. Numerical simulation of model (1) with the parameterized function (2). The curve depicted with blue solid line is obtained with the value of 𝛽𝐹 , fixed at 0.0214 and the
value of 𝐴𝑐 = 1.753, which is estimated with projected data (𝜆𝑝𝑖 , 𝐴𝑖 ). While the curve depicted with red dashed line, the value of 𝛽𝐹 is fixed at 0.1066 and the value of 𝐴𝑐 = 0.6547
is the one estimated with the observed data (𝜆𝑜𝑖 , 𝐴𝑖 ). The fragment area is 0.5, 1, 10, 100 ℎ𝑎 (continuous forest) in panels (a), (b), (c) and (d), respectively.

average observed data comparatively to the matrix model used in Bruna qualitatively the plant density temporal evolution pattern (see Fig. 6).
(2003). Specifically, we compare the temporal pattern obtained with This occurs for all three distinct areas considered. In contrast, the
the model parameterized with the observed data set (𝜆𝑜𝑖 , 𝐴𝑖 ), with the matrix model predictions only capture the observed temporal evolution
observed and projected average plant population density. The results of the plant density in the continuous forest (panel c)). Nonetheless, our
are indicated in Fig. 6. ODE model fits the observed data slightly better than the linear matrix
model.
6. Discussion of results Due to the fact that the data sets are small it is difficult to conduct
robustness test by removing some of the data points and see whether
Our results are obtained with a model parameterized with the one gets similar fitting results. Sensitivity analysis could be used to
best data set available. The collection of data is in the scope of the gain additional insight on how the model respond to uncertainty in the
Brazil’s Biological Dynamics of Forest Fragments Project — BDFFP. This parameters. This is part of future work.
project is one of the largest and longest experimental work to study the
ecological impact of habitat fragmentation (Laurance et al., 2011). It 7. Conclusions
constitutes an incredible effort in the field but there is only three patch
sizes in the data and three time instances (1998, 1999, and 2000). This We explored the effect of fragmentation on plant population size in
information is far from ideal to parametrize the model. In addition to an ordinary differential equation (ODE) based framework. This model
being small, the data sets exhibit high variability and missing values. assume that the growth of the plant population size is of logistic type
Consequently, several approximations were made. For example, there and that the net growth rate (recruitment rate minus mortality rate) is
were missing values of projected and observed asymptotic growth rate affected by the reduction of patch area through a general net growth
𝜆 for transitions years 1998–1999 for 3 sites with area 100 ℎ𝑎. So, for function. We use the model to investigate the minimum patch area
these sites, we assume that the density of Heliconia acuminata in year needed to sustain any plant/shrub system. Additionally, we parametrize
1998 was the same as in the year 1999. The model parameterization the model with a three years census data of Heliconia acuminata in
can be easily changed to accommodate distinct data sets not only both forest fragments and continuous forest provided in the work
concerning Heliconia acuminata that may be produced in the future but by Bruna (2003). We apply the resulting fitted model to determine the
also concerning other plant systems. minimum fragment size necessary to sustain the Heliconia acuminata
The results in Fig. 5 suggests that the model, either parameterized density in a fragmented landscape. For this specific data set, we also
by the projected data (𝜆𝑝𝑖 , 𝐴𝑖 ) or parameterized by the observed data test the hypothesis that the ODE-based model, with the dependence
(𝜆𝑜𝑖 , 𝐴𝑖 ), do not fit well the plant density observed in each individual of plant growth rate on the population density explicitly modeled, fits
site of fragments of same area. This might be due to the constrains the observed Heliconia acuminata density better than the linear matrix
of the data sets mentioned above. However, when using the average demographic models that are common used in the field.
observed plant density instead of the observed density in each indi- Our theoretical analysis of the model indicates that the persistence
vidual site, the model parameterized by the data set (𝜆𝑜𝑖 , 𝐴𝑖 ) captures of Heliconia acuminate is possible when the area of the fragment is

6
M.C.A. Leite et al. Ecological Modelling 457 (2021) 109681

Fig. 5. Simulation of plant density as a function of the patch size 𝐴 using model (1) parameterized with projected data set (𝜆𝑝𝑖 , 𝐴𝑖 ), shown with black solid line, and observed
data set (𝜆𝑜𝑖 , 𝐴𝑖 ), depicted with blue dashed line. The red circle represents the observed data (𝑏𝑜𝑖 , 𝑡𝑖 ). In the 𝑥-axis 1, 2, 3, 4 represent the years 1998, 1999, 2000, 2001 respectively.
(a) Patch size 𝐴 = 1 ℎ𝑎. (b) Patch size 𝐴 = 10 ℎ𝑎. (c) Patch size 𝐴 = 100 ℎ𝑎.

Fig. 6. Simulation of plant density as a function of the fragment size 𝐴 using model (1) parameterized using the observed data set (𝜆𝑜𝑖 , 𝐴𝑖 ). The blue curve is the time series
obtained with the model, the red circle is the average population density of the Heliconia acuminata observed while the black circle represents the projected data points obtained
with the matrix model. At the year 1998 (year 1) the density is the observed one. (a) Patch size 𝐴 = 1 ℎ𝑎. (b) Patch size 𝐴 = 10 ℎ𝑎. (c) Patch size 𝐴 = 100 ℎ𝑎.

7
M.C.A. Leite et al. Ecological Modelling 457 (2021) 109681

Table A.1 needed to verify if this findings are applied to other plant populations in
Values of the projected 𝜆𝑝𝑖 , observed 𝜆𝑜𝑖 asymptotic growth rate, and
fragmented systems. The model in this paper is very simple. It consists
area of the fragment 𝐴𝑖 for each site. The units are: area in ha and 𝛽
in 1/year. of an ordinary differential equation that represent average or mean
Area 𝐴𝑖 𝜆𝑜𝑖 𝛽𝑜𝑖 𝜆𝑝𝑖 𝛽𝑝𝑖 field approximations of the true time and spatial scales of the object
1 1.0375 0.0368 0.978 −0.0222
of study. An alternative formulation could involve an age structured
1 1.221 0.1997 0.987 −0.0131 model and/or including spatial component. Additionally, those models
1 1.075 0.0723 0.988 −0.0121 could incorporate distinct processes in which fragmentation affects, for
1 1.0675 0.0653 0.997 −0.0030 example, pollination, seed dispersal, the seed and seedling establish-
1 1 0.0000 0.988 −0.0121
1 1.175 0.1613 0.989 −0.0111
ment. However, any attempt in this direction would require a multitude
1 1.0125 0.0124 0.978 −0.0222 of parameters and, necessarily, data to validate it. Given the nature
10 1.025 0.0247 0.988 −0.0121 of the phenomena at hand, these requirements may be very difficult
10 1.212 0.1923 0.991 −0.0090
to satisfy. Therefore, the approach of using simpler models may have
10 1.075 0.0723 0.999 −0.0010
10 1.025 0.0247 0.978 −0.0222 advantages over uses of agent based or computational models that are
10 1.2125 0.1927 0.994 −0.0060 more realistic but providing very limited theoretical insights.
10 1.0626 0.0607 0.996 −0.0040 Realistically, we understand that we should consider multi-patch
100 1.05 0.0488 1.028 0.0276
and interactions between the patches; however, since this is an alter-
100 1.1 0.0953 1.079 0.0760
100 1.1 0.0953 0.987 −0.0131 native approach to the use of linear demographic model it is important
100 1.1625 0.1506 1.04 0.0392 to understand the dynamics of a single isolated patch. In the future
100 1.175 0.1613 1.022 0.0218 we will consider a more realistic model with multi-patches and their
100 1.1625 0.1506 0.991 −0.0090
100 1.0875 0.0839 1.014 0.0139
connectedness.
100 1.05 0.0488 1.018 0.0178
100 1.125 0.1178 1.016 0.0159
CRediT authorship contribution statement
𝛽𝐹 𝑜 Observed 𝛽𝐹 𝑝 Projected
0.1066 0.0214
Maria C.A. Leite: Conceptualization, Methodology, Software, In-
vestigation, Data curation, Writing – original draft, Writing – review
& editing, Visualization, Supervision, Funding acquisition. Rebecca
higher than a threshold 𝐴𝑐 , the system switching point. That is, for values Sauchuk: Software, Visualization. Folashade B. Agusto: Conceptu-
of the area greater to 𝐴𝑐 the net population growth rate is greater than
alization, Methodology, Software, Formal analysis, Funding acquisi-
zero which implies persistent plant population whose size will reach
tion, Writing – original draft, writing – review & editing. Orou G.
(asymptotically) the fragment carrying capacity, if no disturbances
Gaoue: Conceptualization, Resources, Investigation, Data curation.
occur. The persistence of the plant is still possible if the forest fragment
Benito Chen-Charpentier: Conceptualization, Methodology, Formal
area is equal to 𝐴𝑐 , but in this case the population stagnates at the
analysis, Investigation, Writing – original draft, Supervision.
initial value.
The switching point 𝐴𝑐 estimated with the proposed net growth
rate function parameterized with the census data from Amazonian Declaration of competing interest
Heliconia acuminata is 0.655 ≈ 1 ℎ𝑎. This result is in agreement with
the findings reported by Bruna and Kress (2002), concerning the same One or more of the authors of this paper have disclosed potential or
Amazonian Heliconia acuminata system, which shows that, among the
pertinent conflicts of interest, which may include receipt of payment,
three area patches studied (1, 10, 100, ℎ𝑎), the herb in fragments of
either direct or indirect, institutional support, or association with an
1 ℎ𝑎 is more vulnerable to fragmentation. The switching point estimated
entity in the biomedical field which may be perceived to have potential
with same function but parameterized with the data calculated using
conflict of interest with this work. For full disclosure statements re-
the asymptotical growth rate 𝜆 obtained from the matrix based model
fer to https://doi.org/10.1016/j.ecolmodel.2021.109681. Benito Chen-
and presented in the work by Bruna (2003) is 1.75 ≈ 2 ℎ𝑎. Thus, the
Charpentier reports financial support was provided by Nimbios, Uiver-
predictions based on the census data indicate that the plant system is
more resilient to fragmentation than the predictions based on the data sity of Tennessee, Knoxville. Fola Agusto reports financial support was
recovered from the matrix-based model. provided by Nimbios, Uiversity of Tennessee, Knoxville. Maria Leite
The mortality decay constant estimated through the function pa- reports financial support was provided by Nimbios, Uiversity of Ten-
rameterized with the data census reveals that the mortality rate of nessee, Knoxville. Orou Gaoue reports financial support was provided
Heliconia acuminata stabilizes at the value of the mortality rate observed by Nimbios, Uiversity of Tennessee, Knoxville. Fola Agusto reports
in an ecosystem without fragmentation when the fragment size is financial support was provided by NSF. Maria Leite reports financial
greater or equal to ≈ 4 ℎ𝑎. Thus, our study suggests that, in the support was provided by University of South Florida St. Petersburg.
particular scenario of Heliconia acuminata Amazonian system, the net
growth rate is less sensitive to fragmentation than the mortality rate. Acknowledgments
Such phenomena was also reported for the same plant system (Bruna
and Kress, 2002). It was reported that fragments of 10 ℎ𝑎 of land
suffered higher mortality than the continuous forest. Leite acknowledges support from USFSP 2019–20 Internal Research
Given our results we observed that the ODE-based model presented Grant, United States that made possible the full support of the under-
in this study can serve as an alternative to the linear demographic graduate student R. Sauchuk’s research project. Agusto acknowledges
model since we have included in our ODE model some assumptions support from NSF, United States under EPSCOR grant number OIA
about how the mortality and recruitment rates behave which are not in- 1656006. The authors are thankful for support from NIMBioS spon-
cluded in the linear demographic model. Furthermore, the ODE model sored by NSF, United States Award # DBI – 1300426 and, partially
is able to fit the given data. This is an interesting result suggesting the by University of Tennessee, Knoxville, United States. The authors ac-
use of models that take in account the dependence of the plant popu- knowledges Emilio Bruna for useful suggestions. The authors are also
lation growth rates on population size when investigating the impact thankful to the anonymous reviewers for their valuable comments that
of fragmentation on plant species. However, further investigations are essentially improve the paper.

8
M.C.A. Leite et al. Ecological Modelling 457 (2021) 109681

Table A.2
Values of the asymptotic growth rates 𝜆𝑝𝑖 , 𝜆𝑜𝑖 , area 𝐴𝑖 , as well as the observed (𝑏𝑜𝑖 ) and projected (𝑏𝑝𝑖 ) values of the population density
corresponding to each site in the years 1998, 1999, and 2000. The symbol ∗ indicates that values of the density were assumed to be equal to
the values of density in the year 1999. The units are: area in ℎ𝑎, and density in units of 𝑖𝑛𝑑𝑖𝑣𝑖𝑑𝑢𝑎𝑙𝑠∕ℎ𝑎..
Area 𝑏𝑝𝑖 𝜆𝑝𝑖 Year 𝑏𝑜𝑖 𝜆𝑜𝑖 Area 𝑏𝑝𝑖 𝜆𝑝𝑖 Year 𝑏𝑜𝑖 𝜆𝑜𝑖
1 217.7 – 1998 219.5 – 100 732.5 – 1998 697.6 –
1 157.4 – 1998 155.2 – 100 514.4 – 1998 467.6 –
1 203.3 – 1998 168.7 – 100 712.2 – 1998 647.5 –
1 247 – 1998 232.6 – 100 112(∗ ) – 1998 112(∗ ) –
1 214 0.983 1999 214 0.975 100 171(∗ ) – 1998 171(∗ ) –
1 161 0.978 1999 161 1.0375 100 235(∗ ) – 1998 235 (∗ ) –
1 206 0.987 1999 206 1.221 100 753 1.028 1999 753 1.05
1 250 0.988 1999 250 1.075 100 555 1.079 1999 555 1.1
1 213.4 0.997 2000 228.4 1.0675 100 703 0.987 1999 703 1.1
1 159.1 0.988 2000 161 1 100 112 – 1999 112 –
1 203.7 0.989 2000 242.0 1.175 100 171 – 1999 171 –
1 244.5 0.978 2000 253.1 1.0125 100 235 – 1999 235 –
10 164.0 – 1998 158.0 – 100 774.1 1.04 2000 875.4 1.1625
10 402.0 – 1998 331.7 – 100 598.8 1.022 2000 652.1 1.175
10 577.0 – 1998 536.7 – 100 693.9 0.991 2000 817.2 1.1625
10 162 0.988 1999 162 1.025 100 113.6 1.014 2000 121.8 1.0875
10 402 0.991 1999 402 1.212 100 174.1 1.018 2000 179.6 1.05
10 577 0.999 1999 577 1.075 100 238.8 1.016 2000 264.4 1.125
10 158.4 0.978 2000 166.0 1.025
10 399.6 0.994 2000 487.4 1.2125
10 574.7 0.996 2000 613.1 1.0626

Appendix Table A.3

Average value of the observed plant density and projected herb density in year 1998
corresponding to areas 1, 10, 100 ha.
A.1. Data preparation
Area Average Observed density Average Projected density
(ha) 𝑏 in 1998 (unit of 𝑖𝑛𝑑𝑖𝑣𝑖𝑑𝑢𝑎𝑙𝑠∕ℎ𝑎.) 𝑏 in 1998 (units of 𝑖𝑛𝑑𝑖𝑣𝑖𝑑𝑢𝑎𝑙𝑠∕ℎ𝑎.)
The data was extracted from Bruna (2003). More precisely, the plant
1 194 206
density at each site corresponding to year 1999 as well as the fragment 10 342 381
area are given in Table 1, (Bruna, 2003). The projected asymptotical 100 388 413
growth rate 𝜆 corresponding to the transition years 1998 − 1999 and
1999 − 2000 for each site is also provided in Table 1. The details on
Table A.4
the computation of these values are given also in Bruna (2003). The
Average values of the plant’s mortality rate 𝜇 in each site with units 1/year. 𝜇𝐹 is the
values of observed 𝜆 are depicted in Fig. 5 in the same work. The average of the mortality rate in patches of area 100 ℎ𝑎..
transition matrices describing the dynamics of the Heliconia acuminata Area Average mortality Area Average mortality 𝜇𝐹
are provided in Appendix B, (Bruna, 2003). (ha) 𝜇𝑖 (1/year) (ha) 𝜇𝑖 (1/year) (1/year)
1 0.0287 100 0.0595
A.1.1. Values of the projected 𝜆𝑝𝑖 , observed 𝜆𝑜𝑖 asymptotic growth rate, and 1 0.017 100 0.0111
area 𝐴𝑖 of the fragment for each site 1 0.0343 100 0.0813
1 0.545 100 0.0427 0.0336
The projected values of 𝜆 (that is, 𝜆𝑝𝑖 ) and area of each frag-
1 0.0886 100 0.0289
ments/site (𝐴𝑖 ) were read off Table 1 (Bruna, 2003) and is provided 1 0.0932 100 0.0482
in Table A.1 for completeness. The resulting data set is denoted by 1 0.0437 100 0.0001
(𝜆𝑝𝑖 , 𝐴𝑖 ). The observed transition values of 𝜆𝑜𝑖 summarized in the Ta- 1 0.0606 100 0.0248
10 0.0186 100 0.006
ble A.1 were read off Fig. 5 and the data set is denoted by 𝜆𝑜𝑖 . The
10 0.0154
corresponding values of the growth rate 𝛽𝑝𝑖 and 𝛽𝑜𝑖 were computed 10 0.0121
using the relation 𝛽 = ln(𝜆). On the right hand-side of Table A.1 is listed 10 0.1052
the values of the projected and observed growth rates corresponding to 10 0.0296
the continuous forest. These values are the average of the values of 𝛽𝑝𝑖 10 0.0301

and 𝛽𝑜𝑖 for sites with area 100 ha and they are denoted by 𝛽𝐹 𝑝 and 𝛽𝐹 𝑜 ,
respectively.
the transition year 1998 − 1999 are not available in Bruna (2003), we
A.1.2. Values of the population density of Heliconia acuminata 𝑏𝑝𝑖 , (pro- assumed that the population density in year 1998 is the same as in year
jected) and 𝑏𝑜𝑖 (observed) in years 1998, 1999, and 2000 for each site 1999. The sets (𝑏𝑝𝑖 , 𝑡𝑖 ) and (𝑏𝑜𝑖 , 𝑡𝑖 ) are given in Table A.2. In Table A.3
The density of the plant in the years 1998, 1999 and 2000 were we present the average values of Heliconia density in sites with area 100
computed from the values of the density registered for year 1999 ha (continuous forest) and corresponding to year 1998.
(Table 1, Bruna, 2003) and using the formula:
A.1.3. Mortality rates
density of Heliconia in year 𝑡 + 1 The mortality rate for each fragment was computed using the sur-
𝜆= . (A.1)
density of Heliconia in year 𝑡 vival probability calculated from the transition matrices describing the
For each area and each year, we obtain two data sets: projected density dynamics of Heliconia acuminata provided in Appendix B (Bruna and
(𝑏𝑝𝑖 , 𝑡𝑖 ) and observed density (𝑏𝑜𝑖 , 𝑡𝑖 ), where the indices 𝑖 = 1, 2, 3 rep- Kress, 2002). The values presented in Table A.4 are the average of the
resent years 1998, 1999, 2000, respectively. The former set is obtained mortality in years 1999 and 2000.
by using the values 𝜆𝑝𝑖 in Eq. (A.1). The values of 𝑏𝑜𝑖 in latter set are The value of the natural death rate 𝜇𝐹 in the continuous forest was
computed using 𝜆𝑜𝑖 in the same equation. For the sites with area 100 computed by averaging over all values of the plant’s mortality in all
ha for which the values of projected and observed 𝜆 corresponding to sites of area 100 ℎ𝑎.

9
M.C.A. Leite et al. Ecological Modelling 457 (2021) 109681

References Gause, G., 1934. Gause GF the Struggle for Existence. Williams & Wilkins, Baltimore.
Hastings, A., Cuddington, K., Davies, K.F., Dugaw, C.J., Elmendorf, S., Freestone, A.,
Allen, L.J.S., 2007. An Introduction To Mathematical Biology. Pearson/Prentice Hall. Harrison, S., Holland, M., Lambrinos, J., Malvadkar, U., Melbourne, B.A., Moore, K.,
Andreazzi, C.S., Pimenta, C.S., Pires, A.S., Fernandez, F.A.S., Oliveira-Santos, L.G., Taylor, C., Thomson, D., 2004. The spatial spread of invasions: new developments
Menezes, J.F.S., 2012. Increased productivity and reduced seed predation favor in theory and evidence: Spatial spread of invasions. Ecol. Lett. 8 (1), 91–101.
a large-seeded palm in small atlantic forest fragments. Biotropica 44, 237–245. http://dx.doi.org/10.1111/j.1461-0248.2004.00687.x, URL http://doi.wiley.com/
http://dx.doi.org/10.1111/j.1744-7429.2011.00782.x. 10.1111/j.1461-0248.2004.00687.x.
Bolker, B., Brooks, M., Clark, S., Poulsen, J., Stevens, M., White, J.-S., 2009. Generalized Jacquemyn, H., De Meester, L., Jongejans, E., Honnay, O., 2012. Evolutionary changes
linear mixed models: A practical guide for ecology and evolution. Trends Ecol. Evol. in plant reproductive traits following habitat fragmentation and their consequences
24, 127–135. http://dx.doi.org/10.1016/j.tree.2008.10.008. for population fitness: Habitat fragmentation and plant trait evolution. J. Ecol.
Broadhurst, L.M., Young, A.G., Forrester, R., 2008. Genetic and demographic responses 100 (1), 76–87. http://dx.doi.org/10.1111/j.1365-2745.2011.01919.x, URL http:
of fragmented Acacia dealbata (Mimosaceae) populations in southeastern Australia. //doi.wiley.com/10.1111/j.1365-2745.2011.01919.x.
Biol. Cons. 141 (11), 2843–2856. http://dx.doi.org/10.1016/j.biocon.2008.08.021, Jacquemyn, H., Rossum, F.V., Brys, R., Endels, P., Hermy, M., Triest, L., Blust, G.D.,
URL https://linkinghub.elsevier.com/retrieve/pii/S0006320708003145. 2003. Effects of Agricultural land use and fragmentation on genetic, demography
Brudvig, L.A., Damschen, E.I., Haddad, N.M., Levey, D.J., Tewksbury, J.J., 2015. The and population persistence of the rare primula vulgaris, and implications for
influence of habitat fragmentation on multiple plant–animal interactions and plant conservation. Belg. J. Bot. 136 (1), 5–22.
reproduction. Ecology 96 (10), 2669–2678. http://dx.doi.org/10.1890/14-2275.1, Kolb, A., Dahlgren, J.P., Ehrlén, J., 2010. Population size affects vital rates but not
URL http://doi.wiley.com/10.1890/14-2275.1. population growth rate of a perennial plant. Ecology 91 (11), 3210–3217. http:
Bruna, E.M., 2002. Effects of forest fragmentation on heliconia acuminata seedling //dx.doi.org/10.1890/09-2207.1, URL http://doi.wiley.com/10.1890/09-2207.1.
recruitment in central Amazonia. Oecologia 132, 235–243. http://dx.doi.org/10. Laurance, W.F., Camargo, J.L., Luizão, R.C., Laurance, S.G., Pimm, S.L., Bruna, E.M.,
1007/s00442-002-0956-y. Stouffer, P.C., Bruce Williamson, G., Bení tez Malvido, J., Vasconcelos, H.L.,
Bruna, E.M., 2003. Are plant populations in fragmented habitat recruitment limited? Van Houtan, K.S., Zartman, C.E., Boyle, S.A., Didham, R.K., Andrade, A., Love-
Test with an Amazonian herb. Ecology 84, 932–947. http://dx.doi.org/10.1007/ joy, T.E., 2011. The fate of Amazonian forest fragments: A 32-year investigation.
s00442-002-0956-y. Biol. Cons. 144 (1), 56–67. http://dx.doi.org/10.1016/j.biocon.2010.09.021, URL
Bruna, E.M., Kress, W.J., 2002. Habitat fragmentation and the demographic structure https://linkinghub.elsevier.com/retrieve/pii/S0006320710004209.
of an Amazonian understory herb (Heliconia acuminata). Conserv. Biol. 16 (5), Lienert, J., 2004. Habitat fragmentation effects on fitness of plant populations – a
1256–1266. http://dx.doi.org/10.1046/j.1523-1739.2002.99494.x, URL http://doi. review. J. Nat. Conserv. 12 (1), 53–72. http://dx.doi.org/10.1016/j.jnc.2003.07.
wiley.com/10.1046/j.1523-1739.2002.99494.x. 002, URL https://linkinghub.elsevier.com/retrieve/pii/S1617138104000068.
Bruna, E.M., Kress, W.J., Marques, F., Silva, O.F.d., 2004. Heliconia acuminata Lucas, E., 2020. A review of trail-related fragmentation, unauthorized trails, and other
reproductive success is independent of local floral density. Acta Amaz. 34 aspects of recreation ecology in protected areas. In: California Fish and Wildlife. pp.
(3), 467–471. http://dx.doi.org/10.1590/S0044-59672004000300012, URL http: 95–125. http://dx.doi.org/10.1007/s00442-002-0956-y, Recreation Special Issue.
//www.scielo.br/scielo.php?script=sci_arttext&pid=S0044-59672004000300012& Matesanz, S., Teso, M.L.R., García-Fernández, A., Escudero, A., 2017. Habitat fragmen-
lng=en&tlng=en. tation differentially affects genetic variation, phenotypic plasticity and survival in
Bruna, E.M., Oli, M.K., 2005. Demographic effects of habitat fragmentation on a tropical populations of a gypsum endemic. Front. Plant Sci. 8, art. 843. http://dx.doi.org/
herb: life-table response experiments. Ecology 86 (7), 1816–1824. 10.3389/fpls.2017.00843.
Cantrell, R., Cosner, C., 2003. Spatial Ecology Via Reaction-Diffusion Equations, John Morris, W.F., Doak, D.F., 2003. Quantitative Conservation Biology: Theory and Practice
Wiley & Sons Ltd. of Population Viability Analysis. Sinauer Associates, Sunderland, Massachusetts.
Cantrell, R.S., Cosner, C., Fagan, W.F., 2011. The implications of model formulation Murray, J.D., 2002. Mathematical Biology: An Introduction, third ed. In: Interdisci-
when transitioning from spatial to landscape ecology. Math. Biosci. Eng. 9 (1), 27– plinary applied mathematics, vol. I, Springer-Verlag, New York.
60. http://dx.doi.org/10.3934/mbe.2012.9.27, URL http://www.aimsciences.org/ Musgrave, J., Lutscher, F., 2014. Integrodifference equations in patchy landscapes: II:
journals/displayArticlesnew.jsp?paperID=6843. Population level consequences. J. Math. Biol. 69 (3), 617–658. http://dx.doi.org/
Caswell, H., 2001. Matrix Population Models: Construction, Analysis, and Interpretation, 10.1007/s00285-013-0715-1, URL http://link.springer.com/10.1007/s00285-013-
second ed. Sinauer Associates, Sunderland, Massachusetts. 0715-1.
Cramer, J.M., Mesquita, R.C.G., Bentos, T.V., Moser, B., Williamson, G.B., 2007. Forest Neubert, M.G., Parker, I.M., 2004. Projecting rates of spread for invasive species. Risk
fragmentation reduces seed dispersal of duckeodendron cestroides, a central amazon Anal. 24 (4), 817–831. http://dx.doi.org/10.1111/j.0272-4332.2004.00481.x, URL
endemic. Biotropica 39 (6), 709–718. http://dx.doi.org/10.1111/j.1744-7429.2007. http://doi.wiley.com/10.1111/j.0272-4332.2004.00481.x.
00317.x, URL http://doi.wiley.com/10.1111/j.1744-7429.2007.00317.x. Pachepsky, E., Levine, J.M., 2011. Density dependence slows invader spread in frag-
Cushing, J., 1998. An Introduction To Structured Population Dynamics, second ed. mented landscapes. Amer. Nat. 177 (1), 18–28. http://dx.doi.org/10.1086/657438,
SIAM, Philadelphia. URL https://www.journals.uchicago.edu/doi/10.1086/657438.
Didham, R.K., Kapos, V., Ewers, R.M., 2012. Rethinking the conceptual foundations of Rybicki, J., Abrego, N., Ovaskainen, O., 2020. Habitat fragmentation and species
habitat fragmentation research. In: From the Students of BI454 At Western Oregon diversity in competitive communities. Ecol. Lett. 23, 506–517. http://dx.doi.org/
University, Habitat Fragmentation and Pollination Dynamics, Plant Ecology. 10.1111/ele.13450.
Fahrig, L., 2017. Ecological responses to habitat fragmentation per se. Annu. Rev. Ecol. Silva Matos, D.M., Freckleton, R.P., Watkinson, A.R., 1999. The role of density
Evol. Syst. 48, 1–23. http://dx.doi.org/10.1146/annurev-ecolsys-110316-022612. dependence in the population dynamics of a tropical palm. Ecology 80 (8),
Ferrari, J.R., Preisser, E.L., Fitzpatrick, M.C., 2014. Modeling the spread of invasive 2635–2650.
species using dynamic network models. Biol. Invasions 16 (4), 949–960. http: Tang, C.Q., He, L.-Y., Gao, Z., Zhao, X.-F., Sun, W.-B., Ohsawa, M., 2011. In-
//dx.doi.org/10.1007/s10530-013-0552-6, URL http://link.springer.com/10.1007/ creased productivity and reduced seed predation favor a large-seeded palm in
s10530-013-0552-6. small Atlantic forest fragments. Mt. Res. Dev. 31, 343–350. http://dx.doi.org/
Flaspohler, D.J., Giardina, C.P., Asner, G.P., Hart, P., Price, J., Lyons, C.K., Cas- 10.1659/MRD-JOURNAL-D-11-00004.1, URL http://www.bioone.org/doi/full/10.
taneda, X., 2010. Long-term effects of fragmentation and fragment properties on 1659/MRD-JOURNAL-D-11-00004.1.
bird species richness in Hawaiian forests. Biol. Cons. 143 (2), 280–288. http: Vaughn, N.R., Asner, G.P., Giardina, C.P., 2014. Centennial impacts of fragmentation
//dx.doi.org/10.1016/j.biocon.2009.10.009, URL http://linkinghub.elsevier.com/ on the canopy structure of tropical montane forest. Ecol. Appl. 24 (7), 1638–
retrieve/pii/S0006320709004509. 1650. http://dx.doi.org/10.1890/13-1568.1, URL http://doi.wiley.com/10.1890/
Gagnon, P.R., Bruna, E.M., Rubim, P., Darrigo, M.R., Littell, R.C., Uriarte, M., 13-1568.1.
Kress, W.J., 2011. Growth of an understory herb is chronically reduced Williamson, 1989. Mathematical Models of Invasion, in Biological Invasions: A Global
in Amazonian forest fragments. Biol. Cons. 144 (2), 830–835. http://dx.doi. Perspective, J. A. Drake and others John Wiley and Sons.
org/10.1016/j.biocon.2010.11.015, URL http://linkinghub.elsevier.com/retrieve/ Zambrano, J., Salguero-Gómez, R., 2014. Forest fragmentation alters the population
pii/S0006320710004854. dynamics of a late-successional tropical tree. Biotropica 46 (5), 556–564. http:
//dx.doi.org/10.1111/btp.12144, URL http://doi.wiley.com/10.1111/btp.12144.

10