Hurricanes commonly disturb and damage tropical forests. Hurricane frequency
and intensity are predicted to change under the changing climate. The
short-term impacts of hurricane disturbances to tropical forests have been
widely studied, but the long-term impacts are rarely investigated. Modeling
is critical to investigate the potential response of forests to future
disturbances, particularly if the nature of the disturbances is changing
with climate. Unfortunately, existing models of forest dynamics are not
presently able to account for hurricane disturbances. Therefore, we
implement the Hurricane Disturbance in the Ecosystem Demography model (ED2)
(ED2-HuDi). The hurricane disturbance includes hurricane-induced immediate
mortality and subsequent recovery modules. The parameterizations are based
on observations at the Bisley Experimental Watersheds (BEW) in the Luquillo
Experimental Forest in Puerto Rico. We add one new plant functional type
(PFT) to the model – Palm, as palms cannot be categorized into one of the
current existing PFTs and are known to be an abundant component of tropical
forests worldwide. The model is calibrated with observations at BEW using
the generalized likelihood uncertainty estimation (GLUE) approach. The
optimal simulation obtained from GLUE has a mean relative error of
Hurricanes are an important disturbance agent in tropical forests. They damage individual trees and reduce aboveground biomass (Zimmerman et al., 1994; Uriarte et al., 2019; Rutledge et al., 2021; Leitold et al., 2021). For example, Hurricane Hugo in 1989 uprooted and snapped 20 % of the trees at El Verde in the Luquillo Experimental Forest (LEF), Puerto Rico (Walker, 1991; Walker et al., 1992; Zimmerman et al., 1994), and reduced the aboveground biomass by 50 % at Bisley in the LEF (Scatena et al., 1993; Heartsill Scalley et al., 2010). Hurricane Katrina in 2005 damaged about 320 million large trees on US Gulf Coast forests, and the damaged trees are equivalent to 50 %–140 % of the net annual US carbon sink (Chambers et al., 2007). In the long term, the recovery from those damages will alter forest species composition and structure (Royo et al., 2011; Heartsill Scalley, 2017).
Hurricane-induced mortality varies with many factors, including hurricane severity (Parker et al., 2018), environmental conditions (Uriarte et al., 2019; Hall et al., 2020), forest exposure to hurricane winds (Boose et al., 1994, 2004), forest structure (Zhang et al., 2022b), and traits and size of individual trees (Curran et al., 2008; Lewis and Bannar-Martin, 2011). Trees with a larger diameter have been found to be more resistant to wind forces but more likely to suffer broken branches (Lewis and Bannar-Martin, 2011). Species with higher wood density tend to suffer less from hurricane disturbances (Zimmerman et al., 1994; Curran et al., 2008). Hurricanes with heavier rainfall and stronger wind generally lead to higher mortality (Uriarte et al., 2019; Hall et al., 2020), and forests that are more exposed to strong winds tend to have higher mortality (Uriarte et al., 2019). However, forests with a more wind-resistant structure and composition experience lower mortality even during a stronger hurricane event or a higher exposure (Zhang et al., 2022b).
The recovery from hurricanes also depends on many factors, such as the disturbance severity (Walker, 1991; Everham and Brokaw, 1996; Cole et al., 2014; Heartsill Scalley, 2017) and traits of individual species (Curran et al., 2008; Lewis and Bannar-Martin, 2011). Species with lower wood density have shorter times to resprout (Paz et al., 2018), higher growth rate (King et al., 2006), and shorter biomass recovery times (Curran et al., 2008). The number of resprouts of some species further varies with time since disturbance (Brokaw, 1998). Less severe disturbances lead to a faster recovery and a higher recovery of stem density and aboveground biomass compared to the level observed prior to the disturbance (Wang and Eltahir, 2000; Parker et al., 2018). For example, observations on a tropical forest canopy in western Mexico after two hurricanes – category 2 Jova and category 4 Patricia – showed that Hurricane Jova destroyed 11 % of the aboveground biomass while Hurricane Patricia destroyed 23 %; the recovery was more rapid after the less intense Hurricane Jova (Parker et al., 2018). Although the immediate mortality and subsequent recovery of tropical forest from hurricane disturbances have been thoroughly studied via observations, the long-term effects of consecutive hurricane disturbances on tropical forests have rarely been studied. Models that can simulate the immediate mortality and subsequent recovery of an ecosystem can play a role in understanding potential mechanisms driving the mortality and recovery of the ecosystems and studying the long-term effects of disturbances, particularly if the nature of the disturbances is changing with climate. Uriarte et al. (2009) implemented hurricane disturbance in a forest simulator and investigated the long-term dynamics of forest composition, diversity, and structure. However, the biological and environmental processes of the forest simulator used are not dynamic and thus the model cannot simulate the adaptation of vegetation to the changes of environment (Jorgensen, 2008). Vegetation dynamics models can account for changes in the ecosystem resulting from a changing environment (Medvigy et al., 2009; Longo et al., 2019a) and further allow us to explore scenarios via synthetic experiments and thus emulate what might happen in forests under novel environmental conditions. For example, Feng et al. (2018) used the Ecosystem Demography model (ED2) (Moorcroft et al., 2001) to study the impact of climate change on the forest studied in Uriarte et al. (2009). The ED2 model is a process-based vegetation dynamics model; it represents the size and age structure of the forest, and thus the model can represent the observed differential impact from disturbances (such as fire, drought, insects, land use change, and natural disturbances) across plants of different functional groups and size classes (Medvigy et al., 2012; Zhang et al., 2015; Miller et al., 2016; Trugman et al., 2016). However, the impacts of hurricane disturbances have not been implemented in vegetation dynamics models, and thus the long-term effects on the forest of a changing hurricane regime have not been investigated.
As mortality and recovery vary with species, the species composition of the forest is affected by hurricane disturbances. In modeling studies, it is impractical to incorporate each and every individual species (tens and hundreds). To address variation in species diversity, there has been a strong effort in the past decades to incorporate functional diversity in vegetation dynamics models (Moorcroft et al., 2001; Sakschewski et al., 2016; Fisher et al., 2018; Fisher and Koven, 2020). This effort acknowledges the variability in traits and trade-offs of species that exist in tropical forests (e.g., Baraloto et al., 2010). Three plant functional types (PFTs) are identified for the species in tropical forests during a secondary succession after a disturbance; they are early, mid, and late successional PFTs (hereafter Early, Mid, and Late PFTs), corresponding to the three successional stages during the secondary succession (Kammesheidt, 2000). Specifically, Early PFT dominates the early successional stage of the recovery; it includes fast growing pioneer species that have low wood density, establish and recruit in open gaps formed after disturbances, and grow rapidly in the high light environment. Mid PFT dominates the mid successional stage after a disturbance and includes species that have intermediate growth and are somewhat shade tolerant. Late PFT dominates the late successional stage and includes species that have slow growth and are shade tolerant. Using three PFTs is also a compromise between representing a range of life strategies while not adding too much complexity in model parameterizations (Moorcroft et al., 2001; Medlyn et al., 2005).
One important and distinct species in tropical forests in the Caribbean
islands is the palm species
In this paper, we describe the implementation of hurricane mortality and recovery modules that account for the variation with disturbance severity, forest resistance state, PFT, and diameter size of individual stems in the Ecosystem Demography model (ED2). The model is then used to study the recovery of a tropical rainforest after hurricane disturbances. The results indicate that a scenario with a single hurricane disturbance has little long-term impact on forest structure and composition but enhances the aboveground biomass accumulation of a tropical rainforest, relative to a scenario of no hurricane disturbance.
Tree censuses were carried out at Bisley Experimental Watersheds (BEW) in
the Luquillo Experimental Forest in Puerto Rico starting in 1989, 3 months before Hurricane Hugo (pre-Hugo 1989), and repeated 3 months
after Hurricane Hugo (post-Hugo 1989), and then every 5 years since then
(1994, 1999, 2004, 2009, 2014). The census recorded the diameter at breast
height (1.3 m) (DBH) and species of each stem with DBH
The Ecosystem Demography model (ED) is a cohort-based model, and it describes the growth, reproduction, and mortality of each cohort in each patch in a forest site. A cohort is a group of stems with the same PFT and similar diameter size and age. A patch is an area with the same environmental condition and disturbance history. A cohort accumulates carbon through photosynthesis, and the net accumulated carbon (i.e., gross primary productivity minus respiration and maintenance of living tissues) will be used for growth and reproduction. When a cohort is mature, reaching the maturity reproductive height (e.g., 18 m), the cohort will allocate a portion of carbon to reproduction (e.g., 30 % of net carbon accumulation to seeds, flowers, and fruits), and the rest of the net accumulated carbon will be used for structural growth. Structural growth is quantified by the increase in DBH through structural-biomass–DBH allometries; stem height, leaf biomass, and crown area are then scaled given the H–DBH, leaf-biomass–DBH, and crown-area–DBH allometries. Each cohort will also experience mortality from multiple factors, including aging, competition, and disturbance, which will be described in detail in Sect. 2.3.2.
The model simulates transient fluxes of carbon, water, and energy during short-term physiological responses and long-term ecosystem composition and structure responses to changes in environmental conditions. The second version of the ED model, ED2, modifies the calculations of radiation and evapotranspiration of the original ED model, leading to a more realistic long-term response of ecosystem composition and structure to atmospheric forcing (Medvigy et al., 2009; Longo et al., 2019a). Details of the ED and ED2 models can be found in Moorcroft et al. (2001), Medvigy et al. (2009), and Longo et al. (2019b). Here we add a new PFT (Palm) and implement hurricane disturbance in the ED2 model, and we name it ED2-HuDi V1.0.
The standard ED2 model represents a variety of broadleaf trees, needleleaf
trees, grasses, and lianas (Albani et al., 2006; Medvidy et al., 2009; Longo et
al., 2019b; di Porcia e Brugnera et al., 2019). Yet, to date, none of the
existing PFTs describe the traits of palms, even though palms are a globally
abundant component of tropical forests (Muscarella et al., 2020). We know
that the palm species that occurs at our study site (
The allometric relationships between stem height (
For other allometric relationships, such as leaf-biomass–DBH, structural-biomass–DBH, and crown-area–DBH relationships, we used the model default for Early, Mid, and Late PFTs and assumed that Palm has the same relationships as Early (Fig. S1 in the Supplement).
The height–diameter (DBH) relationship for the four PFTs:
The ED2 model accounts for several types of disturbances, such as fires,
land use, and logging (Albani et al., 2006; Longo et al., 2019b), but not
hurricane disturbance. To account for hurricane impacts, we implement a
hurricane-induced wind mortality module and a seedling recovery module in
the model. The wind mortality module consists of two parts – the disturbance
rate of the forest area (
The mortality as a function of the size structure of the
forest for each PFT and DBH class. The size structure is represented as the
proportion of large stems (DBH
Hurricanes not only cause immediate stem mortality, but also affect the
establishment of seedlings by opening the canopy (Everham and Brokaw, 1996;
Brokaw, 1998; Uriarte et al., 2009, 2012). Brokaw (1998)
pointed out that hurricanes promote germination and seedling establishment
of the early successional species
The seedling density from seed rain for each PFT as a function of time since disturbance.
The concept of the generalized likelihood uncertainty estimation (GLUE) (Binley and Beven, 1991; Beven and Binley, 1992; Mirzaei et al., 2015) has been widely used to calibrate parameters in complex hydrological models. The steps of GLUE include (1) generating a number of samples of the parameter set from a prior distribution of the parameters, (2) running the simulation for each parameter set, (3) choosing a likelihood function (or weight function) to calculate the weight of each simulation based on observations and the estimated outputs from the model simulation, and (4) selecting the optimal parameter set and estimating the posterior distribution of the parameters and the posterior distribution of the output variables. Here we use GLUE, for the first time, to calibrate the parameters in the ED2 model.
To obtain the prior distribution of parameters, we build on a previous parameter sensitivity analysis using the ED2 model for a nearby forest in Puerto Rico by Feng et al. (2018). They demonstrated that model simulations are sensitive to 10 parameters, listed in Table 1, and provided the posterior mean and 95 % confidence limits of the parameters calibrated from plant traits observations using the Predictive Ecosystem Analyzer (PEcAn; LeBauer et al., 2013). We select the same parameters and use the posterior distribution of those parameters from Feng et al. (2018) as the prior distribution for the GLUE in our study. We cannot just use their parameter distributions as final results because our implementation has a site-specific set of allometric equations, explicitly represents palms as a separate PFT, and considers hurricane disturbances (Sect. 2.2). Feng et al. (2018) reported only the mean and the upper and lower 95 % confidence limits of the parameters (not the entire distribution); we assume that the parameters have lognormal distributions. For the Palm PFT, we assume that it has the same distributions as Late, except that the woody tissue density of Palm has the same distribution as that of Early. From a different study system, Wang et al. (2013) constrained the dark respiration factor from 0.01–0.03 to 0.01–0.016 by assimilating observations of model output variables. Following Wang et al. (2013), we restrict the dark respiration factor to a smaller range with a uniform distribution between 0.005 and 0.0175 for each PFT. Consistent with Meunier et al. (2022), we found that model results are also sensitive to the parameter clumping factor (Fig. S2). Therefore, we add the parameter of clumping to the set being calibrated. Clumping factor is the ratio of effective LAI to the total LAI and affects the transmission of radiation (Chen and Black, 1992); it ranges from zero to one with zero representing leaves clumped in a single point (0-area) and one representing leaves uniformly distributed in the unit area. Because of tree crowns, branches, and subbranches, leaves of the plant canopy are not uniformly distributed per unit area nor clumped at a single point. We assume that the clumping factor is the same for all PFTs, and the distribution of the clumping factor is uniform between 0.2 and 0.8.
We sample 10 000 realizations for the 41 parameters (10 parameters for each
of the four PFTs and the 1 clumping parameter for all PFTs) using the
Latin hypercube sampling method embedded in MATLAB (Stein, 1987). We
initialize the model with the pre-Hugo 1989 observations and run the model
for 29 years, corresponding to 1989–2018. The first 25 years (1989–2014)
are used to calibrate the model with observations and the last 4 years
(2015–2018) for validation. We tested different calibration lengths
(1989–1999, 1989–2004, and 1989–2009). The 1989–2009 calibration period gives
the same optimal simulation as the 1989–2014 calibration period (Fig. 4), but
shorter calibration lengths 1989–1999 (Fig. S3) and 1989–2004 (Fig. S4) throw away critical recovery information and cannot give robust
simulation in the validation period. We calculate the mean squared errors
(MSEs) of each realization (
We select the simulation with the smallest MSE as the optimal simulation and
the corresponding parameter set as the optimal parameter set. To obtain the
posterior distribution of parameters, we first calculate the weight
(likelihood) of each realization following Binley and Beven (1991),
The non-hurricane mortality of Palm is not well represented in the model
(Fig. S7), as initially calibrated. The observed non-hurricane mortality
is an overall mortality regardless of the cause of the death and is
calculated from non-hurricane censuses, whereas the non-hurricane mortality
in model simulations includes aging mortality, competition mortality, and
disturbance mortality. We turned off all disturbances except for hurricane
disturbance and treefall disturbance. The disturbance mortality includes the
background exogenous mortality and treefall disturbance rate. Background
mortality rate is 0.014 yr
With a lower mortality (decreasing aging mortality from
After changing the aging mortality of Palm to zero and the seedling density
to a lower and slowly decreasing value, we did not repeat the GLUE. This is
because Palm has constrained DBH size (between 10 and 25 cm) and decreasing
the aging mortality increases its density while decreasing seedling
reproduction decreases its density, which maintains the overall density of
Palm, without affecting other variables of Palm nor variables of other PFTs
(Fig. S9). Therefore, we use the parameter set found from the GLUE (Table 1) but with 0-aging mortality and a lower seedling density recovery
(
Using a similar approach to PEcAn (LeBauer et al., 2013), we analyze the sensitivity of model simulations to the parameters and the contribution of the parameters to the variances. Specifically, we set up nine experiments for each of the 41 parameters, corresponding to the nine quantiles (10th, 20th, …, 90th) of the posterior distribution of each parameter, while all other parameters remain constant at their optimal. For the total 369 sensitivity experiments, we initialize the model with the pre-Hugo observation and run each experiment for 25 years (1989–2014).
To study the stability of the optimal parameter set, we calculate the MSE of
each experiment and compare it with the MSE of the optimal. To quantitatively
study the sensitivity of output variables to the parameters, we calculate
the standardized cubic regression coefficient (
To quantitatively study the uncertainty of the simulated variables (stem
density, AGB, BA, LAI, etc.) from the uncertainties of the parameters, we
calculate the coefficient of variation (
To study the impact of the initial condition of the forest on the recovery, we set up two experiments with different initial forest states (pre-Hugo state and pre-Maria state) with a hurricane disturbance in the first simulation year (experiment IhugoH1 and experiment ImariaH1, hereafter), and one control experiment with pre-Hugo state and no hurricane disturbance in all simulation years (experiment IhugoHn, hereafter). The three experiments run for 112 simulation years (corresponding to years 1989–2100). The meteorological drivers between 1989 and 2017 are observations from meteorological towers at BEW (González, 2017), and the meteorological drivers between 2018 to 2100 are randomly sampled from the observations between 1989 and 2017. Hurricane disturbance is turned off in all simulation years for experiment IhugoHn and in all but the first simulation year for experiments IhugoH1 and ImariaH1. Thus, experiment IhugoHn represents the succession of the forest without hurricane disturbances for more than a century. Experiments IhugoH1 and ImariaH1 represent the recovery of the forest from a hurricane disturbance given different initial conditions of the forest.
Figure 4 shows the optimal model simulation along with census observations
for years 1989–2018. The simulated stem density of Early increased from
0.0027 individuals m
Time series of variables from observation (dots and error
bars) and the optimal simulation (red lines).
The standard deviation of the estimated variables with
In the verification period between 2015–2018, the simulated overall stem
density, basal area, and aboveground biomass have a relative bias of
Table 1 shows the optimal set of the parameter values. The clumping factor
(0.34) is lower than that from other studies in different locations
(
The optimal parameter set obtained from the GLUE method.
Figure 6 shows the posterior and prior probability distribution functions (PDFs) of the parameters. The most significant differences between the posterior and the prior distributions are for the parameters of clumping factor (Clf) and dark respiration rate (Rdf). The posterior PDFs of some parameters (i.e., carboxylation rate, specific leaf area, leaf width, stomatal slope, and wood density), which are well constrained by observational trait data (Feng et al., 2018), do not change much from the priors (the maximum difference between the prior and posterior CDFs is generally less than 0.1). The posterior PDFs of other parameters (e.g., leaf turnover rate, quantum efficiency, and fine root allocation), especially for the Early and Mid PFTs, with few observational trait data (Feng et al. 2018), changed greatly from the prior distributions (the maximum difference between the distributions is around 0.3).
The prior (solid line) and posterior (dashed line) probability density functions for the four PFTs (colors) of the 11 parameters. The first 10 parameters are PFT-dependent, and the last one leaf clumping factor (Clf) is PFT-independent. Palm has the same prior distribution as Late for all parameters except that the wood density (WDe) of Palm has the same prior distribution as that of Early. The long name of each parameter is shown in Table 1.
Among the 369 sensitivity experiments with different parameter values, 57 of them have slightly smaller-than-optimal MSEs, but the simulated variables (stem density, AGB, PFT composition, and size structure) from those experiments are very close to optimal (Fig. S10), indicating that the optimal simulation we found from GLUE is stable given the uncertainties of the parameters.
In terms of the sensitivity of simulated variables on the parameters, the
magnitude of standardized cubic regression coefficients (
The standardized cubic regression coefficient (
The stem density has a larger variation than LAI, BA, and AGB after 25 years of simulation (Fig. 8). Given that large stems contribute more to LAI, BA, and AGB, larger variation of stem density than LAI, BA, and AGB indicates that small stems are more variable than large stems. The variation of those variables also varies with PFTs. For the stem density, Late PFT has the largest variation, followed by Early, then Mid, and Palm has the smallest variation, indicating that stem density of small Late is the most sensitive to the uncertainty of the parameters. For BA, AGB, and LAI, Early and Mid PFTs show the highest variability, followed by the Palm PFT, and the Late PFT has the lowest variation, indicating that large stems of Early and Mid PFTs are more sensitive to the uncertainty of the parameters than large stems of Late and Palm PFTs.
The coefficient of variation (
The variance explained by each parameter for variables
The variance decomposition analyses reveal that 50 % of the uncertainty of the stem density comes from the quantum efficiency of Late (QefL) (Fig. 9). However, QefL explains less than 10 % of the uncertainty in BA, AGB, and LAI, indicating that QefL has significant effects on the density of small stems, but fewer effects on the density of large stems. In other words, QefL impacts the recruitment and establishment of stems more than the growth of stems. The uncertainty of the growth of stems comes from the growth respiration factor (Rgf), which explains about 10 % of the uncertainty. The interaction among parameters accounts for 21 % of the uncertainty of the stem density, and more than 50 % of the uncertainty of the BA, AGB, and LAI.
Figure 10 shows the 112-year simulations of the forest initialized with
different forest states (pre-Maria state and pre-Hugo state) with or without
hurricane disturbance at the first simulation year. Without hurricane
disturbance (IhugoHn), the forest experiences a decrease (
Time series of eight variables from the simulation of
the three experiments: IhugoHn, IhugoH1, ImariaH1. The dotted lines are the
initial state of the variables for each experiment (IhugoHn and IhugoH1 have
the same initial state). The variables in
After 80 years, the PFT composition reaches a steady state (the change of
30-year moving average is less than 1 % compared to the previous year;
Fig. S13), where the Early, Mid, Late, and Palm PFTs account for 11.8 %,
10.6 %, 65.3 %, and 12.3 % of the total stem density, respectively
(Fig. 10 e, f, g, h). This state is significantly different from the
initial state and exhibits a 16 % reduction in the proportion of the Mid
PFT. It exhibits increases on all other PFT proportions (
Compared with the experiment without hurricane disturbance in the first
simulation year (IhugoHn), the experiments with hurricane disturbance in the
first simulation year (IhugoH1 and ImariaH1) reach higher BA and AGB levels
after 60 years of succession from the hurricane disturbance (Fig. 10c and
d). This is due to the carbon accumulation of large Late PFT in disturbed
forests (Fig. S12g and k). Large Late trees in disturbed forest (IhugoH1
and ImariaH1) have a higher growth rate and lower background mortality rate
compared to those in the undisturbed forest (IhugoHn) (Fig. 11) because of
the decreased competition to reach the open canopy. As the disturbed forest
recovers, the BA and AGB increase to the level of the undisturbed forest
(Fig. 10c and d), the growth rate decreases (Fig. 11a), and the
mortality rate increases to the levels of those in the undisturbed forest,
especially for severely disturbed forest (IhugoH1) (Fig. 11). With lower
mortality and higher growth rate in the first 60 years, there will be more
large Late trees in the canopy at the end of the simulation (12 vs. 8 individuals ha
Times series of
The recovery is different with different initial states. With pre-Hugo state (IhugoH1), the forest takes 25 years to recover to the pre-disturbance BA and AGB levels (Fig. 10c and d), but with pre-Maria state (ImariaH1), it takes only 10 years to recover to the pre-disturbance BA level (Fig. 10c) and 5 years to the pre-disturbance AGB level (Fig. 10d). The succession dynamics are different, too. With pre-Hugo state, the hurricane-induced mortality is very high, and thus the canopy opens, Early and Palm PFTs recruit greatly in the first 20 years (Fig. S11e and h), and then it is taken over by the Late PFT (Fig. S11g). With pre-Maria initial state, the hurricane-induced mortality is low, the canopy is not significantly changed after the hurricane, and Early PFT does not recruit as much as it does in the pre-Hugo state initialized simulation (Fig. S11i and e). The PFT composition after 100 years is similar for the two simulations, but the BA and AGB are not (Fig. 10). The BA and AGB with the pre-Maria initialization are higher than those with the pre-Hugo initialization throughout the 110 years of simulations, even though the initial BA and AGB levels in the pre-Maria state are lower than those in the pre-Hugo state (Fig. 10c and d). This is because of the higher mortality in the first year with pre-Hugo state, leading to a larger reduction in the density of large stems. With the succession following the disturbance, there are more large stems, especially Late and Palm, in the pre-Maria simulation than in the pre-Hugo simulation (Fig. S14), contributing to the higher AGB and BA in the pre-Maria simulation (Fig. S12g, h, k, and l).
We developed a hurricane module (including a mortality module and a recovery module) for the ED2-HuDi model, based on census observations. We then applied a parameter estimation algorithm, GLUE, to calibrate important parameters in the model and selected the optimal parameter set for the final model simulation. However, because the observations are limited to only two hurricane events, the hurricane module may be biased toward the two observations. The simulation results show some discrepancies with observations, and these discrepancies could be in part due to the GLUE approach and parameter uncertainties. Here we discuss the uncertainty associated with the developed hurricane module, the limitations and advantages of the GLUE framework, and the uncertainties of model outputs.
We included a hurricane mortality module and a hurricane recovery module for hurricane disturbance. Crown damage is also an important part of hurricane disturbance and could have important impact on forest structure and carbon accumulation (Leitold et al., 2021), but we did not include crown damage in the hurricane disturbance module because the census data used to develop and calibrate the module do not include crown damage information. The hurricane mortality module was developed based on observations from two hurricane events at the study site. The relationship between mortality and forest size structure (proportion of large stems) was fitted to a logistic function (Fig. 2) for each PFT and DBH class. Generally, Palm PFT has a lower mortality than other PFTs, but Palm mortality was higher (11 % for Palm, 9 % for Mid, and 3 % for Late) when the forest was dominated by large stems (e.g., large stem proportion is 0.6, except for the high mortality of 39 % for Early; Fig. 2b). This was due to the high mortality of Palm during Maria, which was a result of plant pathogens (Zhang et al., 2022b; Heartsill Scalley, 2017). The mortality of large-stem Early PFT is significantly different from other PFTs, and this difference was due to the significantly higher mortality of large-stem Early during Hurricane Maria compared to other PFTs. Such high mortality of large-stem Early may be a result of other factors besides hurricane disturbance, and it could be further studied if there were more observations. Future work could include observations from other study sites to improve the hurricane disturbance module.
There are four critical parameters associated with the hurricane disturbance
module, including disturbance rate of forest area (
GLUE samples from continuous distributions, but the sampled parameter sets are in a discrete space; therefore, the GLUE approach may not lead to the true optimum due to the finite number of samples. To justify the sample size of 10 000 for 41 parameters in this study, we repeated GLUE for a larger sample size (20 000). The optimal simulation from 20 000-sample GLUE (Fig. S15) is very similar to that from the 10 000-sample GLUE (Fig. 4), and the optimal parameter sets from the two GLUEs are similar, suggesting that the two GLUEs found an optimum around the same local optimum and 10 000 samples are sufficient for the 41 parameters. However, given the nature of equifinality, there may be multiple parameter sets that can lead to the same observed state (Beven and Freer, 2001), and thus the optimal parameter set we found from GLUE may be one of many possible solutions.
Although GLUE may not guarantee the global optimum, it implicitly handles any effects of model nonlinearity, model structure errors, input data errors, and parameter covariation (Beven and Freer, 2001). Moreover, GLUE allows us to optimize parameters using any variables of interests in the cost function. For example, in our study, we want to make sure the model captures the size structure and PFT composition of the forest community, and thus we utilized forest stand variables including stem density, growth rate, and BA of each PFT in the cost function. Note that we did not calibrate the parameters using plant trait observations in this study, because the parameters we use are already calibrated with plant traits observations in Feng et al. (2018) and we adopted their calibrated parameters in our study (see Sect. 2.3.1).
To be consistent with census observations, we included stems with DBH
Our results that modeled variables have different responses to parameters in the short term (e.g., first simulation year) and in the long term (e.g., 25th simulation year) agree with a previous study (Massoud et al., 2019). Furthermore, we showed that variables of a specific PFT are most sensitive to the parameters of the same PFT, but also sensitive to parameters of other PFTs. Those interactions between variables and parameters indicate the competition among PFTs. For example, Palm is sensitive to its own parameters, but also to Early SLA. This can be explained by the competition for light between Early and Palm, where a higher SLA of Early PFT leads to a higher LAI of Early allowing Early to photosynthesize more efficiently and thus be more competitive in the community. Those competitions are important for the co-existence of PFTs in model simulations and critical to the PFT composition and succession.
Hurricanes are a major disturbance to tropical forests, but hurricane disturbance had not been implemented in any model of vegetation dynamics. In this study, we implemented hurricane disturbance in the Ecosystem Demography model (ED2) and calibrated the model with forest stand observations of a tropical forest in Puerto Rico. The calibrated model has good representation of the recovery trajectory of PFT composition, size structure, stem density, basal area, and aboveground biomass of the forest. We used the calibrated model to study the recovery of the forest from a hurricane disturbance with different initial forest states and found that a single hurricane disturbance changes forest structure and composition in the short term and enhances AGB and BA in the long term compared with a no-hurricane situation. Forests with wind-resistant initial state will have lower mortality, recover faster, and reach a higher BA and AGB level than forests with a less wind-resistant initial state.
The model developed and results presented in this study can be utilized to
understand the fate of tropical forests under a changing climate. Hurricanes
are likely to become more frequent and severe in the future with global
warming (IPCC, 2021). With frequent hurricane disturbances in the future,
forests will not have enough time to reach a steady state, and the structure
and composition will be constantly changing, which provides different
initial states for future hurricane disturbances and thus different recovery
trajectories. Climate change with changing temperature, precipitation, and
CO
The ED2-HuDi software is publicly available. The most up-to-date source
code is available at
The supplement related to this article is available online at:
RLB conceptualized the work. THS provided field data and contributed the ecological interpretation of the results. RLB and JZ developed the methodology and performed the analyses. JZ and ML interpreted results. JZ wrote the first draft of the manuscript. All authors discussed results and critically revised and edited the manuscript.
The contact author has declared that neither they nor their co-authors have any competing interests.
The findings and conclusion in this publication are those of the authors and should not be construed to represent any official USDA or U.S. government policy.Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
We thank Paul Moorcroft, Xiangtao Xu, Elsa Ordway, Félicien Meunier, and
Erik Larson for discussions on the model implementation and parameter
sensitivity analyses. We acknowledge high-performance computing support from
Cheyenne (
This research has been supported by the National Science Foundation (grant no. EAR1331841).
This paper was edited by Hans Verbeeck and reviewed by three anonymous referees.