Abstract
Background Hurricanes and other wind events are significant disturbances that affect coastal urban forests around the world. Past research has led to the creation of wind resistance ratings for different tree species, which can be used in urban forest management efforts to mitigate the effects of these storms. While useful, these ratings have been limited to species common to urban forestry in Florida, USA.
Methods Drawing on past ratings and data from a global literature review on tropical storm research, we created a machine learning model to broaden both the geographic coverage and the variety of species currently assessed for their resistance to wind.
Results We assigned wind resistance ratings to 281 new species based on the available data and our modelling efforts. The model accuracy and agreement with the original ratings when applied to the testing data set was high with 91% accuracy.
Conclusions Our study demonstrated how a machine learning algorithm can be used to expand rating systems to include new species given sufficient data. Communities can use the expanded wind resistance rating species list to choose wind resistant species for planting and focus risk assessment on low wind resistant trees.
Introduction
Hurricanes profoundly impact coastal communities and their urban forests. Extreme winds and flooding directly damage trees by breaking branches, knocking trees over, snapping trunks, or causing stress from prolonged inundation or salt exposure (e.g., Wang and Xu 2000; Wiersma et al. 2012; Middleton 2016). In turn, broken trees can damage property and infrastructure, contributing to power outages and hindering emergency operations (Yum et al. 2020; Taylor et al. 2022). During cleanup after a hurricane, tree-generated debris removal can increase cleanup costs, significantly increasing the ecosystem disservices generated by the urban forest. For example, in Florida the 2004–2005 hurricane season produced an average of 233 m3 of debris per kilometer of street (Staudhammer et al. 2009). Furthermore, the cleanup process itself can lead to injuries among residents and professionals (Marshall et al. 2018). Ultimately, the loss of urban trees leads to a loss of ecosystem services (Olivero-Lora et al. 2022) and the harm damaged trees can cause to people, infrastructure, and property can all increase the public’s negative perception of trees (Wyman et al. 2012; Judice et al. 2021; Roman et al. 2021).
Observational studies of urban and rural forests demonstrated that many abiotic and biotic factors interact to influence the likelihood and severity of hurricane damage to trees (Salisbury et al. 2023). For example, some studies reported that taller trees were more likely to fail during a hurricane (Francis 2000; Johnsen et al. 2009), while others reported no relationship between height and damage (Wiersma et al. 2012; Landry et al. 2021). Rather than considering height in isolation, additional characteristics such as slenderness and crown width can moderate the impacts of height on susceptibility (Tabata et al. 2020; Torres-Martínez et al. 2021). Additionally, some species are more likely to be damaged by hurricanes than others (Basnet et al. 1992; Saito 2002; Curran et al. 2008). Species characteristics such as wood density, the ability to grow buttresses, and crown to stem ratio can influence susceptibility (Elmqvist et al. 1994; Vandecar et al. 2011; Paz et al. 2018). Though again, these characteristics interact with environmental conditions such as stand density or soil type (Foster 1988; Rutledge et al. 2021). Other researchers have developed empirical and mechanistic-based models to predict hurricane damage to forests (e.g., Blennow and Sallnäs 2004; Gardiner et al. 2008; Dupont 2016). While such models add to our mechanistic understanding of hurricane damage to trees, their development for trees growing in natural or managed forest stands limits their direct application to trees in the built environment.
One of the most studied and consistently significant predictors of tropical storm failure is tree species (Salisbury et al. 2023). To mitigate some of the risks hurricanes pose to the urban forest, Duryea et al. (2007a, 2007b) created a rating system that classified tree species based on their ability to resist hurricane wind damage. They based the rating system on observations of urban tree damage across Florida and Puerto Rico following the 2004–2005 hurricane season, encompassing 9 hurricanes. They also surveyed urban forest professionals in their region and asked them to estimate the wind resistance of common urban tree species (Duryea et al. 2007a, 2007b). They combined the damage observations and expert opinions to rate 137 tree and palm species. This is the most comprehensive set of wind resistance ratings for urban trees, though the list focused on tree species common to Florida’s urban areas. Yet, hurricanes and typhoons pose risks to urban forests in many other regions of the world (e.g., Cole et al. 2021), highlighting the need for an expansion of the Duryea et al. (2007a, 2007b) rating system.
Everham and Brokaw (1996) also compiled a list of tree species damaged by catastrophic winds (hurricanes, gales, severe windstorms) based on a literature review. They reported low, medium, and high damage ratings for 242 tree species. However, ratings were reported as originally published by the cited authors. No attempts were made to standardize these ratings to allow for formal comparisons of wind resistance across the entire collection of species. Their approach resulted in some species receiving multiple ratings; for example, Quercus virginiana Mill. was documented in 4 studies and given ratings of low (twice), medium, and high damage. While a substantial body of work, Everham and Brokaw’s collection only relied on damage observations from a wide range of forested environments and was not tailored to the conditions present in the urban forest. By contrast, Duryea et al. (2007a, 2007b) designed their rating system explicitly for use by urban forestry professionals.
Some studies of hurricane damage in forest ecosystems have compared their results to the Duryea ratings. Negrón-Juárez et al. (2010) observed that study plots with a greater proportion of lower wind resistant species tended to have higher mortality compared to plots with more wind resistant species. Harcombe et al. (2009) ranked species in their long-term study plots by wind resistance. Their ratings of some species aligned with ratings developed by Duryea et al. (2007a, 2007b), though notably Pinus taeda and P. palustris did not, likely because in this study these 2 species grew in dense, mixed stands surrounded by hardwoods. Other researchers have incorporated Duryea’s ratings into scoring systems that combine multiple characteristics into a climate change vulnerability assessment for urban tree species (e.g., Foran et al. 2015; Brandt et al. 2021; Liu et al. 2021). Many urban tree species recommendations lists include information about wind resistance (e.g., City of Melbourne 2011; Fazio 2014), some of which are based on Duryea’s system (City of Dunedin 2016; USF Water Institute 2024; Wilson 2024).
Given the utility of the original work of Duryea et al. (2007a, 2007b) and the limited palette of species they evaluated, our research aims to increase the number of tree species with wind resistance ratings beyond the original 137 species documented in their work. However, this rating system poses a challenge to increasing the number of rated species since Duryea et al. (2007a, 2007b) provided a limited description of the rating process and did not create clear definitions for each wind resistance category. Machine learning methods excel at finding patterns in large, often nonlinear datasets and generating accurate predictions (Olden et al. 2008) and have been used in other forestry applications (e.g., Hart et al. 2019; Correa Martins et al. 2021). Here, we demonstratethe novel use of a machine learning methodology with storm damage data collected from a literature review and species characteristic data to generate wind resistance of previously unrated species. By increasing the number of tree species with wind resistance ratings, this research can be used by urban foresters, arborists, and communities to inform species selection in regions prone to hurricanes.
Methods
Systematic Literature Search
We conducted a systematic literature search to identify peer-reviewed research and dissertations that contained hurricane damage data as a proportion of a population for a given species. We searched for papers and dissertations published between 1900 and 2022 in English, Chinese (Mandarin), French, Japanese, Portuguese, and Spanish. We searched in several search engines and databases in addition to forestry-related journals that may not have been indexed in a particular database (Salisbury et al. 2023). The last search was conducted on 2022 May 5. Our core search string in English was “forest AND (hurricane OR cyclone OR typhoon)”; its translation and synonyms in the 5 additional focal languages can be found in Salisbury et al. (2023).
We screened the results of our search using the following criteria to include papers in the dataset: (1) data collection occurred within 2 years of a tropical cyclone; (2) the only disaster studied was a tropical cyclone or tropical storm as opposed to a study where a forest was impacted by both a tropical cyclone and second disaster (e.g., landslide) and tree mortality could not be attributed to severe wind damage alone; (3) researchers used ground-based methods of data collection as opposed to techniques such as aerial surveys; and (4) the paper reported data at the species level as a proportion of a population or sample and provided the scientific binomial name of the species. We excluded mangrove ecosystems since these species are not typically planted in managed urban habitats.
We rated methodological completeness by answering the following questions for each study: did the study (1) collect data using a randomized study design or by conducting a complete inventory?; (2) report observations of damage based on the type of damage (e.g., broken branches, snapped trunk)?; (3) conduct an assessment of the tree’s condition or potential risk of failure?; and (4) measure tree size? We assigned one point for each question that received a “yes,” for a total potential score of 4.
After screening, we extracted damage data to a species’ population and other relevant information from each study. Each species reported in each study represented an observation in our dataset, i.e., an observation was composed of a species, its study, information about the study site and hurricane, the proportion of the population damaged by the hurricane, and the species’ traits. Consequently, a species reported in multiple studies had multiple observations within our study. When possible, we used Tabula (Aristarán et al. 2018) to extract damage data in table form, otherwise we manually copied the data into spreadsheet form. We extracted data from figures using WebPlotDigitizer (Rohatgi 2024)(Automeris LLC, Frisco, TX, USA). For papers written in Spanish, Japanese, or Chinese, a multilingual team member checked the translation to English made by Google Translate. We also recorded the location of the study site, the tropical cyclone name, and the method details.
We classified damage data into one of 4 categories (Table 1) and each study as urban or rural. We used the proportion of a species’ population damaged or killed by a hurricane as the most direct measure of a species’ ability to resist hurricane damage. Duryea et al. (2007a, 2007b) used the proportion of mortality as one factor when assigning wind resistance ratings to species.
Damage categories and definitions used to classify data extracted from the literature review.
To the best of our knowledge, no study has compared hurricane damage to tree species between urban and rural settings (i.e., trees in highly built environments and trees growing in large forest stands). Nevertheless, considering these are 2 distinct settings, we included urban or rural setting as a model variable to account for these differences among studies. An urban study collected data within a city or town, either in a highly managed environment (e.g., street trees) or in a natural area located within an urban matrix. A rural study collected data within a natural area or timber plantation that had little to no potential impact from urban development. We excluded observations that were only made to the genus or family level.
The broad search terms produced 5,449 studies, of which 58 passed the screening process and had appropriate data for the study (Appendix Table S1 and Table S2). We attribute the low percentage of retained studies to the extremely general search terms we used and the apparent inability of some databases we searched to effectively utilize Boolean operators. The final studies in English, Chinese, Japanese, and Spanish produced 1,094 observations of species-level damage data. The studies took place in 15 countries and examined 42 unique tropical cyclones (Figure 1)(Appendix Table S1). Out of the original collection of observations, 285 observations representing 213 species lacked sufficient trait data to be used in the study (Figure 2).
The location of rural and urban study sites for papers with species-damage data identified in the literature search. Light blue shaded areas indicate the paths of tropical cyclones that have made landfall since 1970.
Wind resistance rating model development process.
Tropical Cyclone and Study Site Characteristics
Since some regions of the world and biomes are more prone to hurricanes than others, we included biome, latitude, and longitude in the model. Hurricane disturbance history may also influence a site’s susceptibility to future hurricane damage in diverging ways: previous storms could remove susceptible trees leaving the population more resistant to future damage or gaps created by previous storms could expose remaining trees to additional turbulence in future storms (Everham and Brokaw 1996; Ostertag et al. 2005).
We used data from the International Best Track Archive for Climate Stewardship (IBTrACS) to determine the maximum sustained wind speed for each tropical cyclone in our dataset (Knapp et al. 2010; Gahtan et al. 2024). This provided a consistent metric to compare studies using one facet of storm intensity. We also used IBTrACS to determine the amount of time that had elapsed between a study’s tropical cyclone and the previous tropical cyclone that had passed within 50 km of the study site. We determined the biome of each study site using the typology developed by Olson et al. (2001). Note: although a territory of the United States, we counted Puerto Rico separately from other USA study sites because of its distinct tropical habitats not found in the continental USA.
Tree Species Characteristics
Prior to extracting species’ predictor traits from several datasets (Table 2), the names of species identified in our literature review and species in the trait datasets were harmonized to the Leipzig Catalog of Vascular Plants (LCVP) taxonomic backbone (Freiberg et al. 2020) using the lcvplants package v.2.1.0 (Freiberg et al. 2020) in R v.4.2.2 (R Core Team, Vienna, Austria). We first harmonized species to the LCVP backbone using exact matching, then we used fuzzy matching for species without an exact fit. All fuzzy matched species were manually checked to ensure a reasonable match. We could not match 3 species (all hybrids) to the LCVP backbone with reasonable certainty (many of the species did not include botanical authorities), so we excluded them from further analysis.
Predictors used in the random forests model. IBTrACS (International Best Track Archive for Climate Stewardship).
Following observations from several hurricanes in Puerto Rico, Lugo (2008) hypothesized that tree growth rate could represent a hurricane response syndrome that includes architecture, elastic modulus (i.e., the ability to return back to its original shape when bent), successional status, and wood density. Of these traits, only wood density is widely and consistently documented. Species with denser wood, greater modulus of rupture (i.e., the ability to withstand bending), and modulus of elasticity can be more resistant to hurricane damage (Francis 2000; Duryea et al. 2007a, 2007b; Curran et al. 2008; Nakamura 2021). Granted, other biotic and abiotic factors can moderate the effects of wood density (e.g. Paz et al. 2018; Uriarte et al. 2019). Wood density also strongly correlates with other wood properties and captures many aspects of wood functions (Chave et al. 2009).
Several researchers have observed greater rates of hurricane damage to early successional or pioneer species which tend to be fast growers (Zimmerman et al. 1994; Ostertag et al. 2005; Canham et al. 2010). Yet, without a consistent definition of early, mid, and late successional species across a range of biomes and continents, successional status did not lend itself to predictive modeling. Instead, we selected leaf mass per unit area as a proxy variable since it tends to correlate with shade tolerance or successional status (Wright et al. 2004; Reich 2014; Lichstein et al. 2021). Generally, species with low leaf mass per unit area tend to be fastgrowing and intolerant of shade, or early successional, while higher leaf mass per unit area species tend to be slow-growing and tolerant of shade—characteristics associated with late successional species.
We used maximum height potential (as reported in the literature) as a predictor since taller trees are often (Foster 1988; Johnsen et al. 2009; Xi 2015), though not always, prone to more damage (Gao and Yu 2021; Landry et al. 2021). Most of our data sources did not include height data, so we used maximum height to generalize results at the species level. Observations of multiple types of catastrophic windstorms suggest gymnosperms (conifers) tend to be less wind resistant compared to angiosperms (Everham and Brokaw 1996; Gardiner 2021). Similarly, deciduous or semi-deciduous trees may have an advantage in high winds compared to evergreen species, though this effect has not been consistently observed (Everham and Brokaw 1996; Van Bloem et al. 2005).
Many of the traits (e.g., leaf type, leaf mass per unit area, maximum plant height, and wood density) came from the TRY Plant Trait database’s publicly available data (Kattge et al. 2020)(Table 2). Prior to the analysis, we removed TRY observations from experimental settings (e.g., growth chambers, glasshouses, etc.). We also removed TRY observations which had an error risk greater than 4, meaning that the trait value was more than 4 standard deviations away from the mean for other close relatives—as suggested by the database creators (Kattge et al. 2022). Occasionally, a species had multiple trait values in a dataset. In those cases, we calculated the mean trait value for the species and used that value in our analysis. If a species had multiple leaf types, we either assigned the leaf type with the most observations or deciduous/semideciduous.
Random Forest Predictive Model
The Duryea et al. (2007a, 2007b) system classified trees and palms into 4 categories: Low, Medium-Low, Medium-High, and High wind resistance. Our objective was to create a predictive model to assign previously unclassified tree species into one of these 4 categories. For the purpose of developing our predictive model, we limited our choice of predictors based on the availability of data within the studies identified in our literature review or within other databases. Some model variables represent generalizations of the species while others try to capture variation among study sites. We used a machine learning technique called random forest (RF) classification to create our predictive model that uses observations of hurricane damage and species characteristics to predict a species’ wind resistance category.
Despite their name, classification trees and RFs are statistical classification techniques that have nothing to do with the biological organisms we know as trees. Classification methods predict what category an observation belongs to, such as using leaf characteristics to predict species (Breiman et al. 1984). A classification tree takes a group of training observations and splits them into subgroups based on input variables so that the observations within each subgroup are more similar to each other (Genuer and Poggi 2020). Once trained, the tree can be used to make predictions with new observations. However, single trees are highly unstable. That is, small alterations to the training data or variables can produce a very different tree and outcome (Genuer and Poggi 2020).
By contrast, an RF method generates a large number of randomly constructed classification trees that are not necessarily optimal, but when their results are combined together, they have better predictive capability (Breiman 2001). Typically, an RF will contain hundreds to thousands of classification trees, called an ensemble. The RF will run a new observation through each tree in the ensemble and then make a final prediction based on the most frequent classification. The advantage of this approach is that it produces stability in the model by essentially averaging many classification trees together (Breiman 2001). They generally produce models with low bias and variance. Compared to other classification techniques, RFs are valuable in ecological studies because they can assess the importance of predictors, use proximities to impute missing data, and be applied to a wide range of research questions (Cutler et al. 2007).
Random Forest Model Setup
Table 2 lists the predictors input into the model. While we observed 4 types of damage in our literature review (mortality, multiple damage types, stem failure, and root failure) some damage types (e.g., stem failure) were present in some studies but not others, leading to severe imbalance in the availability of damage data. Consequently, we consolidated the 4 damage types into a single variable (damage). After testing different permutations of the damage data, when multiple types of damage were reported for a single observation (e.g., mortality and root failure), we would first assign “multiple damage types” to the final damage value. If the observation lacked multiple damage types data, then we would use mortality data, followed by root failure, then stem failure.
For the response variable, the 4 levels of wind resistance ratings, we first collected all observations for the original tree species in Duryea et al. (2007a, 2007b) which had a complete set of wood density, leaf mass per unit area, and maximum height trait data. We then randomly selected 70% of those observations for use as training data, while the remaining 30% were test data. During the random selection process to split the data, we stratified data by wind resistance rating to ensure even representation of each classification group.
We fit an RF model to the training data using the “rf” method of the caret package in R (Kuhn 2008). We set the model to contain 1,000 RF trees and we used 10-fold cross validation with 5 repeats when fitting the model to reduce model variance. Model tuning indicated that the model should test 8 variables at each node in a tree.
We subsequently tested model performance using the test dataset to determine overall accuracy, adjusted Cohen’s Kappa with equal weights to each response category using the DescTools package (Signorell et al. 2024), sensitivity, and specificity. Sensitivity is the ratio of true positives to all positive predictions while specificity is the ratio of true negatives to all negative predictions. These values were calculated for each wind resistance category. For example, the specificity of the High category would be the number of correctly classified High observations to the total number of all observations predicted to have a High rating. We also calculated the multiclass area-under-curve (AUC) for the testing dataset following a procedure created by Hand and Till (2001) and implemented with the pROC package (Robin et al. 2011). We assessed the importance of each predictor using caret’s “varImp” function, which calculates the total decrease in node impurity, measured by the Gini index, that results from splitting data on a given variable and then averages that decrease across all trees.
Model Application
We identified 486 species in our literature review (described above) which did not have wind resistance ratings. Of those species, 39% had a complete set of wood density, maximum height, leaf mass per unit area, and leaf type data (188 species). We investigated the utility of imputing missing data using a pre-processing bagged trees method (m = 10 models) (Kuhn and Johnson 2013). Preliminary testing revealed that observations with imputed wood density or imputed leaf mass per unit area and maximum height did not produce reliable predictions. Consequently, we excluded species with missing wood density or missing leaf mass per unit area and maximum height from further analysis (n = 205 excluded species). We applied imputation with bagged trees to predict missing values to species which were missing only leaf mass per unit area or maximum height (n = 93 imputed species). After this preparation, we applied the trained RF model to 281 species to predict their wind resistance rating based on 440 total observations of population damage with 85 species having data reported in more than one study (Appendix Table S2)(data and original model also available at https://github.com/AllysonS/TreesForHurricanes).
We evaluated the confidence of each classification by examining the predicted probability that an observation was assigned to a given wind resistance rating. The greater the predicted probability, the greater the confidence in the classification. For ease of interpretation by future users, we assigned each species a categorical confidence rating of Low Confidence (predicted probability ≤ 0.33), Moderate Confidence (0.33 < predicted probability ≤ 0.66), or High Confidence (predicted probability > 0.66).
Many species had multiple observations of damage data from different studies, and consequently each observation received a unique predicted rating. Only 18 species with multiple observations received more than one rating. For these cases, the species was assigned the rating with the highest predicted probability and was marked as having Low Confidence. In other cases where multiple observations for a species were all assigned the same rating, we assigned the species the confidence rating from the highest predicted probability.
We then combined the original and new species into a single table that serves as the foundation for the Estimating Tree Community Hurricane Resistance Tool (ETCHR v.01; https://github.com/AllysonS/TreesForHurricanes). We created ETCHR v.01 as an Excel Workbook (Microsoft, Redmond, WA, USA) which can use our database of wind resistance ratings and a community’s inventory data to estimate the proportion of wind resistant species in a tree population.
Results
Model Performance
We trained the RF model using data from 73 species extracted from 39 studies and then tested the model using data from 52 species and 32 studies. Note that some species had multiple observations of damage. The model accuracy and agreement with the original ratings when applied to the testing data set were fairly high; accuracy was 0.91 while the adjusted Cohen’s Kappa was 0.91 (Table 3). The model multiclass area-under-the-curve (AUC) was 0.99. Within the 4 wind resistance ratings (Low, Medium-Low, Medium-High, High), the model performed best for Medium-High and High species and performed more poorly for Low and Medium-Low.
Performance metrics for the testing dataset. Wind resistance rating accuracy across all data was calculated at 0.91 (0.84 to 0.96) with an adjusted Kappa of 0.91 (0.90 to 0.91). Note that an observation is the proportion of damage to a species’ population within a study; some species had multiple observations across different studies.
Wood density, maximum height, and leaf mass per unit area were the most important predictors in the RF model (Figure 3). When one of those variables was included at a node, they were better at splitting the data so that subgroups contained observations with the same classifications. Percent damaged, latitude, and longitude were also moderately important predictors.
Variable importance scores for the model predictors. A greater Mean Gini Decrease indicates a greater importance in the model.
Species in the training and testing data set with a High or Medium-High rating tend to have greater wood density compared to those with Low ratings (Figure 4). By contrast, High species tend to have shorter maximum heights. The rating groups had similar average leaf mass per unit area, though the maximum leaf mass per unit area in the Low group was much greater than the other groups. Unsurprisingly, the average extent of damage decreased with increasing wind resistance rating, though within all groups damage varied substantially. The wide variability of predictor variables within the ratings and lack of linear relationships highlight the value of using a classification-based approach and the difficulty of relying on a single characteristic to predict wind resistance.
Box and whisker plot showing the distribution of percent damaged, leaf mass per unit area (LMA), maximum species height, and wood density among wind resistance ratings. Data came from training and testing sets. The top of the bar indicates the upper quartile of data, the middle line is the median, the bar bottom is the lower quartile. The whiskers extend from the largest or smallest data value within 1.5 × the inter-quartile range, and the dots indicate outliers beyond the 1.5 × inter-quartile range cutoff. L = Low; ML = Medium-Low; MH = Medium-High; H = High.
Ratings for New Species
Species with new wind resistance ratings came from studies in the North Atlantic; Northwest and South Pacific; and North and South Indian tropical cyclone basins. They were studied in temperate conifer forests, tropical and subtropical moist broadleaf forests, tropical and subtropical dry broadleaf forests, and temperate broadleaf and mixed forests. Of these new species, 42% were assigned a Low rating, 30% Medium-Low, 14% Medium-High, and 14% High (Appendix Table S3, Table S4, Table S5, and Table S6). Both Medium-Low and High wind resistance ratings had the greater proportion of species with High Confidence in their predictions (30% and 29%, respectively, within each rating)(Figure 5).
The proportion of confidence levels within each wind resistance rating category.
Twenty-two species were assigned more than one wind resistance rating since those species had data from multiple studies and we allowed the model to assign different ratings to different studies. These multirating species accounted for 24% of species with Low Confidence predictions. The majority of the other Low Confidence species received that classification because one or more of their traits was imputed prior to prediction. Examples of species with multiple ratings include Ficus religiosa L. from China, India, and Sri Lanka (Dittus 1985; Wang and Xu 2000; Sundarapandian et al. 2014; Lin et al. 2017; Zhou and Dong 2018; Guo et al. 2020); Ginkgo biloba L. from Japan (Tabata et al. 2020; Nakamura 2021); and Schefflera morototoni (Aubl.) Maguire, Steyerm. & Frodin. from Puerto Rico (Zimmerman et al. 1994; Francis 2000).
Discussion
Model Performance
Our analysis of the original rated species and new species demonstrated that our RF model is a reasonable approach for predicting wind resistance ratings that align with original work by Duryea et al. (2007a, 2007b). The RF approach allowed us to accommodate many predictor variables which often had non-linear relationships with ratings groups (Figure 4). And importantly, the predictive model can be applied to other new species as trait and tropical cyclone damage data become available. It is possible that adding additional predictors could have further increased the performance of the model with the training data, however, trying to further improve model accuracy could have overfit the model and reduced its predictive capabilities (Kuhn and Johnson 2013).
The high importance value of wood density indicates that our model aligns well with the original Duryea et al. (2007a, 2007b) ratings. Wood density is a commonly reported trait and was one of the key tree characteristics that Duryea et al. (2007a, 2007b) analyzed and considered in their determination of the ratings system. Other wood anatomy traits such as the modulus of rupture and wood fiber width can also predict tropical cyclone tree damage and other windbased tree failures (Xu et al. 2014; Gardiner 2021; Nakamura 2021). However, for the purposes of prediction, wood density is a more widely reported trait and tends to be directly related to other wood characteristics (Chave et al. 2009). That noted, wood density and other mechanical properties and crown traits can vary within species that have broad ranges with varying exposure to windstorms (Plourde et al. 2015; Cannon et al. 2023).
The high importance values for leaf mass per unit area and max height emphasize that our predictive model is primarily driven by intrinsic characteristics and represents generalized predictions about species’ abilities to resist wind damage. The original rating system incorporated expert opinions, which provided substantial value to the rating system by capturing a broader range of experiences beyond the post-storm data collected by researchers (Duryea et al. 2007a, 2007b). However, that approach was challenging to replicate for a large number of new species.
Using Wind Resistance Ratings
There are several ways communities can utilize the wind resistance rating system to increase the resilience of their urban forests in the face of future tropical cyclones. Such activities can be considered mitigation, actions which preemptively eliminate or decrease the potential harm from a natural disaster (FEMA 2023). Many urban forestry and urban greening practices can facilitate recovery after natural disasters, and with careful planning, foster more resilient communities through recovery efforts (Campbell et al. 2019).
Many communities use urban forest management plans to set goals such as the extent of canopy cover or the diversity of tree species (Hauer and Peterson 2016). Communities could also set targets for the proportion of Medium-High and High wind resistance species in their urban forest. We created an interactive spreadsheet to facilitate such a process (https://github.com/AllysonS/TreesForHurricanes). These goals could be achieved by incorporating Medium-High and High species into new planting projects. Wind resistance ratings could be incorporated into forest climate change vulnerability assessments (e.g., Brandt et al. 2016). Many organizations use recommended species lists with details about site tolerances and species characteristics to encourage community members to plant the right tree in the right place (e.g., New York City Parks 2023; USF Water Institute 2024). Adding wind resistance ratings to such lists could help community members consider this characteristic when planting new trees.
Granted, the goal should not be achieving a tree community with 100% Medium-High and High species, as Low rated species are important in urban and rural forests. Indeed, maintaining functional diversity—a collection of species with a broad range of traits or characteristics—in urban forests minimizes vulnerability to changing climate and pest and disease outbreaks (Paquette et al. 2021). And in natural areas, fallen trees create gaps where younger trees establish, playing an important role in the life of the forest (Lugo 2008). Priority could be given to planting Medium-High and High species in locations with high occupancy or high value targets, such as infrastructure or busy streets (Ellison 2005).
Conducting risk assessments and proactively pruning trees also contribute to mitigating hurricane damage to trees (Gilman et al. 2008; Koeser et al. 2020; Nelson et al. 2022). While resources for urban forestry programs can be limited compared to their needs (Hauer and Peterson 2016), wind resistance ratings could be used to complement other high volume risk assessment methods such as windshield surveys to identify trees with a high likelihood of failure (Rooney et al. 2005).
Study Limitations
While our RF classification approach enables us to effectively assign new trees to the Duryea wind resistance rating system, those assignments are only as good as the original classification system. One shortcoming of the original Duryea et al. (2007a, 2007b) research was that the original data collection could have been more statistically rigorous by using a random sampling approach to collect. Though importantly, the system was not based purely on damage data, it integrated other species characteristics and professional opinion as well. Another drawback of the original rating system is the assignment of 4 wind resistance categories. Ideally, risk matrices clearly distinguish between very high and very low risk conditions, but increasing the number of risk categories can muddy such distinctions (Cox 2008). And indeed, the properties of species in the Medium-Low and Medium-High categories tend to be similar (Figure 4). We maintained the original 4 categories to maintain consistency with the original research, though depending on local conditions, practitioners may find utility in combining the Medium-Low and Medium-High categories. Additionally, the original system is not spatially explicit and does not account for environmental conditions such as proximity to infrastructure or other trees that can influence the wind resistance of an individual tree. Considering Duryea et al. (2007a, 2007b) is currently the only widely available wind resistance rating system, it remains a useful foundation for evaluating wind resistance, and professionals can modify both the original and new rating based on their own local experiences.
Research Needs
To collect as many examples as possible to train a robust model, many of the species in this study come from rural settings and are not in nursery production. Nevertheless, the advantage of our modeling approach is that when data becomes available for unrated urban species, the model can use that new data to rate those species. Other research needs on this topic include examining the interaction between species’ wind resistance ratings and pruning techniques, and further evaluating the efficacy of practices intended to mitigate hurricane damage to urban trees.
While our work expands and helps synthesize past research on the wind resistance of trees, there are still gaps in our understanding. The literature referenced remains dominated by research published in English and focused on study sites in the Atlantic Ocean and Caribbean Sea. The South Pacific is currently under-represented, and we were unable to find research that met our criteria from Madagascar (Figure 1).
Finally, Duryea et al. (2007a, 2007b) combined their posthurricane field observations with a survey of the professional experiences of urban tree managers. Similarly, it will be important to continue collecting reports of hurricane damage from the field to further verify model results. Strengthening the connection between model predictions and hurricane observation damage will help improve urban forest management efforts in hurricane-prone areas.
Conclusion
Duryea et al. (2007a, 2007b) developed a wind resistance rating system that arborists and urban foresters have used as a planning tool to improve species selection and identify species at greater risk of failure during hurricanes. In this paper, we demonstrated how an RFs predictive model, a machine learning technique, can extend the original Duryea et al. (2007a, 2007b) rating system to include new tree species not observed in their original study. Based on our test data, the model performed with reasonably high accuracy overall (multiclass AUC = 0.99) with the model producing the lowest accuracy for the Low wind resistance category (Table 3). Our model assigned many new species a rating with Moderate to High Confidence, though ultimately future observations of hurricane damage to these species will support or refute these ratings. Researchers have applied machine learning techniques to data analysis in a variety of fields; here we successfully applied the technique to a species classification problem in urban forestry. While our analysis focused on wind resistance, this approach could be applied to other subjects such as resistance to ice damage, provided a sufficiently sized body of data is available for model training and testing. Sharing the model code can allow it to be adjusted to incorporate new data and further improve the rating system. We intend for our model and its interactive spreadsheet, ETCHR (https://github.com/AllysonS/TreesForHurricanes), to be an additional tool in the toolbox of urban forest hurricane mitigation strategies. As more storms occur in regions previously understudied, our methods can be replicated to continue to expand our understanding of relative wind resistance ratings.
Acknowledgements
Funding for this research was provided by the Florida Forest Service (#20-DG-11083112-009) and the Federal Emergency Management Agency Higher Education Program (#WX01809N2022T).
Appendix.
The quantities of studies found by the literature search and which passed screening criteria, in addition to the number of unique tropical cyclones observed in the studies and the countries or territories where the study took place.
Research studies that documented hurricane damage to tree populations grouped by tropical cyclone basin. We used these studies as data sources for the predictive model.
Tree species with High wind resistance ratings. Bolded species are our additions to the original Duryea et al. (2007a, 2007b) rating lists. These new species come from hurricane studies that met our inclusion criteria (Appendix Table S1), and the ratings were assigned using our random forests model. Confidence indicates the probability a rating was correctly assigned according to the model. More detailed information for the species modeled in this study can be found at https://github.com/AllysonS/TreesForHurricanes. NA (not applicable).
Tree species with Medium-High wind resistance ratings. Bolded species are our additions to the original Duryea et al. (2007a, 2007b) rating lists. These new species come from hurricane studies that met our inclusion criteria (Appendix Table S1), and the ratings were assigned using our random forests model. Confidence indicates the probability a rating was correctly assigned according to the model. More detailed information for the species modeled in this study can be found at https://github.com/AllysonS/TreesForHurricanes. NA (not applicable).
Tree species with Medium-Low wind resistance ratings. Bolded species are our additions to the original Duryea et al. (2007a, 2007b) rating lists. These new species come from hurricane studies that met our inclusion criteria (Appendix Table S1), and the ratings were assigned using our random forests model. Confidence indicates the probability a rating was correctly assigned according to the model. More detailed information for the species modeled in this study can be found at https://github.com/AllysonS/TreesForHurricanes. NA (not applicable).
Tree species with Low wind resistance ratings. Bolded species are our additions to the original Duryea et al. (2007a, 2007b) rating lists. These new species come from hurricane studies that met our inclusion criteria (Appendix Table S1), and the ratings were assigned using our random forests model. Confidence indicates the probability a rating was correctly assigned according to the model. More detailed information for the species modeled in this study can be found at https://github.com/AllysonS/TreesForHurricanes. NA (not applicable).
- © 2025 International Society of Arboriculture
Literature Cited
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.