This study investigated canopy coverage for the 300 largest cities (by population) in Florida (United States). Florida is a peninsular state that is surrounded by the Gulf of Mexico to the west, the Caribbean Sea to the south, and the Atlantic Ocean to the east. The state is predominantly classified as having a humid subtropical climate (Cfa), with the southernmost tip of the state and its keys (islands) being classified as being tropical monsoon, tropical rainforest, or tropical savanna (Am, Af, and Aw, respectively) climates (Beck et al., 2018). Florida has an annual hurricane season that begins June 1 and extends until November 30 (Florida Climate Center, 2021).

In conducting our canopy analysis, we used aerial imagery from the National Agricultural Imagery Program (NAIP; U.S. Department of Agriculture, 2019)—selecting leaf on imagery from 2019 to capture canopy conditions just prior to the passage of a Florida Statute which preempts local oversight of trees found on residential property (The 2021 Florida Statutes, 2021). The spatial resolution of the imagery was 1 m.

A random point sampling method, also known as the “dot method” (Nowak et al., 1996), was conducted to determine canopy coverage for each municipality. Boundary shapefiles for each municipality in the study were obtained from the American Community Survey (ACS; United States Census Bureau, 2015). A geographic information system (ArcGIS v. 10.2.2; ESRI, Redlands, CA, United States) was used to import NAIP aerial imagery and generate random points to evaluate UTC within each city. Each point was interpreted as “no tree,” “tree/shrub,” or “open water” by at least two interpreters per city. A minimum of 2,000, non-“open water” (e.g., lakes, rivers, and retention areas) sampling points were used to determine UTC for each city. To accomplish this objective, 2,500 random points were placed within the city boundaries and were interpreted one by one until the total number of points labeled “no tree” and “tree/shrub” totaled 2,000. Canopy percentage and agreement between interpreters were noted for each municipality.

We derived historic land cover classifications from the natural vegetation map created by Davis (1967). Specifically, we accessed a digitized version of the map through the Florida Geographic Data Library (Table 1) and aggregated land cover types using the categories specified by Volk et al. (2017). While generally limited to description of historic vegetation cover, Davis’s map did delineate the footprint of 11 cities (i.e., Daytona Beach, Jacksonville, Lakeland, Orlando, Miami, Pensacola, St. Augustine, St. Petersburg, Tampa, Vero Beach, and West Palm Beach). For these municipalities (categorized as “urban” in the 1967 map), we used the surrounding land cover for our model predictor.

We accessed hurricane wind field data from the “disasters” repository housed at GeoPlatform.gov. Specifically, we looked at recent hurricanes that made landfall in Florida and were severe enough to become a Federal Emergency Management Agency declared disaster and had wind field data available. Storms included in our analysis were:

To determine the extent to which cities were impacted by past hurricanes, we overlaid our city boundary layers with the wind field layers noted above. For cities with only one weather station, the maximum recorded wind gust speed for that station was used. For larger cities with more than one station within their boundaries, the average of the recorded peak gusts was calculated prior to analysis. For smaller cities where a weather station was not present, the maximum gust values from the nearest station were used. If a city was hit by more than one hurricane event, we used the peak gust values from the more intense storm in our predictive model (Table 1).

Additionally, we sourced population, median income, population density, and percent of houses built 10, 20, and 30 years ago from the US Census Bureau (Table 1). Housing density was calculated manually based on the area of our city boundaries and the 2018 population projections (U.S. Census Bureau, 2018a).

Finally, as noted in detail in Koeser et al. (2021), we used a multi-mode email survey to communities and a systematic search of online resources to determine which municipal tree protection ordinances were in place within the 300 assessed communities prior to our canopy analysis. The online resources referenced included two municipal code databases^1,2 as well as individual municipal websites. The ordinances of interest with regard to their potential impact on canopy coverage are listed in Table 1. We coded each as being present or absent (i.e., yes or no binary) prior to our data analysis.

We analyzed our data using linear regression with percent canopy coverage as the response variable (Table 1). All analyses were conducted in R (R Core Team, 2020). To guide model building, we ran the regsubsets() function from the leaps package (Lumley and Miller, 2017) and plotted (by R² value) the 20 best subsets of our full set of predictor variables (Table 1; see Hilbert et al., 2019). We then ran a maximal model with all predictor variables using the lm() function (R Core Team, 2020). From this, we employed a one-at-a-time simplification strategy—removing non-significant predictors from our initial model based on P-value and the regsubsets() plot. With each iteration of simplification, we compared changes in overall fit between the original and simplified models using the anova() function in R (Crawley, 2013). Once all of the non-significant terms had been removed in this manner, we retested a few variables that had commonly been associated with high R² models in our subset plot, as well as potentially meaningful two-way interactions, to see if improvements to model fit could be detected.

In model building we adopted a P-value of 0.05 as our threshold for statistical significance. The underlying assumptions of our linear model were assessed visually using Q-Q plots (i.e., normality of residuals) and residual scatter plots (i.e., homogeneity) to assure that they were not being violated. Additionally, residual plots were used to assess the presence or absence of high-leverage outliers (based on Cook’s distance). Finding none, we adopted the resulting simplified model for our results and discussion.

The cities included in this study sample ranged from Florida’s largest municipality, Jacksonville (population 903,889) to Sneads, a small community of 1798 (Supplementary Table 1). The median population for our sampled cities was 12,678. The cities in our study (combined population of 10,629,924) were home to approximately 50% of the state’s 2018 population. Cities ranged in area from 2,265.3 km² (Jacksonville) to 0.8 km² (Virginia Gardens).

Population densities ranged from 8 people per kilometer for Bunnell (a small community with a wildlife refuge within its boundary) to 8,618 people per square kilometer for North Bay Village (a chain of highly developed and largely man-made islands in the Miami-Dade metropolitan area; Figures 2, 3). The median population density for our sampled cities was 814 people per square kilometer (Figures 2, 3). Median household incomes ranged from $17,908 in the city of Opa-locka to $154,415 for the village of Pinecrest (Figure 3). Both municipalities are located in the Miami-Dade metropolitan area within 39 km of each other.

The majority (72%) of our cities were built on land that had formerly been wooded (Figure 2). Of these cities, 31% were developed in areas of scrub and sandhill. Another 29% of communities were built in areas that had historically sustained mesic pinelands. A smaller proportion of communities were developed in former upland hardwood forests (5%) or forested wetlands (7%; Figure 2). Of the cities established in non-forested regions, 21% were former uplands and 7% were former wetlands (Figure 2).

Six of the eight tree ordinances evaluated as part of this study were adopted by more than half of the study cities (Table 2). The most commonly adopted tree ordinances in the surveyed cities included Planting requirements—new development (89.3%) and Planting requirements—parking lots (89.3%) followed by Tree preservation ordinance (86.7%) and Ability to fine (84.0%; Table 2). Local licensure was the least commonly adopted ordinance (17.7%; Table 2).

Across the 300 cities sampled in Florida, UTC ranged from 5.9% in Redington Shores to 68.7% in Sanibel, respectively. The median UTC for our sample was 32.3% (Figures 3, 4). Summary statistics for developed UTC (i.e., UTC for the inhabited census blocks in a municipality), which is what we used in our modeling efforts as it excludes forest preserves and wildlife areas that might inflate city-wide canopy levels, were nearly identical to those for straight UTC (Figures 3, 4).

Of the 16 variables initially modeled, we retained five for our final, simplified model (adjusted R² = 0.230). They included historic land cover (aggregated more simply as forested vs. non-forested after initial exploratory modeling showed this was the main delineation), maximum hurricane gust speed, population density, the presence of a tree preservation ordinance, and the presence of a heritage tree ordinance (Table 3). Of these, increases in population density and increases in hurricane gust wind speeds were associated with decreases in developed UTC (Figure 2 and Table 3). Similarly, non-forested historic land cover and the presence of a tree preservation ordinance were negatively associated with developed UTC (Table 3). Only the presence of a heritage tree ordinance (70% of cities) was positively associated with developed UTC. Median income, housing age, home ownership, and the other six ordinances were not included in the final model given non-significance.

Our analysis of 300 cities in Florida demonstrated that both natural and anthropogenic legacies can influence UTC at the city scale. Overall, this finding is in line with observations within other U.S. cities such as Toledo, OH (Berland et al., 2015), Philadelphia, PA (Roman et al., 2021a), Chicago, IL (Fahey et al., 2012), and Baltimore, MD (Grove et al., 2018), among others (Roman et al., 2018). This analysis of Florida’s largest cities is an important addition to this body of research because our study included an examination of hurricanes. Considering that climate change is predicted to increase the intensity and intensification of hurricanes and other tropical cyclones (Bhatia et al., 2018), it is imperative to understand how these extreme weather events can influence UTC in regions vulnerable to this type of extreme weather.

Our analysis demonstrated the impact of hurricanes across Florida on the urban forest, though hurricane impacts can be lessened or exacerbated by other factors. Other research in Florida has observed that wind resistance varies among tree species and is also affected by pruning practices (Duryea et al., 2007a,b). Additionally, in Florida larger trees were more likely to fail during a hurricane while neighboring trees and structures appear to offer little protection from high winds (Landry et al., 2021). Escobedo et al. (2009) observed interacting effects between wind speed, tree cover, and urban land cover on debris generation in Florida following the 2004 and 2005 hurricane seasons. Hurricane related tree losses also varied by land use type following Hurricane Ike in Texas, United States (Staudhammer et al., 2011). However, Thompson et al. (2011) found that urban forest structure was a better predictor of hurricane woody debris generation rather than hurricane related variables such as wind speed. Further research is needed to understand the extent to which hurricane associated tree loss is a direct result of hurricane damage vs. tree removals driven by property owner responses to extreme weather. Clearly, hurricane severity can have a significant impact on the urban forest, though understanding the influence of other factors such as pruning and species selection can help to moderate these impacts (Duryea et al., 2007a,b).

The influence of historic vegetation cover on current UTC reflects both the influence of biogeographical constraints on tree growth and the contributions of remnant forest patches. Florida has fairly high average rainfall across the state ranging from 102 to 178 cm annually (Florida Climate Center, 2021) and a distinct dry season (Misra and Mishra, 2016). This dry season and typically sandy soils with low water holding capacity (Kern, 1995) may constrain tree establishment. Indeed, irrigation can improve tree establishment and condition for tree plantings in the state (Koeser et al., 2014; Blair et al., 2019). In environments with low average annual rainfall, UTC can increase in historical non-forest habitats during urbanization as management activities overcome water limitations (McBride and Jacobs, 1986; Nowak, 2012). Given the presence of treed cities in areas that historically were non-forested, it appears management has played a role in shaping the current urban forest of Florida. That said, cities established in former forested areas did have higher UTC percentages (Fig 2). Similar to our findings, the positive association between historical forest habitats and higher UTC suggests remnant forest habitats likely play a key role in contributing to cover as observed in Chicago, Illinois, United States (Fahey et al., 2012). This finding suggests that protecting remnant forests from encroaching development or densification would be an important strategy for maintaining urban tree cover.

The negative relationship between population density and UTC (Figure 2) in Florida at the coarse spatial resolution used in this study was fairly unsurprising and consistent with other research (Bigsby et al., 2014; Locke et al., 2016; Fahey and Casali, 2017). Presumably as a city’s human population grows the extent of infrastructure such as buildings and roads increase, consequently reducing tree canopy. Similarly, increases in impervious cover (Davies et al., 2008) and housing density (Iverson and Cook, 2000; Hilbert et al., 2019) are also associated with decreased urban tree cover. When examining anthropogenic drivers of urban tree cover within cities, socio-demographic factors such as race, income, public policies, and local history may be more useful predictors at finer resolutions (Gerrish and Watkins, 2018; Roman et al., 2018; Watkins and Gerrish, 2018; Locke et al., 2021). However, Nesbitt et al. (2019) observed education level and income were important predictors of tree cover across cities. For regions with expected population increases such as Florida (Carr and Zwick, 2016), special attention should be paid to protecting tree cover as urban populations and development increase.

While municipal ordinances are one method for protecting the urban tree canopy, the contrasting relationships between different types of ordinances and UTC in this study likely suggests that simply having an ordinance is not sufficient for maintaining or increasing canopy cover. Hilbert et al. (2019) observed that across 43 Florida cities, heritage tree ordinances were also associated with greater urban tree cover. This type of ordinance may be particularly effective because they are designed to protect larger trees which contribute to more cover. Or alternatively, communities which already have substantial urban tree cover may adopt such an ordinance to protect what they already have. Landry and Pu (2010) observed properties in Tampa, FL built after the adoption of a tree protection ordinance had higher UTC compared to properties in nearby communities without such an ordinance. By contrast, our analysis of 300 cities in Florida observed a negative association between the adoption of a tree protection ordinance and urban tree cover. This is not to say that tree protections caused a decrease in canopy coverage. Rather, a protective ordinance could have been enacted as a response to a noticeably diminishing UTC. Moreover, it is unclear how much time must elapse between the adoption of any tree related ordinance and the observation of meaningful improvements in UTC. If a community already has a comparatively lower amount of tree cover to begin with, a tree protection ordinance at best would maintain that level of cover rather than producing dramatic increases. Additionally, an ordinance can only be effective if it is enforced. We lacked the data to differentiate between cities that actively enforced their protection ordinances and those lacking the resources or political will to do so.

Support for and knowledge of municipal tree ordinances can be highly variable. In a survey of Canadian communities, Conway and Bang (2014) observed that while residents had generally neutral to favorable attitudes about the urban forest, they expressed less support for particular policies. Knowledge of and support for municipal tree bylaws in the Greater Toronto Area tended to be higher among residents with higher levels of formal education and residents who were born in Canada (Conway and Lue, 2018). Zhang et al. (2007) also observed differences in attitudes toward financing urban tree programs along socio-demographic lines, where individuals with young families and individuals younger than 56 were more likely to feel that financing urban forestry initiatives is the government’s responsibility. In their survey of Florida municipalities after the passage of a bill that limits municipal implementation of tree protection ordinances, Koeser et al. (2021) observed that some municipalities have positive relationships with the public and tree care companies and enjoy support for their programs while other municipalities have more counter-productive relationships with the local tree care industry. The adoption and enforcement of municipal tree ordinances not only directly impacts urban tree cover, but they can also be another mechanism by which a city’s socio-demographic characteristics and culture can influence the urban forest.

Photointerpretation as a method to evaluate urban tree canopy has its limitations (O’Neil-Dunne et al., 2014; Locke et al., 2017), though it is a proven and affordable approach to land classification used in the urban forestry field (Nowak et al., 1996; Walton et al., 2008; Morgan et al., 2010). To assess accuracy, a 10% subset of points for each city was assed by two photo interpreters. Of the 300 cities in the study, 82% had interpreter agreement greater than 90% with agreement for the remaining cities ranging between 80 and 90% (Supplementary Table 1). UTC is only one metric of the condition of the urban forest, consequently this study’s findings cannot offer insight into drivers of other characteristics such as diversity and age structure (Leff, 2016). However, our findings offer interesting avenues to investigate drivers which may affect these other characteristics of the urban forest that are more time and labor intensive to measure.

Moreover, the overall predictive power of our model was low, accounting for only about a quarter of the variability seen in canopy coverage among the communities studied (adjusted R² = 0.230). While our work does attempt to look at how historic legacies related to landcover and storms can shape the current state of urban forests, we lacked long-term data to assess how management differences over time affected canopy coverage. Moreover, while the state’s dominant urban tree species have largely been spared from the noxious pest and disease outbreaks, several palm species and understory native trees common to wooded remnant forests have not been so lucky. This was not captured in our model.