Environmental and anthropogenic variables influence the distribution of a habitat specialist (Sylvilagus aquaticus) in a large urban forest

Conservation of habitat fragments, including those embedded in an urban matrix which typically support fewer species than those in other landscape contexts, is important for combatting the global extinction crisis. Because urban avoiding species are often absent from habitat fragments within an urban matrix, studies of their distributions in these habitats are rare, and therefore the mechanisms by which species are excluded from habitat fragments within an urban matrix are poorly understood. We investigated the impacts of urbanization in the matrix (e.g., noise pollution, light pollution, invasive species) on the distribution of an urban avoiding species, the swamp rabbit (Sylvilagus aquaticus) in the largest urban forest fragment in the United States. We used the location of swamp rabbit latrines and presence‐background species distribution modeling to reveal the impacts of urbanization in the matrix on swamp rabbit distribution while accounting for spatial heterogeneity in naturally occurring environmental variables. Swamp rabbits used mature forest, heterogeneously structured forest, and moderately inundated forest, and they avoided areas of high (Sus scrofa) hog activity at our study site. Our findings provide novel information to conservation practitioners and urban planners attempting to conserve high priority species in habitat fragments as urbanization continues to spread.


| INTRODUCTION
Urbanization is a major threat to the conservation of biodiversity (McDonald et al., 2008), as it dramatically modifies landscapes and is spreading rapidly (Güneralp & Seto, 2013;Seto et al., 2012). Although urbanization does not impact all species similarly, a common consequence of this process is the loss of native species and a decline in biodiversity (Aronson et al., 2014;Clergeau et al., 2006). Biodiversity declines as urbanization increases because few species can cope with such dramatic environmental change (Kark et al., 2007;Sih et al., 2011).
Urbanization is a major source of habitat fragmentation (Marzluff & Ewing, 2001), and forest fragments surrounded by an urban matrix have less biodiversity than forest fragments in rural landscapes (Miller et al., 2003;Rodewald & Bakermans, 2006). As the surrounding matrix becomes less hospitable, habitat fragments suffer a loss of biodiversity and an increased risk of local extinction (Kennedy et al., 2010;Kennedy et al., 2011). Species that avoid the matrix around a habitat fragment tend to decline or disappear as the matrix urbanizes, while those that tolerate or exploit the matrix often remain stable or increase in abundance (Gascon et al., 1999). Thus, the species lost from forest fragments as the matrix urbanizes are often the same urban-avoiding species that are lost when habitat is directly altered by urbanization.
Two hypotheses explaining the decline of urban avoiding wildlife within suitable habitat fragments in response to urbanization of the matrix can be distinguished based upon either direct or indirect effects of such changes in the matrix. Hypotheses related to direct impacts look to the influence of urbanization on matrix composition and fragment size. Urbanization of the matrix decreases its suitability for foraging and dispersal for urban avoiders, leading to a decrease in the resource base and an increase in population subdivision (Kupfer et al., 2006). Indirect impact hypotheses explain the loss of biodiversity from urban fragments through anthropogenic stressors from the matrix like noise pollution (Francis & Barber, 2013), light pollution (Gaston et al., 2013), or invasive species (Crooks & Soule, 1999). These hypotheses may not be mutually exclusive, and the importance of direct versus indirect impacts of an urbanizing matrix on biodiversity loss in habitat fragments is unknown, as is the relative importance of different indirect anthropogenic impacts.
The Great Trinity Forest (GTF) is the largest urban forest in the United States, encompassing over 4000 ha of unmanaged riparian bottomland hardwood forest within the city of Dallas, TX. The GTF is surrounded by a complex urban matrix, where the intensity of the surrounding urbanization (i.e., percent impervious surface) varies greatly across the study area. Reports from local naturalists (Ben Sandifer Pers. Comm.) and citizen science tools indicate that the GTF, unlike other well-studied urban forests, has diverse and complete wildlife communities including urban-avoiding species. The presence of urban avoiders in the GTF presents a unique opportunity to test hypotheses explaining the loss of urban avoiding species from urban habitat fragments. Because the study location is one exceptionally large forest fragment, as opposed to multiple fragments, the indirect effects of the urban matrix on swamp rabbit distribution can be isolated from changes in distribution associated with variations in fragment size and connectivity. The swamp rabbit (Sylvilagus aquaticus), a habitat specialist restricted to bottomland hardwood forests in the southeastern United States (Robinson et al., 2016), occurs throughout the GTF. The swamp rabbit is considered an indicator species for bottomland hardwood forests (Hillard et al., 2017), meaning that its presence indicates healthy habitat that can harbor other bottomland hardwood forest specialists. Therefore, trends in swamp rabbit occurrence in an urban forest can indicate areas within the forest that are suitable for other bottomland hardwood forest specialists.
The purpose of this study was to understand how swamp rabbits persist in an urban forest fragment and to identify specific impacts of the urban matrix that may limit their distribution. Swamps rabbits are easy to survey because of their use of latrines, locations where they deposit highly visible piles of scat on top of fallen trees or stumps (Zollner et al., 1996). Latrines are used by both males and females and are thought to function as sites for the exchange of olfactory information (Zollner et al., 1996). We used a presence-background modeling framework and latrine locations to assess how urbanization in the matrix and naturally occurring heterogeneity in habitat characteristics influenced the distribution of swamp rabbits in the GTF. This analysis allowed us to test direct versus indirect hypotheses that may explain the loss of urban avoiding wildlife from urban forest fragments. If the indirect impacts of the anthropogenic matrix restrict the distribution of the swamp rabbit in the GTF, then the indirect impacts of the urban matrix may be a key driver of the broadly observed absence of urban avoiding species from urban habitat fragments. If swamp rabbits use all areas of the GTF without regard for the spatial distribution of indirect impacts from the urban matrix, then direct impacts of urbanization on the matrix like limitations to dispersal may be a more important driver of the broadly observed absence of urban avoiding species from urban habitat fragments. The habitat requirements of urban avoiding species in urban forest fragments are poorly understood, as they are often absent from these habitats. Enhanced knowledge of the habitat requirements of urban-avoiding species in urban forest fragments has the potential to improve conservation practice in existing urban forests (i.e., prioritizing mitigation efforts to expand habitat for urban avoiders), and conservation planning as urban expansion fragments previously intact habitats and urbanizes the matrix around already fragmented habitats. The ability to conserve urban avoiding species in urban forest fragments helps combat the worldwide extinction crisis (Foley et al., 2005), and provides social and educational benefits for people in urban areas with few opportunities to experience intact wildlife communities (Dearborn & Kark, 2010;Miller & Hobbs, 2002).

| Study area
The GTF is a large (>4000 ha) urban forest fragment located within the fourth largest metropolitan area in the United States, Dallas-Fort Worth (United States Census Bureau, Population Division, 2020; Figure 1a). The GTF is a bottomland hardwood forest centered on an unlevied portion of the Trinity River, giving it an intact flooding regime, and making it unsuitable for further development. Bottomland hardwood forests are diverse habitats because of the natural patterns of ecological succession driven by disturbance from regular flooding events (Hodges, 1997). Habitat diversity in the GTF is amplified by previous anthropogenic disturbance (e.g., illegal dumping sites and cattle grazing), creating habitats varying in age from early successional grassland/savanna to old growth forest (see supporting information for more information on site vegetation characteristics).
The GTF is surrounded by a complex urban matrix ( Figure 1b). The highest levels of urbanization are in the northwest portion of the forest near city center, where commercial and industrial developments are intermixed with dense residential neighborhoods. Urbanization becomes less intense moving to the southeast, where residential neighborhoods intermix with agricultural land. State, U.S., and interstate highways with high traffic volume intersect the forest in multiple locations. Parts of the forest are also within a major flight path for commercial aircraft.

| Latrine surveys
We located swamp rabbit latrines over a 4-year period from 2018 to 2021. We performed dedicated searches during the winters of 2018/2019, 2019/2020, and 2020/2021, when leaf-off conditions made latrines easier to locate, and swamp rabbits increase the frequency of latrine use (Zollner et al., 1996). We searched for latrines in an unstructured manner, checking all appropriately sized and aged woody debris. During each of these three winters, we searched the entire study area. Because an important purpose of these latrine searches was collecting swamp rabbit feces for a companion landscape genetics study, we did not use structured search methodologies (e.g., transects or plots) to look for latrines. We also spent more time searching in areas where latrine density was highest to maximize the probability of collecting fresh samples. During each winter, we spent additional search effort looking for latrines in large areas of the forest where no latrines were found. We also located latrines incidentally during other field work in the GTF during F I G U R E 1 (a) The location of the Great Trinity Forest within the city of Dallas, TX, USA (32.733043, À96.733511). (b) The location of our study area within the Great Trinity Forest and its spatial relationship to impervious surface in the surrounding matrix, which was identified using supervised, object-based classification in ArcGIS Pro (version 2.2) the spring, summer, and autumn seasons. Although characteristics of swamp rabbit latrines can shift seasonally (Zollner et al., 2000a), we only identified 17 latrine locations in non-winter searches which prevented us from making season-specific models. Additionally, these incidental detections were not spatial outliers as they clustered with latrines detected in winter. When we located a latrine, we mapped the location using ArcGIS Collector and a Bad Elf GNSS Surveyor GPS Receiver with 1-m accuracy.

| Statistical analysis
The unstructured nature of searches we used to located latrines and potential differences between swamp rabbit latrine locations and the full breadth of this species habitat use limited the modeling approaches available for assessing the impact of urbanization on swamp rabbit distribution. Species distribution models exhibit a hierarchy of robustness, with presence-absence models being the most robust, followed by presence-background approaches, and finally presence-only models (Guillera-Arroita et al., 2015). Presence-absence models require either perfect detection or additional information from structured surveys (e.g., repeated surveys of the same location) that inform the detection process and can be used to model the impact of imperfect detection on occurrence (Guillera-Arroita, 2017). Our study undoubtedly suffered from imperfect detection, and false absences may have been spatially biased, as previous studies have shown that latrines are more common in areas with high amounts of downed wood and that latrine locations may not represent the full breadth of swamp rabbit habitat (Zollner et al., 2000a). Our survey techniques were designed to maximize genetic collections and did not provide additional information needed to model imperfect detection, making a presence-background approach the most robust species modeling method applicable for this study.
We used MaxEnt (Phillips et al., 2021), an ecological niche modeling software program that uses presence and background data to model species distributions, to assess the influence of urbanization on swamp rabbit probability of presence (POP). MaxEnt has consistently outperformed other modeling methods that do not require presenceabsence data, and has shown nearly equivalent results when compared to newer ensemble approaches (Elith et al., 2006;Hern andez et al., 2006;Kaky et al., 2020;Valavi et al., 2021). Although MaxEnt is an effective tool for assessing species distributions when there is not reliable information for absence locations, it is still vulnerable to inaccuracy when sampling effort across the study area is spatially biased, especially when spatial sampling bias correlates with environmental covariates (Kramer-Schadt et al., 2013). The potential for bias occurs because information from model parameters at presence locations is compared to information from background or pseudo-absence locations. To account for this, we standardized cells by search effort, which significantly improves model accuracy (Phillips et al., 2009). In areas where multiple visits were made each winter to collect genetic samples, we assigned cells twice as much search effort as other locations (see supporting information for the spatial distribution of search effort). This meant that cells without latrines that were searched twice as often were twice as likely to be selected as background or pseudo-absence locations.
Within their home ranges, swamp rabbits use different microhabitat for latrine locations when compared to other activities (Zollner et al., 2000a). MaxEnt spatially segregates the modeled area into a grid with a userdefined size. Cell size selection in MaxEnt models can influence model performance (Connor et al., 2018); therefore, we selected a 1 ha cell size relevant to our study species which occupies home ranges from <1 to 5 ha in size (Zollner et al., 2000b). Quantifying space use at a broader scale (i.e., one that more closely approximated an entire home range) than what was utilized by Zollner et al. (2000a); 20 Â 20 m grid plots, means that potential spatial and temporal biases of space use within the home ranges in this study were unlikely to drive bias in our results. If a spatial bias between latrine locations and other space use still existed at the scale of our study, it is unlikely that swamp rabbit response to the impacts of urbanization in the matrix would change between latrine use and other use areas.
In our presence-background approach, we developed a suite of model parameters that were spatially segregated into 1 ha grid cells. These included our variables of interest: anthropogenic variables representing both the direct (urbanization intensity) and indirect (noise pollution, light pollution, invasive plants, and invasive animals) impacts of the urban matrix that we hypothesized could influence swamp rabbit distribution, and covariates represented by naturally occurring environmental variables known to influence swamp rabbit distribution (vegetation structure, flooding frequency, access to refugia during flooding events, vegetation composition, and forest width). We assessed correlations among model parameters and did not include correlated variables (r > .6) in the final MaxEnt model (Merow et al., 2013;Smith & Santos, 2020).
We prescreened regularization coefficient values and feature combinations used to fit response curves of POP to model parameters (Merow et al., 2013). Our goal was to create a simple model in which we could easily interpret the impact of model parameters on POP while maximizing gain. We used a regularization coefficient of 1 and only the hinge feature to fit the model. We completed 10k-fold cross validations of the MaxEnt model for model testing, in which we withheld a random 10% of presence locations in each iteration. The mapping of POP across all cells and the response of POP to model parameters was evaluated using the average of these 10 models. We used raw outputs of POP as opposed to logistic ones, as raw outputs do not rely on post processing assumptions (Merow et al., 2013). We used average area under the ROC curve (AUC) for model validation, a threshold independent statistic that represents the probability that a random presence site will be more suitable than a random background point.

| Anthropogenic variables
Anthropogenic noise can displace small mammals from otherwise suitable habitat (Chen & Koprowski, 2015). To assess the impact of anthropogenic noise on swamp rabbits, we used data from the 2018 National Transportation Noise Map (United States Department of Transportation, Bureau of Transportation Statistics, 2018). This dataset allowed us to split noise from transportation into two different types, continuous noise from highways and intermittent noise from airplane and high-speed rail traffic, as continuous and intermittent noise may impact wildlife in different ways (Francis & Barber, 2013). This data set measures a 24-h weighted potential noise exposure across yearly average environmental conditions, making it appropriate for estimating chronic noise exposure for wildlife in the GTF. We quantified and mapped noise exposure as either "at risk for noise exposure" or "not at risk," with areas that met or exceeded average daily noise levels of 45 dBA considered at risk for noise exposure for both continuous and intermittent noise.
Investigations of artificial light at night (ALAN) on wildlife have shown impacts on gene expression, physiology, foraging, daily movements, migratory behavior, reproductive behavior, and mortality (Gaston et al., 2015). More specifically, small mammals can be unwilling to use brightly lit areas (Bliss-Ketchum et al., 2016). We mapped ALAN in the GTF using 2020 data from Visible Infrared Imaging Radiometer Suite Day/Night Band in ArcGIS Pro.
The amount of impervious surface in a landscape is often used as a surrogate for the intensity of urbanization (e.g., Evans et al., 2015;Rodewald et al., 2013). As urbanization in the matrix surrounding habitat fragments intensifies, these fragments often suffer a loss of biodiversity and an increase in the risk of local extinction (Kennedy et al., 2010;Kennedy et al., 2011). We measured the amount of impervious surface in and around the GTF using ArcGIS Pro's image classification wizard. We defined impervious surfaces as any human-made material that is impenetrable to water (e.g., roofs, concrete, building materials, etc.). We quantified urbanization intensity as the amount of impervious surface within 1000 m buffers from any location in the forest. We measured urbanization at this scale because of its relevance to swamp rabbit space use; the most recent home range estimates for swamp rabbits are between 4.7 and 5.9 ha (Dumyahn & Zollner, 2010;Zollner et al., 2000b). A 1000 m buffer encompasses $8 ha around the latrine ensuring that most of the space an individual regularly uses falls within this buffer, even if the latrine location was near the periphery of the individual's home range.
As the matrix around a habitat fragment urbanizes, invasive species become more abundant in those fragments (Johnson et al., 2020). Chinese privet is a rapidly spreading invasive shrub in the southeastern United States and Texas (Arevilca et al., 2016), and has become dominant in the understory of some areas of the GTF. Feral hogs (Sus scrofa) are a destructive invasive species that have spread across the southern United States (Campbell & Long, 2009). Rooting behavior by feral hogs reduces plant diversity and richness as well as total vegetation cover in wetlands (Arrington et al., 1999) and forested habitat (Siemann et al., 2009). Feral hogs are abundant in the GTF, and signs of their damaging rooting behavior are widespread. We assessed the relative abundance of both Chinese privet and feral hogs in the GTF during avian surveys in the summers of 2018, 2019, and 2020. Assessments of invasive species took place in $140 randomly placed 5 ha plots in each summer. In each plot, we classified Chinese privet relative abundance as absent, low, moderate, or high. We classified feral hog activity as absent, low, moderate, or high based on observations of tracks, feces, and disturbance from rooting. With these data, we used empirical Bayesian kriging in ArcGIS Pro (default settings) using information from all three summers (404 plots) to create a continuous map of invasive species relative abundance in the GTF.

| Environmental variables
Swamp rabbits prefer habitat where canopy gaps create increased structure in the understory (Dumyahn et al., 2015). We used LIDAR point cloud data imported into ArcGIS Pro (version 2.2.0) to develop three variables that described vegetation structure: foliage height diversity, canopy cover, and maximum vegetation height (see supporting information for more information on LIDAR data and metrics). These three variables were the most often predictive of wildlife distributions among the many LIDAR metrics (Bakx et al., 2019).
Swamp rabbits prefer permanently wet areas of the forest, but also need access to highland refugia to survive large flooding events (Scharine et al., 2011;Zollner et al., 2000b). Besnard et al. (2013) showed that topographic wetness index (TWI) successfully predicted the occurrence of bird species in a European wetland. TWI was developed to approximate flooding susceptibility and water accumulation capacity. We used a digital elevation model of the GTF and a combination of four TWI indices to model wetness in our study area using the SAGA GIS "Terrain Analysis Hydrology" tools (Böhner & Selige, 2006). Many areas of the GTF do not have access to highland refugia as almost all areas above the flood line are developed. We mapped access to highland refugia across the forest, where access was defined as the ability to disperse to a higher elevation area that was not developed without crossing urban areas or waterways that could be a source of flooding.
Swamp rabbits also select for habitat based on vegetation age, although different studies have found selection for both older and younger habitat (Crawford et al., 2018;Vale & Kissell Jr, 2010). We created a habitat map for the GTF, in which habitat types were based on canopy tree composition which can be linked to both successional state and flooding frequency in bottomland hardwood forests (Hodges, 1997). To identify canopy trees, we made extensive drone flights in the fall of 2021 to obtain aerial images of the forest. We manually piloted drone flights and obtained images at the elevation required to effectively identify trees from images of their crowns to create a continuous habitat map across the entire study area. We extensively verified tree identification on the ground as part of an initial training period to ensure that we could correctly identify trees from aerial images, and to determine the appropriate flight altitude for tree identification. We developed the following six habitat types based on canopy tree composition: cedar elm forest, climax mixed forest, early successional grassland, green ash forest, levee forest, and young mixed forest (see supporting information for drone specifications and habitat type descriptions).
Swamp rabbit distribution is impacted by habitat size, where rabbits preferred larger forest patches (Scharine et al., 2009;Scheibe & Henson, 2003). The dominant paradigm for riparian forest conservation is that forest width is the best available indicator of ecological function (Maure et al., 2018;Shirley & Smith, 2005). Therefore, we measured forest width in ArcGIS Pro as the perpendicular distance across the forest at any point.

| RESULTS
We located 650 latrines during field work from 2018 to 2022 (Figure 2a). MaxEnt spatially segregated the study area into 3060 1 ha cells, with 248 of those cells considered presence locations (i.e., they contained ≥1latrine).
Prescreening for correlation among model parameters revealed significant correlation between all three LIDAR derived variables describing vegetation structure. Therefore, only maximum vegetation height was included in the final model. ALAN was also closely correlated with the amount of impervious surface within 1000 m of each F I G U R E 2 (a) The location of all 650 identified latrines recorded over 4 years (2018)(2019)(2020)(2021) in the Great Trinity Forest. (b) Probability of swamp rabbit presence in our study as determined by MaxEnt. Probability of presence (POP) was determined for 3060 cells that were 1 ha in size each, showing that suitable habitat was heterogeneously distributed within the site cell, causing us to only include the amount impervious surface in the model as a surrogate for both anthropogenic impacts. MaxEnt used the remaining model parameters to map POP across the study site (Figure 2b; see supporting information for the spatial distribution of all model parameters included in the final model). Average AUC for the 10 model replicates was 0.829, showing the model was successfully able to distinguish presence locations from background locations. Vegetation composition, vegetation structure, TWI, and feral hog activity explained the most variation in the model based on their percent contribution and permutation importance ( Figure 3; Smith & Santos, 2020). POP values were higher in areas where the vegetation composition was cedar elm forest or climax mixed forest (Figure 4a), vegetation structure (maximum vegetation height) was at moderate levels ( Figure 4b), and TWI was at moderate levels ( Figure 4c). In contrast, POP values were lower in areas where evidence of feral hog activity was absent or high ( Figure 4d).

| DISCUSSION
Four variables best explained the distribution of swamp rabbits in the GTF, only one of which, feral hog activity, was associated with the urban matrix. Feral hogs are not strictly associated with urbanization, as they are also spreading rapidly in rural and wild areas (Adams et al., 2005). Noise, invasive shrubs, light, and urbanization intensity were not important explanatory variables for the distribution of this urban avoider. The willingness of swamp rabbits to use habitat near loud highways and intense urbanization lends support to hypotheses that urban avoiding species are excluded from urban habitat fragments by the direct impacts of urbanization in the matrix that may limit dispersal and foraging opportunities (Kupfer et al., 2006). Swamp rabbits and other urban avoiders may persist in the GTF because it is a large urban forest (>4000 ha) and may have some connectivity to other tracts of habitat. There is a narrow ($300-700 m) continuous riparian forest corridor along the Trinity River that flows south from the GTF to the next large riparian fragment $25 km to the southeast. Urban forest fragments may not appear to be islands of habitat to species that tolerate urbanization, but for species that cannot adapt to the drastic changes associated with urbanization, urban habitat fragments become isolated islands with reduced genetic connectivity and population viability. In some cases, even urban adapting small mammals become isolated in habitat islands surrounded by dense urbanization, where dispersal abilities are reduced in ways that are consistent with the principles of island biogeography (Richardson et al., 2020).
Feral hog activity had the highest combined contribution to the model of any model parameter. The decrease in swamp rabbit POP in areas of high hog activity was not surprising, but the decrease in swamp rabbit POP in areas where feral hogs were absent is less intuitive. Extensive damage from rooting by feral hogs is common in the GTF, and research has shown that this behavior reduces plant cover and diversity in forest ecosystems (Barrios-Garcia & Ballari, 2012). In addition, feral hogs are significant sources of predation for small mammals (Wilcox & van Vuren, 2009). Thus, not only do feral hogs alter swamp rabbit habitat, they also may create a F I G U R E 3 The percent contribution (range: 0-100) and permutation importance (range: 0-100) of all model parameters included in the final MaxEnt model. The four variables that contributed most to the variation in swamp rabbit probability of presence were vegetation structure, vegetation composition, topographic wetness index, and evidence of feral hog activity landscape of fear in areas where they are abundant (Laundre et al., 2010). The reduction in swamp rabbit POP associated with feral hog absence may be explained by both species responding to other unmeasured variables. For example, feral hogs selected for areas with high primary productivity in another urban forest (Stillfried et al., 2017). In the GTF, areas with low feral hog density, and sometimes a complete absence of hog sign, were concentrated in locations where young monotypic stands of green ash dominated the canopy. Feral hogs may have avoided these areas because of possible lower levels of productivity and a lack of masting trees, which are an import source of food for this species (Taylor & Hellgren, 1997). Swamp rabbits, on the other hand, may have avoided these same areas because of a lack of complex understory cover or appropriate forage (Vale & Kissell Jr, 2010).
Swamp rabbits responded positively to several environmental variables including mature forest habitat, habitats at moderate levels of structure, and habitat with moderate levels of flooding frequency. Swamp rabbits preferred cedar elm and climax mixed forest habitat and avoided early successional habitat. They also preferred areas with moderate levels of structure, where heterogeneity in the canopy created more structure and cover in the understory. Swamp rabbit preference for areas of moderate topographic wetness in the GTF is somewhat surprising but may be explained by the urban matrix. The literature on swamp rabbits consistently shows a preference for habitat that is highly inundated with access to nearby higher and drier areas to survive flooding events (Crawford et al., 2018;Scharine et al., 2011;Vale & Kissell Jr, 2010;Zollner et al., 2000b). Lower swamp rabbit POP in areas where TWI was low is F I G U R E 4 The response of swamp rabbit probability of presence (POP) to the four variables that explained the most variation in the model. (a) Swamp rabbit POP was highest where the vegetation type was cedar elm or climax mixed forest and lowest where the vegetation type was early successional grassland. (b) Swamp rabbit POP was highest where mean maximum vegetation height was between 10 and 30 m and lowest where mean maximum vegetation height was below 5 m. (c) Swamp rabbit POP was highest where levels of topographic wetness were moderate and lowest where levels of topographic wetness were the driest or wettest. (d) Swamp rabbit POP was highest where evidence of feral hog activity was low to moderate, and lowest where evidence of feral hog activity was absent or high consistent with the literature, but their avoidance of areas with high topographic wetness is unexpected when examined in isolation. Despite the importance of access to highland refugia for survival during flooding events, only $8% of the GTF has access to highland refugia. The border between the GTF and the urban matrix is in large part defined by elevation, where areas above the flood line were developed. It is possible that because most swamp rabbits in the GTF do not have access to undeveloped highland refugia, that rabbits in areas that are frequently flooded have lower survival. Preference for wetter habitat and a reduction in survival in the most frequently flooded areas may create this preference for moderate levels of topographic wetness.
Our study indicates that indirect impacts of urbanization from the matrix do not limit swamp rabbit distribution in a large urban forest fragment. Because there are so few studies examining the distribution of urban avoiding species, our findings could serve as a possible hypothesis explaining drivers of urban avoidance in other species. Direct evidence of urbanization limiting dispersal in swamp rabbits or other urban avoiding species would increase evidence for the application of island biogeography concepts to urban habitat islands, especially for species with limited dispersal ability. The control of feral hogs may be critical to the conservation of sensitive species in urban habitat fragments like the swamp rabbit, but also for a broad range of small mammals in urban, rural, and wild habitats (McClure et al., 2018). Our study was the first to characterize the distribution of an urban avoiding species in an urban habitat fragment, providing new and important information that may facilitate conservation in and near urban areas as urbanization spreads, a critical challenge in combatting the worldwide extinction crisis (Foley et al., 2005). Zollner, P. A., Smith, W. P., & Brennan, L. A. (2000b). Home range use by swamp rabbits (Sylvilagus aquaticus) in a frequently inundated bottomland forest. American Midland Naturalist, 143, 64-69.

SUPPORTING INFORMATION
Additional supporting information can be found online in the Supporting Information section at the end of this article.