Estimates of wildlife species richness, occupancy, and habitat preference in a human-impacted landscape in New York State


 Despite the widespread adoption of motion-triggered cameras, studies using camera-traps to characterize wildlife communities in human-impacted, residential areas in North America are limited. To fill this data gap, we placed camera traps over three seasons in 22 residential neighborhoods within Dutchess County, NY. To account for imperfect detection, we applied individual-level and community-level Bayesian site-occupancy models to these data. Overall, we captured 280,686 images over 17,820 camera-trap days. We detected 17–22 mammal and non-passerine bird species in each of the seasons of data collection, with our full-community models estimating an actual diversity of 24–33 species in each season. Small, cryptic species were not accurately detected, limiting our ability to model their occupancy. Our models did not indicate any geographic trends. We identified five species, raccoons (Procyon lotor), eastern gray squirrels (Sciurus carolinensis), red foxes (Vulpes vulpes), Virginia opossums (Didelphis virginiana), and white-tailed deer (Odocoileus virginianus) found in all neighborhoods. The most common variable included in our final occupancy models was the percent of area within each neighborhood that was not habitat, which positively affected occupancy for some species, and negatively affected occupancy for others. The amount of forest, the second most common variable in our final models, negatively affected occupancy for all species. Our estimates characterize a baseline for quantifying species richness and composition in residential areas of Dutchess County, NY and surrounding regions, and offer a comparison to similar studies in natural areas. Overall, the results improve understanding of how human use of landscapes affects individual species and communities.


Introduction
Wildlife provide important ecosystem services to people. For example, wildlife disperse seeds, control pests, and enrich human recreation. Wildlife can also be sources of con ict with humans, for example when they damage vegetation, collide with motor vehicles, or transmit pathogens. As humans continue to encroach on wildlife habitat worldwide (Venter et al. 2016; Williams et al. 2020), understanding the diversity and species composition of wildlife communities, particularly in human-impacted ecosystems, can inform wildlife and landscape management. But many wildlife species are di cult to detect, particularly if they are cryptic, nocturnal, or highly mobile.
One approach to detecting elusive species is to use cameras, an approach that has rapidly grown in popularity. The annual number of published studies of wildlife species or communities using camera traps has increased 5.2-fold in the past decade and 81-fold since 1994 (Delisle et al. 2021). In part, this widespread adoption is likely attributable to the utility and ease of camera-trapping technology. Infrared, motion-triggered camera traps are cost-effective and accurate tools that allow researchers to study species richness, occupancy, distribution, and relative abundance (Wearn and Glover-Kapfer 2019). Since camera traps remain in the eld for extensive periods of time, they allow researchers to capture images of rare species not easily detected by other techniques. Camera-traps have been found to be signi cantly more effective than live traps at detecting certain wildlife (Wearn and Glover-Kapfer 2019) and, unlike live traps, offer a passive method of detection that does not disturb study animals. Importantly, since camera traps yield repeated observations from a single location, camera data are well suited for tting site-occupancy models. Site-occupancy models (Mackenzie et al. 2002) account for imperfect detection (i.e., not detecting a species within a certain area does not necessarily mean the species is absent) by modeling the occupancy process and the detection process separately. This can effectively resolve whether a species is absent because it was not detected, or because it was not present. Failing to resolve the ambiguity of species absence can cause incorrect interpretation of data. For example, when the probability of detecting a species is < 1, failing to account for imperfect detection leads to underestimates of species distributions and estimates of covariate relationships that are biased toward zero. Additionally, factors that affect our ability to detect a species may erroneously be included in models of species occurrence (Kery 2010). Differentiating occupancy and detection allows both covariates to be included in site-occupancy models, providing a more robust statistical framework to study habitat associations of individual species. works]) have deployed camera traps directly on residential properties. To address this data gap, speci cally in the US Northeast, we collected camera-trap data over three seasons in residential neighborhoods within Dutchess County, NY, a predominantly rural-suburban area, with a population of roughly 300,000, and an average housing density of 51 houses / km 2 . We applied Bayesian siteoccupancy models to these data to distinguish detection and occurrence. A rst goal of the study was to describe the wildlife community within this human-impacted landscape, which can serve as a baseline for biodiversity in residential areas of the Northeast. To achieve this goal, we estimated the full species richness of mammal and non-passerine bird species, as well as the geographic distributions of individual species, while accounting for imperfect detection. A second goal was to explore how anthropogenic features may affect biodiversity by identifying which species are most and least present in humanimpacted environments, and which covariates related to human disturbance increase and decrease individual species occupancy.

Study area
The study area was in Dutchess County, NY, located midway between New York City and Albany, and included residential neighborhoods (average area: 0.28 km 2 Moruzzi et al. 2002) have suggested that this distance satis es the necessary assumption of independence of site-occupancy models.
Forest cover of our study neighborhoods ranged from 18% − 63% of the total area (median = 46.3%), and non-habitat cover ranged from 17% − 42% of the total area (median = 24.2%). Features considered as nonhabitat include buildings, sidewalks, parking lots, pools, trampolines, roads (paved and unpaved), and bodies of water. Neighborhoods contained residential lawns and garden areas comprising 19% − 48% of the total area (median = 30.3%) ( Table 1). Neighborhood clumpiness, which measures how aggregated or disaggregated forest patches are ranged from 0.93-0.96 (median = 0.95). Values less than 0 indicate less aggregation than would be expected randomly, while values greater than 0 indicate more aggregation than random. The upper bound of discrete forest patches per neighborhood ranged from 14-60 (median = 33) ( Table 1). We report an upper bound because our arti cial neighborhood boundaries caused forest patches connected outside of the neighborhood boundary to be interpreted as multiple, separate forest patches.

Camera trapping and data storage
We deployed three un-baited cameras on three unique properties within each neighborhood for roughly three months of autumn-winter camera-trapping in each year 2016-2018 (Table 1) Unless a property dropped-out of the study, the camera was placed on the same three properties in each neighborhood each season. Similarly, unless a tree was downed, cameras were placed at the same location on the property each season. Both events happened infrequently (< 10%). Each property that hosted a wildlife camera contained forest cover and was selected based on aerial imagery (Parcel Access in ArcMap 10.3, https://gis.dutchessny.gov/parcelaccess/parcelaccess_map.htm) and in-person observation. We used Bushnell No-Glow Aggressor (model #119776C) cameras set to take three photos, each two seconds apart, when activated by motion. Prior to camera deployment, we identi ed optimal camera settings (Table S1) to capture images roughly 5 m away. As predator species and other larger mammals, such as deer, may preferentially use game trails for movement, we attempted to limit our placement of cameras along observed game trails to one per neighborhood to avoid bias in our detection. In some neighborhoods, we opted to place more than one camera along a game trail when at these locations, game trails co-occurred with brush or dense vegetation within the view of the camera. We reasoned these areas would provide refuge for meso-mammals that do not typically use game trails. We placed cameras at approximately 45-degree angles to game trails to best detect wildlife moving along the trail. We placed cameras in relatively open areas under the canopy to maximize our detection rate and to avoid having moving vegetation trigger the camera. We mounted cameras on trees that were at least 15 cm in diameter at breast height (DBH). The height at which we placed the camera on the tree depended on the conditions present (e.g., slope, brush cover, downed logs, debris) but, in general, we placed cameras about 0.5 m above the base of the tree.
After the initial 2016-2017 camera trapping season, we created species accumulation curves as a function of trap days to determine the appropriate length of time for cameras to be deployed ( Figure S1).
In each season, we checked cameras monthly after deployment to replace batteries and SD cards. After a minimum of 90 days of deployment for each camera, we removed all cameras and materials from the eld.
We used Camelot (Camelot-Project) software on a secure network to identify species and store the data. Images of a particular species from a single camera were considered independent if they were taken at least 20 minutes apart, this assumption prevented us from recounting the same individual multiple times if they stood in the camera's view for an extended period. Statistics Since cameras were placed to optimally capture images from 5 m away, we did not include any passerine bird species in our results because small birds have a very low probability of detection from this distance (Randler and Kalb 2018). Instead, our results include mammal and non-passerine bird species.
The Relative Abundance Index (RAI) for each year was calculated as the number of independent observations divided by the number of camera-trap days, all multiplied by 100 to convert this ratio to a percentage. Camera-trap days were calculated as the number of cameras times the number of days they were deployed.
To evaluate factors that could affect occupancy, we included the following covariates in our model: the percent of the neighborhood that was not habitat, including roads or structures (referred to as nonhab in model outputs), the percent of the habitat area within the neighborhood that was forest (forest), the average distance between each of the three cameras in the neighborhood and the nearest road (DTR), and the average distance between each of the three cameras in the neighborhood and the nearest house (DTH). Similar to (Eaken et al. 2018), we considered area of non-habitat to represent the intensity of human impact because it includes buildings and pavement and is therefore linked to tra c, urbanization, light, and noise pollution. We expected increased areas of non-habitat within a neighborhood to decrease the probability that each species would be present in that neighborhood (McKinney 2008). We considered the percent of habitat in the neighborhood that was forest to serve as a proxy for forest connectivity and expected that a species would be more likely to be present in neighborhoods with higher percentages of forest (Baguette and van Dyck 2007; Kindlmann and Burel 2008). We used the camera distance to the nearest household to serve as a proxy for human presence, and thus expected the presence of humans to decrease the presence of wildlife, which has been demonstrated for white-tailed deer (Odocoileus virginianus) (Keim et al. 2019). Finally, we expected occupancy probability to be negatively correlated with a camera's distance to the nearest road (Fahrig and Rytwinski 2009), given the potential for roads to inhibit movement.
To account for potential detection bias, we included two additional covariates: game trail (gt) and fencing (fence). The game trail variable describes the number of cameras within the neighborhood that were placed on a game trail (0, 1, 2, or 3). Fencing on each property was classi ed as a (1, 2, or 3) where 1 represents complete, or nearly complete fencing near the perimeter of the property on which the camera was placed, such that > 0.75 of the property is enclosed, 2 represents complete, or nearly complete fencing that encloses between half and three-quarters of the property, and 3 represents smaller or less continuous fenced areas than (1) or (2), including yards with no fencing. We then took the mean of this value for the three properties per neighborhood to calculate an average fencing variable. Although neither of these variables are expected to directly affect a species' occupancy within a neighborhood, they are expected to affect our ability to detect each species, with game trails increasing our ability to detect  Individual species site-occupancy models. To determine which mammal and non-passerine bird species occupied each neighborhood while accounting for imperfect detection, we modeled species-speci c detection probabilities (p) and occupancy probabilities (ψ), using single-species occupancy models (MacKenzie et al. 2002) based on a Bayesian model of inference. Within these models, each day constituted a survey, allowing for repeated observations that enabled us to distinguish between a lack of detection and a true lack of occupancy. Our models assume that occupancy comes from a Bernoulli distribution, and our observation data come from a Bernoulli process with a success rate of (ψ * p). Covariates (β i ) were modeled into occurrence and detection models using the logit function. Due to minimum data requirements, we only t site-occupancy models for species with naive occupancies > 0.10.
To select the best subset of covariates for our site-occupancy models for each species, we rst separately ranked both the covariates expected to affect occupancy and the covariates expected to affect detection based on their a priori expected in uence on that species' occupancy and detection probabilities, respectively. We then used a top-down approach, starting with the full model and removing a single covariate, sequentially, to form each candidate model.
For model evaluation, we performed a 70 − 30, strati ed, train-test split on our 2018-2019 presence data using the R function "partition" in the package splitTools (Mayer, 2020). Using our train data, we t siteoccupancy models for each candidate subset of covariates. For these models, we ran 3 Markov chains with non-informative, uniform priors for 12,000 iterations, with a burn in rate of 2000, and a thinning rate of 5. Although utilizing AIC values for model comparison is standard when tting maximum-likelihood site-occupancy models, AIC is not recommended for Bayesian, hierarchical models due to the model's latent parameters (Broms et al. 2016; Gelman and Vehtari 2013). Instead, to assess model performance and select an optimal candidate model for each species, we took 1000 draws from the posterior distributions of our hyperparameters within our estimated occupancy models to calculate expected occurrence and detection probabilities at each site. We then retained the maximum value from n stochastic, Bernoulli simulations for each of our 1000 draws, where n represents the number of surveys (days) in our test data, to estimate observed occupancy (presence/absence) from our test data. In order to choose the most parsimonious model for each species, we calculated the accuracy of each model as the average number of correct predictions divided by the total number of predictions over the 1000 simulations. Due to its out-of-sample design, this methodology naturally penalizes extra parameters, which would lead to over-tting of the training data and poor t among the test data. For convenience, this model makes the simplifying assumption that each species occupying a neighborhood is independent of each other species occupying the neighborhood.

Results
Over three years and from 66 cameras, we accumulated 17,820 camera-trap days (5,940 per season) and captured 280,686 images (Table 1) Table 2). Based on our species accumulation curve from 2016-2017 data, the number of mammal and non-passerine bird species detected plateaued around 3600 camera trap days ( Figure S1), indicating that the camera deployments were of su cient duration each season to capture detectable species.  the county to capture full species richness (Fig. 3d-f), which is ~ 7x larger than the number we deployed.
All site-occupancy models had adequate t as demonstrated by their Bayesian p-values between 0.05 and 0.95 (Table S2). Of the species for which we could create occupancy models, wild turkeys (Meleagris gallopavo) were the rarest when accounting for imperfect detection, with an estimated true occupancy of 0.09 in 2016-2017, 0.29 in 2017-2018, and 0.030 in 2018-2019 (Fig. 4). Eastern chipmunks (Tamias striatus) and woodchucks (Marmota monax) were the most di cult species to detect as demonstrated by the relatively large discrepancies between their naïve occupancies and model-predicted occupancies each year when accounting for imperfect detection (Fig. 4). Table 4 Best subset of occupancy | detection covariates for site-occupancy models based on comparison between train and test data from 2018-2019 camera trapping season and corresponding accuracy values for all species with a naïve occupancy > 0.10. Nonhab is the percent of the neighborhood that was not habitat, including roads or structures; forest is the percent of the habitat area within the neighborhood that was forest; DTR is the average distance between each of the three cameras in the neighborhood and the nearest road; DTH is the average distance between each of the three cameras in the neighborhood and the nearest house; gt is the number of camera in a neighborhood that were placed along a game trail; fencing is an index characterizing how much of the property was fenced (see Methods for details) Overall, the percent of nonhabitat within a neighborhood was the most common occupancy covariate in our site-occupancy models, ultimately included in the nal models for 8 of the 13 species (Table 3). The percent of forested habitat was the second covariate, included for 7 out of the 13 species (Table 3).
For all species for which forest was included as a covariate in the nal model, the relationship was negative, meaning that occupancy declined as the percent of habitat that is forest in a neighborhood increased (Table 4). For four species -eastern cottontails (Sylvilagus oridanus), gray foxes (Urocyon cineroargenteus), red squirrels (Sciurus vulgaris), and striped skunks (Mephitis mephitis) -occupancy increased as the area of nonhabitat increased. In contrast, for coyotes (Canis latrans), eastern chipmunks (Tamias striatus), woodchucks (Marmota monax), and wild turkeys (Meleagris gallopavo), occupancy decreased as nonhabitat area increased ( Table 4).
The average distance between the cameras within a neighborhood and the nearest house, and the average distance between cameras in the neighborhood and the nearest road, were both included in nal models for 6 of 13 species. Occupancy increased with distance from the nearest house for eastern cottontails, gray foxes, woodchucks, and striped skunks, and decreased with distance from the nearest house for coyotes and wild turkeys ( Table 4). As the average distance to the nearest road increased, occupancy increased for eastern chipmunks, woodchucks, striped skunks, and wild turkeys, while occupancy decreased for eastern cottontails and gray foxes (Table 4).
Of our detection covariates, whether or not cameras were placed on game trails was ultimately included in nal models for 6 out of the 13 species (Table 3). For all species for which game trail was included as a covariate in the model, detection decreased as the percent of cameras within the neighborhood located on a game trail increased (Table 4). Fence was only included in the nal models for 3 out of 13 species (Table 3). Fenced properties decreased detection for woodchucks and wild turkeys but increased detection for eastern cottontails (Table 4). Neighborhood biodiversity estimates. Considering only those species for which we were able to create occupancy models, the average number of estimated species in each of our 22 neighborhoods was 11.9 (range: 10-13, Figure S2), when accounting for imperfect detection. The average difference between the number of species detected and the number of species expected to be present based on our models was 1.4 (range: 0-4, Figure S2). We did not detect any geographic patterns.

Discussion
Our study utilized 66 motion-triggered cameras deployed annually for three-month periods, for three consecutive years, to study the wildlife community within residential properties of a human-impacted landscape in New York state. Three-month deployment periods allowed detection of both transient and resident individuals, although we could not distinguish between these categories. Passerine birds and some small mammals (e.g. shrews, voles) known to be present in the properties were not readily detectable and excluded from analyses. Our full species-richness model estimated a maximum of 33 species present, based on our 2018-2019 data. The difference between the number of species estimated by our species-richness model and the number of species directly detected by our cameras highlights the importance of accounting for imperfect detection when using motion-triggered cameras to study biodiversity. We did not detect any strong spatial patterns when modeling estimated distributions of species with naïve occupancies > 0.10, which may re ect homogeneity between our camera clusters at broader spatial scales (~ 1 km).
Eight species that meet the basic requirements for detection by our cameras (body weight > 100g, grounddwelling) and have been identi ed in Dutchess County, NY (Ueda K, 2021) were not captured by our cameras: muskrats (Ondatra zibethicus), American beavers (Castor canadensis), North American river otters (Lontra canadensis) North American porcupines (Erethizon dorsatum), American minks (Neovison vison) long-tailed weasels (Mustela frenata) snowshoe hares (Lepus americanus), and short-tailed weasels (Mustela erminea). These species might occur at densities low enough that we failed to detect them with our cameras, or they might not be present in residential communities, owing to habitat specialization. For example, muskrats, beavers, otters, and minks require aquatic habitat.
A second goal of our study was to examine how individual species respond to anthropogenic landscape features. White-tailed deer, Virginia opossums, red foxes, eastern gray squirrels, and raccoons were detected in every one of our study neighborhoods. Their presence at all study locations prevented us from using covariates within site-occupancy models to understand the varying in uences of anthropogenic landscape features. These species may thrive in human-impacted areas due to human-subsidized resources, or lower predation/hunting risk. While both ground-dwelling and large enough to be captured by cameras, bobcats and shers were rarely detected in our neighborhoods. Two additional studies that utilized cameras in residential yards and natural locations in North Carolina (Kays and Parsons 2014; Hansen et al. 2020), also failed to detect bobcats in residential yards, only nding them in less humanimpacted locations. Infrequent detection of bobcats in residential areas suggests they may be unable to thrive in such fragmented landscapes. Additionally, as obligate carnivores, due to their diet, bobcats may be unable to exploit supplemental food resources (Anderson and Lovallo 2003) found in residential yards.
We were able to create site-occupancy models for 13 species. Based on our results, gray foxes, red squirrels, striped skunks, and eastern cottontails were more likely to be present in neighborhoods with more structures and road coverage. Similar results for gray foxes were found in Kays and Parsons (2014 .
Surprisingly, neighborhoods with more habitat that is forest had lower occupancy for coyotes, eastern chipmunks, eastern cottontails, gray foxes, woodchucks, striped skunks, and wild turkeys. These species might be avoiding residential yards when more forested habitat outside of yards is available.
Alternatively, these species might be attracted to supplemental food in yards that have less forest.
The average distance between cameras and the nearest house positively affected occupancy for eastern cottontails, woodchucks, striped skunks, and gray foxes, potentially suggesting aversion to humans.
Similarly, Gallo et al. (2019) hypothesized that eastern cottontails limit their activity patterns to early and late hours to avoid human activity. Additionally, Lehrer et al. (2012) concluded that woodchucks were not strongly habituated to humans. Occupancy was negatively related to the average distance between the cameras and the nearest house for wild turkeys and coyotes, indicating these species are less humanaverse. In the Chicago Metropolitan Area, human-exploited resources, such as garbage, can comprise up to 25% of coyote's diet during seasons with less prey availability (Morey et al.2007), which may explain why coyotes were more likely to occupy areas close to human homes. Finally, occupancy was positively related to the average distance between the camera and the nearest road for chipmunks, woodchucks, striped skunks, and wild turkeys, indicating road avoidance. Occupancy of eastern cottontails and gray foxes, however, was negatively related to the average distance to the nearest road.
Chipmunks, eastern cottontails, woodchucks, red squirrels, striped skunks, and wild turkeys were all less likely to be detected when more cameras were located along game trails. These species might avoid game trails to limit detection by predators. Finally, fencing negatively in uenced detection for wild turkeys and woodchucks, indicating we were less likely to detect each species in properties with more fencing. Fencing, however, had a positive relationship with detection for eastern cottontails, which was also found in Kays and Parsons (2014). Fenced yards potentially create a safer environment for eastern cottontails by excluding predators.
Interpretations of species occupancy and diversity patterns in residential areas should consider the broader landscape context encompassing the study sites. The regions of Dutchess County, NY in which we conducted our study are largely forested (median 46.3% of the land area), with medians of 30.3% and 24.2% in herbaceous vegetation (including lawns and agriculture) and non-habitat, respectively. Forest connectivity was high (median clumpiness = 0.95), and discrete forest patches were uncommon (median upper bound of discrete patches = 33) ( Table 1). Prior research in this region, focusing on changes in vertebrate community structure as the size of forest patches varied, found that small mammals and deer occupied all size ranges of forest patches but other mammal species were more likely to be detected in larger patches ( Together, these results allow us to describe a biodiversity baseline for residential properties in the US Northeast and contribute to understanding how anthropogenic landscape features affect biodiversity. Understanding biodiversity in residential areas is particularly important because the rate at which natural landscapes are changing (Venter er al., 2016; Williams et al. 2020). Of course, our ndings are based only on data collected within Dutchess County, NY and the response of species occupancy could vary in different geographic areas. Additionally, our cameras were deployed over autumn and winter, so our results may not generalize over all seasons, especially for species that hibernate or undergo periods of torpor during colder months. Finally, as is true of all camera trap studies, motion-triggered cameras are not able to detect small and volant animals. In the future it could be useful to combine camera trapping data with other methods to detect small and non-ground dwelling species to get a more complete picture of overall biodiversity. As landscapes continue to change, it will be important to continue to use camera traps to document biodiversity changes in residential areas. Declarations