Combining community science and MaxEnt modeling to estimate Wild Turkey ( Meleagris gallopavo ) winter abundance and distribution

. Understanding the distribution and abundance of species is a fundamental aspect of conservation biology. Species distribution models aim to predict distributions based on species observations and ecologically relevant information. To understand the contemporary distribution of Wild Turkeys ( Meleagris gallopavo ) in Ontario, we curated and collated Wild Turkey flock observations from eBird and iNaturalist submitted during winter 2018. We combined these with environmental predictors to build distribution models using MaxEnt and evaluated model fit using 10-fold cross validation. We also estimated total population size for this species under different modeling scenarios. The potential presence of unknown spatial bias in community science datasets is a complex problem often requiring context-specific statistical solutions. Data cleaning, sometimes referred to as thinning, filtering, or culling, is often proposed to manage this bias. As such, we tested the effect of data cleaning on model outputs and on subsequent analyses. We evaluated all models using area under the curve (AUC). We found building density to be the most important environmental variable followed by winter severity. We validated our habitat suitability estimates using fine-scale GPS data and found that data cleaning had no effect on habitat suitability estimates inside available Wild Turkey habitat or inside core-use areas, except at one site in 2012 (t = -2.2, P = 0.04, df = 14). Use of community collected data offers a cost-efficient and collaborative method to obtain data for species distribution modeling and management. We discuss implications for Wild Turkey management and present potential contemporary distribution maps for this species.


INTRODUCTION
Understanding the distribution and abundance of species is a fundamental aspect of conservation biology, allowing for informed, endangered, and invasive species management (Livingston et al. 1990, Higgins et al. 1999, Gallagher et al. 2010, Guisan et al. 2013, successful reintroductions (Martínez-Meyer et al. 2006, Massaro et al. 2017, Malone et al. 2018, Riaz et al. 2020, ecosystem restoration (Mladenoff et al. 1995(Mladenoff et al. , 1997, population viability analyses (Akçakaya et al. 1995, Akçakaya and Atwood 1997, Roloff and Haufler 1997, and harvest management (Jonzén et al. 2001). Species distribution models (SDMs) aim to predict the distribution of a given species based on species observations and ecologically relevant environmental features (Hirzel et al. 2002). Absence data are often unavailable or unreliable, so modeling techniques that require presence-only information are becoming increasingly common in species distribution modeling (Razgour et al. 2011, Nazeri et al. 2012, Isaac et al. 2014a, Tran and Vu 2020. MaxEnt, a presence-only machine learning software , has become the most commonly used program to generate SDMs because it has been demonstrated to outperform many other presence-only and presence-absence modeling techniques (Elith et al. 2006, Hernandez et al. 2006. MaxEnt estimates the distribution of a given species by finding the distribution that is closest to geographic uniformity (i.e., maximum entropy) subject to constraints provided by given environmental features at each species occurrence location .
One way to curate species presence information over a large geographic area is to look to community-collected data: data collected or processed by volunteers as part of a scientific inquiry (Silvertown 2009). Community science projects can cover larger geographic and temporal scopes than traditional survey types and data derived from such projects are being increasingly used in modeling landscape-scale movements and distributions (Snäll et al. 2011, Tulloch et al. 2013, Supp et al. 2015, Bradsworth et al. 2017, Brommer et al. 2017. In fact, it has been argued that any project seeking to collect large volumes of data over a wide geographic area can only succeed with the assistance of community members (Silvertown 2009). Such data are often collected without experimental design or set survey methods (Steen et al. 2019). As a result, these datasets may have unknown spatial biases, imprecise spatial or temporal resolutions, or result in the under-or over-reporting of species (Fitzpatrick et al. 2009, Dickinson et al. 2010, Steger et al. 2017, La Sorte et al. 2018. Because community science can yield a large volume of data, even with these limitations it can be very useful in predicting species distributions and informing management decisions (Dickinson et al. 2010, Gallagher et al. 2010, Guisan et al. 2013, La Sorte et al. 2018. For some species, SDMs built from unfiltered community data can match or exceed the performance of SDMs built using systematically collected data (Bird et al. 2014, Isaac et al. 2014a, Steen et al. 2019, Johnston et al. 2021. However, for other species, filtering the dataset can greatly improve model performance , Fourcade et al. 2014, Aiello-Lammens et al. 2015, Kiedrzyński et al. 2017, Steen et al. 2019). As such, it has been suggested that significant time should be invested in determining how to best clean datasets prior to modeling Taylor 2017, Steen et al. 2019) and that cleaning the data should be two-fold, involving (1) thinning the data to manage sampling bias, and (2) filtering the data to manage spatial bias associated with species sociality , Syfert et al. 2013.
Findings from theoretical modeling indicate that cleaned datasets generally produce better models than the full dataset and that filtering based on survey effort yields more powerful models than filtering based on surveyor knowledge (Kramer-Shadt et al. 2013, Steen et al. 2019, Johnston et al. 2021. However, in most cases, the most appropriate data cleaning practices and model specifications are species and context specific and should be determined using researcher knowledge and preliminary statistical analyses (e.g., tests for spatial-autocorrelation or prior knowledge about sociality).
In practice, community data cleaning methods employed by researchers vary substantially depending on the study and the species. For instance, some researchers have filtered community data by year (Bradsworth et al. 2017) or based on variation around the mean (Supp et al. 2015), whereas others have attempted to account for pseudoreplication by removing observations within a predetermined and ecologically informed distance (Razgour et al. 2011, Bradsworth et al. 2017. Biddle et al. (2021), who studied the distribution of an avian species often considered difficult for community members to identify, filtered data based on geographic location and metrics of species identification accuracy, whereas Coxen et al. (2017) conducted filtering based on geographic distance between observations, also known as rarefication. Inconsistencies in data cleaning methodology may lead to inaccurate comparisons between studies, even when researchers are studying the same or closely related species.
The Wild Turkey (Meleagris gallopavo) is a dark, large-bodied bird native to North America. Prior to European colonization, this species was an important component of forest biodiversity and held substantial cultural significance for Indigenous groups (Aldrich 1967). Agricultural and residential development of North America resulted in large-scale forest loss. This habitat loss, paired with unregulated hunting, led to the extirpation of turkeys in many regions across North America by the early 1900s (Blakey 1941, Davis 1949. The Wild Turkey has been successfully reestablished across its range largely because of the implementation of harvest regulation, habitat reclamation following farmland abandonment in the 1920s and 1930s, and coordinated trap and transfer efforts. Wild Turkeys were trapped and transferred across North America beginning in the 1950s and continuing into the 1980s (Mosby 1959, 1973, 1975, Bailey 1980, Kennamer 1986). Contemporary turkey populations occupy all American states, 6 of 13 Canadian provinces or territories, and northern Mexico (Kennamer and Kennamer 1996, Tapley et al 2001. Across their range, Wild Turkey population sizes have been estimated using helicopter surveys (Beason 1970, Thompson and Baker 1981, Kubisiak et al. 1997, Butler et al. 2007b, 2008, roadbased surveys (Butler et al. 2005, Erxleben et al. 2008, brood counts (Schwertner et al. 2003, Butler et al. 2007c), rural mail surveys (Ontario Government 1985Applegate, 1997), harvest rates (Gonnerman 2021), and roost counts (Thomas et al. 1966). However, these traditional survey types are often costly, time intensive, and require substantial effort, all of which act as a barrier to applying these methods over large geographic areas or temporal scales. As such, most population surveys of this species conducted thus far have been limited in scope and are often unstandardized (Healy andPowell 1999, Butler 2006). In addition, many survey types provide indices of population size or growth, rather than an estimate of true population size, and may need to be repeated over subsequent years to yield informative results and understand their management implications.
Turkeys are highly visible birds, especially in autumn and winter when the birds form large flocks, and their dark plumage forms visual contrast with the white snow. Consequently, we considered Wild Turkeys to be a good species to use for exploring the potential for use of community science in generating large-scale maps of distribution and abundance for conservation and management. Using Wild Turkeys as a case study, we investigated the accuracy of SDMs derived from community science datasets for predicting species distribution and abundance. We used the abundancesuitability relationship to investigate Wild Turkey populations in a range of scenarios. Geographic regions with higher environmental suitability should also have larger populations of Wild Turkeys and vice-versa (de la Fuente et al. 2021, Gonnerman 2021. This hypothesis is termed the abundance-suitability relationship and states that environmental suitability derived from species distribution models should explain the spatial variation in abundance over a species' geographical range (Weber et al. 2017).
High environmental suitability does not always indicate high abundance. Indeed, density can be a misleading indicator of habitat quality (Van Horne 1983). However, Wild Turkey natural history suggests that there is typically a positive relationship between abundance and suitability for this species (Porter 1992, Gonnerman 2021, and a recent meta-analysis concluded that occurrence data can be a reasonable proxy for abundance, especially for vertebrates (Weber et al. 2017).
We tested the sensitivity of our analyses to data cleaning by building models and generating population size estimates using both the raw community science dataset (un-thinned) and a heavily filtered version of the same dataset in which we attempt to account for replicate flock detections and spatial bias associated with sampling. It has been suggested that, when possible, researchers should externally validate SDMs against real fine-scale spatial use data (Bradsworth et al. 2017). As such, we validated our predicted probabilities and population size estimates against traditional aerial and road surveys as well as fine-scale GPS tracking data in two regions in Ontario, Canada. We expected that community science data would provide the basis for accurate estimation of distribution and abundance, particularly for a common and easily identifiable species like the Wild Turkey.

Study area
The total land area is 908,699.33 km², and it contains the most populated city in Canada: Toronto (Statistics Canada 2017; Fig.  1). The estimated human population size of Ontario in 2022 was 15,034,547, equalling almost 40% of the country's total population (Statistics Canada 2022).
Ontario contains 3 of Canada's 15 terrestrial ecozones: the Mixedwood Plains in the south, the Boreal Shield, and Hudson Bay lowlands in the north (Crins et al. 2009). Historically, Wild Turkeys only inhabited the Mixedwood Plains ecozone. However, land conversion in the north has allowed Wild Turkeys to expand their range northward and inhabit regions of the Boreal Shield that have been converted for agriculture or other anthropogenic developments (OMNR 2007). Eastern Wild Turkeys (Meleagris gallopavo silvestris), the only subspecies present in Ontario, are found in eastern North America from Ontario to Florida . As such, Ontario encompasses the northern range edge of the Eastern Wild Turkey.
Ideal Wild Turkey habitat has been described as an even mix between primarily deciduous forest and open field (Kurzejeski and Lewis 1985). However, large contemporary turkey flocks can be found in regions that do not meet these criteria, suggesting that Wild Turkeys may display high adaptability to anthropogenic landscapes (OMNR 2007). For instance, although the geographic range of Wild Turkeys is seemingly expanding northward, this species still primarily inhabits the southern portion of the province, which is characterized by a high level of disturbance, habitat fragmentation, and agricultural development (OMNR 2007).
The Wild Turkey was considered extirpated from Ontario in 1909(OMNR 2007. In 1984, efforts began to reintroduce this species to the province, and from 1984 to 1987, 274 Wild Turkeys were trapped and transferred to Ontario from Missouri, Iowa, Michigan, New York, Vermont, New Jersey, and Tennessee (OMNR 2007). In subsequent years, to support the dispersal of this species in Ontario, over 4000 individuals resulting from the initial founder population were trapped and transferred to numerous sites throughout potentially suitable habitats in Ontario primarily within the Mixedwood Plains ecozone. The most recent estimate of Ontario's Wild Turkey population size, published in 2007, was 70,000 individuals. However, this estimate is a rough approximation based on annual harvest and an assumption about harvest rate (OMNRF 2007).

Data collection and cleaning
To compile known Wild Turkey observations from across Ontario, we conducted a community science campaign in which we requested that participants submit Ontario Wild Turkey observations to eBird (Sullivan et al. 2009) or iNaturalist (California Academy of Sciences 2020) from 1 December 2018 to 31 Mar 2019. eBird is a virtual tool that allows users to submit and review observations of bird species anywhere in the world. The community-collected data are then archived and freely accessible to anyone, including researchers (Sullivan et al. 2009). iNaturalist is a similar observation reporting platform to eBird however, it extends to all taxa. iNaturalist also differs from eBird in that it encourages users to submit additional media with each observation. For instance, users may upload a photograph of a plant species or an audio recording of a bird call to aid in species identification.
To advertise our project and encourage Wild Turkey observation submissions to each platform, we made an informative poster with a summary of important details about the project and instructions regarding how to submit observations (Appendix 3, Figure A3.1). This poster was circulated to field naturalist groups across the province and was shared on multiple social media platforms. If participants sought more information about the project, they were encouraged to visit a website we produced with multiple pages presenting further details about the project and updates about our findings.
We ran this campaign during Ontario's winter because turkeys are dark-colored and are thus more visible in the winter months against the contrasting white snow. Turkeys also congregate into large flocks during the winter (Healy 1992), allowing observers to easily detect multiple individuals at one time. We chose to use eBird (Sullivan et al. 2009) and iNaturalist (California Academy of Sciences 2020), because both reporting platforms are free, wellknown, and already widely used (Silvertown 2009). The popularity of these two platforms also allowed us to include observations from individuals that may not have been familiar with our community science campaign because all Wild Turkey observations submitted within the reporting period were included in our analysis.

Removing replicates and thinning the data
To examine how thinning the data may impact SDM performance in this case, we ran models with two versions of our community collected dataset: one containing all observations collected during the study period and another heavily cleaned version of the dataset that we (1) filtered for replicate observations and (2) thinned to account for potential underlying spatial bias. This resulted in a raw, un-thinned version of the dataset (n = 5846) and a heavily filtered and thinned version of the same dataset (n = 492).
To generate the thinned dataset, we first identified and removed potential replicate observations. During the winter, Wild Turkeys congregate and move through the landscape in large flocks and individuals are rarely observed alone (Healy 1992). As such, during this time of year, individual home ranges can be used to approximate the size and distribution of whole flock home ranges. Thus, to identify potential replicate observations of flocks, we first estimated the average winter home range size of Wild Turkeys in two regions of the province: the Bruce Peninsula and Peterborough County ( Fig. 1) with data derived from individuals that had been GPS-tagged for research projects during 2011-2012 on the Bruce Peninsula and during 2017-2019 in Peterborough County. Birds were captured at baited locations using a rocket net (Grubb 1988) following methods outlined by Bowman (2014, 2016). Processing entailed weighing, sexing, collecting a blood sample for DNA extraction, and GPStransmitting (Model PP-VHF-3600L, Lotek, Newmarket, Canada) attachment. The schedule at which GPS locations were recorded varied depending on the capture year and location and ranged from once per hour to once every 4.25 hours.
We estimated Wild Turkey winter home range by calculating 95% kernel home range polygons using GPS points collected during the winter for 45 individuals (24 M, 21 F). Kernels were estimated using the adehabitatHR package in R Studio (Calenge 2006). We selected bandwidth values using an ad-hoc approach in which we reduced bandwidth values from 100% percent of the reference bandwidth (246.95 in Peterborough County and 144.77 on the Bruce Peninsula) to 10%, by 10% decrements, and then selected the lowest bandwidth that resulted in the same number of polygons as the reference bandwidth. Individuals who were tracked for less than 13 days, or approximately 15% of the season, were excluded from the analysis (Niedzielski and Bowman 2016). Across both sites and both sexes, we estimated a mean winter home range size (95% kernel) of 1.8 km², ranging from 0.06 km² to 6.58 km², with a standard deviation of 1.45 (Appendix 2, Table  A2.1).
Observation clusters were identified using the leader algorithm from the leaderCluster package in R Studio (Arnold 2014). This clustering algorithm allows users to set the approximate radius of clusters rather than the desired number of clusters (Arnold 2014). This is useful when attempting to assign observations to clusters based on geographic distance. We selected a radius of 756 m because this value results in circular clusters with approximately the same area as the average Wild Turkey mean winter home range (1.8 km²; see Appendix 2, Table A2.1). For each cluster, only one observation was retained. Observation locations were then converted to the centroid of each cluster. This differed from the un-thinned dataset in which observation locations were kept at the same location as reported by participants.
In addition to filtering for replicate observations, we also thinned this dataset, using the spThin package (Aiello-Lammens et al. 2019) in R Studio (R Core Team 2013), to achieve an approximate uniform density of one flock per 10 km². This was done to account for potential unknown spatial bias in the dataset, e.g., observer density (Aiello-Lammens et al. 2015).
Multiple users may sometimes submit identical information. For instance, eBird users may submit shared checklists: lists submitted by multiple users that contain identical observations. We did not intentionally remove shared eBird checklists. However, if the observations shared identical GPS locations for the observed individuals, then the observations would have been identified as replicate observations in geographic space and removed. We did not remove observations associated with incidental checklists (Sullivan et al. 2009) but we did compare the effect of removing observations associated with checklists that exceeded five hours, five kilometers in length, or both, as is recommended under eBird's best practices . Upon comparing models built with datasets including observations from long (spatially and temporally) checklists with models built without these observations, we found no substantive effect on any model estimates including area under the curve (AUC), variable contributions, and threshold values. As such, we chose to include observations associated with long (spatially or temporally) checklists in our analyses.

Species distribution modeling
To estimate the distribution of Wild Turkeys in Ontario, we collected and analyzed six environmental variables and modeled their relationship to community-collected Wild Turkey observations. We modeled this relationship using MaxEnt, a machine learning process that determines the spatial probability distribution of a species based on presence-only records and relevant environmental variables . Winter is the season that most limits Wild Turkey abundance at their northern range edge Bowman 2014, Gonnerman 2021) and the time when turkeys are the most visible because of their dark color and flocking behavior. Therefore, we focused on estimating the distribution and abundance of Wild Turkeys in winter. We evaluated winter habitat suitability across the Wild Turkey range in Ontario (Phillips and Dudik 2008).
We selected environmental variables based on turkey ecology and relationships among variables. For instance, turkeys are a generalist, forest-dependent species (Healy 1992, Porter 1992. Wild Turkeys are also heavily reliant on supplemental food sources, particularly during the winter months when their natural food sources are less available (Vander Haegen et al. 1989, Nguyen et al. 2003, Kane et al. 2007, Restani et al. 2009). Thus, we included both forest and agricultural land cover in our model. Using ArcMap (ESRI 2011), we created the forest and agriculture variables by combining land cover classes within the OMNR Provincial Land Cover Dataset (2000). Cells representing coniferous forest and deciduous forest were combined and exported as a new data layer to create our forest variable, and cells representing pasture and cropland were combined and exported to create our agriculture variable. Both variables were converted to binary, categorical variables prior to modeling; cells that contain the respective landscape cover type (1) and cells that do not (0).
In addition to plant and animal agricultural operations, bird feeders may also represent a winter food source for Wild Turkeys in Ontario (Niedzielski and Bowman 2014). Thus, two anthropogenic variables were included to represent human development as a proxy for human-maintained supplemental food sources: road density (OMNR 2010-2013) and building density (OMNR 1977(OMNR -2014. Wild Turkey presence may also have a negative association with road density because individuals closer to roads may be more likely to experience increased road mortality as a confounding effect of hunting pressure (Holbrook and Vaughan 1985) and/or increased predation (Thogmartin and Schaeffer 2000). Road density was estimated by calculating the cumulative road length, in kilometers, per pixel. Building density was estimated by calculating the number of buildings per pixel.
It is known that Wild Turkeys are affected by deep, powdered snow in that it restricts their movement across the landscape during the winter months, reducing their ability to forage and evade predators (Austin and DeGraff 1975, Wunz and Hayden 1975, Porter 1977, Porter et al. 1980, Nguyen et al. 2003, Kane et al. 2007, Gonnerman 2021. There has also been one reported instance of a Wild Turkey dying from frostbite in Ontario (MacDonald et al. 2016). Thus, to represent differences in winter severity across their range, minimum winter temperature, and average snow depth were included in the model. Climate data were retrieved from Environment and Climate Change Canada weather stations in Ontario using the weathercan R package (LaZerte and Albers 2018). We interpolated values between the stations using kriging (ESRI 2011). Kriging is a form of spatial interpolation that uses mathematical formulas to estimate a continuous surface of values. Kriging assumes that there is a structural component present (in this case, observations) and that the local trend varies from one location to another (ESRI 2011).
To standardize the cell size and spatial extent across layers, each layer was rasterized, and values were assigned to 2.27 x 2.27 km square pixels (5.15 km²) covering the extent of the survey area (ESRI 2011). This cell size was selected for multiple reasons. First, to allow for multiple flock observations to occur within the same cell because we estimated a mean winter home range size (95% kernel) of 1.8 km². Furthermore, all data layers were resampled to a resolution of 5.15 km² because this value allowed for multiple Wild Turkey flock observations to occur within the same cell for the un-thinned dataset prior to MaxEnt modeling and allowed for the standardizing of cell sizes across data layers of different spatial scales. We then converted all layers to data format *.asc for import into MaxEnt. In MaxEnt, model restrictions are applied as feature types. Hinge features represent piece-wise linear functions in that they behave like linear functions with thresholds allowing the steepness and direction of the linear relationship to differ below and above each threshold (Phillips 2017). As such, hinge features tend to make linear and threshold features redundant (Elith et al. 2010). The default setting in MaxEnt is to employ auto features, which allows the software to tune parameters based on model performance (Phillips and Dudík 2008). To reduce the potential of overfitting, our final models included linear, quadratic, product, and threshold feature types but not hinge (Phillips andDudík 2008, Elith et al. 2010).
Prior to setting the user-specified parameters in MaxEnt, we used the "ENMeval" R package to identify the most appropriate regularization parameter value for our dataset (Muscarella et al. 2014). Forty-eight models with combinations of restrictions (feature types) and regularization multipliers were compared to select the most appropriate multiplier value. The regularization multiplier is a parameter that helps to prevent model overcomplexity and/or overfitting (Elith et al. 2010). The MaxEnt default value is centered at 1.0. A regularization parameter less than 1.0 will produce estimates with a more localized output distribution with a closer fit to the presence records provided. A regularization parameter greater than 1.0 will produce estimates with a less localized prediction (Phillips 2017). We developed models with multiplier values increasing from 0.5 to 4 by increments of 0.5.
Akaike information criterion scores corrected for sample size (AICc) were generated by ENMeval for all models, with the lowest score indicating the model with the highest maximum likelihood estimate. This analysis indicated that a regularization multiplier of 1.5 would be most appropriate to model Wild Turkey distribution because this value yielded the lowest AICc (Isaac et al. 2014b).
The climate variables "snow depth" and "minimum seasonal temperature" were highly correlated (R² = -0.92; R Core Team 2017). Research shows that Wild Turkey movement is limited during the winter by both deep snow and cold temperatures resulting in higher mortality from starvation and predation (Austin and DeGraff 1975, Wunz and Hayden 1975, Porter 1977, Porter et al. 1980, Nguyen et al. 2003, Kane et al. 2007). As such, two global models were generated: (1) minimum temperature as a climate variable and (2) snow depth as a climate variable. Both global models were run using the thinned dataset and the un-thinned dataset. All models were run using 10-fold cross validation, which allowed us to calculate summary statistics. By running the models using 10-fold cross validation, we have parsed the larger dataset into 10 subsets and calculated summary statistics, such as mean AUC from the outputs. For each model, we also examined variable contributions and habitat suitability threshold values. For each of the 10 runs, 90% of the dataset was used to train the model, and 10% was used to test the model. To evaluate each model's goodness-of-fit, we referred to the AUC of the receiver operator characteristics (ROC; Pearson et al. 2006. Area under the curve measures the ability of a probabilistic model output to correctly distinguish presence from random locations . When assessing the importance of each of the variables in the models, we selected the jackknife method: a resampling technique that provides an estimate of variance by analyzing subsamples of size (n -1) obtained by omitting one observation during each run (Wu 1986). This method is effectively sampling without replacement.
The MaxEnt software, version 3.4.4.,  generated a raster layer with pixels representing the probability of presence from 0-1, following a user-specified conditional loglog (clog-log) transformation. It is important to note that in models relying on presence-only records, such as MaxEnt, the probability of presence only refers to the probability of presence based on the information available to the model. So, in this case, the probability of presence refers to the probability of Wild Turkey presence given the observations used by MaxEnt's algorithm and their association with the layers included. As such, from MaxEnt's probability of presence, we can only infer habitat suitability rather than true probability of a species' presence in an area.

Population size estimation
To account for uncertainty in our modeling process and in the dataset, we generated a range of potential population sizes using our thinned (n = 492) and un-thinned (n = 5846) datasets (Steen et al. 2019) under different habitat suitability threshold scenarios (Liu et al. 2016) and under different population density scenarios.
To distinguish between suitable and unsuitable habitat, we used four habitat suitability threshold values, two thresholds identified by MaxEnt: minimum training presence (MTP) and 10th percentile training presence, and two thresholds assigned posthoc based on a visual assessment of the spread of the data (Appendix 3, Fig. A3.1). The MTP threshold finds the lowest predicted suitability value for an occurrence point and is thus the least conservative threshold value resulting in the largest predicted range. The 10th percentile training presence omits all regions with habitat suitability estimates lower than the suitability values for the lowest 10% of occurrence records . This is a more conservative value and is commonly used in SDM studies (Raes et al. 2009, Rebelo andJones 2010). Prior to generating population size estimates, raster cells with habitat suitability values less than the defined threshold for each scenario were removed from the dataset resulting in four potential geographic ranges of Wild Turkeys in Ontario.
To estimate potential densities of turkeys in Ontario, we first calculated the mean density of Wild Turkeys reported for each U. S. state by Erikson et al. (2014). Researchers compiled estimates of total population size by surveying state wildlife agencies (  Table A2.5). We divided the estimated total population size by the estimated occupied range for each state, then calculated the mean and standard deviation. To generate values representative of low-and high-density estimates, we subtracted and added 1 standard deviation to the mean, respectively (2.07 +/-0.817).
To examine the sensitivity of our analysis to various user-specific parameters, we also generated estimates using un-thinned data and an inflated mean density of 4.02: the mean density of Wild Turkeys in Alabama and the highest density reported in the United States (Erikson et al. 2014). We used a standard deviation of 2.07 to generate low-and high-density estimates around the inflated mean.
For each habitat suitability threshold scenario, the suitability values were binned to identify the proportion of values contained within each bin and to examine the overall distribution of the data. Using a "for loop" in R (R Core Team 2017), we generated a range of population size estimates by selecting 1000 habitat suitability values proportional to the underlying distribution of the data. For instance, the SDM built using thinned data predicts that 21.4% of the habitat suitability values fall between 0.7 and 0.79. Thus, we randomly selected 214 (21.4% of 1000) samples from this bin.
We then assigned these values as thresholds between low-and high-quality habitat and categorized raster cells below this value as likely containing habitat suitable for low-density populations (1.25 turkeys per km²), and raster cells above this value as likely containing habitat suitable for high-density populations (2.89 turkeys per km²). We then calculated the geographic area of the low-and high-density habitat and multiplied the areas by our estimated low and high turkey densities to generate a distribution of potential population size estimates from which we could calculate summary statistics.

Validation of SDM against fine-scale tracking data
It has been suggested that more researchers should consider external validation of species distribution models against real, fine-scale spatial data (Bradsworth et al. 2017). Thus, we attempted to validate the results of our SDM against Wild Turkey GPS tracking data collected within two regions in Ontario: Peterborough County and the Bruce Peninsula (Fig. 1).
Peterborough County is a region in south-central Ontario, characterized by mixed forest, plant, and animal agricultural operations. The Bruce Peninsula also contains a mix of land cover types but is characterized primarily by coniferous forest, mixed forest, fields, and deciduous forest (Niedzielski and Bowman 2016). The dominant industries in both regions are agriculture and tourism. The main crops grown in Peterborough County are soybean, corn, and winter wheat. Pasture and hay fields are also significant components of the landscape to support animal agriculture. Animal agriculture is the predominant type of farming in the Bruce Peninsula in the form of small-scale beef farming (Niedzielski and Bowman 2016). Both regions are located nearby publicly accessible provincial parks: Kawartha Highlands Provincial Park in the northern half of Peterborough County and Bruce Peninsula Provincial Park on the north-western tip of the Bruce Peninsula.
Using GPS data collected in the Bruce Peninsula (Niedzielski and Bowman 2014) and Peterborough County, we compared the mean habitat suitability within four hierarchical levels of Wild Turkey space use: (1) annual available habitat, i.e., a 100% minimum convex polygon (MCP) around pooled Wild Turkey GPS locations; (2) winter available habitat (a 100% MCP around pooled Wild Turkey winter GPS locations); (3) Wild Turkey winter home ranges (95% kernel home range polygons); and (4) Wild Turkey winter core use areas (50% kernel home range polygons). See supplemental material in Appendix 3, Fig. A3. 3, and Koen et al. 2014. We calculated the 50% kernel home range polygons using the same percent bandwidth as when calculating the 95% kernel home range polygons. That is, the bandwidth selected using our ad-hoc approach. We defined winter from December 21 to March 20, the calendar winter season in North America, because calendar seasons align with seasonal changes in turkey ecology and behavior (Kurzejeski et al. 1987, Badyaev et al. 1996, Humberg et al. 2009, Niedzielski and Bowman 2014. Wild Turkeys were captured and fitted with transmitters in the winter, in both Ontario study areas. As such, we defined each year from December 21 (the first day of winter) to December 20 (the last day of fall) the following year. For instance, year one begins on 21 December 2016 and ends on 20 December 2017.
To the 100% MCPs representing annual available and winter available habitat, we added buffers based on the annual and winter Wild Turkey home range sizes. We estimated the mean annual and mean winter Wild Turkey home range size using 95% kernels and then calculated the diameter of a circle equivalent in area to the average home range size for each time period. We then added this value, in meters, as a buffer to the respective available habitat MCPs (Appendix 2, Table A2.8).
Using t-tests, we explored whether thinning the data had a significant effect on habitat suitability values inside Wild Turkey home ranges and core use areas. If our habitat suitability models accurately predict Wild Turkey probability of presence, then we expected that suitability predictions would be higher in known areas of Wild Turkey space use than in surrounding areas.

Validation of density estimates against survey data
In addition to validating our results against fine-scale GPS tracking data, we also compared our provincial density estimates against density estimates derived for our Peterborough County and Bruce Peninsula study sites using systematic survey techniques.

Peterborough County
We employed three Wild Turkey survey methods during the winter of 2017 (1 December 2017 to 31 March 2018) to detect and estimate Wild Turkey density in a 422 km² region within a smaller survey region in Peterborough County (Fig. 1). Within this region, we employed an aerial survey, a community science survey, and an opportunistic road survey. For a complete description of each survey method and data processing, including replicate detection, see Appendix 1.
We investigated the accuracy of our model-predicted density estimates by comparing them with an estimate we derived for the survey region by combining population counts across survey types while also accounting for replicate observations.

Bruce Peninsula
A comprehensive survey was conducted on the Bruce Peninsula in 2012 during which time most roads along the peninsula were driven numerous times throughout the season, and the locations of all observed turkey flocks were recorded. This included a turkey capture and telemetry program. The Bruce Peninsula is a geographic region on the eastern shore of Lake Huron dividing the lake from Georgian Bay (Fig. 1). The peninsula lends itself to a road survey as a census technique for this species because the roads provide good visibility of the agricultural land and operations on the peninsula: an important resource for contemporary Wild Turkeys (Vander Haegen et al. 1989, Roberts  , Nguyen et al. 2003, Kane et al. 2007, Restani et al. 2009).
Again, we investigated the accuracy of our province-wide density estimates with an estimate of turkey density calculated for a 1140 km² region on the Bruce Peninsula (Fig. 1).

Community-collected data
We received 5846 observations of Wild Turkeys in Ontario from 1 December 2018 to 31 March 2019. The vast majority, 97%, were submitted through eBird, with only 2% submitted through iNaturalist, and 1% received through direct email (Fig. 2a). After filtering to account for potential replicate observations and spatial bias, our thinned dataset included 492 flock locations (Fig. 2b).

Species distribution models: un-thinned data
Model 1 (climate variable: minimum winter temperature), which employed linear, quadratic, product, and threshold feature types with a regularization value of 1.5, yielded a mean test AUC of 0.73 (0.01). Model 2 (climate variable: snow depth), which employed the same feature types and regularization value as model 1, also yielded a mean test AUC of 0.73 (0.01). For both models, variable jackknifing indicated that building density had the strongest contribution to the model (model 1: 67.70%, model 2: 58.24%). For model 1, minimum annual temperature had the next strongest contribution to the model (27.55%), and for model 2, snow depth had the second strongest contribution (38.10%). See Table 1.
For model 1, habitat suitability trended in a positive direction as the minimum temperature increased with a steep decrease in suitability when minimum temperatures dropped below -13°C (Fig. 3a). For model 2, the habitat suitability decreased with snow accumulation beginning at a snow depth of approximately 30 cm (Fig. 4a). For both models, suitability and building density were positively correlated until an apparent density threshold of approximately 250 buildings per pixel when the slope began to flatten (Figs. 3b and 4b).   For model 3, there appeared to be no relationship between habitat suitability and minimum annual temperature until the temperature dropped below -14°C at which point the suitability values also begin to decrease (Fig. 5a). For model 4, the habitat suitability decreased as snow depth increased beginning at a snow depth of approximately 10 cm (Fig. 6a): 20% lower than was predicted using un-thinned data. For both models, the suitability values increased linearly as building density increased until an apparent density threshold of 1500 buildings per pixel at which point the suitability dropped from 0.92 to 0.55 (Figs. 5b and 6b). Following this threshold, building density and habitat suitability were again positively linearly related.

Contemporary range map and density estimates
The models with temperature as the climate variable (1 and 3) reported almost identical AUC scores to models with snow depth as the climate variable (2 and 4). There is more empirical evidence to support the relationship between snow depth and winter turkey survival (Austin and DeGraff 1975, Wunz and Hayden 1975, Porter 1977, Porter et al. 1980, Nguyen et al. 2003, Kane et al. 2007, Gonnerman 2021 than there is to support the relationship between temperature and winter turkey survival (MacDonald et al. 2016). Therefore, we chose to estimate current Wild Turkey density and distribution based on habitat suitability predictions from models 2 and 4 using snow depth as the climate variable (Figs. 7 and 8).
When using un-thinned data and assuming low and high density estimates of 1.261 and 2.896, respectively, our population size estimates ranged from 0 to 40,000 individuals with mean population size estimates ranging from 22,444.24 to 31,898.36 depending on the threshold scenario (Table 2). When using thinned data, our population size estimates again ranged from 0 to 40,000 individuals with mean population size estimates ranging from 24,190.6 to 30,216.02 depending on the threshold scenario (Table 2).
Using un-thinned data and inflating the mean density to 4.02 unsurprisingly resulted in much higher population size estimates ranging from 0 to 120,000 individuals and a mean population size estimate of 51,821.78 ( Table 2).

Validation of SDM against fine-scale tracking data
Mean habitat suitability of available habitat (winter and annual) was higher than inside core-use areas and home ranges in Peterborough County in both 2018 and 2019 (Appendix 3, Fig.  A3.4). In the Bruce Peninsula in 2012, we observed the opposite pattern with higher mean habitat suitability estimates inside coreuse areas and home ranges than in available habitat (Appendix 3, Fig. A3.4).
To account for regional and annual differences, we explored the effect of thinning on habitat suitability estimates for each region and each year separately. The results of our t-test indicate that, in Peterborough County, thinning had no significant effect on the mean habitat suitability values inside known turkey home ranges in 2017, 2018, or 2019 (P > 0.05; Appendix 2, Table A2.7). Thinning also had no significant effect on habitat suitability values inside core use areas, winter available habitat, or annual available habitat in Peterborough County (Appendix 2, Table  A2.7). Our results were consistent on the Bruce Peninsula with thinning having no significant effect on the mean habitat suitability inside Wild Turkey home ranges, core use areas, winter available habitat, or annual available habitat in 2012 (P > 0.05; Appendix 2, Table A2.7).

Validation of density estimates against survey data
In Peterborough County, using a combination of three different survey types, we detected 905 turkeys, resulting in an estimated density of 2.14 individuals per km² within our 422 km² survey area. We detected 53 flocks resulting in a flock density of 0.   Habitat suitability values derived with un-thinned data were much lower in the Bruce Peninsula survey area ranging from 0.09 to 0.63 (mean = 0.40, SD = 0.22; Fig. 1; Niedzielski and Bowman 2016). Our SDM built with thinned data predicted a slightly higher mean value in the Peterborough County survey area (mean = 0.71, SD = 0.06), but a slightly lower mean value in the Bruce Peninsula survey area (mean = 0.36, SD = 0.22) than the SDM built with un-thinned data.

DISCUSSION
We found that community science in combination with MaxEnt modeling could be used to estimate the size and distribution of populations. We found that our SDMs were not sensitive to data thinning because the mean probability of predicted presence did not differ significantly when MaxEnt models were built using thinned versus un-thinned data, nor did the range in population size estimates. We also found that range size estimates derived from SDMs using thinned data were only slightly larger than range size estimates derived from SDMs using un-thinned data ( Table 2). Our study provides some support for using community science in building species distribution models, which might be helpful given that community science has substantial potential as a cost-effective alternative to systematic sampling across large geographic areas (Pimm et al. 2014).
Our population size estimates were unsurprisingly most sensitive to the mean population density value we used to set the low-and high-turkey densities. Population size estimates were much higher when mean turkey density was set to an inflated value of 4.02: the density of Wild Turkeys reported in Alabama and the highest reported across their range.
It is important to note that in maximum entropy modeling, all observations are weighted equally, although the reported flock size in our dataset ranged from 1 to 200 individuals with a mean of 14.14. In the winter in Ontario, Wild Turkeys form large social groups and remain in these groups during daily behaviors such as foraging and roosting (Korschgen 1967, Healy 1992; J. Baici, personal observation). Turkeys are rarely observed alone for extended periods of time during this time of year. Most observations (92.1% of the un-thinned dataset) were of more than one individual, with 20.1% of observations reporting more than 20 individuals. Only 409 observations (6.9% of the un-thinned dataset) reported one individual. However, in these instances, it is likely that a larger group was simply out-of-sight because Wild Turkeys often seek cover in woodlots or behind tall vegetation.
Habitat suitability estimates inside core-use areas and home ranges were higher than in available habitat in the Bruce Peninsula region whereas the opposite pattern was observed in Peterborough County (Appendix 3, Fig. A3.4). Furthermore, our results predicted lower mean habitat suitability in the Bruce Peninsula region than in Peterborough County overall regardless of the modeling scenario. This indicates that there may be evidence of habitat selection occurring within available habitat in the Bruce Peninsula where conditions are harsher and habitat suitability is predicted to be lower overall. In southern populations, where habitat suitability values are higher, Wild Turkeys may not be selecting habitat based on resource availability because resources are more abundant and homogeneous in this region of the province. Furthermore, the regional population surveys estimate that the Wild Turkey population density in Peterborough County is over 10 times the population density in the Bruce Peninsula region.
It is important to note that the Bruce Peninsula population survey was conducted in 2012, whereas the Peterborough County surveys were conducted in 2018. The climate data included in our models were also collected in 2018, forest and agricultural cover were mapped in 2000, roads were mapped in 2013, and buildings were mapped in 2014. We validated our predicted probabilities against annual home range estimates for 2017, 2018, and 2019 even though the climate variables may not have necessarily reflected conditions in 2017 or 2019. The available turkey habitat in both regions was estimated using all available GPS data for the region, regardless of the year they were collected. In reality, the strength and direction of the effect of each of our predictor variables are likely to differ between years. We have modeled the broad-scale relationships between Wild Turkey occurrences and each of our predictor variables, but more work should be done to better understand the inter-annual fluctuations of these dynamic relationships. In addition, we echo Bradsworth et al. (2017) in stating the importance of externally validating SDM predictions using fine-scale tracking data because relatively few examples of this exist (Fonderflick et al. 2015, Bradsworth et al. 2017, Coxen et al. 2017 We exported the results of our SDM following a conditional loglog transformation (clog-log), because this process is derived from the interpretation of MaxEnt as an inhomogeneous poisson process (IPP), which has stronger theoretical justification and allows for more robust interpretations of predicted probabilities (Aarts et al. 2012, Fithian and Hastie 2013, Renner and Warton 2013). An assumption of the IPP is that each observation is independent. Our un-thinned dataset violates this assumption because observations that are close together in geographic space may be replicate observations of the same flock or may be observations of unique individuals that belong to the same flock and are thus influenced by one another's behavior. However, one of the goals of this study was to determine the effect of data cleaning on model outputs. We determined that, in this case, data cleaning to address unknown sampling and spatial bias did not substantially affect model outputs or subsequent analyses.
Our habitat suitability models indicate that in Ontario during the winter, turkeys appear to be positively associated with buildings and negatively associated with deep snow and cold temperatures. These results remain consistent when modeling with thinned and un-thinned data. It is important to note that because our predicted habitat suitability values were derived using observations collected during the winter, our estimates represent those of winter population size and distribution only. We also note however, that Wild Turkeys do not make a seasonal migration, and we expect large-scale distribution and abundance patterns to be quite similar among seasons.
A known challenge associated with employing community science data to estimate species distributions is the unknown spatial bias associated with the data. For instance, in our dataset, regions with a high density of buildings (e.g., residential neighborhoods, city centers, etc.) may represent regions with a higher density of observations because of higher sampling effort (Phillips et al. 2009, Ruiz-Gutiérrez andZipkin 2011). However, residential neighborhoods also contain a high density of bird feeders and supplemental feeding is common at many high-traffic birding and outdoor leisure areas in the province (Jones and Reynolds 2008). Today, Wild Turkeys are heavily reliant on supplemental food sources, such as bird feeders and agricultural operations, during the winter in their northern range (Vander Haegen et al. 1989, Nguyen et al. 2003, Kane et al. 2007, Restani et al. 2009). As such, in this case, any underlying spatial bias in the dataset resulting from supplemental food sources and uneven sampling may assist in identifying regions that represent high-quality Wild Turkey habitat.
Turkeys are considered deciduous forest habitat generalists (Healy 1992, Porter 1992. However, forest cover contributed minimally to our models, with coefficients ranging from 0.3 to 0.89 depending on the modeling scenario. It has been reported that the ideal habitat composition for Wild Turkeys is an equal ratio of forest (primarily deciduous) to open land (Kurzejeski and Lewis 1985) but thriving turkey populations are found in landscapes with much less forest. For instance, some of the highest harvest densities of turkeys in Ontario are found in Wildlife Management Units (WMUs) with < 25% forest cover (OMNR 2007).
We evaluated model performance using the area under the receiver operator curve (AUC). The AUC provides a single measure of overall model accuracy and is not dependent on a predetermined habitat suitability threshold like other model performance evaluation techniques (Deleo 1993, Fielding andBell 1997). This statistic is represented as a value between 0.5 and 1. A value of 0.8 indicates that for 80% of the time, the model will correctly predict Wild Turkey presence and that for 20% of the time, the model will predict Wild Turkey presence where there were no observations. Researchers suggest that an AUC value > 0.75 is necessary for SDMs to accurately model species distribution (Elith 2000, Nazeri et al. 2012). Our AUC values did not quite meet this threshold but were close, with mean test AUCs ranging from 0.723 to 0.73 depending on the climate variables used and whether observations were represented by thinned or un-thinned data. A known limitation of ROCs is that certain algorithms result in a high potential for commission errors: errors that result in false positives (Peterson et al. 2008). As such, one's interpretation of the AUC is dependent upon a subjective understanding of the cost of these false positive errors versus the cost of false negative errors. One of this study's goals was to identify a potential contemporary range map for this species and another was to estimate the current size of Ontario's Wild Turkey population. As such, AUC values < 0.75 may have resulted in an overestimation of the geographic range of this species and the overall size of the population because low AUC values can be associated with higher rates of commission (Peterson et al. 2008). However, Wild Turkeys are a highly adaptable and generalist species, known to thrive in a wide variety of habitats (Healy 1992). Therefore, it is reasonable to assume that geographic regions predicted by the modeling process to contain Wild Turkeys that do not, may represent regions that could be exploited by Wild Turkeys in the future as the population continues to expand their range over time.  (Kurzejeski et al. 1987, Badyaev et al. 1996, Humberg et al. 2009, Niedzielski and Bowman 2014, Gonnerman 2021, meaning that detection ability may differ as well. We were unable to find an empirical field study that reported a Wild Turkey density as high as four Wild Turkeys per km². Thus, although we derived total population size estimates by using a mean Wild Turkey density similar to those reported in this species' most recent management plan (OMNRF 2007), it is unlikely that this density of Wild Turkeys exists in Ontario given that this region represents their northern range limit, and species densities are typically lowest at range limits (Caughley et al. 1988, Hampe andPetit 2005).
This is not the first time that community-collected data have been used to inform the distribution of Wild Turkeys. For instance, Donohoe et al. (1983:189) incorporated "reliable reports from interested citizens" into estimating the occupied range of Wild Turkeys in Ohio, and Ontario's provincial wildlife branch relied on personal interviews with area residents to determine the success of the trap-and-transfer efforts in the years directly following reintroductions (Ontario Government 1985, unpublished manuscript). The reporting platform eBird (2019) also provided relative abundance and geographic range estimates for many species, including Wild Turkeys, using a generalized additive modeling approach. In this case, relative abundance is defined as the count of individuals of a given species detected by an expert eBirder on a one hour, one kilometer traveling checklist at the optimal time of day . This approach estimates a mean relative Wild Turkey abundance of 0.04 for the province of Ontario 2021 (Fink et al. 2022). Among other parameters, models tend to include a subset of variables selected from 60 environmental descriptors and 12 hourly weather variables . This estimate of relative abundance is useful in understanding broad-scale geographic patterns of distribution but may be less informative of regional predictors of density and abundance. For instance, eBird provides year-round relative abundance estimates, however, Wild Turkey behavior differs dramatically with the seasons, and as such, the species detectability differs as well. Our research investigates Wild Turkey distribution and population size during the winter months only because this is when they are easiest to detect in Ontario. Our models also include fewer predictor variables and only include those with evidence from the literature to suggest their usefulness in addressing our research question. For instance, among the 60 environmental descriptors included in eBird's models, there may be numerous variables that are useful for predicting robust estimates of relative abundance across their range but are not necessarily useful for predicting regional Wild Turkey population size.
An informative addition to our community science project could be to request information regarding the presence of supplemental food sources near the turkey flock observations. Anthropogenic food sources (e.g., agricultural operations and residential bird feeders) are difficult to map because they are often located on private property and are ephemeral in nature. Collecting information about the presence of these food sources at the time of the flock observation would allow researchers to investigate this important relationship directly, rather than relying on other anthropogenic variables as proxies. Although we did not explicitly request information about nearby supplemental food sources, 177 flock observations (3%) included notes about nearby residential bird feeders. Based on what we know about the frequency of bird feeding in residential areas (Jones and Reynolds 2008), it is likely that the true proportion of turkey sightings associated with bird feeders is much higher.
Our results demonstrate that community science can be a useful tool in estimating the distribution of reintroduced species across a large geographic scale. For generalist species with high adaptability to anthropogenic landscapes, filtering observations based on sampling effort may not be necessary to develop informative SDMs. Because Wild Turkeys exploit a wide variety of habitat types and food resources, especially anthropogenic food resources like agricultural operations and bird feeders, their detectability among community scientists was high. Furthermore, community scientists most commonly sample in and around anthropogenic landscapes (Tulloch and Szabo 2012), overlapping with this species niche. For other species, like those that have neutral or avoidant relationships with anthropogenic landscapes, thinning the observations based on sampling effort may be a more important step in elucidating the true relationships between species presence and habitat variables.
Future research could involve applying our approach to species with similar distinctive morphologies, high visibility, and seasonal flocking behavior, although there are a variety of factors to consider in identifying appropriate species. For instance, Wild Turkeys do have powered flight, but are primarily observed foraging or walking on the ground, which likely has a positive effect on their detectability. Additionally, virtual species reporting platforms are commonly used in North America, making community science data like Wild Turkey observations readily available and easily accessible. Most of our observations were derived through eBird (97%), which is significantly underutilized in regions like northern Africa and Russia, for instance .
We chose not to include topographic parameters, such as a topographic index because there is little evidence from the literature to suggest that they are important in predicting eastern Wild Turkey habitat suitability. There is also little variation in topography in most of the Wild Turkey's range in Ontario, suggesting that topographic variables are unlikely to inform Wild Turkey ranges within our study area. However, topography can be a limiting factor for Wild Turkey subspecies in habitats with higher topographic variability (Bakner et al. 2022).

Management implications
Our work provides empirical estimates of the size and distribution of Ontario's reintroduced Wild Turkey population. Prior to this, the most recent estimate of Ontario's Wild Turkey population size, published in 2007, was 70,000 individuals (OMNR 2007). However, this estimate was based on limited information and is not informative of how the population may continue to grow and change over time. We present our findings in a provincial context. However, regional densities could also be estimated using this methodology. For instance, in Ontario, regions are designated as Wildlife Management Units (WMUs) and harvest rates are measured and managed within each unit. Estimating turkey densities inside each WMU based on SDM-informed predicted probabilities of presence would allow for finer-scale management (Francis et al. 2009).
In Ontario, spring turkey hunting is assumed to remove approximately 30% of the male population (OMNR 2007). Since 2008, there have generally been ~ 10,000 to 20,000 adult males harvested per year (OMNRF 2021). Researchers in Texas found that the brood sex ratio (BSR) of Wild Turkey populations at two different study sites both skewed male with approximately 56% of sampled eggs being biologically male (Collier et al. 2007). If we assume a similar BSR in Ontario, and that ~ 10,000 to 20,000 individuals represent 30% of the male population, we can infer that Ontario should have a population of ~ 65,000 to 130,000 individuals. None of our modeling scenarios produced population size estimates this large except when estimating population size using an inflated mean density of 4.02 Wild Turkeys per km² (Table 3).
Survival rates are likely to differ between the sexes, which may cause the sex ratio in the larger population to differ from the BSR. There is little published information regarding Wild Turkey survival rates at their northern range limit. One study estimates the mean annual survival rate of hens in Sudbury, Ontario (46°1 0′ 0.0012" N, 80° 25′ 00.0012″ W) to be 0.28 (Nguyen et al. 2003), while another conducted in the municipality of Northern Bruce Peninsula (45° 0′ 0″ N, 81° 19′ 0.0012″ W) estimated the mean annual survival rate of hens to be 0.37 (Niedzielski and Bowman 2014). However, neither study reported the survival rates of males in the same population. This is also true of survival rates estimated in other regions of the Wild Turkey's range. Researchers tend to estimate survival of one sex or the other, making comparisons across the sexes challenging (Keegan and Crawford 1999, Holdstock et al. 2006, Restani et al. 2009). One study conducted on both sexes in northern Indiana did find that there were substantial differences in survival between the two. Researchers estimated mean male and female annual survival rates to be 0.257 and 0.777, respectively (Humberg et al. 2009).
When using a mean density value representative of the average density across the Wild Turkey's USA range, we estimated an average of fewer than 30,000 individuals in Ontario (Table 3). Only when our mean was inflated to match the highest recorded density across their range did we estimate population sizes greater than 40,000 (Table 3). These results indicate that some regions in Ontario may exhibit a higher density of Wild Turkeys than expected based on densities reported for other geographic regions or previously for Ontario. Turkey populations can remain stable when up to 35% of the male population is harvested during the spring season (Vangilder 1997). However, our findings indicate that more than 30% of the male population may be being harvested during the spring season, particularly during the last several harvest years. This is a notable finding given the historical context of Wild Turkeys in Ontario. We recommend that the distribution, abundance, and harvest pressure of Wild Turkeys in Ontario continue to be monitored to gain a better understanding of their anthropogenic-centered habitat selection patterns and long-term population trends.
Ongoing research efforts should continue to work toward a comprehensive understanding of Wild Turkey population density and distribution across the province while considering any habitat changes that are likely to occur because of climate change. Special attention should be paid to regions experiencing any changes in harvest pressure. Wild Turkey populations are most sensitive to hen survival and recruitment . However, established populations can fluctuate in size annually by as much as 50% of the long-term mean (Mosby 1967). So, monitoring changes over time is important because population densities are unlikely to remain constant.

Author Contributions:
Both authors contributed equally to the manuscript. manuscript. This work would not be possible without the cooperation of several private landowners who provided access to trap and track Wild Turkeys throughout the study period.