Our study was conducted on the Albemarle Peninsula in the northeastern region of North Carolina (Fig 1 ). The study area included approximately 6,000 km2 of federal, state, and private lands comprising a row-crop agricultural-bottomland forest matrix with little change in elevation (<50 m). Agricultural crops (i.e., corn, cotton, soybean, and winter wheat) and managed pine (Pinus spp.) composed of approximately 30% and 15% of the land cover, respectively. Other prominent land-cover types were coastal bottomland forests and pocosin (peatlands with a low [1–4 m] and dense evergreen shrub layer; 35%), herbaceous wetlands and saltwater marshes (5%), open water (5%), and other minor land-cover types (10%). The climate was typical of the mid-Atlantic: 4 distinct seasons, nearly equal in length, with an annual precipitation averaging between 122 to 132 cm. Summer climate was typically hot and humid with daily temperatures ranging from 27°C to over 38°C and winters were relatively cool with daily temperatures ranging between -4° to 7° C.
As part of long-term monitoring and management of red wolves and coyotes on the Albemarle Peninsula, the Recovery Program conducted annual trapping during autumn and winter to capture and fit individual red wolves and coyotes with radio collars. Our field study assisted annual trapping efforts from 2009 through 2011 to capture coyotes and red wolves. Coyotes were not a listed or protected species and the permitting authority for their capture and release was the North Carolina Wildlife Resources Commission. However, red wolves were listed as critically endangered by the International Union Conservation of Nature’s (IUCN) red list of threatened species and we operated under a cooperative agreement with the USFWS that permitted us to trap under special handling permits issued to the Recovery Program to trap and handle red wolves. This study, including all animal handling methods, was approved by the Louisiana State University Agricultural Center Institutional Animal Care and Use Committee (Protocol Number AE2009-19) and meets the guidelines recommended by the American Society of Mammologists [26 (link)]. Permission to access private lands for trapping occurred under memorandum of agreements (MOAs) between individual landowners and the Recovery Program. We access private lands of landowners without existing MOAs by contacting those individuals to receive permission to trap their lands.
We captured coyotes using padded foot-hold traps (Victor no.3 Softcatch, Woodstream Corporation, Lititz, Pennsylvania, USA) from October through May, 2009–2011. Coyotes were typically restrained using a catchpole, muzzle, and hobbles. Although most coyotes were not anesthetized, several were chemically immobilized with an intramuscular injection of ketamine HCl and xylazine HCl to inspect inside the mouth for injuries. Coyotes were sexed, measured, weighed, and aged by tooth wear [27 ], and a blood sample was collected. We categorized coyotes >2 years old as adults, 1–2 years old as juveniles, and <1 year old as pups. Coyotes on the Albemarle Peninsula were reproductively sterilized by the USFWS to prevent introgression into the red wolf population [24 ,25 (link)]. Coyotes were taken to a local veterinary clinic for surgical sterilization where males and females were reproductively sterilized by vasectomy and tubal ligation, respectively. This process keeps hormonal systems intact to avoid disrupting breeding and territorial behavior [28 (link),29 (link)]. Prior to release at the original capture sites, we fit coyotes with a mortality-sensitive GPS radio collar (Lotek 3300s, Newmarket, Ontario, Canada) scheduled to record a location every 4 hours (0:00, 04:00, 08:00, and so on) throughout the year.
The Recovery Program monitored radio-collared red wolves and coyotes 2 times a week from aircraft to identify red wolf and coyote territories on the Albemarle Peninsula. Resident pairs of coyotes were identified as radio-collared individuals of breeding age (≥2 years old) who were temporally and spatially associated with one another and defending a territory for ≥4 months. When trapping was not feasible after radio-collared coyotes established territories, we confirmed the presence of a mate via field inspection for sign (i.e., visual observations and tracks) of another individual over the course of several weeks. To avoid autocorrelation, we only fit one coyote in each pair of residents with a GPS radio-collar. We classified radio-collared coyotes as transients when they were solitary and not associated with other radio-collared coyotes and displayed extensive movements throughout the Albemarle Peninsula.
To reflect the anthropogenic effects of agricultural practices on the landscape, we divided each year into 2 6-month seasons based on agricultural activity: growing (1 March–31 August) and harvest (1 September–28 February). We estimated space use of resident and transient coyotes by fitting dynamic Brownian bridge movement models (dBBMMs) to the time-specific location data to estimate the probability of use along the full movement track of each coyote [30 ], using R package moveud [31 ] in Program R [32 ]. Brownian bridge movement models use characteristics of an animal’s movement path among successive locations to develop a utilization distribution of an animal’s range. Because many factors influence telemetry error and recent studies suggest telemetry error for GPS radio collars range between 10–30 m [33 (link)], we used an error estimate of 20 m for all locations. Our error estimate was calculated based on recommendations and assumptions outlined in Byrne et al. [34 (link)]; we chose a moving window size of 7 locations (equivalent to 14 hours) with a margin of 3 locations for full tracks of each animal to reflect temporal shifts in coyote movements related to photoperiods. For residents, we considered 95% and 50% contour intervals as home ranges and core areas, respectively. Because transients do not maintain and defend territories, we did not refer to transient space use as home ranges and core areas. Instead, we considered 95% and 50% contour intervals for transients as transient ranges and biding areas [20 (link)], respectively. We used t-tests to investigate changes in the area of space use among seasons.
We estimated predominant landscape features from a digitized
landscape map of vegetative communities developed by the North Carolina Gap Analysis Project [35 ]. We collapsed vegetative communities estimated by McKerrow et al. [35 ] into 4 general habitat classes with a 30-m resolution. For the habitat selection analysis, we divided the landscape into agriculture, coastal bottomland forest, pine forest, and wetlands (e.g., herbaceous wetlands, marshes, and pocosin). Because coyotes are known to use roads and forage along edges, we also developed road and agricultural-forest edge layers [36 (link)]. We created distance raster maps for habitat classes, roads, and agricultural-forest edges (hereafter edges) using the ‘Euclidean Distance’ tool in the Spatial Analyst toolbox in (ArcGIS 10; Environmental Systems Research Institute Inc., Redlands, California) to calculate the distance from every 30 m pixel to the closest landscape feature [37 (link), 38 (link)]. We used analysis of variance (ANOVA) and Tukey tests [39 ] for multiple comparisons to determine if habitat composition of home ranges, core areas, transient ranges, and biding areas differed.
We used RSFs to examine relationships between landscape features and coyote establishment of home ranges on the landscape (2nd-order selection) [40 (link)] and to examine relationships between landscape features and coyote use within their home ranges (3rd-order selection) following Design II and III approaches suggested by Manly et al. [41 ]. For 2nd-order selection, we used individual animals as our sampling units and measured resource availability at the population level. For 3rd-order selection, we used individual animals as our sampling units and resource availability was measured for each animal. Despite the presence of territorial red wolves on the Albemarle Peninsula and active management by the Recovery Program to reduce red wolf-coyote hybridization, coyotes were found throughout the entire peninsula. We used distance-based variables to assess habitat selection to eliminate the need to base inference on subjectively chosen reference categories [37 (link)]. Therefore, we inferred “selection” when known (used) locations were closer to resource features than were random (available) locations and “avoidance” was inferred when known locations were farther from resource features than random locations. We used a binomial approach to estimate resource-selection functions by comparing characteristics of known locations to an equal number of random locations within the Albemarle Peninsula study area (2nd-order selection) and within home ranges and transient ranges (3rd-order selection) of coyotes [41 ]. We used generalized linear mixed models with a logistic link to compare habitat selection between resident and transient coyotes. We included random intercepts for individual coyotes in each model to account for correlation of habitat use within individuals and the unbalanced telemetry data. We modeled resource selection using the R package ‘lme4’ [42 ] with a binary (0 = available, 1 = used) response variable. Prior to modeling, we rescaled values for all distance-based variables by subtracting their mean and dividing by 2 standard deviations [38 (link),43 (link)].
We designed 5 candidate models for coyote occurrence guided by 4 a priori general hypotheses to develop RSFs: (1) Coyotes require cover and shelter found primarily in forests. (2) Coyotes favor linear landscape characteristics, such as edges and roads. (3) Coyotes prefer open, treeless habitats, such as agricultural fields. (4) Coyotes avoid wetland habitats. We used an information-theoretic approach to assess models by calculating Akaike’s information criterion for small sample sizes (AICc) [44 ,45 (link)] and used ΔAICc to select which models best supported habitat selection. First, we used all resident and transient locations from our telemetry data, included main effects for all fixed predictor variables, and considered interactions between a coyote status variable (resident = 1, transient = 0) and each landscape feature variable to investigate potential differences in selection between resident and transient coyotes. Second, we subsetted resident and transient locations and constructed separate models to derive 2nd- and 3rd-order selection coefficients for each landscape feature without interactions. We included all landscape features described above in our global models sets because correlation between individual predictor variables was low or modest (all r < 48%).We conducted model validation of the best model using k-fold cross-validation and then tested for predictive performance using area under the curve (AUC) [46 (link)–49 (link)]. This cross-validation is based on partitioning the data into k bins and performing k iterations of training and validation in which a different bin of the data is held out for validation, while remaining k–1 bins are used for the training set. We used 10 folds (k = 10) to estimate performance of RSF models. Area under the curve of a receiver operating characteristic (ROC) curve represents the relative proportions of correctly and incorrectly classified predictions over a range of threshold levels by plotting true positives versus false positives for a binary classifier system.
As part of long-term monitoring and management of red wolves and coyotes on the Albemarle Peninsula, the Recovery Program conducted annual trapping during autumn and winter to capture and fit individual red wolves and coyotes with radio collars. Our field study assisted annual trapping efforts from 2009 through 2011 to capture coyotes and red wolves. Coyotes were not a listed or protected species and the permitting authority for their capture and release was the North Carolina Wildlife Resources Commission. However, red wolves were listed as critically endangered by the International Union Conservation of Nature’s (IUCN) red list of threatened species and we operated under a cooperative agreement with the USFWS that permitted us to trap under special handling permits issued to the Recovery Program to trap and handle red wolves. This study, including all animal handling methods, was approved by the Louisiana State University Agricultural Center Institutional Animal Care and Use Committee (Protocol Number AE2009-19) and meets the guidelines recommended by the American Society of Mammologists [26 (link)]. Permission to access private lands for trapping occurred under memorandum of agreements (MOAs) between individual landowners and the Recovery Program. We access private lands of landowners without existing MOAs by contacting those individuals to receive permission to trap their lands.
We captured coyotes using padded foot-hold traps (Victor no.3 Softcatch, Woodstream Corporation, Lititz, Pennsylvania, USA) from October through May, 2009–2011. Coyotes were typically restrained using a catchpole, muzzle, and hobbles. Although most coyotes were not anesthetized, several were chemically immobilized with an intramuscular injection of ketamine HCl and xylazine HCl to inspect inside the mouth for injuries. Coyotes were sexed, measured, weighed, and aged by tooth wear [27 ], and a blood sample was collected. We categorized coyotes >2 years old as adults, 1–2 years old as juveniles, and <1 year old as pups. Coyotes on the Albemarle Peninsula were reproductively sterilized by the USFWS to prevent introgression into the red wolf population [24 ,25 (link)]. Coyotes were taken to a local veterinary clinic for surgical sterilization where males and females were reproductively sterilized by vasectomy and tubal ligation, respectively. This process keeps hormonal systems intact to avoid disrupting breeding and territorial behavior [28 (link),29 (link)]. Prior to release at the original capture sites, we fit coyotes with a mortality-sensitive GPS radio collar (Lotek 3300s, Newmarket, Ontario, Canada) scheduled to record a location every 4 hours (0:00, 04:00, 08:00, and so on) throughout the year.
The Recovery Program monitored radio-collared red wolves and coyotes 2 times a week from aircraft to identify red wolf and coyote territories on the Albemarle Peninsula. Resident pairs of coyotes were identified as radio-collared individuals of breeding age (≥2 years old) who were temporally and spatially associated with one another and defending a territory for ≥4 months. When trapping was not feasible after radio-collared coyotes established territories, we confirmed the presence of a mate via field inspection for sign (i.e., visual observations and tracks) of another individual over the course of several weeks. To avoid autocorrelation, we only fit one coyote in each pair of residents with a GPS radio-collar. We classified radio-collared coyotes as transients when they were solitary and not associated with other radio-collared coyotes and displayed extensive movements throughout the Albemarle Peninsula.
To reflect the anthropogenic effects of agricultural practices on the landscape, we divided each year into 2 6-month seasons based on agricultural activity: growing (1 March–31 August) and harvest (1 September–28 February). We estimated space use of resident and transient coyotes by fitting dynamic Brownian bridge movement models (dBBMMs) to the time-specific location data to estimate the probability of use along the full movement track of each coyote [30 ], using R package moveud [31 ] in Program R [32 ]. Brownian bridge movement models use characteristics of an animal’s movement path among successive locations to develop a utilization distribution of an animal’s range. Because many factors influence telemetry error and recent studies suggest telemetry error for GPS radio collars range between 10–30 m [33 (link)], we used an error estimate of 20 m for all locations. Our error estimate was calculated based on recommendations and assumptions outlined in Byrne et al. [34 (link)]; we chose a moving window size of 7 locations (equivalent to 14 hours) with a margin of 3 locations for full tracks of each animal to reflect temporal shifts in coyote movements related to photoperiods. For residents, we considered 95% and 50% contour intervals as home ranges and core areas, respectively. Because transients do not maintain and defend territories, we did not refer to transient space use as home ranges and core areas. Instead, we considered 95% and 50% contour intervals for transients as transient ranges and biding areas [20 (link)], respectively. We used t-tests to investigate changes in the area of space use among seasons.
We estimated predominant landscape features from a digitized
landscape map of vegetative communities developed by the North Carolina Gap Analysis Project [35 ]. We collapsed vegetative communities estimated by McKerrow et al. [35 ] into 4 general habitat classes with a 30-m resolution. For the habitat selection analysis, we divided the landscape into agriculture, coastal bottomland forest, pine forest, and wetlands (e.g., herbaceous wetlands, marshes, and pocosin). Because coyotes are known to use roads and forage along edges, we also developed road and agricultural-forest edge layers [36 (link)]. We created distance raster maps for habitat classes, roads, and agricultural-forest edges (hereafter edges) using the ‘Euclidean Distance’ tool in the Spatial Analyst toolbox in (ArcGIS 10; Environmental Systems Research Institute Inc., Redlands, California) to calculate the distance from every 30 m pixel to the closest landscape feature [37 (link), 38 (link)]. We used analysis of variance (ANOVA) and Tukey tests [39 ] for multiple comparisons to determine if habitat composition of home ranges, core areas, transient ranges, and biding areas differed.
We used RSFs to examine relationships between landscape features and coyote establishment of home ranges on the landscape (2nd-order selection) [40 (link)] and to examine relationships between landscape features and coyote use within their home ranges (3rd-order selection) following Design II and III approaches suggested by Manly et al. [41 ]. For 2nd-order selection, we used individual animals as our sampling units and measured resource availability at the population level. For 3rd-order selection, we used individual animals as our sampling units and resource availability was measured for each animal. Despite the presence of territorial red wolves on the Albemarle Peninsula and active management by the Recovery Program to reduce red wolf-coyote hybridization, coyotes were found throughout the entire peninsula. We used distance-based variables to assess habitat selection to eliminate the need to base inference on subjectively chosen reference categories [37 (link)]. Therefore, we inferred “selection” when known (used) locations were closer to resource features than were random (available) locations and “avoidance” was inferred when known locations were farther from resource features than random locations. We used a binomial approach to estimate resource-selection functions by comparing characteristics of known locations to an equal number of random locations within the Albemarle Peninsula study area (2nd-order selection) and within home ranges and transient ranges (3rd-order selection) of coyotes [41 ]. We used generalized linear mixed models with a logistic link to compare habitat selection between resident and transient coyotes. We included random intercepts for individual coyotes in each model to account for correlation of habitat use within individuals and the unbalanced telemetry data. We modeled resource selection using the R package ‘lme4’ [42 ] with a binary (0 = available, 1 = used) response variable. Prior to modeling, we rescaled values for all distance-based variables by subtracting their mean and dividing by 2 standard deviations [38 (link),43 (link)].
We designed 5 candidate models for coyote occurrence guided by 4 a priori general hypotheses to develop RSFs: (1) Coyotes require cover and shelter found primarily in forests. (2) Coyotes favor linear landscape characteristics, such as edges and roads. (3) Coyotes prefer open, treeless habitats, such as agricultural fields. (4) Coyotes avoid wetland habitats. We used an information-theoretic approach to assess models by calculating Akaike’s information criterion for small sample sizes (AICc) [44 ,45 (link)] and used ΔAICc to select which models best supported habitat selection. First, we used all resident and transient locations from our telemetry data, included main effects for all fixed predictor variables, and considered interactions between a coyote status variable (resident = 1, transient = 0) and each landscape feature variable to investigate potential differences in selection between resident and transient coyotes. Second, we subsetted resident and transient locations and constructed separate models to derive 2nd- and 3rd-order selection coefficients for each landscape feature without interactions. We included all landscape features described above in our global models sets because correlation between individual predictor variables was low or modest (all r < 48%).We conducted model validation of the best model using k-fold cross-validation and then tested for predictive performance using area under the curve (AUC) [46 (link)–49 (link)]. This cross-validation is based on partitioning the data into k bins and performing k iterations of training and validation in which a different bin of the data is held out for validation, while remaining k–1 bins are used for the training set. We used 10 folds (k = 10) to estimate performance of RSF models. Area under the curve of a receiver operating characteristic (ROC) curve represents the relative proportions of correctly and incorrectly classified predictions over a range of threshold levels by plotting true positives versus false positives for a binary classifier system.
Full text: Click here