Using dynamic Brownian bridge movement ... - Semantic Scholar

5 downloads 0 Views 710KB Size Report
Rio Grande wild turkey (Meleagris gallopavo intermedia) initiated laying as well as when dur- ing the day ... Journal of Animal Ecology 2014, 83, 1234–1243.
Journal of Animal Ecology 2014, 83, 1234–1243

doi: 10.1111/1365-2656.12205

Using dynamic Brownian bridge movement modelling to measure temporal patterns of habitat selection Michael E. Byrne1*, J. Clint McCoy2, Joseph W. Hinton1, Michael J. Chamberlain1 and Bret A. Collier3 1

Warnell School of Forestry and Natural Resources, University of Georgia, Athens, GA 30602, USA; 2School of Forestry and Wildlife Sciences, Auburn University, Auburn, AL 36849, USA; and 3School of Renewable Natural Resources, Louisiana State University, Baton Rouge, LA 70803, USA

Summary 1. Accurately describing animal space use is vital to understanding how wildlife use habitat. Improvements in GPS technology continue to facilitate collection of telemetry data at high spatial and temporal resolutions. Application of the recently introduced dynamic Brownian bridge movement model (dBBMM) to such data is promising as the method explicitly incorporates the behavioural heterogeneity of a movement path into the estimated utilization distribution (UD). 2. Utilization distributions defining space use are normally estimated for time-scales ranging from weeks to months, obscuring much of the fine-scale information available from high-volume GPS data sets. By accounting for movement heterogeneity, the dBBMM provides a rigorous, behaviourally based estimate of space use between each set of relocations. Focusing on UDs generated between individual sets of locations allows us to quantify fine-scale circadian variation in habitat use. 3. We used the dBBMM to estimate UDs bounding individual time steps for three terrestrial species with different life histories to illustrate how the method can be used to identify fine-scale variations in habitat use. We also demonstrate how dBBMMs can be used to characterize circadian patterns of habitat selection and link fine-scale patterns of habitat use to behaviour. 4. We observed circadian patterns of habitat use that varied seasonally for a white-tailed deer (Odocoileus virginianus) and coyote (Canis latrans). We found seasonal patterns in selection by the white-tailed deer and were able to link use of conifer forests and agricultural fields to behavioural state of the coyote. Additionally, we were able to quantify the date in which a Rio Grande wild turkey (Meleagris gallopavo intermedia) initiated laying as well as when during the day, she was most likely to visit the nest site to deposit eggs. 5. The ability to quantify circadian patterns of habitat use may have important implications for research and management of wildlife. Additionally, the ability to link such patterns to behaviour may aid in the development of mechanistic models of habitat selection. Key-words: animal movement, behaviour, Brownian bridge movement model, GPS, habitat, space use, utilization distribution

Introduction Quantifying space use represents a vital component in animal habitat selection studies and provides the foundation for a range of applications important to research and management. A wide variety of approaches have been developed to quantify space use, increasing in sophistication *Correspondence author. E-mail: [email protected]

and utility concomitant with technological and computational advances. Many current methods centre on calculation of utilization distributions (UDs) which can be used to estimate animal home ranges (Burt 1943) and quantify intensity of use within a landscape. Traditionally, UDs were most commonly estimated using kernel-based methods (Worton 1989). However, kernels may perform poorly on data sets with large sample sizes when common methods of determining the smoothing parameter (h) are used

© 2014 The Authors. Journal of Animal Ecology © 2014 British Ecological Society

Temporal patterns of habitat selection (Getz & Wilmers 2004; Hemson et al. 2005) and do not account for temporal structure of locational data. The advent of GPS tracking increased both accuracy and acquisition of locational data, thus providing great detail on the movements of tracked animals. Analyses that make use of combined spatial and temporal information have the potential to provide more precise estimates of space use and habitats sampled. The Brownian bridge movement model (BBMM) introduced by Horne et al. (2007) represents an improvement over other UD methods by explicitly basing UD estimation of an animal’s range on characteristics of the movement path. The driving parameter is the Brownian motion variance (r2m ), which is a measure of how irregular the path of an animal is between successive locations. The BBMM described by Horne et al. (2007) assumes a constant r2m along an entire movement path. However, animals are known to transition between a number of different movement behaviours over time (Morales et al. 2004; Jonsen, Flemming & Myers 2005; Gurarie, Andrews & Laidre 2009; McClintock et al. 2012). Thus, Kranstauber et al. (2012) introduced an improvement to the BBMM, termed the dynamic Brownian bridge movement model (dBBMM), which allows r2m to vary along a path in response to changes in the underlying behaviour of the animal. To detect shifts in movement behaviour, an adjustment to the behavioural change point analysis introduced by Gurarie, Andrews & Laidre (2009) is implemented via a sliding window along the path, producing multiple estimates of r2m for each time step, which are then averaged to produce a final, independent r2m for each path step (Kranstauber et al. 2012). The ability of r2m to vary along the path allows for more accurate UD estimates and provides a measure of behavioural change, related to movement trajectory, along a movement path. Typically, UD models are used to define a home range or core-use area based on aggregated locations collected over the course of weeks or months. However, habitat use may change through time as an animal responds to biological needs and transitions between activities such as resting, foraging, mating, predator evasion and thermoregulation. Thus, common application of UDs fails to capture circadian variation in habitat selection. Because the dBBMM uses values of r2m associated with each time step, the resulting UD bounding each individual time step provides an estimate of space used between sequential relocations. Using the individual UDs bounding, each time step can offer insight into the temporal structure of habitat use and thus provide a means to extract more information from GPS-derived telemetry data sets. Here, we illustrate how using UD estimates for individual time steps based on a dBBMM provides information on circadian variation in habitat use by applying our approach to three terrestrial animals with distinctly different life histories: a white-tailed deer (Odocoileus virginianus), an eastern coyote (Canis latrans) and a Rio Grande wild turkey (Meleagris gallopavo intermedia). For the deer, we

1235

compare circadian patterns of habitat use during breeding and non-breeding periods and illustrate how comparing time-indexed measures of habitat use to habitat availability within ranges can be used to infer fine-scale temporal variation in habitat selection. For the coyote, we compare circadian patterns of habitat use between a winter and summer period, as well as incorporate measures of r2m to elucidate relationships between habitat use and behaviour during each period. For the wild turkey, we identify movement processes relative to nest site selection and egg laying behaviour based on female movement patterns.

Materials and methods white-tailed deer As part of a larger study evaluating movement ecology and harvest management of white-tailed deer, we captured and placed GPS collars on 37 adult male white-tailed deer during Fall of 2009 (n = 14), 2010 (n = 13) and 2011 (n = 10). The study area was Brosnan Forest, a privately owned, 5830 ha tract of lower coastal plains forest habitat near Dorchester, South Carolina, USA and is managed intensively for timber production and recreational hunting. Vegetation on Brosnan Forest was comprised primarily of interspersed stands of mature longleaf pine (Pinus palistris), bottomland hardwood drains and mixed pine–hardwoods. Habitat types included natural pine (mature stands of longleaf pine), planted pine (actively managed and harvested stands of loblolly pine [Pinus taeda]), creek hardwood (stands of mixed hardwoods along creek drainages), planted hardwood (5to 7-year-old stands planted in either cherrybark oak [Quercus pagoda] or shumard oak [Q. shumardii]) and numerous small ( x = 11 ha) wildlife food plots interspersed within the forest matrix. We used a combination of aerial imagery, ground truthing and stand-level information from the property’s certified forester to create a digital land cover map of the area by delineating habitat types in ArcGIS 10 (ESRI, California, USA). Deer hunting season on Brosnan Forest was open from 15 August through 1 January. For our example, we used movement data from one adult (>25 years old) male white-tailed deer obtained from a programmable drop-off GPS collar (Advanced Telemetry Systems, Minnesota, USA). From August 23–November 23 of each year, we programmed collars to record 48 locations/day at half hour intervals (0:00, 0:30, 1:00, etc.). Based on a > 15 year history of foetal ageing on Brosnan Forest as part of the deer population management plan, we estimated that >80% of conceptions occurred between September 25 and October 25. Thus, we categorized our study period into three relatively equal time periods and compared circadian patterns in habitat use across seasons: pre-breeding (August 23 – September 24), breeding (September. 25 – October 25) and post-breeding (October 26 – November 23).

coyote As part of a study evaluating the spatial interactions between coyotes and red wolves (Canis rufus) in North Carolina, we captured and fit 41 coyotes with GPS collars during 2009 – 2011. The Red Wolf Recovery Area is located in the coastal plain region of northeast North Carolina and was an intensively

© 2014 The Authors. Journal of Animal Ecology © 2014 British Ecological Society, Journal of Animal Ecology, 83, 1234–1243

1236 M. E. Byrne et al. farmed agricultural matrix, with c. 30% of the area in row crops and c. 15% of the area consisting of managed pine plantations. Other habitat types included non-forested wetlands, non-riverine bottomland hardwood forests and mixed upland forests. We created a digital land cover image of the area by condensing the North Carolina GAP land cover map (McKerrow, Williams & Collazo 2006) into eight habitat types: roads, agricultural fields, conifer forests (primarily managed pine plantations), bottomland forests (seasonally wet hardwood forests), upland forests (upland hardwood and mixed forest), wetlands, open water and residential/urban. For our example, we used movement data from an adult female coyote tagged with a Lotek 3300s GPS collar (Lotek, Ontario, Canada) in 2009. The collar was programmed to collect one location every 2 h (0:00, 2:00, 4:00, etc.) throughout the year. For the purpose of comparing daily variation in habitat use between the summer growing season and the winter period, we extracted the movement data for the months of July and December 2009.

rio grande wild turkey Information for identifying pre-incubation movements and habitat requirements for gallinaceous birds is limited in the ecological literature. For wild turkeys, research suggests that during preincubation, movements of females represent habitat sampling directed towards locating and selecting an optimal nest location (Badyaev, Martin & Etges 1996; Chamberlain & Leopold 2000) However, the spatial-temporal process associated with pre-nesting search behaviours is largely unknown (Chamberlain & Leopold 2000), and current dogma has been questioned (Collier & Chamberlain 2011). Using data collected as part of an ongoing research project focused on identifying drivers of temporal variation in habitat selection of Rio Grande wild turkeys in Texas, we used data collected from a female Rio Grande wild turkey to show how using a dBBMM can be used to delineate pre-nesting and nesting behaviours. We conducted our work on the MT7 ranch, a 5000 ha privately owned property located approximately 10 km east of Breckenridge, Texas, USA. MT7 is managed actively for wildlife, with a primary focus of native grassland restoration and riparian corridor development and maintenance. Management actions on MT7 range from prescribed fire and mesquite (Prosopis glandulosa) control with high-intensity, short-duration seasonal cattle grazing. Movement data from a female Rio Grande wild turkey were collected using a backpack style GPS unit (Guthrie et al. 2010) programmed to record locations every 30 min from 06:00 to 19:00 daily, with four additional locations taken at 2-h intervals each night from 30 March 2011 to 16 May 2011.

here but encourage readers to consult Kranstauber et al. (2012) for details. Within each window, r2m is estimated via maximum likelihood for the entire path within the window. Then, the path within each window is iteratively split into two sections at each location within the window, and r2m estimated for each of the two path sections. Bayesian information criterion (BIC) is used to compare the fit of the model that estimated a single r2m value within the window and all models that estimated two r2m values for each sectioning of the path within the window. If the lowest BIC value is associated with the model consisting of a single r2m estimate, then that r2m is assigned to all steps between the margins within the window. If however the lowest BIC value is associated with a model consisting of two r2m values, then that is considered evidence of a behavioural change, and the appropriate r2m values are assigned to the steps on each side of the break point. As the window moves along the animal’s trajectory, individual steps are assigned a new r2m value for all the times in which they fall within the window. These r2m estimates are averaged to obtain a final r2m for each step. This requires the researcher to define size of the window. Because r2m is estimated using the leave-one-out method, the window needs to encompass an odd number of locations. Additionally, a margin of ≥3 locations on each end of the window in which no break points can occur must be specified by the researcher. After a r2m value has been obtained for each step, UDs can be estimated for any portion of the track as described by Horne et al. (2007). To fit the dBBMM to the white-tailed deer track, we specified a moving window size of 13 (equivalent of 65 h) and a margin of 3, for the coyote a window size of 7 (equivalent of 14 h) with a margin of 3, and for the Rio Grande turkey, a window size of 21 (equivalent to 105 h) and a margin of 5. Following the recommendations of Kranstauber et al. (2012), we chose window sizes based on the temporal resolution of each track and our a priori assumptions of the time-scale of major behavioural shifts. We then used the R package moveud (Collier 2013) to create UD contours indexed to individual time steps based on the r2m estimate for each step derived from the fitted dBBMM (Fig. 1). We intersected the 50% UD contour for each time step of the white-tailed deer and the coyote with the digital land cover of their respective study areas in ArcGIS 10. We indexed each time step by the time of the first location in each pair of locations and

data analysis We fit a dBBMM to the full movement track of each animal using the R package move (Kranstauber & Smolla 2013) in the R statistical computing environment (R Core Team 2013). Similar to the standard BBMM, the dBBMM requires a time-stamped series of animal locations and the estimated telemetry error associated with each location. We used an error estimate of 20 m for all locations. The dBBMM estimates Brownian motion variance (r2m ) by moving a sliding window encompassing n number of locations along a path. We provide a brief synopsis of the process

Fig. 1. Twelve-hour portion of the movement path of a male white-tailed deer in South Carolina with locations taken at 30min intervals. Each time step is bounded by the 50% UD contour as estimated using the dBBMM. Each step is colour coded based on its associated estimated Brownian motion variance parameter (r2m ) scaled from blue (low) to red (high). This figure illustrates how behavioural state influences estimates of space use for individual time steps.

© 2014 The Authors. Journal of Animal Ecology © 2014 British Ecological Society, Journal of Animal Ecology, 83, 1234–1243

Temporal patterns of habitat selection calculated the mean proportion of each habitat type contained within all 50% UD contours for each 30-min time period for the deer (00:00–0:30, 0 0:30–01:00, etc.) and 2-h time period for the coyote (00:00–02:00, 04:00–06:00, etc.). Missed GPS fixes result in longer time steps; as such for purposes of making consistent temporal comparisons, we limited our analysis only to time steps bounded by locations collected at scheduled intervals (30-min and 2-h intervals, respectively). This allowed us to compare variation in habitat use during the diel cycle between seasons. There was no evidence of habitat-induced fix-rate bias (Frair et al. 2004) in either species. To provide a comparison between diel variation in habitat use and the single point estimate of seasonal habitat use as would be obtained in a typical home range analysis, we used the dBBMM to estimate the 95% UD for the deer each season (pre-breeding, breeding and post-breeding, respectively) and calculated the proportion of three habitat types available within each seasonal home range: natural pines, food plots and creek hardwoods. By considering the proportion of each habitat within home ranges as available, and considering the mean proportion of each habitat within 50% contours for each time period as used, we may draw inferences on how habitat selection changes during the day. To illustrate, we investigated temporal patterns of habitat selection for mature pines seasonally by calculating a selection ratio of the mean proportional use for each 30-min time period and the proportion available within the home range. Selection ratios 1 indicate selection for natural pines; the strength of selection can be inferred from how much the ratio deviates from 1. The Brownian motion variance provides a measure of the animal’s behaviour with higher values being associated with irregular paths and/or increased activity and lower values being associated with more regular paths and/or decreased activity (Kranstauber et al. 2012). Therefore, a relationship between r2m and habitat can indicate a relationship between behaviour and habitat type; for instance, if the use of a particular habitat type is positively associated with r2m , then it could be interpreted that use of that habitat type is associated with an active behavioural state. Thus, we investigated the relationship between habitat use and coyote behaviour. To accomplish this, we first calculated the mean r2m for time steps in each 2-h time period. For each season, we examined the correlation between the mean r2m and the mean proportional habitat use of agricultural fields and conifer forests (the two most used habitat types) for each time period. For the wild turkey, our goals were to quantify at what point during the pre-incubation period, the turkey first visited the nest site, relative amount of time each day the turkey spent in the vicinity of the nest prior to the onset of continuous incubation, and times of each day in which the turkey was most likely to visit the nest prior to continuous incubation. Because we were concerned specifically with pre-incubation behaviour for this example, we limited our analysis to the time period from 30 March 2011 through 19 April 2011 (the first day of continuous incubation as identified through daily VHF-telemetry tracking). We extracted the 50% and 95% UD contours bounding each 30-min time step from 06:00 to 19:00 (06:00–06:30, 06:30–07:00, etc.). We considered any case in which the nest fell within the bounds of the 95% contours for any step as evidence that the turkey could have visited the nest within that 30-min time period. We assume that the probability of the turkey visiting the nest was greater if the nest fell within any 50% UD contour, as the 50% UD

1237

contours bound the locations and the straight-line path between locations more closely than the 95% contours. For each day, we quantified the number of time steps (30-min periods) in which the nest was located within the 50 and 95% UD contours, respectively. To quantify the times of day the turkey was most often in the nest vicinity, we created a probability distribution function based on the frequency of times in which the nest was located within the 95% contour bounding the time step for each 30-min time period from 30 March 2011 to 18 April 2011. All data and R code to create bounding UDs have been incorporated into an R data package (ByrneDBBMM) and are available as a Supplemental File (Appendix S1).

Results white-tailed deer We collected 3898 GPS locations (fix success rate = 874%) from the white-tailed deer, yielding 1265-, 1151-, and 1118 30-min time steps for the pre-breeding, breeding and post-breeding seasons, respectively, after accounting for missed fixes. Plots of the mean proportion of habitat types within 50% contours revealed circadian patterns of habitat use that varied seasonally (Fig. 2). Differential use of habitats appeared to oscillate in accordance with the daily cycle. For example, use of wildlife food plots was generally greatest during nocturnal periods, whereas creek hardwoods were used during diurnal periods (Fig. 2). Circadian variation in habitat use was most pronounced during the breeding season and least pronounced during the post-breeding season (Fig. 2). When comparing natural pine availability within seasonal ranges to that within time step UD bounds, we observed that circadian patterns of selection for natural pines varied across seasons (Fig. 3). During the pre- and postbreeding seasons, no clear pattern of selection was evident, as the use of natural pines oscillated around general availability across the diel period (Fig. 3a, c). During the breeding season, however, a clear pattern of selection emerged, in which the deer consistently selected against natural pine stands during the diurnal period and moved into them during nocturnal periods (Fig. 3b).

coyote We collected 337 GPS locations from the coyote in July (fix success rate = 906%) and 333 GPS locations in December (fix success rate = 895%), yielding 302- and 295 2-h time steps for each respective month after accounting for missed fixes. Similar to the deer, the coyote exhibited patterns of variation in daily habitat use (Fig. 4). We observed differences in the pattern of habitat use between July and December. In July, agricultural fields were used during both diurnal and nocturnal periods, whereas in December, the coyote used agricultural fields at night but abandoned them during daylight hours (Fig. 4). The general pattern of use for conifer forests was

© 2014 The Authors. Journal of Animal Ecology © 2014 British Ecological Society, Journal of Animal Ecology, 83, 1234–1243

1238 M. E. Byrne et al. (a)

(b)

(c)

Fig. 2. Mean proportional habitat use of three habitat types within 50% UD contours bounding 30-min time steps for a male white-tailed deer during three seasons; pre-breeding (a), breeding (b) and post-breeding (c). Dashed vertical lines represent the mean times of sunrise and sunset for each season. Dashed horizontal lines represent the proportional availability of creek hardwood, food plot and mature pines within 95% home range estimates for each season.

similar between months, with greater use during the day and reduced use at night. This pattern was more pronounced in December, when the coyote used conifer forests nearly to the exclusion of all other habitat types during the day (Fig. 4). The correlation between habitat use and behaviour varied considerably between December and July. In December, a strong negative relationship existed between the use of conifer forests and r2m (r2 = 092), whereas use of agricultural fields was positively correlated with r2m (r2 = 092, Fig. 5). In July, the use of conifer forests was negatively related to r2m ; however, the correlation was not as strong as that observed in December (r2 = 024, Fig. 5). There was no clear relationship between r2m and use of agricultural fields in July (r2 < 001, Fig. 5). These results suggest that in December, the coyote used agricultural fields when active and conifer forests when less active and resting. In July, the coyote tended to use conifer forests less when active; however, the relationship was not as strong as in December, and use of agricultural fields was not related to behaviour during this time.

wild turkey We collected 679 locations between 30 March 2011 and 19 April 2011, with 555 day-time locations (06:00–19:00) that yielded 526 individual 30-min time steps. Fix success rate for day locations was 984%, which provided a detailed record of the turkey’s behaviour during the pre-incubation period. Based on when the nest site was

first located within the 95% UD contour for any 30-min time step, the turkey appeared to have first visited the nest site on 10 April 2011, 9 days before the onset of continuous incubation (Fig. 6a). Time spent in the vicinity of the nest decreased during the next 6 days relative to the first visit and included 1 day when the turkey likely did not visit the nest at all (14 – April). Time spent near the nest increased considerably during the last 2 days leading up to 19 April, when continuous incubation began (Fig. 6a). The turkey was most often in the vicinity of the nest during mid-day, with the greatest probability of nest visits – and potential egg laying – occurring between 12:30 and 13:30 (Fig 6b).

Discussion Using a dBBMM to estimate UDs around individual time steps, we were able to identify distinct temporal patterns of habitat selection and use. The primary advantage of using a dBBMM in this fashion is that it provides a rigorous, probability-based estimate of the area used between successive relocations conditional on the behaviour of the animal (based on speed and tortuosity) at that time. As such, it is possible to account for the space use of the animal, and thus the habitats that may have been sampled, during the time between fixes. The use of UDs is a major difference between our approach and that of step selection functions (Fortin et al. 2005), which define habitat use based on straight lines connecting pairs of locations. Lewis et al. (2011)

© 2014 The Authors. Journal of Animal Ecology © 2014 British Ecological Society, Journal of Animal Ecology, 83, 1234–1243

Temporal patterns of habitat selection (a)

(b)

(c)

Fig. 3. Habitat selection of mature pines by a male white-tailed deer during the breeding season. Dashed horizontal lines represent the baseline availability of mature pines within the 95% home range estimate for the pre-breeding (a), breeding (b) and post-breeding (c) seasons. Habitat selection is measured as the ratio between availability in the range and mean proportional use for 30-min time steps – values >1 indicate selection for mature pines, while values