Are polar bear habitat resource selection functions developed from 1985–1995 data still useful?

Abstract Greenhouse‐gas‐induced warming in the Arctic has caused declines in sea ice extent and changed its composition, raising concerns by all circumpolar nations for polar bear conservation. Negative impacts have been observed in three well‐studied polar bear subpopulations. Most subpopulations, however, receive little or no direct monitoring, hence, resource selection functions (RSF) may provide a useful proxy of polar bear distributions. However, the efficacy of RSFs constructed from past data, that is, reference RSFs, may be degraded under contemporary conditions, especially in a rapidly changing environment. We assessed published Arctic‐wide reference RSFs using tracking data from adult female polar bears captured in the Beaufort Sea. We compared telemetry‐derived seasonal distributions of polar bears to RSF‐defined optimal sea ice habitat during the period of RSF model development, 1985–1995, and two subsequent periods with diminished sea ice: 1996–2006 and 2007–2016. From these comparisons, we assessed the applicability of the reference RSFs for contemporary polar bear conservation. In the two decades following the 1985–1995 reference period, use and availability of optimal habitat by polar bears declined during the ice melt, ice minimum, and ice growth seasons. During the ice maximum season (i.e., winter), polar bears used the best habitat available, which changed relatively little across the three decades of study. During the ice melt, ice minimum, and ice growth seasons, optimal habitat in areas used by polar bears decreased and was displaced north and east of the Alaska Beaufort Sea coast. As optimal habitat diminished in these seasons, polar bears expanded their range and occupied greater areas of suboptimal habitat. Synthesis and applications: Sea ice declines due to climate change continue to challenge polar bears and their conservation. The distribution of Southern Beaufort Sea polar bears remained similar during the ice maximum season, so the reference RSFs developed from data collected >20 years ago continue to accurately model their winter distribution. In contrast, reference RSFs for the ice transitional and minimum seasons showed diminished predictive efficacy but were useful in revealing that contemporary polar bears have been increasingly forced to use suboptimal habitats during those seasons.


| INTRODUC TI ON
Since at least the mid-1990s, polar bears (Ursus maritimus) have contended with an environment that is rapidly changing due largely to greenhouse-gas-induced climate warming (Atwood et al., 2016).
Polar bears evolved as specialized predators of sea ice-associated seals (e.g., ringed seals [Pusa hispida] and bearded seals [Erignathus barbatus]) and because of this occur only where the surface of Northern Hemisphere marine waters is >50% sea ice for a substantial portion of the year (Stern & Laidre, 2016). Since 1980, global atmospheric CO 2 has risen from ~340 ppm to >410 ppm (https :// www.esrl.noaa.gov/gmd/ccgg/trend s/global.html; accessed 13 April 2019). Concurrently, average global air temperature rose > 1°C above preindustrial levels for the first time in 2015 (Hawkins et al., 2017).
Arctic sea ice declines have been directly related to greenhouse gas emissions (Notz & Stroeve, 2016), and since 1978 ice extent declines have ranged from 3.0% decade −1 in March to 12.8% decade -1 in September, (http://nsidc.org/arcti cseai cenew s/; accessed 13 April 2019). Across the Arctic, the presence of seasonal sea ice decreased 5 days decade −1 between 1979 and 2013 (Parkinson, 2014). Declines in the spatial and temporal distribution of sea ice have been accompanied by reductions in ice age and thickness. During 1985During to 2015 first year ice in March increased from 50% to 70% of total pack ice extent while ice >4 years old decreased from 20% to 3% (Tschudi, Stroeve, & Stewart, 2016). The twelve lowest summertime sea ice extents ever recorded occurred during 2007 to 2018 (http://nsidc. org/arcti cseai cenew s/2018/09/; accessed 13 April 2019). Since the early 2000s, synergistic interactions between decreasing ice thickness, increased mobility and fracturing, and reduced surface albedo have rendered the ice pack more vulnerable to melting (Kashiwase, Ohshima, Nihashi, & Eicken, 2017). These changes have negatively impacted polar bear sea ice habitat (Stern & Laidre, 2016). Since 1979 within the region including the Southern Beaufort Sea polar bear subpopulation (SB; Figure 1), the ice retreat date has occurred 9.0 days decade −1 earlier and the ice advance date 8.8 days decade −1 later (Stern & Laidre, 2016). A significant change in ice retreat (4.0 days decade −1 earlier) and advance (5.3 days decade −1 later) has also occurred in the adjacent region that includes the Chukchi Sea polar bear subpopulation (Figure 1; Stern & Laidre, 2016).
Because polar bears live in an environment that is rapidly changing, they are the focus of conservation efforts by all nations whose jurisdictions overlap their range (Durner, Laidre, & York, 2018).
Nineteen subpopulations of polar bears are recognized (Durner et al., 2018), and all have experienced loss of sea ice habitat since 1979 (Stern & Laidre, 2016). Of those subpopulations with sufficient monitoring, three have undergone population declines (Bromaghin et al., 2015;Lunn et al., 2016;Obbard et al., 2018), two have undergone range contractions  or shifts in habitat use (Laidre et al., 2015), and five have shown no apparent negative impacts (Peacock, Taylor, Laake, & Stirling, 2013;Regehr et al., 2018, Stapleton, Peacock, & Garshelis, 2016, Stirling, McDonald, Richardson, Regehr, & Amstrup, 2011, SWG, 2016. Responses of the remaining nine polar bear subpopulations are unknown (Durner et al., 2018); hence, information to assist conservation decisions is currently unavailable in much of the range of polar bears. As revealing that contemporary polar bears have been increasingly forced to use suboptimal habitats during those seasons.

K E Y W O R D S
Arctic, climate change, habitat, polar bear, resource selection functions, sea ice, Ursus maritimus F I G U R E 1 Study area defined by a rasterized 99% minimum convex polygon (mcp) of satellite telemetry locations from adult female polar bears radio-tagged between 123° and 157° W longitude, 1985-2016. Also shown is the overlap between the study area and polar bear subpopulations (SP) as defined by the International Union for Conservation of Nature (Durner et al., 2018). 5° of latitude = 556 km degradation of the sea ice habitat used by polar bears continues through the 21st century (Durner et al., 2009), understanding how polar bears distribute themselves relative to sea ice composition may serve as a useful proxy for assessing their conservation status (Vongraven et al., 2012).
Resource selection functions (RSF) model the relationship between animal locations and environmental covariates with the goal of estimating the relative probability of a resource being used (Manly, McDonald, Thomas, McDonald, & Erickson, 2002). As an index of habitat suitability, RSFs can quantify the likely distribution of a population across its range (Boyce et al., 2017) and may be used to estimate changes in habitat quality resulting from environmental change (Durner et al., 2009). RSFs are often structured as discrete choice models, where the set of resources available to an individual is allowed to vary between individuals and across choice sets (Arthur, Manly, McDonald, & Garner, 1996;McDonald, Manly, Nielson, & Diller, 2006). As such, discrete choice RSFs are relatively robust to variation in habitat composition (Arthur et al., 1996). Nevertheless, the efficacy of RSFs applied at a future time should be verified, especially if environmental conditions have markedly changed (Garshelis, 2000). This is true for polar bears whose sea ice habitats have changed (Stern & Laidre, 2016) and, in the Beaufort Sea, may now use larger areas because movement rates have increased to compensate for faster sea ice drift in recent years (Auger-Méthé, Lewis, & Derocher, 2015;Durner et al., 2017). Verification is especially important when the target species is of conservation concern and efforts to monitor populations are either deficient, intermittent, or nonexistent; all of which are factors in the international attempts to monitor the 19 subpopulations of polar bears (Vongraven et al., 2012).
In 2007, to inform a decision by the U.S. Secretary of the Interior on whether to list polar bears as a threatened species throughout their range, the U.S. Geological Survey (USGS) launched a suite of comprehensive studies to assess the long-term impact of a changing Arctic on polar bears. One component of that work involved an international collaboration to assess observed and predicted changes in polar bear sea ice habitat throughout the Arctic Ocean and adjoining peripheral seas (Durner et al., 2009). Using satellite telemetry locations of adult female polar bears between 1985-1995, Durner et al. (2009) developed seasonal RSFs that showed, Arcticwide, optimal habitat declined within the ranges of most polar bear subpopulations in the following decade (1996)(1997)(1998)(1999)(2000)(2001)(2002)(2003)(2004)(2005)(2006) and was projected to decline in all subpopulations by the end of the 21st century. Further, Durner et al. (2009)  Philopatry is evident in polar bear subpopulations (Amstrup, McDonald, & Durner, 2004;Bethke, Taylor, Amstrup, & Messier, 1996;Mauritzen et al., 2002), and sea ice changes have regional specificity (Stern & Laidre, 2016). Because governmental polar bear conservation efforts are typically focused on the regions within their jurisdictions, a regional and seasonally focused approach to inform those efforts is needed. Additionally, steep rates of Arctic summer sea ice loss have continued (Stroeve & Notz, 2018) since the period that Durner et al. (2009) used to assess their RSF, raising questions about the model's efficacy for understanding the spatial ecology of polar bears in the contemporary Arctic and for informing present-day conservation actions. In this paper, we present an assessment of the RSFs developed by Durner et al. (2009)

| Polar bear location data
From 1985 to 2016, we captured adult female polar bears in the southern Beaufort Sea and equipped them with either collar, ear tag, or glue-on platform transmitter terminals (PTT) that provided Doppler-derived or GPS locations. We restricted our analysis to data from adult female polar bears because adult male and subadult polar bears cannot wear radio collars, and the vast majority of tracking data has been of adult females. We retained data from a PTT if it was deployed on a bear captured between 123° and 157° west longitude ( Figure 1). We performed an initial filter to remove implausible Doppler locations using the Douglas Argosfilter (DAF) algorithm (Douglas et al., 2012). DAF retained all standard-quality locations (Argos location classes 3, 2, and 1), and auxiliary location classes (0, A, and B) when corroborated by a consecutive location within a 10 km radius, or when movement rates were <10 km/hr and the internal angles (α, in degrees) formed by preceding and subsequent vectors (of lengths d1 and d2 km) were not suspiciously acute (α > −25 + β × ln[minimum (d1,d2)], where β = 15). We assigned β = 15 because it performed well for our specific tracking data across seasons and regions. We excluded locations of bears that were on land or in maternal dens, and from PTTs that had become detached as evidenced by invariant activity sensor data, temperature sensor data emulating ambient conditions, or location data either persistently stationary or persistently following the prevailing ice drift.

| Reference RSFs and environmental data
For the reference RSFs, we used the four seasonal RSFs (winter, spring, summer, and autumn), the four environmental covariates (sea ice concentration [SIC], distance to the 15% SIC interface, ocean depth, and distance to land), and the original coefficients ( Table 1)  accessed 15 August 2017) and converted them to Arc/Info grids (ver. 9.3; ESRI). Although polar bears may respond to many sea ice metrics (e.g., see Ferguson, Taylor, & Messier, 2000), NSIDC estimates of sea ice concentration derived from 25 km resolution passive microwave sensors were the only data consistently available throughout the Northern Hemisphere during the years examined by Durner et al. (2009). Likewise, ocean depth and distance to land are both invariant across years and therefore provided a consistent source of environmental data during the entire study period.
We obtained ocean depth from the International Bathymetric Chart of the Arctic Ocean (IBCAO) Version 3.0 (pixel size: 0.5 × 0.5 km; Jakobsson et al., 2012). A GIS line coverage of the coast was derived by contouring the depth grid at cell values equal to zero.
We created a distance to land grid by calculating the distance from each SIC pixel to its nearest point on the coastline coverage. All grids were converted to Lambert equal area projection with 25 × 25 km cell size.

| Study area
We defined our study area as the 99% minimum convex polygon (MCP) encompassing all polar bear locations, 1985-2016. One location per day per bear was selected for the MCP analysis based on the best location quality (GPS, then Argos quality classes 3, 2, 1, 0, A, and B); ties were decided randomly. A single MCP was derived from the daily polar bear locations using the mcp function in the R package adehabi-tatHR (Calenge, 2017). The MCP polygon was converted to a grid with the same map projection and pixel size as the grids of environmental data. Land cells within the MCP were omitted from the study area. The resulting study area included the extent of four of the 19 recognized subpopulations of polar bears, that is, the SB, Arctic Basin, Chukchi Sea, and Northern Beaufort Sea (Durner et al., 2018). We used the same reference period to establish a minimum RSFvalue threshold for optimal polar bear habitat in each season within the MCP study area. To assess the reference RSFs during conditions of sea ice decline, we divided the remaining years into two decadal periods: 1996-2006, years that had substantially lower sea ice extent relative to 1985-1995(Ogi & Wallace, 2007), and 2007 which included 10 of the 12 lowest annual minimum Arctic sea ice extents in the satellite record (NSIDC, http://nsidc.org/arcti cseai cenew s/2018/09/arctic-sea-ice-extent-arriv es-at-its-minim um/, accessed 26 February 2019).

| Season assignments
We assigned monthly SIC grids to one of four temporally dynamic seasons (i.e., melt, minimum, growth, and maximum) using the methods from Durner et al. (2009). Season length was allowed to vary across years to accommodate the Arctic's changing seasonality. During any given year, a month was assigned to the maximum season when sea ice extent within our study area was greater than the annual maximum extent minus 15% of the respective year's maximum-minimum amplitude. Conversely, a month was assigned to the minimum season if its sea ice extent was less than the annual TA B L E 1 Coefficients and standard errors (in parentheses) of covariates in four seasonal resource selection functions for polar bears in the polar basin, 1985basin, -1995basin, , reported Durner et al. (2009

| Calculating RSF grids and equal area zones
Monthly RSF grids were derived using the appropriate seasonal model (Table 1)

| Assigning RSF zones to polar bear locations
We calculated the percentage of polar bear locations within each equal area RSF zone using only higher-accuracy locations (GPS and Argos classes 3, 2, and 1) that were no less than 72 hr apart (to reduce autocorrelation and standardize interannual sampling intensity). Locations that occurred outside of RSF zones were excluded.
For each decadal period, the seasonal mean percentage of bear locations within each equal area interval and 95% confidence intervals were calculated by bootstrapping. For each of 25 bootstrap iterations, we set the sample size to the number of bears within the respective period and season. Durner et al. (2009)  To assess optimal habitat within an area overlapping the spatial distribution of polar bears, we developed kernel utilization distributions (UD) of polar bear locations for each decadal period and season.

| Assessing polar bear responses to changes in optimal habitat
We defined an overall area of polar bear occupancy as the region that encompassed 95% of the UD, and a "core" area that included 50% of the UD. We used the same higher-quality, temporally restricted subset of polar bear locations described above to generate period-season UDs with function kernelUD in the R package adehabitatHR (ver. 0.4.15;Calenge, 2017). Total area for each period-season UD that was within the study area was calculated. Period-season UD grids were then matched with their respective monthly RSF optimal habitat grids to extract the area of optimal habitat within the UDs for each month.
The percent of optimal habitat within UDs for each month was derived by dividing UD optimal habitat area by the total UD area and multiplying by 100. Season-specific percentages of optimal habitat were compared between periods with boxplots and ANOVAs followed by post hoc Tukey HSD tests. We used chi-square tests to examine pro- We used centroids of optimal habitat within 95% kernel UDs to examine whether the distribution of optimal habitat has changed spatially across periods. Centroids for each period-season were derived by averaging the coordinates of all pixels in each monthly RSF that occurred within the respective 95% kernel UD. We then transformed those coordinate averages to longitude and latitude and used the distHaversine and bearing functions in R package geosphere (ver. 1.5-7; Hijmans, 2017) to, respectively, calculate the great-circle-distance and initial bearing from each centroid to the centroid in the next period (e.g., melt 1985-1995 to melt 1996-2006, and melt 1996-2006 to melt 2007-2016).

| RE SULTS
After imposing filters, 56,977 locations from 301 bears were used to define a 99% MCP study area which encompassed 2,298,125 km 2 after rasterizing and excluding cells over land (Figure 1)

| Ice melt
The mean length of the ice melt season from 1985 to 2016 was 2.0 months (range: minimum = 0, maximum = 3). A smaller proportion of polar bear locations in 1996-2006 and 2007-2016, relative to 1985-1995, often occurred in the highest RSF equal area zones during the ice melt season (Figure 2a,b). During 1985During -1995 the proportion of bear locations in the upper 50% and 20% of the RSF-valued habitat was 0.87 and 0.36, respectively (

| Ice minimum
The mean length of the ice minimum season from 1985 to 2016 was  Table 2).

| Ice growth
The mean length of the ice growth season from 1985 to 2016 was  Table 2).

| Ice maximum
The   From 1985From -1995From to 2007From -2016, an increase in space use by polar bears was evident across the study area, as was a reduction in the proportion of the study area that was comprised of optimal habitat. From 1985From -1995From to 2007From -2016, area of seasonal polar bear 50%

| Assessing the relative distributions of polar bears and optimal habitat
kernel UDs increased across decades by 48%-118% in all seasons except the ice maximum ( Figure 3, Table 3). Spatial area of the 95% UDs increased in all seasons by as much as 16%-88% (Figure 3, Table 4).
In contrast, the proportion of the UD area that was comprised by TA B L E 2 Proportion of polar bear locations in the Beaufort and Chukchi seas occurring in the upper 20% and upper 50% of RSF-valued habitat for 1996RSF-valued habitat for -2006RSF-valued habitat for and 2007RSF-valued habitat for -2016, compared with a chi-square test of proportions to the proportion of locations during the reference period (1985)(1986)(1987)(1988)(1989)(1990)(1991)(1992)(1993)(1994)(1995)  optimal habitat generally showed a declining trend. From 1985From -1995From to 2007From -2016, the average percent of monthly optimal habitat during ice melt within 50% UDs declined (F 2,60 = 9.06, p < 0.001;  Table 3). Like the F I G U R E 3 Spatial distribution and month frequency of optimal sea ice habitat presence, in the Beaufort and Chukchi seas, as modeled by reference (1985)(1986)(1987)(1988)(1989)(1990)(1991)(1992)(1993)(1994)(1995) polar bear resource selections functions (Durner et al., 2009) for three periods and all months within each season, overlaid with 50% (black) and 95% (purple) kernel utilization distribution polygons of polar bears derived from satellite telemetry locations during each period and season. Number of months is provided in the upper right corner of each panel. Map extent and orientation as in Figure 1, north arrow provided for reference, AK = Alaska TA B L E 3 Seasonal 50% kernel utilization distribution areas (km 2 ) and mean monthly amount (%, SD, n) of each UD comprised of optimal polar bear sea ice habitat in four seasons and three decadal periods in the Beaufort and Chukchi seas, 1985-2016  (Table 3).
The decrease in the average percent of monthly optimal habitat between periods was even greater within the 95% UDs (Figure 3).
Optimal habitat within the 95% UD during the melt season declined across periods (F 2,60 = 30.77, p < 0.001; Table 4), with all periods significantly different from the others (1985-1995:25.5%; 1996-2006:11.9%; 2007-2017:5.2%; Figure 4a, Table 4). A similar pattern occurred in the ice minimum season (F 2,39 = 32.45, p < 0.001) as all periods were significantly different from the others (1985-1995:22.5%; 1996-2006:13.7%; 2007-2017:8.0%; Figure 4b, Table 4), and in the ice growth season (F 2,42 = 51.59, p < 0.001), with, again, all periods were significantly different from the others (1985-1995:27.1%; 1996-2006:18.9%; 2007-2017:13.6%; Figure 4c, Table 4). Differences in the percent of optimal habitat across periods occurred in the ice maximum season (F 2,231 = 37.01, p < 0.001), however, this was a result of a greater percentage of optimal habitat   In -2006In and 2007, and during seasons of ice melt, ice minimum, and ice growth, the distribution of polar bears and the distribution of optimal habitat showed relatively little overlap. During those seasons, as optimal habitat decreased and shifted north, polar bears expanded their range and used more sub-optimal habitat. Utilization of suboptimal habitats can be maladaptive for wildlife (Hollander, Dyck, Martin, & Titeux, 2017).  Similarly, polar bear habitat selection in the Chukchi Sea also remained unchanged between periods that experienced sea ice loss (Wilson, Regehr, Rode, & St. Martin, 2016 Additionally, the performance of the RSF declined across periods in the sea ice melt, minimum, and growth seasons. But during the ice maximum season, the availability of optimal habitat has remained largely unchanged across the years and continues to comprise a large proportion of the area typically used by polar bears.

| D ISCUSS I ON
During spring ice melt, summer ice minimum, and autumn ice growth seasons, there was a diminishing amount of optimal habitat in our study area and within the kernel UDs of polar bear distribution ( Figure 4). Also, what optimal habitat remained during the latter two decades became further displaced from the areas normally used by SB polar bears. Regions in our study area that consistently had optimal habitat during each period represented a small and decreasing proportion of the area typically used by polar bears after 1985-1995 ( Figure 3). Optimal habitat in the southern Beaufort Sea during the recent decade (2007-2016) has all but disappeared in F I G U R E 4 Boxplots (median, quartiles, 5% and 95% whiskers, and outliers) showing the annual proportion of polar bear 95% kernel utilization distributions (UDs) comprised of optimal sea ice habitat in the Beaufort and Chukchi seas within decadal periods during each of four seasons: ice melt (a), ice minimum (b), ice growth (c), and ice maximum (d).
Significant differences between periods within panels are indicated by different letters (below the lower left corner of each box; Tukey HSD, p < 0.05). Diamonds represent the mean. Sample sizes (number of months) are shown above the mean. UD areas, and mean fractions of optimal habitat are provided in Table 4 all seasons except the ice maximum ( Figure 3). Optimal habitat has persisted more reliably in the ranges of the Northern Beaufort Sea and Chukchi Sea subpopulations of polar bears. We suggest that the dearth of optimal habitat during the non-winter seasons is at least partly responsible for the increasing range of Beaufort Sea polar bears; there is simply little optimal habitat for bears to find and occupy. Most of the non-winter sea ice habitat is now sub-optimal, so any sea ice, regardless of composition or location that can provide a stable substrate to walk on may be a likely candidate for polar bear occupancy for lack of a better choice and that choice may be maladaptive (Hollander et al., 2017).
A changing Arctic has diminished the abundance of optimal habitat for SB polar bears during the ice melt, minimum, and growth seasons. The loss of optimal habitat in these seasons, coupled with increasing duration (Stern & Laidre, 2016), is likely a contributing factor to the observed declines in survival and abundance of SB polar bears (Bromaghin et al., 2015). However, we found that when sea ice attained its average winter extent, optimal habitat was consistently present during all 32 years of our study. We also found that the distribution of polar bears during the ice maximum season reflected the distribution of optimal habitat across all 32 years of study, conveying predictability in how bears select optimal habitat when it is persistently available. This predictability also suggests that efforts to estimate the abundance and trend of SB polar bears may be most effective during the ice maximum season, as optimal habitat and bear distribution coincides with the distribution of research efforts for this subpopulation (Bromaghin et al., 2015).
Our findings elevate concerns for the future status of SB polar bears as the transitional seasons of sea ice lengthen and the extent of optimal polar bear habitat during those seasons declines.
Indeed, the first conservation criterion of the U.S. Marine Mammal Protection Act is that the health and stability of marine ecosystems F I G U R E 5 Polar bear optimal sea ice habitat centroids across 3 decades. Centroids for ice melt, minimum, and growth seasons generally moved north and east across decades. Centroids remained relatively unchanged during the ice maximum season across decades. Data values for this figure are provided in Appendix 2 be maintained at a level such that polar bears may persist as "significant functioning elements" within those systems (USFWS, 2016).
Consequently, actions directed toward slowing global warming and preserving the duration of the ice maximum season will also preserve optimal habitat, and thereby benefit polar bears. and pilots whose efforts were essential for obtaining radio telemetry data. We thank Sharon Poessel (USGS) and two anonymous referees for their constructive input on prior versions of this manuscript.

ACK N OWLED G M ENTS
This report has received U.S. Geological Service bureau approval for publication. Any use of trade, firm, or product names is for descriptive purposes only and does not imply endorsement by the U.S. Government.

CO N FLI C T O F I NTE R E S T
The authors claim that they have no conflict of interest.

DATA ACCE SS I B I LIT Y
The data used in this research are available through the U.S.