Walrus Haulout and In-water Activity Levels Relative to Vessel Interactions in the Chukchi Sea, 2012-2015

We collected Argos satellite telemetry locations and behavior monitoring sensor data from a sample of 218 female walruses between 14 July 2012 and 14 October 2015 (Jay et al. 2022).

The Argos system estimated locations of each transmitting tag during satellite overpasses whereby multiple transitions were received by the satellite. These Argos location estimates are based on the known geography of the satellite's path and the Doppler principal applied to frequency shifts in the tag transmissions. The frequency of tag transmissions are measured at the time they are received by the Argos radio antenna on-board the satellite. The time of the Doppler shift in transmission frequency is estimated using a least squares method. Based on this Doppler shift time and the known orbital location of the satellite at that time, Argos estimates two locations on the planet's surface where from the transmission could have originated. The Argos system then selects one of the two locations based on the geography of previously estimated locations from each tag. The Argos system also classifies location quality of each location estimate based on the number of messages received from the transmitter during the overpass. Because this method of determining tag locations is prone to substantial error, we filtered Argos locations with the Douglas Argos-filter algorithm (DAF), which is based on the Argos system location quality classifications, spatial redundancy, movement rates, and turning angles (Douglas et al. 2012). We set the DAF to retain (1) all Argos standard class locations (LC3, LC2, LC1); (2) non-standard class locations (LC0, LCA, LCB) within 2 km of the previous or subsequent location; and (3) any remaining locations based on a distance-angle-rate filter that accepted a maximum walrus speed of 10 km/h and rejected locations at the apex of highly acute angles by setting the DAF coefficient "RATECOEF" to 25.

We transformed the filtered geographic locations to an azimuthal equidistant projection centered on the study area (70°N, 170° W) to minimize geospatial distortion in subsequent geographic operations (proj4string = "+proj=laea +lat_0=70 +lon_0=-170 +x_0=0 +y_0=0 +datum=WGS84 +units=m +no_defs +ellps=WGS84").

We derived chronologies of hourly haulout and foraging states of radio-tagged walruses from data collected by a conductivity sensor and pressure transducer on the tags (Fischbach and Jay 2016). We collected this data with either Telonics tags within the "LPT" series (n = 200, deployed by USGS during 2012-2015), or with Wildlife Computers SPLASH tags (n =18, deployed by the Alaska Department of Fish and Game during 2014-2015) configured to have identical sensor and data processing algorithms. Both sets of tags had a salt-water conductivity sensor that determined immersion of the tag in sea water and a pressure transducer to determine the depth of the tag when in sea water. Both sensors were queried every one or two seconds (depending on the tag model and deployment year) and summarized at the top of every hour to characterize the behavior during that hour. If >= 90% of the conductivity measurements within a 1-h interval indicated the tag was out of water, the walrus was considered to be hauled out during that interval. If the majority of the pressure measurements within a 1-h interval indicated the tag was > 10 m deep, the walrus was considered to be foraging during that interval. These behavior states were logged on-board the tag and transmitted as a digitally encoded message. Encoding of the hourly behavior log differed between the two tag manufactures. Telonics tags allocated two bits to each logged behavior hour: one for haulout state and one for foraging. This resulted in four possible behavior states, including the possible nonsensical behavioral state of being hauled out but foraging. Although the onboard logic could not have allowed this state, barring a sensor failure, parity errors in transmitting the digital message to the satellite and then again from the satellite to an Argos ground station may have resulted in this nonsensical state. To guard against inclusion of such parity error data, Telonics implemented an 8-bit cyclical redundancy check sum, and we only retained behavioral logs from hours supported by at least two Argos messages that passed this check sum. Wildlife Computers tags implemented a proprietary compression algorithm to encode the three possible hourly behavior states into an Argos message, which was also transmitted with a check sum to guard against parity errors that may arise during transmission. We only retained messages that passed this check sum and relied on the Wildlife Computers proprietary software to decode and retain behavior log records.

We excluded behavioral intervals obtained within the first 24 hours after tag deployment to guard against the possibility that the tagging process altered behavior during this period. We retained continuous bouts of telemetered walrus behavior > 72 hours in length with at least 10 Douglas-filtered Argos locations and no location gaps > 24 hours.

We assigned a location and an estimate of location error to each retained hour of walrus telemetered behavior by applying a continuous-time correlated random walk model (CRW, Johnson et al. 2008) using the R package 'crawl' (https://github.com/NMML/crawl accessed 2017 November 1). The CRW model includes four components: a movement process sub-model, an activity sub-model, a drift indicator, and an observation error component. For the movement sub-model, we used log links to estimate one parameter each (i.e. intercept only) for movement velocity variation (σ) and for velocity autocorrelation (β). We disabled the drift option, but we applied the activity sub-model to stop movement during hours when we determined that a walrus was hauled-out on land. For the error component, we fixed error values to the log of the corresponding 68th percentile error values for the Douglas-filtered Argos locations specific to each Argos location class: LC3 to log(400 m), LC2 to log(1000 m), LC1 to log(2500 m), LC0 to log(4600 m), LCA to log(1700 m), LCB to log(2800 m), and LCZ to log(1400 m), (Douglas et al. 2012). We then generated a predicted location (and standard error for the x- and y-coordinates) for every telemetered hour based on the fitted CRW model at the half-hour midpoint of each telemetered walrus behavior hour.

We assigned each hour of walrus behavior as having land available if the hourly location was within 5 km of land, based on the shoreline from the Global Self-consistent, Hierarchical, High-resolution Geography Database (Wessel and Smith 1996; data accessed through R package Ptolemy https://github.com/jmlondon/ptolemy accessed 2016 November 01).

We assigned each hour of walrus behavior as having sea ice available if the hourly location was within 5 km of sea ice based on the National Ice Center's daily Marginal Ice Zone (MIZ) product (http://www.natice.noaa.gov/products/ accessed on 2017 November 01). The Marginal Ice Zone represents regions of pack and marginal sea ice as digitally encoded geospatial polygons based on imagery from satellite-borne sensors acquired near local noon each day and interpreted by experienced analysts for the purposes of supporting safety at sea for marine vessels. Marginal ice is defined as areas with approximately 10–80% ice concentration, whereas pack ice is defined as areas with greater than 80% ice concentration.

We characterized air temperature and wind speed coincident with walrus behavior records based on estimates available through the North American Regional Reanalysis (NARR), which provides weather metric estimates at 3 h intervals on an approximately 32.5 km by 32.5 km grid (Mesinger et al. 2006). We acquired NARR estimates of air temperature at 2 m above sea level, and wind speed at 10 m above sea level from the National Oceanic and Atmospheric Administration. (NOAA Physical Sciences Laboratory 2015, accessed online through ftp://ftp.cdc.noaa.gov/Datasets/NARR/monolevel 2015 March 11.) We assigned weather covariates to the 1-h behavioral intervals. Each metric was interpolated by inverse distance weighting, first in space based on the 4 NARR spatial grid points closest to the walrus's hourly location, and then in time.

We assigned a covariate of benthic biomass (i.e., biomass of potential walrus prey) to the hourly walrus behaviors based on predicted benthic macrofaunal biomass (ln[g carbon/m^2]), which was generated by kriging data from van Veen grab samples collected at multiple benthic sampling stations from 2008 - 2012 (Beatty et al. 2016, Jay et al 2017). Because the benthic sampling stations cover a distinct sub-region of the Chukchi Sea, extending north and east from the Bering Straits, but omitting regions to the west, and we lacked a defensible way to extrapolate beyond that region, we assigned NAs to benthic biomass values for walrus locations outside this region. Details of data collection and estimating the distribution of benthic macrofaunal biomass are described in Beatty et al. (2016, Appendix A).

Because walrus behavior occurs in bouts, and bouts may consist of consecutive hours in the same behavioral state, we indicated what the telemetered walrus behavior in the previous hour was and for how many consecutive hours the previous behavior had been maintained. We refer to the latter as Bout-hours of Previous Behavior in the accompanying manuscript.

We obtained vessel tracking locations through a marine automated information system (AIS) that relies on digitally encoded VHF transmissions from compliant vessels to monitor marine traffic. During our study, cargo vessels over 500 gross tonnage, vessels over 300 gross tonnage on international voyages, and passenger vessels were required to transmit AIS data (International Maritime Organization 2009). Other vessels such as industrial support vessels, yachts, and skiffs operated by researchers or coastal residents also voluntarily transmitted AIS data. We purchased AIS data for 2012-2014 through Marine Exchange of Alaska (https://www.mxak.org/) and obtained 2015 data through Harris Corporation, Geospatial Maritime Solutions (Melbourne, FL).

We applied a speed filter (McConnell et al. 1992) to remove implausible locations that implied ground speeds > 60km/hour, and we filtered the resulting sequences to retain each consecutive set of vessel locations with <= 30 min gap between any two filtered locations. We then linearly interpolated vessel locations within each sequence to predict a vessel location for every minute.

We linked the by-minute vessel locations with hourly walrus locations to characterize whether or not each hourly walrus location had vessel presence. For this, we determined the number of by-minute vessel locations for which each hourly walrus location was within threshold distances of the vessel locations. We then assigned walrus hours as having vessels present if there were at least 10 by-minute vessel locations within 15 km, and as having vessels absent if there were no by-minute vessel locations within 20 km. Hourly walrus locations with between 1 and 9 by-minute vessel locations within a 15 km threshold were removed from the analysis. Likewise, hourly walrus locations with any by-minute vessel locations between the 15 km and 20 km thresholds were removed from the analysis.

Because we sought to measure behavior when a walrus first encountered a vessel, we only retained walrus-hours where the previous hour had no vessels within 20 km. The variable "vessel exposure" refers to vessel presence or absence in these hours of first encounter.