Data from: Quantifying impacts of recreation on elk (Cervus canadensis) using novel modeling approaches

Description

Notes

Methods

We used an elk habitat suitability model hereafter referred to as "the westside model" (Rowland et al. 2018) to characterize habitat suitability throughout the full North Rainier Elk Herd range in western Washington, identifying areas of high quality habitat for camera trap deployments. We defined high quality habitat as areas that the westside model ranked as the 40% most suitable elk habitat in the herd range (Fig. 1). The goal of this site selection criterion was to better isolate the impacts of recreation on elk detections while reducing the confounding effects of habitat quality. We then clipped these areas to land parcels managed by Washington Department of Natural Resources (WA DNR) and the United States Forest Service (USFS), as these two agencies provided permission for us to deploy cameras on their managed lands. The WA DNR and USFS properties encompassed by the North Rainier Elk Herd range span a gradient of recreational use intensity, including areas that consistently see over a thousand visitors per week in the summer and areas that are closed to the public via a 24-hour guarded security gate. We randomly selected 80 camera locations along trails or restricted access roads within these areas, with a minimum of 1 km spacing between randomized points (Fig. 1). Beginning on June 2, 2022, we navigated to each random point and deployed a Reconyx Hyperfire (HC600/PC900) camera (Reconyx, Holmen, WI) approximately 1 m from the ground. We set cameras at a 90˚ angle to the anticipated direction of travel (i.e., perpendicular to the trail) when the trail was restricted to hikers or horseback riders only, and at a 45˚ angle to the direction of travel when the trail or road allowed mountain bikers and/or vehicles. These differences in camera set orientations were undertaken to account for difficulty in detecting faster moving targets such as mountain bikes and vehicles (Miller, Leung, and Kays 2017). We also recorded the camera height (m), trail or road width (m), and the distance (m) from the camera lens to the anticipated path of the target (i.e., the nearest edge of the trail or road) for use as covariates in subsequent modeling. We set cameras to take one photo per trigger, with no delay period, on high trigger sensitivity. We performed final camera take-downs on September 18, 2022, with most cameras being deployed for approximately 6-8 weeks.

We pre-processed camera images using the artificial intelligence software MegaDetector (Beery, Morris, and Yang 2019), binning photos into "human", "animal", "vehicle", or "blank" (i.e., false trigger) categories. We then blurred all human images using a publicly available photo-blurring software (WildCo Lab 2021) to protect the privacy of people detected on our cameras (Sharma et al. 2020). We used the automated tags generated by MegaDetector to sort images for rapid batch identification (Fennell, Beirne, and Burton 2022), and we manually tagged all images using the open source software Timelapse (Greenberg 2021). We classified photos of humans into three activity categories: hiking, mountain biking, and horseback riding, and we tagged all animal detections by species, including domestic dogs and cats. We then corrected any erroneous identifications provided during the automated detection process, and exported all detection data for statistical analysis in R (v. 4.2.2; R Core Team 2022).

From our raw detection data, we used an independence threshold of one minute to determine independent detection "events" for all human-related activities, and a threshold of 30 minutes for all wildlife detection events. A 30-minute independence threshold is common in most wildlife studies (Burton et al. 2015). We used the one minute threshold for human detections because numerous independent recreationists are likely to use high-traffic trails within the larger 30-minute window (Procko et al. 2022; 2023). This method may slightly under-estimate visitors, as multiple individuals are recorded as a single detection event if they all pass the camera within one minute of each other. To remedy this, recent works have recommended utilizing raw image counts instead of a time-to-independence method (Peral, Landman, and Kerley 2022), but this method would over-estimate visitation given most people are photographed multiple times when passing in front of a camera, and preliminary modeling showed worse model performance when we removed independence thresholds. We used these independent detection events to calculate detection rates (number of detection events per unit of sampling effort) of all human-related activities and all wildlife species, and we used these detection rates in subsequent modeling either as our response variables (elk detection rates) or predictor variables (detection rates of other species including humans participating in various activities). We did not include any wildlife species which were detected in fewer than 30 unique events.