christianson_idohm_jagsdata_README.txt
TITLE:
Double-observer satellite imagery data for hierarchical modelling of abundance
AUTHOR:
Dr. David Christianson, Ecosystem Science and Management, University of Wyoming, Laramie, Wyoming, USA. david.christianson@uwyo.edu
ORCID: 0000-0002-9601-9595
DATE CREATED:
22 January 2022
VERSION INFORMATION:
JAGS version 4.3.0 https://sourceforge.net/projects/mcmc-jags/files/JAGS/4.x/Windows/
R Version 4.1.0
General Description
This readme file outlines the collection of elk (Cervus elaphus) observations from Google Earth satellite imagery. Data were collected using independent double observers. This dataset can be called directly by a JAGS code to estimate abundance of elk (see christianson_idohm_jagscode_README.txt).
Study Area
The southern end of the Gallatin Mountain range in northwestern Yellowstone National Park, Wyoming and Montana, USA (Figure 1) is summer range (June-October) for some unknown portion of migratory elk groups that inhabit winter ranges along the Madison, Gallatin, and Yellowstone rivers. These migrations have long been observed (Brazda 1953, Lovaas 1970, Craighead et al. 1972). Elk are overwhelming the most common ungulate in this area and aerial counts of elk are often more than 20 times greater than counts of other ungulate species including moose (Alces alces), mule deer (Ococoileus hemionus), bighorn sheep, (Ovis canadensis), and mountain goats (Oreamnos americanus). Other large mammals include grizzly bear (Ursus arctos), black bear (Ursus americanus), grey wolves (Canis lupus), and mountain lion (Puma concolor) that occur at low densities.
We limited our area of interest to elevations above 2750 masl south of Mulherin Creek intersecting the park’s northern boundary based on prior information regarding elk movements in summer (Brazda 1953, Singer 1990). This delineated a mostly contiguous 114.81-km2 search area distributed 20 km east to west and 30 km north to south (Figure 1). Our initial search revealed a large number of elk (see Results) and to ensure we did not exclude a significant portion of the population, we also searched peripheral terrain >2590 masl (an additional 92.38 km2), but this secondary subalpine area circumscribing our primary alpine search area revealed much lower elk densities (see Results). Therefore, we report results on the primary 115-km2 core alpine area >2750 masl but we used all survey data from the combined 207 km2 area >2590 masl in our hierarchical model to better estimate parameters.
Satellite imagery
We accessed satellite imagery of the search area from the Google Maps server (Google Maps, 2021, https://mt1.google.com/vt/lyrs=s&x={x}&y={y}&z={z}, Maxar Technologies 2021, “Greater Yellowstone Ecosystem” 19 July 2014). The Google server provides access to satellite imagery from several dates, but the imagery from 19 July 2014 covered the entire area of interest and offered the greatest clarity of ground features. The time of day of the imagery could not be ascertained but shadow lengths suggested approximately 0930 in the morning which we confirmed by aligning shadows in the imagery with a hillshading algorithm for a digital elevation model. We viewed tiles of the satellite imagery in Quantum GIS (QGIS.org 2021). We also accessed satellite imagery from 16 August 2011 on that ArcGIS web server (ESRI Inc. World Imagery (Clarity), 2021, https://www.arcgis.com/apps/mapviewer, “Greater Yellowstone Ecosystem” 16 August 2011) to provide a reference image to distinguish between elk and background features by switching between the two images at each observer location (Figure 2).
We spatially organized counts of ungulates detected in the imagery by site. We overlaid our study area with a 250-m grid with each 6.25 ha plot (sites) further divided into four 125-m square subplots to enforce a maximum search altitude within each site. We set the viewing altitude such that a single subplot (a 1.5625 ha area) occupied the entire height of a 76-cm monitor. The first observer (DC) tallied elk in all 3315 plots (searching 13260 subplots). A second observer (JW) tallied all elk in 370 randomly chosen plots (1472 subplots totaling 23 km2) without knowledge of the first observer’s detections.
Land cover covariate
We classified land cover in the imagery as sparsely vegetated rock and soil (including snow and ice patches), fully vegetated meadows and shrublands, burned forest with standing and fallen dead trees, and conifer forest using a random-forest machine learning algorithm (Breiman 2001). We used bands 1, 5,6, & 7 from Landsat 8 Collection 2 Analysis Ready Data from the closest cloud-free date (20 July 2014) along with elevation and terrain ruggedness as classification features. We fit 1500 decision trees in R (randomForest package) constructed progressively with one randomly chosen variable from five randomly chosen feature traits. We trained the decisions tress on a sample of 91 polygons containing the four land cover classes.
Hierarchical model of abundance
We developed a hierarchical model using a Poisson formulation of a multinomial model of independent double-observer counts (Kery and Royle 2016). In the double-observer protocol, the two observers are assumed to have separate detection probabilities, p1 and p2 (Caughley 1974). Three-hundred seventy random plots were searched by both observers. The remaining plots were searched solely by observer 1. The counts from plots surveyed by both observers could be used to estimate p1 for the purpose of estimating abundance conditional on the counts obtained by observer 1.
Data Structure
The structure of the .RData file consists of a R list (called ‘data’) that includes the following R objects:, a y matrix of counts with one row per site and columns for unique counts by observer 1, unique counts by observer 2, non-unique counts by both observers, and counts conducted solely by observer 1. The variable ‘nonforest’ is a vector (one observation per site) containing the number of hectares in each site that was not forested. The variables ‘cover_meadow’ and ‘cover_deadfall’ are vectors containing the additively transformed (log of the ratio) proportions of meadow vegetation and burned forest in each plot using barren rock and soil as the common denominator. The vector ‘firstsearch’ is a list of the sites by their index that are contained with the core, alpine search area >2750 masl. Much of the symbology and structure of this code follows Kery and Royle (2016). M is a single integer, specifying the total number of sites. Spatial explicitness of sites is not included.
Literature Cited:
Brazda, A. R. 1953. Elk migration patterns, and some of the factors affecting movements in the Gallatin River drainage, Montana. Journal of Wildlife Management 17:9–23.
Breiman, L. 2001. Random forests. Machine Learning 45:5–32.
Caughley, G. 1974. Bias in Aerial Survey. The Journal of Wildlife Management 38:921–933.
Craighead, J. J., G. Atwell, and B. W. O’Gara. 1972. Elk Migrations in and Near Yellowstone National Park. Wildlife Monographs:6–48.
Kery, M., and J. A. Royle. 2016. Applied hierarchical modelling in ecology. Academic Press.
Lovaas, A. L. 1970. PEOPLE AND THE GALLATIN ELK HERD.
MacKenzie, D. I., and L. L. Bailey. 2004. Assessing the fit of site-occupancy models. Journal of Agricultural, Biological, and Environmental Statistics 9:300–318.
Plummer, M. 2017. JAGS: A program for analysis of Bayesian graphical models using Gibbs sampling. https://sourceforge.net/projects/mcmc-jags/.
QGIS.org. 2021. QGIS Geographic Information System.
R Core Team. 2021. R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria.
Singer, F. J. 1990. The ungulate prey base for wolves in Yellowstone National Park. Pages 2-3-2–37 Wolves for Yellowstone? A report to the United States Congress. Yellowstone national Park, Mammoth, Wyoming, USA.