Identifying drivers of spatial variation in occupancy with limited replication camera trap data

Abstract

Occupancy models are widely used in camera trap studies to analyze species presence, abundance, and geographic distribution, among other important ecological quantities. These models account for imperfect detection using a latent variable to distinguish between true presence/absence and observed detection of a species. Under certain experimental setups, parameter estimation in a latent variable framework can be challenging. Several studies have issued guidelines on the number of independent replicated observations (surveys) needed for each unchanging occupancy field (season) to ensure reliable estimation. In this paper, we present a spatio-temporal occupancy model, and show through a simulation study that it can be fit to data obtained from a single survey per season, so long as the number of seasons is sufficiently large. We include an application using camera-trap data on the Thomson's gazelle in the Serengeti in Tanzania.

A spatiotemporal model for multivariate occupancy data

“…For modeling detection, we used the number of days each time period that the camera was functioning (1–8), an indicator for whether the pixel area is forested (defined as more than 10 trees), and an indicator for the camera being located on a tree as opposed to a pole. These external covariates are substantially similar to those used in Anderson et al (2016) and Hepler et al (2018).…”

“…Hepler et al (2018) used the data to demonstrate that both occupancy and detection parameters could be estimated, despite single observations of the changing occupancy field. Anderson et al (2016) considered T = 23 adjoining 16‐day periods; here, as with Hepler et al (2018), we considered the T = 46 adjoining 8‐day periods beginning on 1/1/2012 (i.e., 1/1/2012–1/8/2012, 1/9/2012–1/16/2012, etc.). There is K = 1 survey per time period, with the survey being conducted over the entire length of time the camera is on within that 8‐day period.…”

“…In a multivariate ecological application though, Warton (2005) showed that zero‐inflated models can yield a worse goodness‐of‐fit than other models, and Arab, Holan, Wikle, and Wildhaber (2012) argued in favor of hierarchical Bayesian models to capture multivariate and nonlinear effects, rather than trying to extend zero‐inflated models to the complicated and high‐dimensional setting. In our setting, imperfect detection is a documented issue with the camera trap study (Anderson et al, 2016; Hepler, Erhardt, & Anderson, 2018) and corrections to imperfect detection have been studied extensively in occupancy and abundance applications (Madsen, Dalthorp, Huso, & Aderman, 2020; Mengersen et al, 2017). Since our explicit goal is parameter estimation for the drivers of both occupancy and detection, we focus on occupancy models.…”

“…Time dependence within‐species can be seen by comparing any plot to others in the same row—for instance, Thomson's gazelle appear to be more likely to occupy a site if they were previously in that site. Not shown are environmental covariates, such as measures of vegetation density and quality, distance to rivers, and so forth, which are known drivers of landscape use in the ecosystem (Anderson et al, 2016; Hepler et al, 2018). The additional dependence present is cross‐species dependence, best seen by comparing any two plots from the same column—for example, zebra (middle row) and wildebeest (bottom row) show strong correspondence at any of the five time points.…”

“…Most of the existing work on occupancy models involves univariate, spatial‐only applications, which reduces the overall computational cost of the approach. An exception is the univariate spatiotemporal occupancy model proposed by Hepler et al (2018), which used the logit link to relate the probabilities to external covariates with a centered spatiotemporal autologistic model assumed for the occupancy probability. That model is computationally expensive for large data sets and not easily extended to two or more species.…”