Addressing challenges in camera-trap studies: Survey designs for multiple species, serial dependence, and site-to-site variability when estimating activity patterns

Abstract

Camera traps are widely used to collect information on the behavior, abundance, and occurrence of wild animals living in all parts of the world. Their low cost and ease of use makes it possible to collect information on many species simultaneously over large areas and long periods of time (Burton et al. 2015). These devices record the presence and activity of the animals that travel in front of their sensors. Most modern cameras rely on passive infrared sensors that are automatically triggered when a change in temperature is detected (Welbourne et al. 2016). They also have very low energy requirements and can be active for weeks without battery replacement, recording thousands of images along with the date and time each picture was taken. These characteristics allow users to continuously survey an area for long periods of time with minimal or no disturbance to wildlife, and make cameras extremely useful for surveying remote locations or studying animals that are particularly wary of human presence (e.g., carnivores). Decrease in cost per unit, technological advancements (e.g., switch from film to digital format, increase in trigger speed and storage capability), and development of statistical methods for estimating (relative) abundance and density have resulted in an exponential increase in the number of camera trap studies, especially those targeting secretive and elusive terrestrial vertebrates (Rovero and Zimmermann 2016). Additionally, the educational and outreach potential of the images collected, paired with the ease of use of camera devices, have led to the development of citizen science programs aimed at engaging the public in image processing and data collection (e.g., Snapshot Safari: snapshotsafari.org; eMammals: https://emammal.si.edu/). Camera traps were initially used for species checklists (Tobler et al. 2008) and for estimating population density for species in which individuals could be individually identified (Karanth and Nichols 1998), but they are now routinely used to quantify relative abundance (Rovero and Marshall 2009), occupancy (Pettorelli et al. 2010, Rich et al. 2016, Scotson et al. 2017), and density (Sun et al. 2017), to study animal behavior (Caravaggi et al. 2017) and diel activity patterns (Wang et al. 2015, Frey et al. 2017, 2020), and to investigate species richness and community composition at a global scale (Ahumada et al. 2011). Despite their widespread use, the field of camera trapping is still young and evolving, and several questions related to the use of this technology require further investigation (Meek et al. 2015). In this dissertation, I focus on two aspects related to survey design and analysis of camera-trap data: differential response across species to survey design strategies in studies aimed at simultaneously collecting data on multiple species (chapter 1), and statistical approaches to minimize or account for correlation when analyzing camera-trap data (chapter 2 and 3). In Chapter 1, I focus on differential responses of North-American carnivores to survey-design strategies often used in studies that target multiple species (i.e., multi-species studies). Cameras allow investigators to collect massive amounts of fine-resolution information on a large number of species simultaneously and have been described as an ideal, cost-effective tool for large-scale multi-species monitoring programs (Steenweg et al. 2017, Ahumada et al. 2019). Nevertheless, researchers have only recently started to quantify consequences of species-specific responses to survey-design strategies when the aim of the study goes beyond estimating species richness (Pease et al. 2016, Evans et al. 2019, Mills et al. 2019, Buyaskas et al. 2020, Holinda et al. 2020). Species’ characteristics, such as movement patterns, behavior and home range size affect their response to survey design; hence, in multi-species studies, a survey design that is appropriate for one species, might not work for another. This chapter was motivated by the Minnesota Department of Natural Resources’ (MN DNR) need for an approach capable of simultaneously monitoring multiple species of carnivores for informing conservation and management decisions. In collaboration with MN DNR biologists, I carried out a multi-year camera-trap study to assess species-specific responses to different survey strategies for 10 species of carnivores. I evaluated responses to two different survey-design frameworks (random- versus road-based), two different lure types (salmon versus fatty acid scent oil), two different placement strategies (completely random versus randomly-selected sites with feature-based placement), and survey timing (spring versus fall); I also assessed temporal trends in daily encounter probabilities within each season. I collected data at 100 locations in northern Minnesota during each of five 6-week long surveys between spring 2016 and spring 2018, accumulating more than 2 million pictures. Using Generalized Linear Mixed Models (GLMMs), I quantified species-specific responses to the different design strategies and found that differences were particularly strong for choice of survey-design framework and lure type. In Chapter 2, I illustrate how the lorelogram, a statistical tool for quantifying correlation in binary responses (Heagerty and Zeger 1998), can be used in ecological contexts to explore spatial and temporal correlation in binary data. Camera traps collect images near-continuously, resulting in data that are temporally correlated (Meek et al. 2014). A common strategy to minimize dependence among observations of a species collected at a certain site is data aggregation; images of a species are assumed to belong to the same encounter event whenever they occur at the same site and within some temporal threshold (e.g., 30 min apart). The temporal interval between subsequent pictures collected at the same site is often chosen arbitrarily (Burton et al. 2015) because methods to explore correlation in binary data (e.g., camera-trap detection/non-detection) are rare in the ecological literature. In this chapter, I show how the lorelogram, a statistical method commonly applied in the biomedical literature, can be used to describe correlation in camera-trap data and help with selecting appropriate time intervals to minimize serial correlation. Although the initial motivation behind this section of my dissertation was to provide a statistical tool for robustly selecting temporal intervals to minimize serial dependency in camera-trap data, the lorelogram can also be applied to other binary data exhibiting spatial or temporal correlation (e.g., distributional data). I demonstrate how the lorelogram can quantify spatial/temporal correlation using data from chapter 1 and data from the North-American Breeding Bird Survey (Pardieck et al. 2018). To facilitate the use of the lorelogram by analysts, I created an R-package that is freely available for download at https://github.com/FabiolaIannarilli/lorelogram. In Chapter 3, I illustrate how GLMMs can be used to describe diel activity patterns from camera-trap data. Understanding how individuals change their patterns of activity in response to the presence of competitors and natural and anthropogenic stressors is the goal of many ecological studies, and is important for predicting how species will adapt to future changes (e.g., climate change and land-use modifications). Camera-trap data collected at multiple sites are likely to be affected by site-to-site variability due to animals’ responses to differences in the biotic (e.g., presence of competitors; levels of human disturbance) and abiotic (e.g., proximity to roads; sampling treatments) characteristics of the sites sampled. Minimizing serial correlation by aggregating the data, as described in chapter 2, does not remove dependencies from repeated measures or account for site-to-site variability in frequency of site-use or in the timing of peak activity levels. Kernel Density Estimators (KDEs) are the most used approach to describe diel activity patterns from camera-trap data; circular KDEs, in particular, can be used to account for periodicities in diel activity patterns. This method, however, assumes independence among observations and, thus, ignores site-to-site variability common in camera-trap studies. In contrast, the GLMM approaches presented in this chapter can account for several types of variability (e.g., in frequency of site-use and correlation among repeated measures) using covariates and random effects; they also provide measures of the level of activity at sites characterized by different conditions. Like KDEs, the GLMM approach can accommodate the periodic nature of activity-pattern data, and can test hypotheses about changes in diel activity patterns due to biotic and abiotic factors, but they do so in a more direct way than KDEs (e.g., using AIC or Likelihood Ratio Tests). Using a simulation study, I compare the accuracy of GLMMs and KDEs when estimating activity patterns from camera-trap data. Using data from chapter 1, I also provide examples of how GLMMs can be used to explore diel activity patterns under different conditions (e.g., sampling treatments, seasons, or presence of competitors). Throughout the rest of the dissertation, I will use the first-person plural voice, ‘we’, to reflect the collaborative nature of the work presented. The second chapter of this dissertation is published in Methods in Ecology and Evolution (Iannarilli et al. 2019) and the material in chapter 1 is currently under review for publication in Wildlife Biology.