Spatiotemporal hierarchical modelling of species richness and occupancy using camera trap data

Abstract

Summary Over the last two decades, a large number of camera trap surveys have been carried out around the world and camera traps have been proposed as an ideal tool for inventorying and monitoring medium to large‐sized terrestrial vertebrates. However, few studies have analysed camera trap data at the community level. We developed a multi‐session multi‐species occupancy model that allows us to obtain estimates for species richness and occupancy combining data from multiple camera trap surveys (sessions). By estimating species presence at the session‐level and modelling detection probability and occupancy for each species and sessions as nested random effects, we could improve parameter estimates for each session, especially for species with sparse data. We developed two variants of our model: one was a binary latent states model while the other used a Royle–Nichols formulation for the relationship between detection probability and abundance. We applied both models to data from eight camera trap surveys from south‐eastern Peru including six study sites, 263 camera stations and 17 423 camera days. Sites covered protected areas, a logging concession and Brazil nut concessions. We included habitat (terra firme vs. floodplain) as a covariate for occupancy and trail vs. off‐trail as a covariate for detection. Among‐camera heterogeneity was a serious problem for our data and the Royle–Nichols variant of our model had a much better fit than the binary‐state variant. Both models resulted in similar species richness estimates showing that most of the sites contained intact large mammal communities. Detection probabilities and occupancy values were more variable across species than across sessions within species. Three species showed a habitat preference and four species showed preference or avoidance of trails. Synthesis and applications. Our multi‐session multi‐species occupancy model provides improved estimates for species richness and occupancy for a large data set. Our model is ideally suited for integrating large numbers of camera trap data sets to investigate regional and/or temporal patterns in the distribution and composition of mammal communities in relation to natural or anthropogenic factors or to monitor mammal communities over time.