Density estimation of non-independent unmarked animals from camera traps

Abstract

Population density estimation is essential for wildlife management and conservation. Camera trap (CT) surveys have become popular as non-invasive monitoring, and several methods have been developed to estimate population densities of individually unrecognizable animals exclusively from CT surveys, including the random encounter model (REM), the random encounter and staying time (REST) model, the instantaneous sampling (IS) model, and the space-to-event (STE) model. However, these models assume that individual detections are independent events. This assumption is violated when animals live in groups and behave correlatedly. Thus, this study aims to compare the performances of the REM, the REST, the IS, and the STE when animals move correlatedly in groups. Their extensions to account for overdispersion of detection count data and an application of group encounters in the REM were also examined. Seven group sizes and two cohesive patterns were simulated using correlated random walk moving biased to the group centers mimicking red deer, and the detection data were fitted to the models. A goodness-of-fit procedure was used to evaluate the fit of the models. It was found that the original model of the IS and the REST, and the REM of the individual encounters to were robust to the violation of independence in detections while the STE underestimated the density; they were, however, poor at data fitting. Although their extensions to the overdispersion improved the model fitting, the negative binomial (NB) models biased the estimates upward in the IS and the REST due to the limitation of sample size, the zero-inflated Poisson in the IS did not bias the estimates but was still poor at fitting in large groups, and the Lomax in the STE did not resolve the underestimation. Applying the group encounters in the REM also biased the density upward. The overestimation in the NB model occurs when the Bayesian p-value exceeded 0.5. This relationship would be informative in evaluating the possibility of overestimations in the NB model.