Simulation-based assessment of the performance of hierarchical abundance estimators for camera trap surveys

The need for knowledge about abundance to guide conservation and management strategies in combination with low detectability of many species has led to a widespread use in ecology and management of a range of hierarchical models (HMs) for abundance. These models also appear like good candidates for inference about local abundance in nature reserves studied by camera traps. However, the best choice among these models is unclear, particularly how they perform in the face of several complicating features of realistic populations that include: (i) movements relative to sites, (ii) multiple detections of unmarked individuals within a single survey, and (iii) low probabilities of detection. We conducted a simulation-based comparison of three HMs (Royle-Nichols, binomial N-mixture and Poisson N-mixture model) in the context of small populations of elusive animals in a single study area, where animals cannot be distinguished individually and hence double counting occurs. We generated count data by simulating camera traps monitoring individuals moving according to a Gaussian random walk. Under the simulated scenarios none of the three HMs yielded accurate abundance estimates. Moreover, the performance of each HM depended on the interpretation of abundance. By pooling abundance estimates for trend estimation, each models performance markedly improves. Overall, the Royle-Nichols and Poisson N-mixture models outperform a binomial N-mixture model. This emphasizes the importance of choosing the appropriate HM for the data problem.