Assistant Professor University of Pittsburgh Pittsburgh, Pennsylvania, United States
Automated acoustic monitoring is an increasingly popular means of assessing populations, particularly for species that are rare or difficult to detect. These recordings can be readily evaluated by human listeners, sometimes in concert with automated classification models, to determine occupancy, the presence or absence of a species at a particular location and time. However, since methods to uniquely identify the songs or calls of individual organisms are very rarely available, estimating density from acoustic recordings represents a substantially harder challenge. This constitutes an important shortcoming of automated acoustic monitoring because estimates of density are needed to answer questions regarding population size. Here, we use simulations to test the ability of two existing hierarchical modeling frameworks to estimate avian abundance directly from acoustic recordings. The Bernoulli-Poisson model takes as input presence/absence surveys, while a variation of distance sampling takes presence/absence in discrete distance bins. Beyond an estimate of the distance to an individual, neither method requires individuals to be distinguishable in audio recordings. We perform simulations covering a range of realistic values for species density and detection probability, then evaluate the performance of each model in light of typical densities for songbirds in an eastern deciduous forest community.
Simulation results reveal that a Bernoulli-Poisson model recovers accurate estimates of abundance for low to moderate densities (< 5 individuals per survey) even when detectability is low, but substantially underestimates abundance for higher densities. When detection probability and density are both high, this model often fails to converge. The variation on distance sampling performs similarly but avoids convergence failures. In both models, very rare species with low detection probabilities occasionally led to inaccurate estimates. In the context of an eastern deciduous forest songbird community, our results suggest that these models could produce reliable abundance estimates for all but the rarest (e.g., Cerulean Warbler) and most abundant (e.g., Chestnut-sided Warbler) species. These conclusions depend on several underlying assumptions about population and song rate distributions that are difficult to test empirically in the field. Validating these models with human-conducted abundance surveys will be critical to their use as predictive tools. We discuss localizing individuals in synchronized audio recordings as an alternative approach for estimating abundance. Finally, we conclude by discussing the application of these and other statistical models to large-scale acoustic recording datasets in concert with automated birdsong recognition.