Performance of unmarked abundance models with data from machine-learning classification of passive acoustic recordings

The ability to conduct cost-effective wildlife monitoring at scale is rapidly increasing due to the availability of inexpensive autonomous recording units (ARUs) and automated species recognition, presenting a variety of advantages over human-based surveys. However, estimating abundance with such data collection techniques remains challenging because most abundance models require data that are difficult for low-cost monoaural ARUs to gather (e.g., counts of individuals, distance to individuals), especially when using the output of automated species recognition. Statistical models that do not require counting or measuring distances to target individuals in combination with low-cost ARUs provide a promising way of obtaining abundance estimates for large-scale wildlife monitoring projects but remain untested. We present a case study using avian field data collected in the forests of Pennsylvania during the spring of 2020 and 2021 using both traditional point counts and passive acoustic monitoring at the same locations. We tested the ability of the Royle-Nichols and time-to-detection models to estimate the abundance of two species from detection histories generated by applying a machine-learning classifier to ARU-gathered data. We compared abundance estimates from these models with estimates from the same models fit using point-count data and to two additional models appropriate for point counts, the N-mixture model and distance models. We found that the Royle-Nichols and time-to-detection models can be used with ARU data to produce abundance estimates similar to those generated by a point-count-based study but with greater precision. ARU-based models produced confidence or credible intervals that were on average 31.9% (+/- 11.9 SE) smaller than their point-count counterpart. Our findings were consistent across two species with differing relative abundance and habitat use patterns. The higher precision of models fit using ARU data is likely due to higher cumulative detection probability, which itself may be the result of greater survey effort using ARUs and machine-learning classifiers to sample significantly more time for focal species at any given point. Our results provide preliminary support for the use of ARUs in abundance-based study applications, and thus may afford researchers a better understanding of habitat quality and population trends, while allowing them to make more informed conservation recommendations and actions.