Modelling misclassification in multi-species acoustic data when estimating occupancy and relative activity

Abstract

Surveying wildlife communities provides data for informing conservation and management decisions that affect multiple species. Autonomous recording units (ARUs) can efficiently gather community data for a variety of taxa, but generally require software algorithms to classify each recorded call to a species. Species classification errors are possible during this process and result in both false-negative and false-positive detections. Available approaches for analysing ARU data do not model the species classification probabilities, meaning erroneous detections are attributed to an omnibus source instead of the presence of another species. Additionally, counts of call recordings for each species are often summarized to binary detection data for analyses. Expanding statistical models to capture these nuances of ARU data would allow for improved inferences about occupancy and relative activity. Motivated by bat acoustic surveys, we developed a model to analyse counts of call recordings from multiple species simultaneously while accounting for species classification errors. Our model expands on previously developed false-positive occupancy models to better describe acoustic data. We used simulations to compare our model to other false-positive occupancy models for an example scenario with ARU data from two species. We also analyse acoustic data for eight bat species in Montana using our model. In simulations, single-species models resulted in biased estimates of occupancy and relative activity because they failed to associate false positives with the presence of the second species. Models analysing binary observations ignored available information on relative activity and led to less precise estimates. Applying our model to bat acoustic data from Montana allowed for species-specific estimates of occupancy and relative activity. This analysis illustrates the flexibility in our model framework while also highlighting the assumptions and data requirements for implementation. Specifically, additional information on the species classification probabilities is needed and we discuss considerations for reliably estimating these parameters. Directly modelling the species classification probabilities allows for improved ecological inferences for both occupancy and relative activity using community ARU data. Our statistical framework helps address the challenges posed by acoustic data, allowing ecologists to better utilize this technology to monitor wildlife communities.