Accounting for noise improves the generalization of acoustic indices when estimating biodiversity

Abstract

The use of Passive Acoustic Monitoring (PAM) for estimating biodiversity often entails the use of acoustic indices that summarize the acoustic properties of the recordings. Although acoustic indices have been shown to successfully track biodiversity in limited contexts, the recurrent failure of acoustic indices generalization across domains has raised doubts regarding the reliability and scalability of PAM. Variation across ecological communities is often cited as a principal limit to the generalizability of acoustic indices-based biodiversity estimation, while environmental noise variability is often overlooked. Background noise can change between recordings because of environmental (e.g., wind) and system (e.g., microphone) variability that make noise equalization across studies difficult. In this study, we use both synthetic soundscapes and field recordings to assess the effects of background noise on acoustic index generalization. We reveal a dramatic effect of background noise on the ability of acoustic indices to generalize biodiversity estimates across datasets. We demonstrate that careful selection of noise-invariant acoustic indices, accounting for the signal-to-noise ratio, training across noise types and segmentation of the recordings to concentrate biophony all improve acoustic index-based biodiversity generalization. We also demonstrate the use of soundscape simulation training for acoustic index-based biodiversity estimations and generalization. We conclude that acoustic indices may have more potential as a simple and cost-effective biodiversity monitoring tool than previously thought and we highlight several possibilities for mitigating background noise variability during training to improve generalization.