Deep learning for passive acoustic monitoring: how to study changing phenology in remote areas

Abstract

Understanding how species adjust to seasonality is fundamental in ecology, especially with rapidly increasing global air temperatures. Bioacoustic monitoring offers promise for tracking shifts in seasonal timing of vocal species, as recent automated sound recorders enable large-scale and long-term data collection. Yet, analyzing vast datasets necessitates automation and innovative detection methods. Here, we introduce BioSoundNet, a deep learning model designed for bird vocalization detection. Trained on field data and open-access databases, BioSoundNet achieved AUC scores of 0.88-0.93 and average precisions of 0.87-0.97 across five datasets spanning various ecosystems, and effectively captured the temporal patterns of avian acoustic activity at different time scales. Our findings underline the importance of evaluating models in ecological contexts and to address the potential consequences of missing detections. Operating efficiently on standard computers, BioSoundNet is a robust tool for automated bird vocalization detection, providing a valuable resource for ecological phenology studies and acoustic dataset analysis.