Assessing the potential of BirdNET to infer European bird communities from large-scale ecoacoustic data

• DOI: 10.1016/j.ecolind.2024.112146
• Publication Date: 2024-05-20T23:22:56Z
• Authors: David Funosas; Luc Barbaro; Laura Schillé; Arnaud Elger; Bastien Castagneyrol; Maxime Cauchoix
• Keywords: Identification; Soundscape; Citizen Science

Passive acoustic monitoring has become increasingly popular as a practical and cost-effective way of obtaining highly reliable data in ecological research projects. Increased ease collecting these means that, currently, the main bottleneck ecoacoustic projects is often time required for manual analysis passively collected recordings. In this study we evaluate potential current limitations BirdNET-Analyzer v2.4, most advanced generic deep learning algorithm bird recognition to date, tool assess community composition through automated large-scale data. To end, 3 datasets comprising total 629 environmental soundscapes 194 different sites spread across 19° latitude span Europe. We analyze using both BirdNET listening by local expert birders, then compare results obtained two methods performance at level each single vocalization entire recording sequences (1, 5 or 10 min). Since provides confidence score identification, minimum thresholds can be used filter out identifications with low scores, thus retaining only ones. The volume did not allow us estimate species-specific taxa, so instead evaluated global selected optimized when consistently applied all species. Our analyses reveal that if sufficiently high threshold used. However, inevitable trade-off between precision recall does obtain satisfactory metrics same time. found F1-scores remain moderate (<0.5) studied, extended duration seem currently necessary provide minimally comprehensive picture target community. estimate, however, usage species- context-specific would substantially improve benchmarks study. conclude judicious use AI-based provided represent powerful method assist assessment data, especially duration.