Meine Cookie EinstellungenDetermining species-specific thresholds to improve precision in passive acoustic monitoring
Scanferla J, Brambilla M, Brambilla G, Hilpold A, Marchetti AE, Puff F, Tappeiner U, Anderle M
Andreas Hilpold, Ulrike Tappeiner, Jarek Scanferla, Matteo Anderle
Passive acoustic monitoring is growing in popularity and utilization worldwide. Automatic and AI-driven species identification is a powerful tool for analysing acoustic data, but its precision varies widely between species, posing a major challenge for reliable monitoring. Species-specific thresholds offer a solution by optimizing model performance for each species individually. Using Alpine birds as a case study and an expert-annotated dataset of 25 h and 17,737 bird sounds, we provide 72 species-specific thresholds for filtering detections from BirdNET (a widely used sound identification software). The thresholds were calculated with a binomial logistic regression, using BirdNET identification scores as a predictor and the binary annotation (species present or not within the recording) as outcome. They significantly enhanced precision of the identification model (mean increase of 0.30 SD = 0.21), making them efficient and practical tools for passive acoustic monitoring. Moreover, our subsampling approach revealed that for several species, the number of annotated recordings and therefore the time effort can be substantially reduced. This approach could greatly benefit similar monitoring schemes, as it can be replicated and applied over different systems and taxa, leading to a significant reduction in the time taken to process passive acoustic data. We demonstrated the applicability and advantages of using thresholds in mountainous regions, where data collection is particularly difficult due to complex topography and harsh conditions, and where passive methods may therefore play a crucial role. Finally, our first list of 72 species-specific thresholds could serve for future species identification in similar European areas.
DOI: 10.1016/j.ecoinf.2025.103423
