Doi: https://doi.org/10.13157/arla.72.2.2025.fo1
Authors: Cristian PÉREZ-GRANADOS
E-mail: cristian.perez@ctfc.cat
Published: Volume 72.2, July 2025. Pages 149-159.
Language: English
Keywords: convolutional neural network, detection radius, machine learning, model confidence and threshold
Summary:
BirdNET is a machine learning tool capable of identifying 6,500 bird species worldwide. It is available on different platforms but with varying versions and algorithms, which may lead to misleading results. For this Forum, I reanalysed recordings from Pérez-Granados (2023) using BirdNETAnalyzer, the research-focused platform of BirdNET, instead of the demo platform (BirdNET-Api) used in the original study. BirdNET-Analyzer outperformed BirdNET-Api by detecting 28% more vocalisations and maintaining high detection probabilities over greater distances (75 metres vs. 50 metres). Unlike in the original study, there were significant differences in detection probabilities among recorder models and between the species more often detected. BirdNET-Analyzer confidence scores also decreased significantly with increasing distances from the bird to the recorder, and varied between species, and recorders – relationships not observed in the original study. unlike in the original study, significant differences were found in detection probabilities among recorder models and in which species were more frequently detected. These findings may offer insights into future passive acoustic bird surveys using BirdNET, highlighting a trade-off between precision and detection radius and aiding in the modelling of detection radii for improved bird density estimation and comparability between studies. Overall, the variability in confidence scores across distances, recorders and species, emphasises the need for further research to optimise large-scale monitoring programmes using BirdNET, which often require identifying optimal species-specific confidence score thresholds. Finally, this Forum highlights the importance of preferring BirdNET-Analyzer for research purposes to avoid misleading interpretations.
Supplementary Material:
Downloadable 1
