Framework for consistent BirdNET detection labelling to prevent biases in downstream community predictions

Passive Acoustic Monitoring (PAM) has become a cost-effective tool in ecological research but generates large volumes of data requiring extensive processing. BirdNET v2.4 is a machine-learning tool that facilitates soundfile analysis by automatically detecting a large number of species within acoustic datasets. However, the ability of BirdNET to correctly detect species is inconsistent. Recent studies recommend the use of Species-Specific Confidence Thresholds (SSCTs) to optimize the trade-off between detection performance and data loss. However, SSCTs require preliminary manual annotations of BirdNET detections to be inferred and the pool of labelled samples is likely to have an impact on SSCT predictions. Using 31 European bird species detected across two independent study areas, we propose a framework to evaluate how many labelled samples are needed to ensure consistent ecological predictions. For each species, a benchmark of 100 manually annotated detections was used to simulate a wide range of annotation pool sizes through bootstrap resampling. It reveals a constant overestimation of predicted call abundance with small annotation pools. Given our set of 31 species, abundance estimations along temporal scales stabilize on average with 57 annotated samples and at worst with 90 samples. Using Euclidean distance enabled to compare occurrence patterns across temporal scales with different size of annotation pools and assess above which number of labelled samples predicted trends stabilize. As PAM surveys are growing for bird conservation policies, this scalable framework looks crucial to ensure the relevancy of SSCTs to accurately predict bird temporal and spatial ranges.