Scalable Bird Diversity: Mastering High-Chorus Soundscape and Species Richness with Time-frequency YOLOv5
Year: 2026
Type: preprint
Abstract: Bioacoustic monitoring of birds plays an important role in avian monitoring and can be valuable in developing conservation strategies. In this study, we developed an automatic detection model for 27 bird species recorded along a 1300-km north-south gradient in Québec, Canada. One of the main challenges addressed in our research is the problem of chorus inherent in neural network detection, in which the overlap of vocalisations from multiple species constitutes a significant obstacle for accurate detection. We demonstrate that the YOLOv5 visual detection model provides accurate identification of individual species in complex acoustic environments using bounding boxes. The model achieved a cmAP of 63% in cases of intense chorusing, compared to 54% for BirdNET, and a cmAP of 60% in all chorus cases, compared to 55% for BirdNET. These results confirm the e!ciency of our approach and the relevance of class-specific thresholds. Furthermore, the model predicted avian species richness fairly accurately (Spearmans rho=62%). This neural network model will therefore be used to monitor distributional changes of these 27 species across Québec.
Source: HAL
Language: English
Topic: Music and Audio Processing
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
SDG: Life on land
Open Access status: green
OpenAlex ID: W7204283463