With data from Norway, Brazil, Taiwan and Costa Rica, a team of researchers show that large-scale monitoring of avian vocalization can deliver immediate applied impact. The results are now published in PNAS.
(A) More than 150,000 hours of acoustic data were recorded. (B) A state-of-the-art convolutional neural network model was used to detect and classify bird vocalizations. (C) Experts manually labeled a subset of the detections. (D) Filtered detections were used to derive reliable avian biodiversity insight across spatial and temporal scales.
Monitoring trends and dynamics of biodiversity is crucial if we are to address steadily increasing global environmental challenges. Using animal vocalizations in already existing audio data offers an inexpensive and taxonomically broad way to monitor these trends. However, expertise required to label new data and fine-tune algorithms has been a barrier.
Unlocking the potential of acoustic monitoring
A team of researchers have now shown that a single vocalization detection model can deliver community- and species-level insights across diverse datasets, unlocking the scale at which acoustic monitoring can deliver immediate applied impact.
The declines in cost and increased accessibility of robotics platforms and electronic sensors have transformed our ability to survey ecosystems at larger scales, says Carolyn Rosten, Researcher at the Norwegian Institute for Nature Research NINA.
In this study, the team applied a pretrained bird vocalization detection model (BirdNET) to more than 150,000 hours of audio datasets from Taiwan, Costa Rica, Brazil, and Norway. They manually listened to a subset of detections for each species in each dataset and found precisions of over 90% for 109 of a total of 136 species.
Growing libraries of sound
Bird vocalizations have been recorded by hobbyists and scientists for decades culminating in rich libraries of annotated data which span the globe.
If training datasets are able to grow in size and accessibility while addressing systematic taxonomic and geographic biases, the performance of machine learning models will continue to improve, says Rosten. This could unlock further opportunities for fully autonomous acoustic monitoring to be used at scale and deliver impact around the world, she concludes.
Read the article: Sarab S. Sethi et al.: Large-scale avian vocalization detection delivers reliable global biodiversity insights, PNAS, 2024.
Contact: Carolyn Rosten