UU Birdsong
AI & Data
Semester programme:Minor AI for Society
Research group:Sustainable Data & AI Application
Project group members:Cédric Berden
Jimmy Bogers
Rik van Bilsen
Stijn Hoeks
Chantal Maas
Anish Thakur
Mees Vaessen
Project description
How can the Birdnet system be developed to identify bird species from environmental audio recordings from the netherlands, and to what extent can such a system be used to monitor and validate biodiversity conservation efforts?
Context
This research focuses on recognizing birds that visit Dutch farmlands using audio recordings, with an emphasis on prioritized species. By automatically identifying bird calls, we aim to gain a comprehensive overview of the bird populations frequenting agricultural landscapes. Understanding which species are present and how their numbers change over time is critical for assessing the effectiveness of biodiversity efforts.
The project leverages bird sound data from open sources such as XenoCanto and Macaulay, alongside machine learning tools like BirdNET, to develop an automated recognition system. By processing and analyzing environmental audio, the system can provide real-time insights into bird diversity, reducing the need for labor-intensive manual surveys.
The potential impact on nature is significant. Automated bird monitoring supports the protection of ecosystems and helps maintain biodiversity. Long-term, it enables efficient tracking of species populations, informing conservation strategies and promoting greener farmlands. By integrating technology into ecological research, this approach contributes to making the Netherlands richer in birdlife while supporting sustainable farming practices. Ultimately, automated recognition of bird sounds not only aids in biodiversity monitoring and research but also strengthens efforts to safeguard farmland ecosystems for the future.
Results
The results of the project are mixed but predominantly positive. For many of the prioritized bird species, the system delivers reliable recognition, surpassing the baseline performance of BirdNET. This demonstrates that automatic sound recognition has strong potential for monitoring bird populations on Dutch farmlands.
However, there are notable limitations. For some species, there is simply too little audio data available to train the model effectively, resulting in lower accuracy for these species. Additionally, hardware constraints present a challenge: downloading the training data and, more importantly, the training process itself is slow, with some sessions taking up to 12 hours. This limits the scalability of the system and its ability to process larger datasets efficiently.
Despite these challenges, the results indicate that automated bird sound recognition has significant potential for biodiversity monitoring. Improvements in dataset size and computational resources could further increase accuracy and make the system suitable for long-term, large-scale ecological surveys.
About the project group
We all have a background in software engineering most from Fontys and one from India.
We've worked on this project since last September till this January 2 times a week.