New article: Domain-specific neural networks improve automated bird sound recognition already with small amount of local data

Sebastian Andrejeff

@andse
| he/him

Coordinator of the citizen science project Bird Sounds Global (BSG). Its aim is to produce training and testing material for machine learning induced automated bird species recognition.

Groups

Full article here: Lauha, P., Somervuo, P., Lehikoinen, P., Geres, L., Richter, T., Seibold, S., & Ovaskainen, O. (2022). Domain-specific neural networks improve automated bird sound recognition already with small amount of local data. Methods in Ecology and Evolution, 00, 1–12

Species-specific model comparison for test data

A neural network was fitted with global data and fine-tuned with local data from Finland that significantly improved the performance for recordings from the same region and even outperformed the current state-of-the-art tool for automatic bird sound classification, BirdNET. Already a small amount of data can improve the model performance. The data were labelled in the citizen science platform Kerttu, which served as a pilot project for the ongoing Bird Sounds Global (BSG).

BSG is about developing automated bird sound identification by combining machine learning and citizen science. It's part of the LIFEPLAN research programme, which aims to improve understanding of the global biodiversity and its driving factors.

The end products of BSG are an expert-validated sound-feature library for birds of the world and locally refined species-identification models for bird vocalisations. The end products will be openly available to everyone for further use. Open-access recognition models provide opportunities for other applications and their further development, such as the current state-of-the-art software applications, Merlin & BirdNET. Both Merlin and BirdNET are working well for targeted recordings, but since the LIFEPLAN project is producing data audio data on soundscapes (i.e., passive recordings of all noises in the environment at a locality, rather than recording a single species), further method development is needed. Fine-tuning the models with data annotated by ornithological experts is a crucial and important part of this development. This will increase not only possible applications for research, monitoring and protection of birds but also applications for birdwatchers and other naturalists.

You can check out the portal at https://bsg.laji.fi/identification/instructions, and harness your expertise to advance automated bird-sound identification, understanding of global biodiversity and enhance your own sound-recognition skills. Also, feel tree to spread the word your birding friends.

Current selection of recording sites. Can you find a site near you?

https://twitter.com/BirdSoundGlobal

https://www.facebook.com/people/Bird-Sounds-Global/100075763524968