Two new papers on avian biodiversity monitoring

I’m excited to share two new papers related to acoustic monitoring of bird bird communities!

1. ‘BirdNET: A deep learning solution for avian diversity monitoring‘ by Stefan Kahl, me, Maximilian Eibl, and Holger Klinck was published in Ecological Informatics.
2. ‘Survey coverage, recording duration, and community composition affect observed species richness in passive acoustic surveys‘ by me, Stefan Kahl, Phil Chaon, Zach Peery, and Holger Klinck was published in Methods in Ecology and Evolution.

The Ecological Informatics paper (available HERE) provides a technical description of a new machine learning algorithm, BirdNET, which can identify 984 bird species by sound – over 95% of the species found in North America and Europe. It was trained with ~1.5 million weakly labeled samples and validated with focal recordings, fully annotated data, and over 33,000 hours of soundscape data recorded at an eBird hotspot. The key performance metric is “mean average precision”, which was 0.79 – meaning that most predictions for most species are correct. If you have any questions about BirdNET, I encourage you to contact Stefan. I’d also like note that there is a free BirdNET app available for iOS and Android, and the current species count is about 3,000.

The MEE paper (available HERE) draws on simulated bird communities as well as soundscape data from the Sierra Nevada (1,000 hours) and the Cornell Lab of O’s Sapsucker Woods (750 hours) which we analyzed with BirdNET to evaluate how different passive acoustic survey designs affected observed species richness. In short, rarefaction curves indicates that longer recordings are better, but by 28 hours (4 hrs at dawn per day for 7 days) species richness is increasing very slowly. Species composition and environmental heterogeneity have important implications for survey design as well. An ancillary finding with exciting implications for our upcoming work is that we identified over 100 species in the Sierra Nevada with just a week’s worth of data at 28 locations. If you have any questions about this paper, please contact me.