Abstract

Using Machine Learning Models for Biodiversity Monitoring:
A Case Study with Gibbon

Thinh Tien VU^1,3, Hoa Thi NGUYEN², Dena Jane CLINK³

¹Department of Wildlife, Faculty of Forest Resource and Environment Management, Vietnam National University of Forestry, vtthinhvnuf@gmail.com
²Institute for Tropical biodiversity and Forestry, nguyenhoa94vfu@gmail.com
³K. Lisa Yang Center for Conservation Bioacoustics, Cornell Lab of Ornithology, Cornell University, djc426@cornell.edu

, Currently, there are 20 gibbon species worldwide. Monitoring usually involves placing multiple recorders in the forest to capture gibbon calls over several days, generating large volumes of audio data that require extensive manual processing. Machine learning tools, such as BirdNET, have recently enabled automatic sound recognition. Deep learning and automatic detection methods have yet to be widely applied in real-world studies of gibbon ecology and behavior. In this study, we used BirdNET to detect gibbon calls from recordings made at Chu Mom Ray National Park, Vietnam, in 2019. We evaluated how detector efficiency varied by band-pass frequency and sample size to select the optimal configuration. Predictions by BIRDNET were verified, and the number of gibbon call segments was calculated for the entire park. We examined how these numbers varied by time of day and environmental variables. The optimal detector range was 600–1800 Hz. With as few as 15 samples per sex, precision and recall exceeded 80%, and with 30 samples, both metrics reached about 90%. In total, we detected 2,162 female and 8,826 male call segments. Most calls (>97%) occurred between 5:15 and 7:45. The strongest associations with environmental variables were found for Rich and medium forest, Tree canopy height, Tree canopy cover, and Distance to village. This approach provides valuable data for occupancy models and facilitates long-term comparisons of gibbon abundance and ecological research.