Machine learning pipeline extracts biologically significant data automatically from avian monitoring videos

Alex Hoi Hang Chan, Liu Jing Qi, Terry Burke, Will Pearse, Julia Schroeder

Measuring parental care behaviour in the wild is central to the study of animal life-history trade-offs, but is often labour and time-intensive. More efficient machine learning-based video processing tools have recently emerged that allow parental nest visit rates to be measured using video cameras and computer processing. Here, we used open-source software to detect movement events from videos taken at the nest box of a wild passerine bird population. We show that visit numbers from our automatic data collection pipeline strongly correlate with human observations and predicts an increase in brood fitness. Using a machine learning-assisted annotation approach on a subset of 18 videos, we show that the accuracy largely increased and cut annotation time by an average of 5.5x compared to that of a cohort of undergraduate students. Since our automatic pipeline collected biological-meaningful data that would have taken approximately 800 days by human observers, we encourage more researchers to apply existing open-source tools to assist data collection in animal behaviour studies.