Animal-worn sensors, especially accelerometers, are increasingly used with machine-learning models to identify animal behaviors. These tools often struggle with uneven training data, uncertain predictions, and noisy results. To address these issues, Dr. Rafiq and Dr. Abrahms, with their collaborators, developed an open-source method that combines machine learning and statistical techniques to improve behavior classification and to provide “prediction sets,” which give a range of likely behaviors with a known level of certainty. Using simulations and data from free-ranging African wild dogs in Botswana, they show that this approach greatly improves accuracy after applying quality checks and balancing the data, especially for rare but important behaviors like feeding. This method can be used for many species and represents an important step toward better, continuous monitoring of animal behavior.
Authors: Medha Agarwal, Kasim Rafiq, Ronak Mehta, Briana Abrahms, Zaid Harchaoui
Journal: Methods in Ecology and Evolution
DOI: 10.1111/2041-210x.70206
Photo credit: Marie-Pier Poulin
