U‑Net Model Detects Lion Roars from Collar Motion Data

Scientists have long relied on microphones to capture lion roars, but the equipment drains batteries and fills storage quickly. Accelerometers in GPS collars record tiny movements in three directions, producing acceleration data (ACC) that can reveal behavior without audio. Machine learning can turn those motion patterns into recognizable signals. The GAIA Initiative sought a solution that works for both sexes and for lions that move while vocalizing.

The team recorded 1,333 lion roaring events during fieldwork in Etosha National Park. They used seven lions that wore both a GPS collar with an accelerometer and an audio logger for several months. The audio logger provided the reference labels needed to train the algorithm. Researchers built a U‑Net, a convolutional neural network originally designed for image segmentation, to process the ACC time series. The model learned to distinguish the subtle motion signature of a roar from other movements such as walking or running. When tested, the U‑Net identified roars with 90% to 96% accuracy. It performed equally well for male and female lions and for calls produced while the animals were stationary or in motion.

Removing the need for microphones cuts power use and storage demands, enabling longer monitoring periods. Researchers can now link roaring to precise GPS locations and movement patterns, improving studies of pride coordination and territorial behavior. The approach may also be adapted to other species that communicate through low‑amplitude vibrations, such as elephants or wolves. Conservation managers could receive near‑real‑time alerts when a lion roars near human settlements, helping to prevent conflict. Future work will test the model in different habitats, integrate it with satellite‑based alert systems, and evaluate its performance across seasons and social contexts.