This study investigated the accuracy of animal identification based on herd behaviour from drone camera footage. We evaluated object detection algorithms and point pattern analysis, using footage from drone altitudes ranging from 15 m to 130 m. We applied transfer learning to state-of-the-art lightweight object detection algorithms (Tensorflow and YOLO) based on feature extraction. In the point pattern analysis, we treated each animal as a point and identified them by the behavioural pattern of those points. The five animal species investigated were African elephant (Loxodonta africana), giraffe (Giraffa camelopardalis), white rhinoceros (Ceratotherium simum), wildebeest (Connochaetes taurinus) and zebra (Equus quaggas). As we increased the altitude of the drone camera, the detection algorithms using features significantly lost accuracy. Animal features are harder to detect at higher altitudes and in the presence of environmental camouflage, animal occlusion, and shadows. The performance of lightweight object detection algorithms (F1 score) decreased with increasing drone altitude to a minimum of 29%, while the point pattern algorithms produced an F1 score above 96% across all drone altitudes. Using point pattern analysis, the accuracy of animal identification is invariant to drone camera altitude and disturbances from environmental conditions. Animal social interactions within herds follow species-specific hidden patterns in their group structure that allow for reliable species identification.
All Science Journal Classification (ASJC) codes
- Ecology, Evolution, Behavior and Systematics
- Modelling and Simulation
- Ecological Modelling
- Computer Science Applications
- Computational Theory and Mathematics
- Applied Mathematics