BEHAVE-UAV: A Behaviour-Aware Synthetic Data Pipeline for Wildlife Detection from UAV Imagery

Unmanned aerial vehicles (UAVs) are increasingly used to monitor wildlife, but training robust detectors still requires large, consistently annotated datasets collected across seasons, habitats and flight altitudes. In practice, such data are scarce and expensive to label, especially when animals occupy only a few pixels in high-altitude imagery. We present a behaviour-aware synthetic data pipeline, implemented in Unreal Engine 5, that combines parameterised animal agents, procedurally varied environments and UAV-accurate camera trajectories to generate large volumes of labelled UAV imagery without manual annotation. Each frame is exported with instance masks, YOLO-format bounding boxes and tracking metadata, enabling both object detection and downstream behavioural analysis. Using this pipeline, we study YOLOv8s trained under six regimes that vary by data source (synthetic versus real) and input resolution, including a fractional fine-tuning sweep on a public deer dataset. High-resolution synthetic pre-training at 1280 px substantially improves small-object detection and, after fine-tuning on only 50% of the real images, recovers nearly all performance achieved with the fully labelled real set. At lower resolution (640 px), synthetic initialisation matches real-only training after fine-tuning, indicating that synthetic data do not harm and can accelerate convergence. These results show that behaviour-aware synthetic data can make UAV wildlife monitoring more sample-efficient while reducing annotation cost.