Rapid UAV-based detection of crane (Grus grus) mortality hotspots during an H5N1 avian influenza outbreak using deep learning

Rapid detection of wildlife mortality during infectious disease outbreaks is essential for ecological risk mitigation, zoonotic-risk management, and timely intervention, yet field surveys are often slow, labor-intensive, and spatially limited. This study evaluates unmanned aerial vehicle (UAV) imagery combined with deep learning as an operational surveillance tool for mapping live and deceased Common Cranes ( Grus grus ) during a highly pathogenic avian influenza (H5N1) outbreak in the Hula Valley, Israel. High-resolution RGB imagery was collected over aquatic roosting and terrestrial feeding habitats and annotated to distinguish living and dead individuals. A YOLOv8-based detection framework was tested using habitat-specific training, combined multi-habitat training, and transfer learning from a large avian dataset, together with selective data augmentation, sliding-window inference, and Weighted Boxes Fusion. Living cranes were detected reliably across habitats, with AP@50 reaching 0.89 in mosaicked feeding-area imagery and 0.98 in raw feeding-area imagery. Dead cranes were detected accurately in water bodies (AP@50 ≈ 0.90), but detection in terrestrial feeding areas remained more limited (AP@50 ≈ 0.64–0.70), based on aerial annotations without in situ confirmation. Oversampling, undersampling, and focal loss provided limited improvement, suggesting that class imbalance alone does not fully explain the lower terrestrial detection performance, and that carcass scarcity, limited visual diversity, and ambiguity against heterogeneous ground backgrounds may also contribute. Joint multi-habitat training and transfer learning improved generalization for the rare dead-crane class on land. These findings demonstrate the value of UAV-based deep learning for rapid mortality assessment and hotspot identification during wildlife disease outbreaks.