Overcoming Extreme Sample Scarcity: Deep Learning-Based Real-Time Detection of Cattle Mounting Behavior From UAV Imagery and Video

Abstract

Early detection of mounting behavior in feedlot cattle is essential for preventing injuries and reducing stress. However, manual observation is inefficient and covers only a limited area. To enable large-scale monitoring, we propose an automated detection system that employs a fine-tuned you only look once version 11 (YOLOv11) model on both uncrewed aerial vehicle (UAV) imagery and videos to identify mounting events in real time. A 112-day field experiment was conducted in Mato Grosso do Sul, Brazil, monitoring 110 Nellore cattle. Because this abnormal behavior is rare, the dataset exhibits extreme sample scarcity: from approximately 10 000 UAV images and 12 pieces of video, only 38 frames contained true mounting events. To address this challenge, we developed a hybrid dataset strategy that integrates real samples, augmented samples, and segment anything model-based synthetic data. With this approach, the model achieved good precision, recall, F1-score, and mean average precision (mAP)@0.5 on the small-scale test set, demonstrating strong generalization performance and confirming the feasibility of UAV-based monitoring for animal-welfare applications. The proposed technique showsx high potential for embedded UAV deployment, enabling autonomous and continuous surveillance in precision livestock systems.