Cattle dung detection in pastures from drone images using YOLOv5

Abstract Livestock excretions are crucial for nutrient cycling in pasture ecosystems. However, conventional methods based on field observations require significant human power and are time‐consuming. This study developed a model, ‘Dung Detector (DD)’, for detecting cattle dung in pastures from drone images using the You Only Look Once (YOLO) v5 algorithm. The DD model was trained using our custom dataset including 1,504 split images from drone orthomosaic images in five paddocks: Obihiro (OBH), Shintoku (STK), Minokamo (MNO), Miyota (MYT), and Yatsugatake (YGK). The detection accuracy was evaluated using ground‐truth data acquired in two quadrats within paddocks. The DD model performed well for OBH and STK ( F ‐score = 0.861 and 0.835) paddocks with simple grass species and low surface sward height (SSH). Although the MNO and MYT, with complex vegetation and high SSH, showed few false positives (precision >0.9), some cattle dung pats were undetectable, presumably due to grass height (Recall = 0.500 and 0.276).