203. Automated Detection and Scoring of Feces in Beef Cattle Feedlots Using Unmanned Aerial Vehicles.

Abstract Manure scoring is a widely used tool for evaluating cattle health and nutrition in beef cattle feedlots. However, manual scoring is time-consuming, subjective, and limited by restricted ground-level visibility across large pens. This study aimed to develop an automated computer vision system capable of detecting and classifying feces scores in large-scale feedlots. High-resolution imagery was collected from 80 Brazilian feedlots using a UAV hovering at 20 meters. A YOLOv11 object detection model was trained and validated on spatially distinct subsets of these feedlots to identify individual feces. Subsequently, an Xception Convolutional Neural Network (CNN) was trained on 3,451 individual images to classify feces into scores 1 to 5, plus a class (0) for misidentifications. Because abnormal scores (1 and 5) are most actionable for management decisions, a second CNN model was developed that combined the intermediate classes 3 and 4. The YOLOv11 detection model achieved 64% accuracy across all environments and 77% under optimal pen conditions. The initial 6-class CNN achieved F1 scores of 76%, 49%, 46%, 42%, and 78% for classes 1–5, respectively. The refined 5-class model (combining classes 3 and 4) yielded F1 scores of 79%, 52%, 73%, and 73% for classes 1, 2, 3-4, and 5. Both models demonstrated consistently high F1-score in the most critical extreme scoring ranges. These preliminary findings indicate that combining UAV imagery and deep neural networks creates an efficient and objective framework for large-scale fecal scoring, improving data-driven feedlot management.