77. Automated Behavior Classification for Modeling Dry Matter Intake in Cattle.

Abstract Enteric methane contributes substantially to greenhouse gas (GHG) emissions from cattle, and Dry Matter Intake (DMI) is often used as a proxy for estimating methane production. However, current methods for measuring feed intake are expensive, labor-intensive, and difficult to scale, complicating efforts to manage methane emissions effectively. To address this challenge, this study presents a scalable and non-intrusive approach for generating behavioral ground truth data for future methane emissions modeling using a computer vision pipeline. While prior work in precision livestock farming has demonstrated that deep learning–based vision systems can monitor animals at scale, robust end-to-end pipelines that integrate detection, classification, and persistent tracking remain limited, restricting the generation of structured behavioral outputs for modeling applications. In this study, we implemented a two-stage framework in which a YOLO11n (You Only Look Once) model was used for object detection and combined it with the BoT-SORT tracking algorithm to enable “per-cow” behavioral time budgeting from a freely available online dataset containing 25,324 annotated crops from seven overhead cameras. The objective is to calculate feeding and rumination time to be used for DMI calculations in the future. The second stage employed a pre-trained Vision Transformer model (google/vit-base-patch16-224-in21k) for behavior classification. Behaviors were categorized into five classes, drinking, lying down, foraging, ruminating and standing which was further split into training, testing and validation sets in a 70/15/15 ratio. The YOLO11n detector achieved 90.1% [email protected] with precision and recall values of 87.0% and 84.7%, respectively, while maintaining inference speeds exceeding 100 frames per second on an RTX 4080, demonstrating real-time capability. The Vision Transformer classifier reached 92.42% test accuracy with a weighted F1 score of 92.42%. Tracking performance evaluated over a 50-frame sample successfully tracked 10 unique cows with an average of 8.7 cows detected per frame. Analysis of classification errors revealed confusion primarily between lying down and ruminating (14%), drinking and standing (7%), and foraging and standing (3–4%), likely due to the absence of temporal modeling where frame similarity challenges frame-based classification approaches. Overall, the results demonstrate that a detection–tracking–classification pipeline can effectively handle behavioral complexity and class imbalance while maintaining real-time performance, providing a promising foundation for scalable behavioral monitoring and improved estimation of enteric methane emissions through feed intake proxies.