From video-derived feeding behaviour to cow-level nutritional deviation signals: A dairy digital-twin decision-support framework

Abstract

Continuous video offers a dynamic view of dairy-cow behaviour, but its value for precision nutrition depends on alignment with physiological context. We developed a dairy digital-twin framework that fuses identity-associated behavioural records from an established video-analytics layer with body weight, milk production, milk fat, parity and days in milk. The 16-cow analytical cohort was monitored for up to 11 days in one tie-stall barn, yielding 153 cow-days after quality exclusions applied before model fitting. NRC (2001) expected dry matter intake provided a transparent physiological reference compatible with the daily records. In matched leave-one-cow-out analysis, adding video-derived feeding duration to body weight, fat-corrected milk and days in milk reduced RMSE from 2.015 to 1.763 kg DM/day, MAE from 1.267 to 1.159 kg DM/day and MAPE from 4.5% to 4.2%, while R{superscript 2} increased from 0.500 to 0.617. A secondary reduced index achieved RMSE 2.362 kg DM/day and MAPE 5.8% across held-out cows. Cow-level analysis delineated the operating domain: median per-cow MAPE was 4.15%, whereas the sole cow at 15 days in milk had MAPE 28.8%. Two independently recorded veterinary events were temporally concordant with unusual feeding trajectories, providing descriptive biological context. Because the endpoint was NRC-derived, these metrics quantify reference reconstruction rather than accuracy against observed intake. By converting continuous behavioural events into auditable cow-day states, the framework links physical animals to physiologically contextualized digital counterparts and establishes a scalable foundation for operator-focused dairy digital-twin decision support.