From Manual Observation to Automated Monitoring: Space Allowance Effects on Play Behaviour in Group-Housed Dairy Calves

Play behaviour serves as a positive welfare indicator in dairy calves, yet the influence of space allowance under commercial conditions remains poorly characterized, particularly at intermediate-to-high allowances (6-20 m2 per calf). This study investigated the relationship between space allowance and play behaviour in 60 group-housed dairy calves across 14 commercial farms in the Netherlands (space range: 2.66-17.98 m2 per calf), and developed an automated computer vision pipeline for scalable monitoring. Video observations were analyzed using a detailed ethogram, with play expressed as percentage of observation period (%OP). Statistical analysis employed linear mixed models with farm as a random effect. A computer vision pipeline was trained on manual annotations from 108 hours on 6 farms and validated on held-out test data. The computer vision classifier achieved 97.6% accuracy with 99.4% recall for active play detection. Calves spent on average 1.0% of OP playing reflecting around 10 minutes per 17-hour period. The space-play relationship was non-linear, with highest play levels at 8-10 m2 per calf (1.6% OP) and lowest at 6-8 m2 and 12-14 m2 (<0.6% OP). Space remained significant after controlling for age, health, and group size. In summary, these findings suggest that 8-10 m2 per calf represents a practical target balancing welfare benefits with economic feasibility, and demonstrate that automated monitoring can scale small annotation projects to continuous welfare assessment systems.

💡 Research Summary

This study investigated how space allowance influences play behaviour in group‑housed dairy calves under commercial conditions and introduced an automated computer‑vision system for large‑scale welfare monitoring. Data were collected from 14 Dutch farms, encompassing 60 calves (64 calves after exclusions) housed in pens offering 2.66 to 17.98 m² per calf. Continuous video recordings (06:00‑23:00) were obtained for a full 17‑hour observation period on each farm. Trained observers annotated the footage using a detailed ethogram that distinguished locomotor, social, object, and straw‑related play, recording behaviours at 0.1‑second resolution. Play time was expressed as a percentage of the observation period (%OP), yielding an average of 1.0 %OP (≈10 minutes per 17 hours).

Linear mixed‑effects models (farm as a random effect) revealed a non‑linear relationship between space allowance and play. Play peaked at 8‑10 m² per calf (1.6 %OP) and was markedly lower in the 6‑8 m² and 12‑14 m² bands (<0.6 %OP). The effect remained significant after adjusting for age, health status, daily milk intake, group size, and bedding cleanliness, indicating that physical space per se drives the observed behavioural changes. These findings suggest that 8‑10 m² per calf balances welfare benefits (maximised play) with economic feasibility, offering a realistic target that lies between the EU minimum (1.5‑1.8 m²) and the EFSA recommendation (20 m²).

To overcome the scalability limits of manual observation, the authors built a computer‑vision pipeline that integrates three state‑of‑the‑art foundation models: YOLOv12 for calf detection, SAM2 for segmentation and multi‑object tracking, and DINOv2 for extracting 1024‑dimensional visual embeddings. From 108 hours of high‑quality video (≈1.5 million frames) across six farms, they generated a labelled dataset by aligning manual annotations with detected frames using Apache Spark for distributed processing and a 0.5‑second temporal tolerance. Because play events are rare, the dataset was balanced by stratified down‑sampling to create equal numbers of “Active Playing”, “Non‑Active Playing”, and “Not Playing” samples (7,609 each).

A multilayer perceptron (input = 1024, hidden layers = 512 and 256 units, ReLU activation, 50 % dropout) was trained on the embeddings with Adam optimisation (learning rate = 0.001) and early stopping. On a held‑out test set, the model achieved 97.6 % overall accuracy, 99.4 % recall for active play, and 96.8 % precision for non‑play, demonstrating that deep visual features combined with lightweight classification can reliably detect calf play in real‑world farm environments despite variable lighting, occlusions, and diverse pen layouts. Misclassifications were mainly confined to the boundary between non‑active play and non‑play, reflecting the inherent ambiguity of some behaviours.

The technical contribution lies in the seamless coupling of detection, segmentation, and self‑supervised feature extraction, which dramatically reduces the need for extensive manual labelling while preserving high performance. The use of Spark for metadata integration ensures reproducibility and scalability to thousands of hours of footage.

From a practical standpoint, the automated system can provide daily or even real‑time play metrics, offering farmers an early‑warning indicator of stress, disease, or sub‑optimal housing. Because the pipeline is modular, it can be extended to monitor other behaviours (e.g., feeding, rumination) or other species, forming the backbone of a comprehensive animal‑welfare monitoring platform. Future work should explore longitudinal links between play frequency, growth, immune function, and milk production, and conduct cost‑benefit analyses to inform policy and farm‑level decision‑making.