Deep learning approach for classifying grazing behavior in yearling horses using triaxial accelerometer data: A pilot study.
Authors: Kamiya Uta, Kakiuchi Kasumi, Kawamura Kensuke, Ueda Koichiro, Kawai Masahito, Matsui Akira, Negishi Natsuko
Journal: Journal of equine veterinary science
Summary
# Editorial Summary Monitoring grazing behavior in horses has traditionally relied on time-consuming direct observation, which fails to capture the detailed temporal patterns essential for optimising pasture management and assessing individual welfare in field conditions. Researchers equipped four yearling Thoroughbreds with triaxial accelerometers positioned under the jaw and recorded over 230,000 data points during a 19-hour free-grazing period, using synchronized video footage to manually annotate each movement as either grazing or non-grazing behaviour. A hybrid deep learning model combining convolutional neural networks with long short-term memory algorithms achieved 98% classification accuracy with near-perfect discrimination between behaviours (F1 scores of 0.99 for grazing and 0.97 for non-grazing), and identified that yearlings spent approximately 58% of their time grazing, with spatial clustering around paddock edges rather than central areas. This non-invasive approach operates at 100 millisecond resolution across varying time windows and sampling rates, enabling practitioners to automatically track grazing patterns without observer bias or labour-intensive video analysis. For farm managers and veterinary professionals, jaw-mounted accelerometer data analysed through this framework could become a practical tool for detecting welfare issues (such as reduced grazing associated with pain or illness), optimising pasture rotation, and generating objective grazing metrics that currently require hours of manual observation.
Read the full abstract on PubMed
Practical Takeaways
- •Jaw-mounted accelerometers with deep learning analysis provide automated, objective monitoring of grazing behavior without labor-intensive video observation, enabling better pasture management decisions
- •The high classification accuracy (98%) suggests this technology could reliably detect behavioral changes indicating health or welfare issues in pastured horses
- •Spatial grazing patterns revealed by this method could help optimize paddock design and rotational grazing strategies to improve pasture utilization and horse welfare
Key Findings
- •CNN+LSTM deep learning model achieved 98.0% accuracy and AUC of 1.00 for classifying grazing versus non-grazing behavior in yearling horses using jaw-mounted accelerometers
- •Grazing behavior represented 58.3% (±2.1%) of observed behavior during the 19-hour free-grazing period
- •Spatial analysis demonstrated grazing concentrated along paddock peripheries while non-grazing behavior was more frequent in central zones
- •The combined CNN+LSTM architecture outperformed individual CNN and LSTM models across varying sampling rates (100-10,000 ms) and time windows (5-60 s)