Monitoring horse behaviour with deep learning models.

Overview

• This study used deep learning to automatically identify common horse behaviors—standing, lying, and drinking—from continuous video recordings in a stall.
• The performance of a convolutional neural network (CNN) was compared to manually labeled data to assess how well it could monitor horse well-being through behavior analysis.

Background and Motivation

• Monitoring horse behavior is valuable for assessing their health and welfare status, traditionally done through direct observation or manual video analysis.
• These traditional methods can be invasive, time-consuming, and subject to observer bias or error.
• Recent advances in deep learning, particularly convolutional neural networks, allow for automated, non-invasive behavioral analysis using video data.
• The goal is to develop a reliable and efficient approach to continuously monitor horse behavior for welfare purposes.

Methods

• A horse housed in a wooden stall was continuously filmed for 29 consecutive days to capture natural behavior over an extended period.
• A convolutional neural network (CNN) model was trained to classify three behaviors from the video frames:
  • Standing
  • Lying
  • Drinking
• Manual annotation was performed by humans on the video to provide ground truth labels against which to validate the CNN’s predictions.
• Model performance metrics used included precision (how many predicted behaviors were correct) and recall (how many actual behaviors were detected).

Results

• The CNN was highly effective at detecting standing behavior:
  • Precision: 97.5% (very few false positives)
  • Recall: 89.2% (most standing events detected)
• Lying behavior was detected with:
  • High precision: 92.8% (predictions were reliable)
  • Lower recall: 63.1% (some lying events were missed)
• Drinking behavior metrics were not specifically reported but was tracked in temporal patterns.
• Activity pattern analysis revealed:
  • Standing was the dominant daily behavior, taking up more than 85% of the horse’s time budget.
  • Lying accounted for 5-10% of the daily time.
  • Drinking mainly occurred in the late afternoon to evening (4:00 pm – 10:00 pm).

Discussion and Implications

• This study demonstrates that deep learning models can reliably classify common equine behaviors from video footage without the need for invasive sensors or extensive manual analysis.
• Such automated monitoring can provide continuous welfare assessments and detect changes in behavior that may signal health or well-being issues.
• High precision ensures that the behaviors identified by the model are trustworthy, though lower recall in some cases indicates room for improvement in detecting all behavior events.
• Temporal behavior patterns identified reflect typical daily routines and could help establish normal baselines for individual horses.
• Future work is needed to:
  • Improve detection accuracy, especially for less frequent or more subtle behaviors.
  • Expand behavior categories beyond standing, lying, and drinking to capture a fuller spectrum of equine activities.
  • Test the approach in different stabling environments and with other horses to generalize results.

Conclusion

• Deep learning, specifically CNNs, offers a promising tool for automated, welfare-focused behavior monitoring in stabled horses using video data.
• This method can provide detailed, objective daily activity budgets that could support better management and health care in equine facilities.