Deep learning approach for classifying grazing behavior in yearling horses using triaxial accelerometer data: A pilot study.

Research Overview

• This study developed and tested a deep learning method to classify grazing behavior in young horses using data from jaw-mounted triaxial accelerometers.
• The combined CNN+LSTM model achieved very high accuracy in distinguishing grazing from non-grazing behavior during free grazing on pasture.

Introduction and Objective

• Monitoring grazing behavior in horses is important for managing pastures and assessing animal welfare.
• Traditional observation methods are laborious and do not provide continuous, fine-grained data over long periods.
• The objective was to create and validate an automated system using deep learning that uses accelerometer data to classify when horses are grazing versus not grazing.
• This pilot study focused on yearling Thoroughbred horses under natural grazing conditions.

Methods

• Subjects: Four yearling Thoroughbred horses grazing freely in a 4.0 hectare paddock.
• Data Collection:
  • Each horse wore a jaw-mounted triaxial accelerometer recording motion data at 10 Hz (one reading every 100 milliseconds).
  • Data were recorded for a 19-hour continuous grazing period.
  • Simultaneous video recordings were used to manually annotate each data point as grazing (G) or non-grazing (NG).
  • A total of 230,286 annotated data points were used for model training and validation.
• Deep Learning Models Tested:
  • One-dimensional Convolutional Neural Network (CNN) which captures local patterns in the accelerometer data.
  • Long Short-Term Memory (LSTM) network designed to model temporal dependencies and sequences.
  • A hybrid CNN+LSTM model combining both approaches to capture spatial and temporal features.
• Model Training and Evaluation:
  • Different data sampling rates between 100 ms to 10,000 ms and time windows between 5 and 60 seconds were explored to optimize performance.
  • Performance metrics used included accuracy, F1 score, precision, recall, and area under the ROC curve (AUC).

Results

• The combined CNN+LSTM model outperformed the individual CNN and LSTM models.
• Key performance indicators for CNN+LSTM:
  • Test Accuracy: 98.0% — meaning 98% of behavior classifications were correct.
  • AUC: 1.00 — perfect discrimination between grazing and non-grazing behaviors.
  • F1 Score: 0.99 for grazing, 0.97 for non-grazing — indicating very high balance between precision and recall.
• Behavioral Observations from the Data:
  • Horses spent about 58.3% (±2.1%) of the time grazing.
  • Spatial analysis showed grazing mostly took place around the edges of the paddock, with non-grazing more common in the central areas.

Conclusions and Implications

• The study demonstrates that a deep learning approach combining CNN and LSTM can reliably classify horse grazing behavior based on accelerometer signals from the jaw.
• Using such jaw-mounted accelerometers offers a non-invasive, automated, and high-resolution method to monitor animal behavior continuously.
• This approach can aid pasture management by providing accurate temporal and spatial grazing patterns without the need for constant human observation.
• It can also contribute to welfare assessments by quickly identifying deviations from normal grazing behavior.
• Further studies with more horses and varied environments would be beneficial to validate and generalize these findings.