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

Area of Science:

• Equine Science
• Animal Behavior
• Machine Learning

Background:

• Accurate monitoring of horse grazing behavior is crucial for pasture management and welfare.
• Traditional observation methods are labor-intensive and lack detailed temporal data.

Purpose of the Study:

• To develop and validate a deep learning model for classifying horse grazing and non-grazing behaviors.
• Utilize jaw-mounted accelerometer data for automated behavioral analysis.

Main Methods:

• Four yearling horses were fitted with jaw-mounted triaxle accelerometers.
• Data were collected over 19 hours, with 230,286 points annotated via video.
• Trained and evaluated deep learning models (CNN, LSTM, CNN+LSTM) on varying data parameters.

Main Results:

• The combined CNN+LSTM model achieved 98.0% test accuracy and an AUC of 1.00.
• High F1 scores (0.99 for grazing, 0.97 for non-grazing) indicate robust classification.
• Grazing behavior was concentrated on paddock peripheries, while non-grazing occurred more centrally.

Conclusions:

• A deep learning framework integrating CNN and LSTM accurately classifies horse grazing behavior using accelerometers.
• This non-invasive, high-resolution method enables automated behavioral monitoring in pasture systems.