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

Abstract

Background

Accurate monitoring of grazing behavior in horses is essential for pasture management and welfare evaluation; however, conventional observation methods are labor-intensive and lack temporal resolution.

Aims/objectives

This pilot study aimed to develop and validate a deep learning model using jaw-mounted accelerometer data to classify grazing and non-grazing behaviors in yearling horses under various pasture conditions.

Methods

Four yearling Thoroughbred horses were equipped with triaxle accelerometers mounted under their jaws. Data were recorded at 10 Hz (100 ms) during a 19 h free-grazing period in a 4.0 ha paddock. A total of 230,286 data points were annotated as grazing (G) or non-grazing (NG) using synchronized video observation. Three deep learning models—one-dimensional convolutional neural network (CNN), long short-term memory (LSTM), and combined CNN+LSTM—were trained and evaluated under varying sampling rates (100–10,000 ms) and time windows (5–60 s). Model performance was assessed using accuracy, F1 score, precision, recall, and area under the curve (AUC).

Results

The CNN+LSTM model demonstrated the highest classification performance with a test accuracy of 98.0 % and an AUC of 1.00. F1 scores were 0.99 for G and 0.97 for NG behavior. Across the full observational period, the proportion of grazing behavior was 58.3 % (±2.1 %). Spatial analysis revealed that grazing was concentrated along paddock peripheries, whereas non-grazing was more frequent in central zones.

Conclusion

A deep learning framework that combines CNN and LSTM can accurately classify grazing behavior in horses using jaw-mounted accelerometers. This non-invasive, high-resolution method offers a promising tool for automated behavioral monitoring in pasture-based systems.