ActBeCalf: Accelerometer-based multivariate time-series dataset for calf behavior classification

Area of Science:

• Animal Science
• Machine Learning
• Data Science

Background:

• Improving pre-weaned calf welfare is crucial for dairy farms.
• Automatic behavior monitoring using accelerometers requires accurately labeled data.
• Time synchronization and data alignment are significant challenges in this field.

Purpose of the Study:

• To introduce ActBeCalf, a novel dataset for calf behavior classification.
• To address the challenge of aligning accelerometer time-series data with behavioral labels.
• To provide a reliable resource for developing machine learning models for calf behavior analysis.

Main Methods:

• Thirty pre-weaned calves were equipped with 3D-accelerometer sensors for 13 weeks.
• Calf behaviors were manually annotated from video recordings using BORIS software.
• Accelerometer data was synchronized and aligned with behavioral annotations using an external clock.

Main Results:

• ActBeCalf dataset contains 27.4 hours of aligned accelerometer data from 30 calves.
• Behaviors annotated include lying, standing, walking, running, sniffing, scratching, social interaction, and grooming.
• Machine learning models developed using ActBeCalf achieved high predictive performance (balanced accuracy: 92% and 84%).

Conclusions:

• ActBeCalf is a reliable and comprehensive dataset for research on pre-weaned calf behavior.
• The dataset facilitates the development of advanced machine learning models for animal behavior classification.
• ActBeCalf supports initiatives aimed at enhancing animal welfare in dairy farming.