Classification of cow behavior patterns using inertial measurement units and a fully convolutional network model

Area of Science:

• Animal Behavior
• Machine Learning
• Wearable Technology

Background:

• Accurate classification of cow behavior is crucial for animal welfare and management.
• Inertial Measurement Units (IMUs) offer a non-invasive method for collecting animal movement data.

Purpose of the Study:

• To automatically classify seven distinct cow behavior patterns using IMU data and a Fully Convolutional Network (FCN).
• To evaluate the impact of data dimensionality (6-axis vs. 9-axis) and window size on classification accuracy.

Main Methods:

• Collected behavioral data from 12 cows using neck-mounted IMUs.
• Processed 9-axis IMU data, reducing it to 6-axis for comparative analysis.
• Trained and compared FCN models with different window sizes (64 and 128) and data configurations (6-axis and 9-axis).

Main Results:

• The FCN model with a 128-sample window size (12.8 seconds) using all IMU data achieved the highest overall accuracy of 83.75%.
• Individual behavior classification accuracies ranged from 45.2% (social licking) to 98.1% (rub scratching - leg).
• Lower accuracy was observed for behaviors involving varied, intensive movements, suggesting a need for further model refinement.

Conclusions:

• Automated cow behavior classification using IMUs and FCNs is feasible, with optimal performance achieved using comprehensive data and larger window sizes.
• Future research should incorporate multi-modal data (audio, video) and adaptive window sizes to enhance classification accuracy, particularly for complex behaviors.