Uncovering Patterns in Dairy Cow Behaviour: A Deep Learning Approach with Tri-Axial Accelerometer Data

Area of Science:

• Animal Science
• Machine Learning
• Biotechnology

Background:

• Early disease detection in dairy cows is crucial for herd health and productivity.
• Monitoring behavioral changes offers a non-invasive method for identifying health issues.
• Accelerometry data combined with advanced algorithms can provide insights into animal welfare.

Purpose of the Study:

• To evaluate the efficacy of deep learning (DL) models in classifying dairy cow behaviors.
• To compare the performance of DL algorithms against classical machine learning (ML) methods.
• To assess the potential of single-point accelerometry for comprehensive behavior monitoring.

Main Methods:

• Collected tri-axial accelerometry data from 12 dairy cows over 136 ± 29 minutes.
• Extracted 15 metrics per 8-second interval, creating a dataset of 211,720 observations.
• Trained and tested an 8-layer convolutional neural network (CNN) and compared it with classical ML models.

Main Results:

• The 8-layer CNN achieved an overall accuracy and F1 score of 0.96.
• Individual behavior classification (precision, sensitivity/recall, F1 score) ranged from 0.93 to 0.99.
• The CNN model significantly outperformed all tested classical ML algorithms.

Conclusions:

• Deep learning, specifically CNNs, demonstrates high performance in classifying dairy cow behaviors from single-point accelerometry data.
• This approach provides a reliable and efficient method for monitoring individual cow behavior and potentially detecting early signs of disease.
• Single-sensor accelerometry coupled with DL offers a scalable solution for improving dairy herd management and animal welfare.