Classification of Cattle Behaviours Using Neck-Mounted Accelerometer-Equipped Collars and Convolutional Neural...

Area of Science:

• Animal Science
• Machine Learning
• Agricultural Technology

Background:

• Monitoring cattle behavior is crucial for detecting health issues and optimizing fertility in large herds.
• Accelerometer-based sensors are widely used on farms to track individual animal activity.
• Current systems need efficient processing for on-farm management decisions.

Purpose of the Study:

• To develop a Convolutional Neural Network (CNN) for classifying cattle behavioral states (rumination, eating, other) using accelerometer data.
• To optimize the CNN for deployment on low-power, memory-constrained micro-controller architectures.
• To balance model performance with computational complexity and memory footprint.

Main Methods:

• Utilized data from neck-mounted accelerometer collars on 18 steers during three UK farm trials.
• Employed muzzle-mounted pressure sensor halters for ground truth behavioral data.
• Explored various CNN architectures and performed hyper-parameter searches, followed by progressive model compression.

Main Results:

• Achieved a 14.30 compression ratio for the CNN model compared to the unpruned version.
• Maintained accurate cattle behavior classification with an overall F1 score of 0.82 (FP32 and FP16 precision).
• The optimized model supports a battery lifetime exceeding 5.7 years.

Conclusions:

• The developed compressed CNN effectively classifies cattle behaviors from accelerometer data.
• The methodology addresses practical deployment challenges on low-power devices for efficient farm management.
• This approach enhances the potential for scalable, real-time monitoring of cattle health and welfare.