Evaluation of sampling frequency, window size and sensor position for classification of sheep behaviour

Area of Science:

• Precision Livestock Farming
• Animal Behavior Analysis
• Sensor Technology

Background:

• Automated behavioral classification using sensors can enhance animal health and welfare.
• Previous studies have not simultaneously evaluated sensor position, sampling frequency, and window size for activity recognition in livestock.

Purpose of the Study:

• To evaluate the impact of sensor position (ear vs. collar), sampling frequency (8, 16, 32 Hz), and window size (3, 5, 7 s) on sheep behavior classification accuracy.
• To determine optimal sensor configurations for accurate and energy-efficient real-time behavioral monitoring in sheep.

Main Methods:

• Utilized triaxial accelerometer and gyroscope sensors placed on sheep (ear and collar).
• Classified behaviors (lying, standing, walking) using a random forest approach with 44 features.
• Analyzed the effects of varying sampling frequencies and window sizes on classification performance.

Main Results:

• Highest accuracy (95%) and F-scores (91%-97%) for behavior classification were achieved with 32 Hz sampling and 5s or 7s window sizes for both ear and collar sensors.
• Comparable accuracy (91%-93%) and F-scores (88%-95%) were obtained using 16 Hz sampling with a 7s window.
• Optimal energy efficiency was observed with a 7s window size.

Conclusions:

• A sampling frequency of 16 Hz combined with a 7s window size provides a practical balance between high classification accuracy and reduced energy consumption for real-time sheep behavior monitoring.
• Sensor configuration parameters significantly influence the effectiveness of automated behavioral classification systems in precision livestock farming.