Machine Learning Methods and Visual Observations to Categorize Behavior of Grazing Cattle Using Accelerometer...

Area of Science:

• Animal behavior analysis
• Machine learning applications
• Wearable sensor technology

Background:

• Accelerometers on animals generate distinct behavioral data.
• Machine learning, like random forest decision trees, can classify these signals.
• Understanding accelerometer signal separation is key for accurate behavior classification.

Purpose of the Study:

• To differentiate accelerometer signals among basic animal behaviors.
• To optimize signal pre-processing window size for behavior classification.
• To determine the minimum observations needed for accurate machine learning models.

Main Methods:

• Utilized tri-axial accelerometers (40 Hz) and GPS collars on 10 crossbred steers (Bos taurus indicus).
• Analyzed functional differences in accelerometer signals corresponding to discrete behaviors, especially grazing.
• Assessed the impact of smoothing window size and observation count on classification accuracy.

Main Results:

• Distinct behavioral signatures, particularly for grazing (head-down posture), were identified from accelerometer signals.
• A 10-second smoothing window size significantly improved classification accuracy (p < 0.05).
• Reducing observations below 50% decreased accuracy, while in-pasture observation enhanced accuracy and precision.

Conclusions:

• Accelerometer data, when processed optimally, can accurately classify distinct animal behaviors.
• Signal pre-processing parameters, such as window size and observation quantity, are critical for model performance.
• In-pasture observations provide superior accuracy and precision for behavior classification compared to collar-mounted video.