Real-Time Classification of Cattle Behaviour using Accelerometer Sensors

In this study, we develop and validate a supervised machine-learning algorithm to monitor grazing and ruminating behaviours of cattle using accelerometer sensors. The method is specifically designed for performing real-time classification on resource-constrained sensor nodes. Twenty multiparous Holstein cows were used for this study. Each cow was wearing an AX3 accelerometer sensor attached to a neck-collar. The cows had daily access to a pasture between 7:30 AM and 2 PM for three weeks. Direct observations of the cows' behaviours were made to validate the sensor data. A new decision-tree algorithm (DT) was developed to classify the raw data. The decision-tree algorithm was selected for its low computational costs, which makes it implementable on the on-cow nodes. The DT presented an overall accuracy of 91% with a sensitivity and precision between 89-94% for ruminating and grazing behaviours. The hourly difference between the predicted and the observed (total) ruminating and grazing times (in min/h, meanstandard error) were 1.90.09 min/h (3.1% of the observed time) and 2.20.07 min/h (3.7%) respectively. This validation illustrates the potential of the collar-mounted accelerometer to classify grazing and ruminating behaviours.

Jaar van publicatie:2022