Behavioral classification of data from collars containing motion sensors in grazing cattle

• Data from electronic cattle monitoring collars were classified into 5 activities.
• Collars contained GPS and accelerometers.
• Mixed distributions were used to obtain threshold values for a decision tree.
• The 5 activities were classified with 85.5% and 90.5% precision in 2 trials.
• Grazing and ruminating were classified most accurately.

Remote monitoring of animal behavior offers great potential to improve livestock management however technologies able to collect data at high frequency and accurate data classification methods are required. The objective of this study was to develop a methodology capable of performing unsupervised behavioral classification of electronic data collected at high frequency from collar-mounted motion and GPS sensors in grazing cattle. Two independent trials were conducted, one for developing the classification algorithm (4 groups of 11 steers) and a second for its evaluation (14 steers). Each steer was fitted with a collar containing GPS and a 3-axis accelerometer that collected data at 4 and 10 Hz, respectively. Foraging, ruminating, traveling, resting and ‘other active behaviors’ (which included scratching against objects, head shaking, and grooming) were observed and recorded continuously at the nearest second in animals wearing collars. Collar data were aggregated to 10-s intervals through the mean (indicative of the position of the neck and travel speed) and standard deviation (SD; indicative of activity level) and then log-transformed for analysis. The histograms of travel speed showed 3 populations and observations revealed these populations represented stationary, slow and fast travel behaviors. The histograms of the accelerometer X-axis mean showed populations corresponding with behaviors of head down or head up. The histograms of the accelerometer X-axis SD showed 3 populations representing behaviors with high, medium and low activity levels. Mixture models were fitted to data from each animal in both trials to calculate threshold values corresponding to where behaviors transitioned between different states. These thresholds from the 3 sensor signatures were then used in a decision tree to classify all 10-s data where behaviors were unknown into 5 mutually exclusive behaviors. The algorithm correctly classified 85.5% and 90.5% of all data points in the development and evaluation datasets, respectively. Foraging showed the greatest sensitivity (93.7% and 98.4%) and specificity (94.6% and 99.4%) followed by ruminating (sensitivity 97% and 87%, and specificity 90% and 95%) for development and evaluation trials, respectively. Major advantages of mixture models include computational efficiency suitable for large data sets (e.g. >2 million data lines), minimal requirement for training datasets, and estimation of threshold values for individual animals under unknown and varying environmental conditions. The technology and methodology allows for the automatic and real-time monitoring of behavior with high spatial and temporal resolution which could benefit livestock industries beyond the research domain for improved animal and ecological management.