Classification of ingestive-related cow behaviours using RumiWatch halter and neck-mounted accelerometers

• Published in: Applied animal behaviour science Vol. 211; pp. 9 - 16
• Main Authors: Benaissa, Said, Tuyttens, Frank A.M., Plets, David, Cattrysse, Hannes, Martens, Luc, Vandaele, Leen, Joseph, Wout, Sonck, Bart
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier B.V 01.02.2019; Elsevier BV
• Subjects: Accelerometer; Accelerometers; Accuracy; Algorithms; animal behavior; Behaviors classification; Classification; Dairy cattle; Dairy cows; decision support systems; Decision trees; Feeding; Feeding behavior; Internet-of-animals; Machine learning; Neck; RumiWatch; Sensors; Support vector machines; RumiWatch; Internet-of-animals; Accelerometer; Dairy cows; Behaviors classification; Machine learning
• ISSN: 0168-1591; 1872-9045
• DOI: 10.1016/j.applanim.2018.12.003

•Simple decision-tree (DT) algorithm to classify feeding and ruminating behaviours.•The DT performs similar to support vector machine and a RumiWatch noseband.•The use of a simple DT would help implementing the algorithm on the on-cow sensor.•It would enable online measurements of the ingestive-related cow behaviours. A new simple decision-tree (DT) algorithm was developed using the data from a neck-mounted accelerometer for real-time classification of feeding and ruminating behaviours of dairy cows. The performance of the DT was compared to that of a support vector machine (SVM) algorithm and a RumiWatch noseband sensor and the effect of decreasing the sampling rate of the accelerometer on the classification accuracy of the developed algorithms was investigated. Ten multiparous dairy cows were used in this study. Each cow was fitted with a RumiWatch halter and an accelerometer attached to the cow’s collar with both sensors programmed to log data at 10 Hz. Direct observations of the cows’ behaviours were used as reference (baseline data). Results indicate that the two sensors have similar classification performances for the considered behavioural categories (i.e., feeding, ruminating, other activity), with an overall accuracy of 93% for the accelerometer with SVM, 90% for the accelerometer with DT, and 91% for the Rumiwatch sensor. The difference between the predicted and the observed ruminating time (in min/h) was less than 1 min. h (1.5% of the observed time) for the SVM and less than 2 min. h (2.8%) for both DT and the RumiWatch. Similarly, the difference in feeding time was 1.3 min. h (2.1%) for the SVM compared to 2.5 min. h (4.3%) and 2.4 min. h (4.1%) for both RumiWatch and DT, respectively. These preliminary findings illustrate the potential of the collar-mounted accelerometer to classify feeding and ruminating behaviours with accuracy measures comparable to the Rumiwatch noseband sensor.