Using a three-axis accelerometer to identify and classify sheep behaviour at pasture

• Published in: Applied animal behaviour science Vol. 181; pp. 91 - 99
• Main Authors: Alvarenga, F.A.P., Borges, I., Palkovič, L., Rodina, J., Oddy, V.H., Dobos, R.C.
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.08.2016
• Subjects: Activity; algorithms; animal behavior; animal production; data collection; decision support systems; Decision-tree; ewes; Grazing; heritability; meat; methane; pastures; Ruminant; Ruminantia; Sensor; walking; Decision-tree; Activity; Grazing; Ruminant; Sensor
• DOI: 10.1016/j.applanim.2016.05.026
• ISSN: 0168-1591

•We used tri-axial accelerometers to discriminate sheep behaviours at pasture.•Three time epochs were tested: 3, 5 and 10s.•The 5s time epoch had the highest precision to predict grazing behaviour.•Natural log transformed X-axis mean can identify grazing and non-grazing behaviour. Identifying and classifying feeding behaviour in free-ranging ruminants will help improve efficiency of animal production. Another potential benefit would be in understanding the role behaviour has in determining heritability of methane measurement. The aim of this study was to determine the accuracy, sensitivity, specificity and precision with which tri-axial accelerometers can identify sheep behaviour at pasture. Two studies, the first over six days and the other over two days were conducted using South African Meat Merino×Merino ewes averaging 55 (±5)kg and 22 months of age, respectively. The animals were located in either a semi-improved pasture (0.3ha) or in a small (30m2) area with access to water to observe five mutually exclusive behaviours, grazing, lying, running, standing and walking. A tri-axial accelerometer was attached to a halter on the under-jaw of each animal. Three epochs (3s, 5s and 10s) with forty-four features calculated from acceleration signals were used to classify behaviours. The five most important features for each epoch were determined using random forest and the five behaviours were classified using a decision-tree algorithm to determine model accuracy, sensitivity, specificity and precision. The decision-tree algorithm correctly classified 90.5, 92.5 and 91.3% of the evaluation data set for grazing behaviour for the 3, 5 and 10s epochs, respectively. There was no difference in the accuracy between the evaluation and validation data sets for grazing behaviour at each epoch. The model predicted grazing and running behaviour highly accurately and with the highest precision, sensitivity and specificity for the validation data set for the 10s epoch. The 5s epoch for both the evaluation and validation data sets was selected as the most suitable epoch based on the Kappa values. We successfully identified from the distribution of component populations that the natural log-transformation of the mean of X-axis accelerations for each epoch could identify grazing and non-grazing states. Therefore, this methodology will be useful in identifying sheep activity for research applications such as before methane measurement using portable accumulation chambers or other applications addressing temporal grazing patterns.