Discrimination of biting and chewing behaviour in sheep using a tri-axial accelerometer

• Published in: Computers and electronics in agriculture 2020-01, Vol.168, p.105051, Article 105051
• Main Authors: Alvarenga, F.A.P., Borges, I., Oddy, V.H., Dobos, R.C.
• Format: Article
• Language: English
• Subjects: Acceleration; Accelerometers; Accuracy; Algorithms; Behavior; Classification; Datasets; Decision trees; Evaluation; Machine learning; Mastication; Micro-swards; Model accuracy; Pasture plots; Precision; Sensitivity; Sheep; Studies; Time synchronization
• ISSN: 0168-1699; 1872-7107
• DOI: 10.1016/j.compag.2019.105051

•Discrimination of biting and chewing behaviours from sheep while grazing.•Three time intervals were tested: 1, 3 and 5 s.•The 5 s window was the best to identify and classify Bite and Chewing behaviours.•Tri-axial accelerometer can discriminate Bite and Chewing behaviours in sheep. The aim of the current studies was to evaluate the capability of a tri-axial accelerometer, attached to the under-side of a halter and positioned on the under-jaw of a sheep, to discriminate biting and chewing activities of sheep during grazing. Two studies were conducted, the first study evaluated the effect of two diverse pasture species on feeding behaviour using micro-sward boxes: forage oats (Avena sativa cv Eulabah) and perennial ryegrass (Lolium perenne cv Wimmera). Two, 4-year old Merino ewes grazed each species for approximately four, two minute sessions over two separate days, one week apart. In the second study, the effect of sward height was investigated using nine plots of ryegrass with three different sward heights (mean ± se 4.0 ± 0.15, 6.2 ± 0.17 and 10.3 ± 1.05 cm; P = 0.005) grazed by three 3-year old Merino ewes for 10 min each. Video recordings of behaviours from both studies were visually assessed and annotated into Bite, Chewing and Other. They were then manually synchronised in time with accelerometer output to create annotated data files which were partitioned into three time intervals (1 s, 3 s and 5 s). Forty-four features were calculated from the acceleration signals and used to classify behaviours using a decision tree to determine model accuracy, sensitivity, specificity and precision. For the micro-sward study, Bite activity was classified with a precision of 90.5% for the evaluation data set, whereas for the validation data set it was classified with a precision of 98.1% for the 5 s time interval. Accuracy of the decision-tree model increased as time interval increased for both data sets. For the sward height study, as time interval increased model sensitivity for Bite and Chewing activity improved from 91.2% to 95.5% and from 75.0% to 93.0%, respectively, while model specificity improved from 88.1% to 98.2% and from 92.1% to 95.9%, respectively in the evaluation data set. The same pattern occurred when the model was applied to the validation data set. The accuracy of the decision-tree algorithm to classify Bite, Chewing and Other activities increased as time interval length increased for both data sets. These two studies have shown that tri-axial