Evaluation of proper sensor position for classification of sheep behaviour through accelerometers

Abstract Automated classification and identification of feeding behaviour through sensors, has the potential to allow improvements in animal health and welfare. Sensor position, as well as epoch setting of segmented signal data, affects classification accuracy in behavioural activity discrimination. The aim of this study was to evaluate the effects of position, at different time epoch setting, of a tri-axial accelerometer-based device (BEHARUM) on the classification of behaviour types such as grazing, ruminating and other activities in sheep grazing a chicory-based sward. Sheep were video recorded, during accelerometer deployment, by fixed camera. Mean, variance and inverse coefficient of variation (i.e., mean/ standard deviation), of the raw acceleration data for each axis, as well as the resultant variance and inverse coefficient of variation values of the three axes, were calculated for the following positions: mouth, nape, collar and epoch settings: 5 s, 10 s, 30 s, 60 s, 120 s, 180 s, 300 s. Video recordings were coded manually assigning to each position and epoch the prevailing behaviour. Multivariate discriminant analysis (DA) was used to distinguish between the three behaviour types. To evaluate the performance of DA in discriminating the three activities, overall accuracy and Coehn’s k coefficient were calculated, based on the error distribution in assignment. Mouth position recorded the highest accuracies and Coehn’s k coefficient in 300 s time epoch setting (88 % and 0.8 respectively), as well as collar position (90 % and 0.8 respectively), while nape position showed the best performance in 180 and 300 s (80 % and 0.6 respectively). Overall, the best performance of DA was obtained using collar position in particular at 300 s time epoch.