Contents lists available at ScienceDirect Computers and Electronics in Agriculture journal homepage: www.elsevier.com/locate/compag Original papers Categorising sheep activity using a tri-axial accelerometer Jamie Barwick a,b, ⁎ , David W. Lamb a , Robin Dobos a,c , Mitchell Welch a , Mark Trotter a,1 a Precision Agriculture Research Group, University of New England, Armidale NSW 2351, Australia b Sheep CRC, University of New England, Armidale, NSW 2351, Australia c NSW Department of Primary Industries, Livestock Industries Centre, University of New England, Armidale NSW 2351, Australia ABSTRACT An animal’s behaviour can be a useful indicator of their physiological and physical state. As resting, eating, walking and ruminating are the predominant daily activities of ruminant animals, monitoring these behaviours could provide valuable information for management decisions and individual animal health status. Traditional animal monitoring methods have relied on labour intensive, human observation of animals. Accelerometer technology offers the possibility to remotely monitor animal behaviour continuously 24/7. Commercially, an ear worn sensor would be the most suitable for the Australian sheep industry. Therefore, the aim of this current study was to determine the effectiveness of different methods of accelerometer deployment (collar, leg and eartag) to differentiate between three mutually exclusive behaviours in sheep: grazing, standing and walking. A subset of fourteen summary features were subjected to Quadratic Discriminant Analysis (QDA) with 94%, 96% and 99% of grazing, standing and walking events respectively, being correctly predicted from ear acceleration signals. These preliminary results are promising and indicate that an ear deployed accelerometer is capable of identifying basic sheep behaviours. Further research is required to assess the suitability of accelerometers for behaviour detection across different sheep classes, breeds and environments. 1. Introduction Behaviour can provide a useful indication of the physiological state of livestock (Frost et al., 1997). Observation of the individual animal’s posture and locomotion is often the first step in determining its overall health and welfare (Moreau et al., 2009; Weary et al., 2009). If beha- viour could be monitored continuously it would provide an objective measure of individual activity from which animal health and welfare could be inferred. However, behavioural assessment is difficult (Martiskainen et al., 2009) and until recently, animal activity has only been quantifiable through direct observation or video monitoring, both of which are labour and time consuming (Müller and Schrader, 2003; Trénel et al., 2009). In a commercial context, the inspection of grazing ruminants needs to be conducted in a short period of time and as infrequently as possible to ensure operational efficiency (Edwards, 2007). One limitation with the conventional approach of direct observation is that it only provides a behavioural assessment over the period in which the animals are actually observed. The consequence of this is that the behavioural states being relied upon for health and welfare assessment (for example grazing activity or travelling) may not be actually observed during inspection. Furthermore, grazing livestock are often farmed under conditions where regular human monitoring is either physically im- possible (e.g. inaccessible terrain) or not cost effective, making constant daily observation impractical (Moreau et al., 2009). As a consequence, animals can go uninspected for extended lengths of time. Sensors which can automatically measure the behaviour of animals have the potential to alleviate some of these issues (Blokhuis et al., 2010), and as such, interest in developing automated measures of animal behaviour has increased (Rushen et al., 2012). One form of animal-borne sensing which has the potential to mea- sure behaviour autonomously is accelerometer technology (Howell and Paice, 1989). Some accelerometers use the piezoelectric effect in which stresses on microscopic crystals (caused by acceleration) generate a measurable voltage. Alternatively the piezo-resistive effect produces a change in the material resistance due to stresses effected by accelera- tion, again which can be measured in an appropriate circuit (via a generated voltage). Another approach is where the acceleration may change the capacitance that exists between two microstructures in close proximity to one another other. All of these (and similar approaches) generate an electrical signal which may be converted into a measure of the accelerative activity along defined axes. This signal is generated by https://doi.org/10.1016/j.compag.2018.01.007 Received 20 September 2017; Received in revised form 6 January 2018; Accepted 10 January 2018 ⁎ Corresponding author at: Precision Agriculture Research Group, University of New England, Armidale NSW 2351, Australia. 1 Present address: Central Queensland University, CQIRP, Rockhampton QLD 4702, Australia. E-mail address: jbarwic2@une.edu.au (J. Barwick). Computers and Electronics in Agriculture 145 (2018) 289–297 0168-1699/ © 2018 Elsevier B.V. All rights reserved. T