An automated sensor-based method of simple behavioural classification of sheep in extensive systems

Abstract

Automated sensor-based behaviour classification systems are already developed and used for cattle in an agricultural context. To develop a similar tool for sheep we investigated the use of pitch and roll tilt sensors to provide information about sheep behaviour. The aim was to determine if different behaviour types associated with grazing can be accurately identified using tilt sensor data. We collected data from two adjacent enclosures at two contrasting sites, one in the Southern Uplands grazed by Cheviot ewes and one in the Western Highlands of Scotland grazed by Scottish Blackface ewes. In addition to our observations of ewes on hill pasture we also observed collared ewes in flat and gently sloped fields and in a shed to evaluate the tilt data. The data was analysed using three classification methods, a linear discriminant analysis, a classification tree method, and a manually developed decision tree consisting of four “if then” loops. We found that we can distinguish between the two behaviour categories “active” and “inactive”, even if only pitch tilt data is used. All three methods provide very good classification predictions with more than 90% correct results. However, the classification tree method was less robust than the other methods. Using the manually developed decision tree, we produced results of activity that were robust and credible. By using the behaviour classification we are able to collect data on different genotypes, systems or management options in sheep farming, and in combination with GPS we will be able to improve management strategies and gain information about grazing ecology. It can also be seen as a first step to support farmers with a viable system to comply with welfare regulations in the light of European Union Common Agricultural Policy and as an integral part for future developments regarding virtual fencing technology for sheep.