Can an automated labelling method based on accelerometer data replace a human labeller? - Postural profile of farrowing sows

The objective of this research was to develop a method for the automated classification of sow's postures based on accelerometer data. The second objective was to evaluate if the technique can be used in a reliable way instead of human labelling to automatically create postural profiles of sows. The experiment took place at the research farm of The University of Veterinary Medicine Vienna. The farm had a herd of 120 Large White sows. Data were collected from sows housed in three types of farrowing pens: wing, trapezoid and SWAP (Sow Welfare and Piglet Protection). The behaviour of 18 sows was video recorded and labelled for a period of 24 h before the start of farrowing until the end of farrowing. The focus of labelling was on three types of postural behaviours: active, resting in lateral position (RLP) and resting in sternal position (RSP). Each sow had a specific ear tag with a 3 axis accelerometer sensor. Acceleration was measured at a frequency of 10 Hz. Linear discriminant analysis (LDA) was used to classify the three labelled postural behaviours on the basis of acceleration data. To evaluate how the LDA classifier would generalise to an independent dataset an 18-fold cross-validation method was used. The overall classification accuracy of postural behaviours with the developed method was 70% in cross-validation and 73% on the training set. Comparison of labelling and classification results revealed that the accuracy is not good enough to detect the effect of pen type on the behaviour of the animals. However, the influence of crating on time spent resting in either a lateral or sternal position was correctly recognised with the automated method. The developed method could be further tested for the automated monitoring of the health and welfare status of sows.