Main content area

Automatic detection of parturition in pregnant ewes using a three-axis accelerometer

Smith, Daniel, McNally, Jody, Little, Bryce, Ingham, Aaron, Schmoelzl, Sabine
Computers and electronics in agriculture 2020 v.173 pp. 105392
accelerometers, algorithms, automatic detection, dystocia, ewes, field experimentation, lambing, lambs
An on-animal sensor based research tool is currently being developed to support research studies of dystocia in lambing ewes. Given dystocia is associated with a prolonged period of parturition, the tool is being developed with the intention to automatically estimate the duration of parturition using motion data from a 3-axis accelerometer fitted to the ewe. In this study, novel algorithms were proposed to detect the start of parturition for pregnant ewes, post-partum. This is the first step towards the ultimate goal of estimating the parturition duration.The algorithm operated on the basis that the start of parturition could be identified by a significant change in a pregnant ewe’s behaviour. Changes in an individual ewe’s activity levels were measured across time by computing the distance between an activity distribution and distribution of baseline activity, using either the Maximum Mean Discrepancy (MMD) or the Earth Mover’s Distance (EMD). The start of parturition was estimated from time periods with the most significant differences to baseline activity. The performance of six different detection algorithms were investigated for 76 pregnant ewes that were fitted with our proposed collar system for up to 17 days and whose births were recorded during the field trials. The study showed the best performing algorithm variant had a Mean Absolute Error (MAE) of 5.33 h between the estimated and human observed start of parturition.The MAE was of a similar duration to the parturition process itself, which suggests the current approach was not capable of estimating the start of parturition for a majority of the ewes in the trial. Whilst it was evident that a majority of ewes exhibited salient changes in activity during parturition, activity needed to be modelled over sufficiently long periods (2 ⩽ time period ⩽6 h) to capture the association. The detection algorithms were shown to be capable of estimating the birth date of lambs with 84% of the ewe’s parturition estimates falling within 12 h of the actual birth time.