Classification of sheep urination events using accelerometers to aid improved measurements of livestock contributions to nitrous oxide emissions

We attached tri-axial accelerometer sensor tags to thirty Welsh Mountain ewes for thirty days to assess if we could identify urination events. We used random forest models using different sliding mean windows to classify behaviours. Urination had a distinctive pattern and could be identified from accelerometer data, with a 5 s window providing the best recall and a 10 s window giving the best precision. 'State' behaviours considered (foraging, walking, running, standing and lying down) were also identified with high recall and precision. This demonstrates the extent to which the identification of discrete 'event' behaviours may be sensitive to the window size used to calculate the summary statistics. The method shows promise for identifying urination in sheep and other livestock, being minimally invasive compared to other methods, and has clear potential to inform agricultural management practices and policies.