Predicting pen fouling in fattening pigs from pig position

•Pen fouling behavior of slaughter pigs can be predicted from changes in the pigs’ positioning behavior within the pen.•Distinguishing between lying and standing behavior does not add additional information value, compared to just knowing the pigs’ position.•Morning behavior is more informative for fouling predictions than afternoon behavior.•A naïve Bayesian ensemble of all final models yields better performance than any one model achieves alone.•The improved performance of the Bayesian ensemble is seen, even though most of the final models did not perform very well on their own. Pen fouling is an undesired behaviour of fattening pigs, where they excrete in their designated resting area and rest in their designated excretion area. This causes problems with health due to poor hygiene, and requires laborious efforts for the farmer to clean the pen and correct the behaviour. A review of the existing literature suggests that changes in lying behaviour may precede an event of fouling. Furthermore, observing the lying patterns of fattening pigs in the morning before entering the fattening unit, as a means of assessing the risk of imminent pen fouling, is known to be a common strategy among Danish farmers. In this study, we show that machine learning methods, specifically random forests and artificial neural networks, can be made to predict pen fouling in the days leading up to the event, based on the position of the pigs within the pen at specific times of the day. We could not show any added information value from distinguishing between standing/lying behaviour within a given area of the pen, as opposed to simply knowing the pigs’ position. We found that the most information value, relevant for training a method for predicting fouling events, are located in the last 2–3 days before the event occurs and when the pigs are observed during the morning hours before any disturbance. Lastly, we demonstrate a Bayesian ensemble strategy for combining multiple different prediction models, which yield higher performances than the best performing models do on their own.