Automated detection of piling behaviour in UK laying hen flocks

Abstract Piling is an understudied group behaviour where chickens gather in dense clusters. It appears to be widespread in commercial laying hens, leads to mortalities (smothers) and impacts production. Automated detection of piling could support: (i) stockpersons to reduce the behaviour and (ii) scientific efforts to study the behaviour. Computer vision-based classification is increasingly used in animal behaviour research to accelerate annotation of datasets. Here we aimed to automatically detect piling using the well-established residual network (ResNet) architecture convolutional neural network (CNN) to classify video frames of piling and non-piling chickens in commercial free-range laying farms. Data from 10 flocks (8 for training, 2 for testing) were used. The model achieved high classification accuracy, correctly identifying 89.6% of individual frames and 90.9% of aggregated events, though misclassifications at both levels were driven primarily by false negatives (16.2% for frames and 16.8% for events). Cases of misclassification could arise from not accounting for the temporal aspect of piling, human annotation error, image quality, flock-specific characteristics or ambiguous cases. ResNet demonstrated good performance for classifying piling from non-piling from still frames in brown laying hens in commercial houses but would need further training data to be robust to different housing types and chicken breeds.