Automated identification of social dominance status in group-housed pigs using AI

Abstract

In commercial pig farming, regrouping pigs during different production stages often leads to increased agonistic behaviors, one of the essential indicators of pig welfare, which has negative effects on their performance and overall well-being. It has been known that this behavior is closely related to dominance status and monitoring its dynamics is essential for improving pig welfare and managing the well-being of pigs. Traditionally, identifying the dominance status required manual observation, which is not feasible for continuous monitoring. The availability of low-cost video surveillance provides a partial solution but still requires extensive data analysis. Our study addresses this challenge by introducing a two-stage deep learning framework for automating: 1) the detection of agonistic interactions, 2) the identification of initiators or receivers, and 3) the classification of winners or losers of the agonistic interactions. Our CNN-RNN network, combining ResNet18 and GRU, achieves over 93% accuracy in each task. Particularly, our model outperforms existing pig agonistic detection models by using global average pooling for spatial feature downsampling, incorporating bounding box coordinates as additional features, and a temporal attention layer for more accurate predictions. We also demonstrate the proposed two-stage framework’s ability to automatically identify each pig’s dominance status based on the models’ predicted outcomes. Our findings show the framework’s accuracy in predicting the dominance status, particularly for dominant individuals, underscoring its effectiveness in capturing the social dominance of pigs. This research not only advances the process of behavioral monitoring and analysis in pigs but also contributes significantly to improving welfare assessments and practices in smart livestock farming (SLF).