Toward the automatic detection of social interactions in gestating sows using image analysis data

The identification of social interactions for group-housed gestating sows is crucial in order to monitor their welfare. They are key to understanding the social structure of the group and ensuring that disturbances, such as an abnormal frequency of aggression, are promptly detected. Assessing social interactions can be done through direct observation, manually recording the number and valence of interactions (positive or negative). This approach requires substantial human and financial resources. The study assesses the feasibility of using data produced by an image analysis software identifying sows' postures and key points to automatically detect and classify social interactions. Two pens of 19 and 20 gestating sows each were video recorded from 00:00 to 02:00 on days 30 and 103 of two consecutive gestations. The videos were analyzed manually using the annotation system of the software in development. In total, 120 moments of positive interactions and 120 moments of negative interactions were annotated, including images before, during, and after the interaction. In addition, 120 moments without interactions (two sows not in contact) were annotated. The valence of the interaction (i.e., positive or negative), and for each interacting sow, the posture (i.e., standing, sitting, lying), and the coordinates of 3 key points (nose, neck, tail) were identified. Relative distances between sows, movement speeds, and individual distances traveled were calculated. Decision trees were performed to assess the relevance of these variables to detect an interaction. The start and end of an interaction were reliably detected with 88% accuracy using distances and postures. Interactions were characterized by a convergence of key points, spatial proximity, and at least one sow adopting a standing posture. The valence was effectively determined using the movement speeds and individual movements of sows during and after the interaction (respectively 74 and 80% accuracy). Negative interactions were characterized by faster movements and avoidance behavior. The performance in detecting sows' orientation during interaction varied across categories (nose-nose, nose-neck, nose-tail, from 33% to 100%), and was particularly good for nose-tail interactions (94% F-score). This study shows the potential of the automatic image analysis software under development to detect and classify social interactions through post-processing analysis of key points and posture data.