Aim and Objectives:
Imagine the challenge of identifying the sows with the best mothering abilities for breeding. These are the sows that will raise thriving offspring that grow into productive, robust and calm pigs. The maternal behavior of sows is a cornerstone of piglet survival, directly influencing the health and development of the next generation. Understanding and identifying the best mothering traits in sows, those that ensure high piglet survival, require meticulous observation.
However, manually observing maternal behaviors presents considerable challenges. It is a labor-intensive, time-consuming process that is often prone to human error, limiting its effectiveness on a larger scale.
This is where automation steps in. By leveraging technology to analyze maternal behaviors, we can extract relevant traits with precision and efficiency. Automation not only enhances the accuracy of observations but also offers unparalleled scalability, allowing for data-driven insights across larger datasets.
Dataset:
The dataset forms the backbone of this innovative approach to maternal behavior analysis. It contains 24h videos analysis of sow's pens that capture detailed behavioral data. The data material is divided into three critical phases of their reproductive cycle. The pre-farrowing phase, spanning 2 to 7 days, captures key behaviors leading up to the parturition process. The farrowing phase, lasting 6 to 12 hours, provides real-time insights into maternal actions during labor and delivery. Finally, the post-farrowing phase, which extends for approximately 32 days, focuses on the sow's interactions with her piglets, highlighting nurturing behaviors essential for piglet survival. This comprehensive and time-segmented dataset is instrumental in understanding and identifying the traits that define exceptional mothering abilities.
Video observations and maternal traits shortlist:
After thorough video observations and discussions on sow’s position and negative maternal traits, we shortlisted the key maternal traits of interest. These traits fall into three distinct categories: behavioral traits observed through video analysis alone, behavioral traits derived from a combination of video and audio analysis, and traits related to the sow’s exterior (physical) characteristics. The behavior traits from video include udder availability when the sow is laterally lying, nest building behavior, restlessness during the farrowing time, the time interval between the newborns, the average time to first suckling, and disturbances in suckling patterns among piglets and the nursing frequency during the first ten days. The behavior traits from both video and audio recordings encompass piglet screams, the sow's reaction to piglet screams, and nursing initiation (grunting sounds) by the sow. Lastly, the sow exterior traits from video cover the time spent in posture changes, the sow not getting up when feed comes, and efforts to maximize udder access when lying down.This comprehensive approach ensures a holistic understanding of the maternal qualities essential for optimal piglet care and survival.
Annotation and training:
After identifying the targeted maternal traits, the next step was to outline the required annotation tasks. Annotation is the process of labeling the data to help the model understand what to focus on, providing a foundation for effective model training. It also ensures consistency and precision in identifying specific behaviors. We have annotated sufficient volume of data to train deep learning models to automate the detection of sows' maternal behaviors, paving the way for scalable and precise behavior analysis. We have trained models for sow detection, sow posture classification, piglet detection, newborn piglet classification, and sow nursing behaviour detection. Additionally, we have developed methods to analyse sow nestbuilding and farrowing restlessness by post-processing the trained models.
Towards conclusion:
While the initial journey has presented its share of challenges, such as data collection and annotation complexities, our progress underscores the potential of innovative solutions in overcoming these hurdles. As we continue to refine our methods and systems, we remain optimistic about the lasting impact this project will have on automating the sow’s maternal traits detection to enhance piglet survival rates.
Animal welfare traits have been included in Norsvin’s breeding programme since 2001, and maternal ability was included in 2010. Currently, mothering ability traits, including piglet survival and exterior score, have a low heritability. In addition, certain traits are scored by employees, with risk of subjectivity and being high cost when recording at scale. This results in a low response to selection as observed in the last years.
This project aims to develop multi-modal analysis methods for obtaining, objectively and automatically, phenotypes describing positive mothering abilities. By doing so, we expect to get improve specificity of the phenotypes, and thereby improve the heritability, genetic progress, and piglet survival.
We will focus on measuring traits that have been identified in the literature as positively correlated with piglet survival, like gait-related traits and farrowing related traits. We will acquire data with multiple modalities to better capture the complexity of the traits: sow communication will require combining audio and video of the free-farrowing pen, whereas leg conformation will require lateral capture of the animal. By using existing and developing new analysis methods based on machine learning, we will extract high quality phenotypes that will reflect the traits of interest. Finally, we will be analysing these phenotypes against genomic data for these animals, with the final objective to include these phenotypes in our breeding programme to directly increase piglet survival and thus animal welfare.
