Automated Welfare Assessment of Gestating Sows: Machine Learning Unveils Individual Variability

By Soham NandiReviewed by Susha Cheriyedath, M.Sc.Dec 4 2023

In an article published in the journal Nature, researchers aimed to estimate gestating sow welfare components by leveraging machine learning and behavioral data. The approach involved clustering behaviors, linking clusters to sow attributes and environmental factors, and ultimately employing machine learning for welfare classification.

Background

Assessing animal welfare at an individual level on farms is increasingly important for sustainable and ethical livestock management. While existing protocols, like the Welfare Quality program, evaluate welfare at the group level, understanding individual variability is crucial, considering animals' unique perceptions and responses to stimuli. Previous studies often overlooked individual welfare assessment, focusing more on group dynamics.

Technological advancements in precision livestock farming, particularly integrating sensors and connected devices, provide an avenue for individual animal monitoring. Radiofrequency identification (RFID) recognition enables real-time data collection on behaviors and activities. However, analyzing the vast datasets generated by these technologies requires sophisticated tools, such as machine learning.

This study focused on gestating sows and aimed to categorize their welfare based on behavioral data, specifically social interactions and physical activity. Initial clustering of behaviors into three groups (scapegoat, gentle, aggressive) was a key step, offering insights into individual responses to environmental perturbations. Interpreting clusters involved examining sow characteristics, health status, and environmental conditions, providing a comprehensive understanding of welfare determinants.

The subsequent application of machine learning, specifically a decision tree, enabled the classification of individual sows into welfare categories. The study drew on continuous variables related to feeding behavior and postures, extracted from sensors and cameras during 2-day intervals over 6-10 weeks.
This innovative approach not only addressed the current gap in individual welfare assessment but also demonstrated the potential for automated decision support systems in categorizing welfare based on behavioral components.

Method

The study aimed to automatically estimate the welfare status of farm animals using machine learning and behavioral data. The approach involved clustering gestating sows based on behavioral data collected through manual and automatic video analysis, with a total of 69 sows observed over 6 to 10 periods. Clustering utilized the K-medoids algorithm, identifying three subgroups of sows. Offline learning involved interpreting clustering results, considering individual sow characteristics and experimental conditions. A decision tree, known for interpretability, was then applied for supervised classification based on feeder and automatic video analysis data.

Throughout the study, sows were housed in groups, and various environmental events were induced, such as competitive feeding, sound events, thermal variations, and enrichments. Animal-based measures and electronic feeders provided additional data on health, skin lesions, and feeding behavior. The decision tree was trained on a subset of the data and evaluated for its ability to classify sows into the welfare clusters determined by the clustering step.Results from Cochran's Q test, Friedman tests, and post hoc tests were used to assess the impact of factors like week, group, and events on the clusters. The decision tree's interpretability was emphasized for understanding algorithmic decision rules.

The study, conducted from July 2020 to April 2021, presented a novel approach to individual welfare assessment in farm animals. By integrating behavioral data, clustering, and a decision tree, the research demonstrated a potential automated decision support system for categorizing welfare based on real-time observations. The comprehensive methodology involved both offline learning and online forecasting processes, showcasing the adaptability of machine learning techniques for enhancing animal welfare monitoring on farms.

Results

The study employed clustering to identify three distinct behavioral patterns ('gentle,' 'aggressive,' and 'scapegoat') in gestating sows based on activity and social interactions. Clusters exhibited associations with welfare, with 'gentle' sows predominant during certain events, and 'aggressive' sows more prevalent in thermic and impoverishment events. Notably, sow health influenced cluster dynamics, as unhealthy sows tended to be more 'aggressive.' Group variations and temporal dynamics in cluster assignments were observed.

A decision tree, trained on feeder and video analysis data, demonstrated 80% accuracy in classifying sows into behavioral clusters. 'Gentle' and 'aggressive' sow classifications exhibited higher accuracy (F1-scores of 0.86 and 0.80) than 'scapegoat' (F1-score of 0.70). The decision tree's interpretability highlighted key variables influencing classification, such as daily eating time and non-nutritive feeder visits. Additionally, the combination of feeder and video analysis data improved overall performance.

Discussion

The study's innovative approach, utilizing clustering and classification on automatically recorded on-farm data, offered promising prospects for estimating the welfare status of gestating sows. Behavioral features, such as physical activity and social interactions, used for clustering, have known associations with welfare components. The method's adaptability to farm data without requiring prior knowledge or annotation was a notable advantage. However, challenges in interpreting cluster meanings underscored the need for further validation on additional datasets. The correlation between the identified clusters and health status, particularly the link with 'aggressive' behavior, aligned with the broader concept that behavioral changes can signal illness, emphasizing the importance of behavioral indicators in welfare assessment.

Surprisingly, sow characteristics had minimal influence on clustering, suggesting the method's potential to estimate individual welfare components at specific times. The study also highlighted the impact of group behavior on individual welfare, emphasizing the social nature of pigs. The decision tree's success in predicting welfare status, especially for 'gentle' and 'aggressive' sow classifications, underscores the significance of feeding behavior and activities in welfare determination. Further validation on diverse farms was essential to ensure the method's robustness across varying environmental conditions.

Conclusion

In conclusion, this study represented a significant advance in estimating the welfare status of gestating sows by integrating machine learning techniques. Combining clustering for data labeling and a decision tree for interpreting sensor data, the approach showed promise. While further interpretation of clusters was needed, the method could be integrated into a decision support system for proactive animal welfare management, alerting farmers to potential issues based on comprehensive monitoring of living conditions.