Accelerometer-Based Multivariate Time-Series Dataset for Calf Behavior Classification

Overview

• Getting insights on how pre-weaned calves adapt their behavior to challenges and diseases can improve calf welfare on dairy farms.
• Monitoring calf behavior using accelerometers and machine learning is a promising approach, but requires labeled data, which is time-consuming to obtain.
• To address this challenge, the researchers present the ActBeCalf dataset, which contains 27.4 hours of accelerometer data aligned with manually annotated calf behaviors.
• The dataset was used to develop machine learning models for classifying calf behaviors, demonstrating its reliability.

Plain English Explanation

Observing how young dairy calves change their behavior when faced with challenges, like transport or illness, can help farmers improve the animals' well-being. One way to monitor calf behavior is by using accelerometers (motion sensors) attached to their neck collars and then using machine learning to analyze the sensor data.

However, to train the machine learning models, researchers need "labeled" data - information that clearly identifies what the calves were doing when the sensor data was recorded. Obtaining these labels is difficult, as it requires someone to closely watch and manually record the calves' behaviors, which is time-consuming and labor-intensive.

To solve this problem, the researchers created the ActBeCalf dataset. They equipped 30 young dairy calves with accelerometer sensors and filmed them for 13 weeks. Then, they carefully watched the videos and annotated the calves' behaviors, such as lying down, standing, walking, and running, as well as less obvious behaviors like sniffing, social interactions, and grooming.

This comprehensive dataset, with both the accelerometer data and the labeled behaviors, can now be used to train machine learning models to automatically recognize and classify calf behaviors from the sensor data alone. The researchers demonstrated the reliability of the dataset by training two different behavior classification models, which achieved high accuracy.

Technical Explanation

The researchers used a 3D-accelerometer sensor attached to a neck collar to record the movements of 30 pre-weaned dairy calves (Holstein Friesian and Jersey) from one week of birth for 13 weeks. Simultaneously, the calves were filmed in their pens.

After the trial, three observers manually annotated the calves' behaviors from the videos using the Behavioral Observation Research Interactive Software (BORIS) and an ethogram with 23 different behaviors. This resulted in the ActBeCalf dataset, which contains 27.4 hours of accelerometer data aligned with the annotated calf behaviors.

The researchers then used the ActBeCalf dataset to develop and evaluate two machine learning models for classifying calf behaviors. The first model classified behaviors into two classes: active and inactive. The second model classified behaviors into four classes: running, lying, drinking milk, and an "other" class.

The first model achieved a balanced accuracy of 92%, while the second model achieved an accuracy of 84%. These results demonstrate the reliability of the ActBeCalf dataset for training machine learning models to classify pre-weaned calf behaviors from accelerometer data.

Critical Analysis

The researchers acknowledge that the manual annotation of calf behaviors from video recordings, while necessary to create the ActBeCalf dataset, is a time-consuming and labor-intensive process. This may limit the scalability of the dataset and the approach, as obtaining labeled data can be a significant bottleneck for developing and deploying behavior classification models in real-world dairy farm settings.

Additionally, the dataset is limited to 30 calves, which may not capture the full diversity of individual calf behaviors and adaptations. Expanding the dataset with more animals and diverse farming environments could improve the generalizability of the machine learning models.

The researchers also note that the ActBeCalf dataset does not include data on calf health or performance, which could be valuable for understanding the relationship between behavior and overall calf welfare. Integrating such data could further enhance the utility of the dataset for dairy farm management and calf care.

Conclusion

The ActBeCalf dataset provides a comprehensive and reliable resource for classifying pre-weaned calf behaviors from accelerometer data using machine learning. This tool has the potential to help dairy farmers gain insights into calf adaptation and welfare, which could lead to improved management practices and better outcomes for the animals.

While the manual labeling process is a limitation, the dataset demonstrates the feasibility of using accelerometer-based behavior monitoring to support calf care and welfare on dairy farms. Future research could explore ways to streamline the data collection and annotation process, as well as integrate additional data sources to provide a more holistic understanding of calf health and well-being.