ActBeCalf: Accelerometer-based multivariate time-series dataset for calf behavior classification

Getting new insights on pre-weaned calf behavioral adaptation to routine challenges (transport, group relocation, etc.) and diseases (respiratory diseases, diarrhea, etc.) is a promising way to improve calf welfare in dairy farms. A classic approach to automatically monitoring behavior is to equip animals with accelerometers attached to neck collars and to develop machine learning models from accelerometer time series. However, this accelerometer time series data must be aligned with labels describing behaviors (gold standard) to be used for model development. Obtaining these labels requires annotating behaviors from direct observations or videos, which is highly time-consuming and labor-intensive. In addition, accurate alignment between accelerometer data and behaviors is always challenging due to time synchronization issues. However, the procedure used for the timestamp alignment has not been described in other studies. We introduce <b>ActBeCalf</b>: <i>Accelerometer Time-Series Dataset for Calf Behavior classification</i> to address this challenge. Thirty pre-weaned dairy calves (Holstein Friesian and Jersey) housed in 4 group pens at Teagasc Moorepark Research Farm (Fermoy, Co. Cork, Ireland) were equipped with a 3D-accelerometer sensor (sampling rate: 25 Hz) attached to a neck-collar from one week of birth for 13 weeks. The calves were simultaneously filmed with a high-up camera in each pen. Every 15 days, accelerometers were removed from the collars to recharge the battery, collect the data, and reattached to the neck collar again. At the end of the trial, behaviors were manually annotated from the videos using the Behavioral Observation Research Interactive Software (BORIS) by three observers (Cohen's Kappa = 0.72 ± 0.01) using an ethogram with 23 pre-weaned dairy calves' behaviors. Videos were carefully selected to annotate calves' behaviors. Observations were synchronized with the accelerometer timestamps using an external clock and aligned to the corresponding accelerometer time series. Thereby, ActBeCalf contains 27.4 h of accelerometer data from 30 pre-weaned calves (age 23.7 ± 10.7 days) aligned adequately with calf behaviors. The dataset includes the main behaviors of the calf time-budget, like lying, standing, walking, and running, as well as less prominent behaviors, such as sniffing, scratching, social interaction, and grooming. The reliability of ActBeCalf was validated by developing two machine learning models designed to classify behaviors into two and four classes, respectively. Good predictive performance was achieved for both models (balanced accuracy: 92 % and 84 %, respectively), thereby confirming ActBeCalf's reliability for model development in the field. The code utilized for the classification is publicly available in the dataset repository. ActBeCalf is a comprehensive, ready-to-use dataset, ideal for advancing research in two key areas: classifying pre-weaned calf behavior to support animal welfare initiatives and developing innovative time-series classification models in machine learning.