Method: Using a commercial precision livestock farming activity collar to automatically record and classify dairy cow activity at pasture

Precision livestock farming technologies are increasingly being implemented on farms to enhance the management of key processes such as reproduction and feeding. Accelerometer technologies are the most spread and are able to provide a large quantity of data on animal activity. However, these data need to be validated against gold standards before being used further in research. We aim at validating the output from Axel Medria® device, a three-axis accelerometer sensor that automatically processes the raw data and classifies the main activity by 5−min epoch, for which the manufacturer does not disclose the classification algorithm. Two groups of six cows were observed during 30 h each, grazing on pasture, during two trials. The objective was to compare the agreement between sensor data and visual observations at different time windows. We used a confusion matrix analysis to assess the correspondence between visual observation and the output of the Medria algorithm and linear regressions associated along with a Bland-Altman analysis to compare the time budgets retrieved from the two sources. We focused on three activities (grazing, ruminating and resting) and on the posture of the animal (standing/lying). Sensitivity was >73.5% for all activities except for resting (48.8%). Specificity reached 87.6–91.9% for all activities but posture showed a poorer result (67.0%). Nevertheless, accuracy was above 80% for the three activities and the posture and precision were more variable, the best results being obtained for posture (88.3%) and for grazing (93.6%). Linear regressions showed slopes between 0.73 and 0.99 for all activities and of 0.81 for posture, but differences between observers across the two trials have been observed for resting. R² were more variable, ranging from 0.30 (for resting in second year) to 0.84 for grazing. The Bland-Altman analysis showed good results despite significant bias for grazing, rumination and resting (only the first year). Due to the technology embedded in Axel Medria ® sensors, their performances were slightly lower than that of other devices which technologies are more precise for estimating specific behaviour (e.g. recording jaw movements is more precise to detect rumination). Nevertheless, Axel Medria ® sensors can provide indicators on different activities and over longer periods of time. The tested device, largely applied on commercial farms, showed good agreement with visual observation. Data can thus be used as a proxy to study dairy cow behaviour at pasture, on large cow groups over a long time, in experimental or commercial farms.