Tracking performance in poultry is affected by data cleaning method and housing system

Sensor-based behavioural observation methods improve our understanding of individual behaviour and welfare in large commercial groups, including poultry. Validating automatically generated data is essential to account for potential sources of error. Our study aimed to validate a sensor-based tracking system for broiler breeders (BB) and laying hens (LH) in commercially relevant housing systems. The BB study was conducted in 10 pens with 33 females and three males (Ross 308) per pen. Half of the pens contained a raised slatted area and two raised group nests (Raised), while in the remaining five pens, the nests and slats were on the floor (Floor). For the LH study, six pens with a commercial aviary were used, with 225 Dekalb White hens housed per pen (Aviary). Focal hens (BB, 10/pen; LH, 18/pen) were equipped with backpacks containing tracking devices that registered transitions between four (BB) or five (LH) resource-related zones covering all accessible areas within each housing system. The tracking data was compared against video observations for 20 focal BB on two days and 18 focal LH on three days (3 × 20 min/day). Three data cleaning methods tested with 30 values of a duration parameter were evaluated for reliability and stability with a cross-validation approach. Initial and post-cleaning performance were assessed with accuracy, precision, and sensitivity of recorded transitions and by calculating the reliability for two aspects of movement: total transitions (Lin’s Concordance Correlation Coefficient) and locations (mean proportion of matching duration). A mixed model was applied to evaluate the duration of stay after false and true tracking registrations. Initial location reliability was high (> 0.949) in all housing systems, while reliability of total transitions was low ( 0.832) while reliability of locations remained high (> 0.949) in Aviary and Raised. The duration between registrations was affected by housing system (p < 0.001) and was longer for true compared to false registrations (p < 0.001). Initial tracking performance varied between movement aspects and housing systems. The difference in duration between true and false registrations allowed for the application of simple yet effective data cleaning in Aviary and Raised, ensuring that the generated data better represented the animal's actual movement with reduced error associated with the tracking system