Determining the onset of heat stress in a dairy herd based on automated behaviour recognition

Dairy cows have various strategies for dealing with heat stress, including a change in behaviour. The aim of this study was to propose a deep learning-based model for recognising cow behaviours and to determine critical thresholds for the onset of heat stress at the herd level. A total of 1000 herd behaviour images taken in a free-stall pen were allocated with labels of five behaviours that are known to be influenced by the thermal environment. Three YOLOv5 architectures were trained by the transfer learning method. The results show the superiority of YOLOv5s with a mean average precision of 0.985 and an inference speed of 73 frames per second on the testing set. Further validation demonstrates excellent agreement in herd-level behavioural parameters between automated measurement and manual observation (intraclass correlation coefficient = 0.97). The analysis of automated behavioural measurements during a 10-day experiment with no to moderate heat stress reveals that lying and standing indices were most responding to heat stress and the test dairy herd began to change their behaviour at the earliest ambient temperature of 23.8 °C or temperature-humidity index of 68.5. Time effects were observed to alter the behavioural indicators values rather than their corresponding environmental thresholds. The proposed method enables a low-cost herd-level heat stress alert without imposing any burden on dairy cows.