Assessment of thermal imaging to objectively body condition score mature horses and multiparous gestating beef cows

This study explored whether thermal imaging could provide an objective means of body condition scoring (BCS) horses and multiparous, gestating beef cows. This study consisted of two parts: one part assessed BCS in horses of the Quarter Horse or Thoroughbred breed types while the other evaluated BCS of gestating beef cows. Ground truth BCS were assigned by five to eight trained scorers for each animal. Thermal images were also collected from one or both sides of the body and analyzed for surface temperature. Surface temperature and BCS were evaluated with the whole body, and for five (cows) or seven (horses) specific body regions. Covariates were monitored, including breed, ambient temperature, cloud coverage (%), housing conditions, blanketing status, individual scorer, and scorer's location. Considerable between-scorer variation in BCS assigned was apparent in both species, with 58% to 65% of observations falling within 0.5 BCS of the median. Much of the variability was attributable to scoring the degree of BCS deviation from normal, and 72% to 82% of scores agreed in directional accuracy. Scaled, centered scores within 0.5 units of the median BCS were defined as high-agreement data. Random forest regressions were derived to explore how well thermal-camera-obtained body surface temperature data could be used to estimate BCS of horses and cows using either all available data or the subset of high-agreement data. In both datasets, 15% of observations were used for hyperparameter tuning, 55% for model-training, and 30% for independent evaluation. Unique models were created for each body area assigned a BCS for horses (the neck, shoulder, withers, ribs, loin, tailhead, and overall) or beef cows (the brisket, shoulder, ribs and spine, hooks and pins, and overall). Models estimating horse BCS had lower RMSE (6.71% to 10.4%) and higher accuracy (62.7% to 90.3%) than those estimating cow BCS (RMSE: 6.39% to 15.9%; accuracy: 46.8% to 82.5%). Models derived from all data had higher RMSE (7.76% to 15.9%) and lower accuracy (53.8% to 82.3%) compared with those derived from only high-agreement data (RMSE: 5.80% to 8.18%; accuracy: 66.7% to 90.3%). Overall, thermal imaging shows promise as an objective tool for BCS assignment in both horses and beef cows. However, future refinement of the method for assigning BCS ground truth is important.