On-farm 3D images of beef cattle for the prediction of carcass classification traits and cold carcass weight

For beef cattle, subjective methods tend to be used on-farm for assessing readiness for slaughter. This means that the target classification grades cannot be accurately estimated, leading to over- and under-finished animals being sent to slaughter. This leads to financial losses and potentially to increased greenhouse gas emissions. To counteract this, objective technologies are needed to select slaughter cattle at the optimal point. Video image analysis systems have proven their suitability for the estimation of carcass traits postslaughter. There is potential for this technology to support on-farm assessment of the live animal. This study explored 3D measurements extracted from images of live beef animals and their ability to predict key carcass traits when coupled with additional animal or carcass details. The images were captured from four commercial beef finishing farms across the UK and the cattle were processed at commercial abattoirs. Data for 762 animals were used to build models for the prediction of EUROP conformation and fat class, and cold carcass weight (CCW), using either traditional statistics (multiple linear regression using stepwise feature selection) or machine learning techniques (random forest models). Various model inputs were combined and tested, including breed, sex, the camera unit that the images were captured on, the number of days before slaughter the images were captured, and the month of image capture as fixed effects, the 34 3D measurements, and CCW. The best linear models predicted EUROP conformation class with moderate accuracy (R² = 0.37), EUROP fat class with low accuracy (R² = 0.24) and CCW with moderate accuracy (R² = 0.38). Moderate accuracies were also found when using the machine learning methods, with the best random forest models predicting EUROP conformation and fat classes with moderate accuracy (58 and 45% of classes, respectively, predicted correctly) and CCW with moderate accuracy (R² = 0.41). The results indicate that there is potential for imaging systems on farm to predict key carcass traits currently assessed in the abattoir, providing a tool for farmers to objectively select cattle at the optimum conditions for slaughter.