Three-dimensional (3D) measurements extracted from beef carcass images were used to predict the weight of four saleable meat yield (SMY) traits (total SMY and the SMY of the forequarter, flank, and hindquarter) and four primal cuts (sirloin, ribeye, topside and rump). Data were collected at two UK abattoirs using time-of-flight cameras and manual bone out methods. Predictions were made for 484 carcasses, using multiple linear regression (MLR) or machine learning (ML) techniques. Model inputs included breed type, sex, and abattoir as fixed effects, and cold carcass weight, visually assessed EUROP fat and conformation classes, and 3D measurements as covariates. Machine learning techniques were only used for models including 3D measurements. The CCW and fixed effects resulted in high accuracy (SMY R2=0.72-0.90, RMSE=2.12-3.96kg, primal R2=0.56-0.67, RMSE=0.36-0.91kg), and including the EUROP covariates increased accuracies (SMY R2=0.75-0.96, RMSE=2.00-3.11kg, primal R2=0.58-0.79, RMSE=0.36-0.79kg). The 3D measurement covariates and abattoir resulted in moderate accuracy (SMY MLR R2=0.39-0.58, RMSE=3.26-10.31kg, primal MLR R2=0.33-0.52, RMSE=0.44-1.14kg) and high accuracy when combined with CCW and all fixed effects (SMY MLR R2=0.72-0.95, RMSE=1.81-3.42kg, primal MLR R2=0.52-0.74, RMSE=0.40-0.81kg). The best ML models resulted in similar accuracies to the MLR models. Models including 3D measurements produced similar accuracies to models built using conventional data recorded at the abattoir, indicting the potential for automated prediction.
Read full abstract