Non-destructive monitoring of excessive-fat pork belly using VCS-2000 and machine learning algorithms

This study investigates a non-destructive classification framework for identifying excessive-fat pork bellies at the 10th and 11th thoracic vertebrae (TV10 and TV11, respectively) using automated carcass traits extracted from the VCS 2000 system coupled with machine learning. Reference fat percentages of 117 pork carcasses were determined via hyperspectral imaging, with consumer acceptability thresholds established at 45% for TV10 and 40% for TV11 to segregate normal and excessive-fat groups. Six distinct classification models were optimized using a stratified grid search validation approach. Principal component analysis revealed severe spatial overlap and linear indistinguishability between the normal and excessive-fat groups in both anatomical sections. Overall classification performance for the TV10 section remained limited across all evaluated frameworks. In contrast, the optimized Random Forest model demonstrated promising predictive performance at the TV11 position, achieving a test set accuracy of 80.00%, an F1-score of 85.00%, and a recall of 94.44%. Shapley Additive exPlanations analysis identified the engineered carcass fatness index and the first lumbar vertebra backfat thickness as the primary variables of model classification. Given the strong biological correlation observed between these adjacent sections (r_p = 0.883), strategically deploying the validated Random Forest model at the TV11 section provides a robust, real-time proxy for monitoring excessive fat trends across the thoracic belly region during commercial slaughter operation without compromising primal carcass value, pending validation across larger, multi-abattoir datasets.