Advancing bovine tissue discrimination with Vis–NIR spectroscopy coupled with machine learning methods

Increasing the accuracy of segmentation of tissue types during automated carcass breakdown would contribute to optimising automated cutting decisions and reducing waste, relative to red-green-blue (RGB)-based vision systems. This study investigates the development of tissue-classification models using visible and near infrared (Vis-NIR) spectra and spectral features combined with machine learning (ML). A total of 310 samples (up to 25 samples from eight tissue categories) were collected from a commercial beef-processing line and cleaned. Principal Component Analysis (PCA) of pre-processed spectra clearly separated fat, marrow, and muscle, with varying degrees of overlap among remaining tissue categories (ligament, tendon, bone, cartilage), using the first two PCs. Five ML algorithms, i.e. Random Forest (RF), Support Vector Machines (SVM), K-Nearest Neighbours (KNN), Linear Discriminant Analysis (LDA), and Partial Least Squares Discriminant Analysis (PLS-DA) were used to construct classification models for tissues. Models were trained and tested using the full spectral range and using selected feature wavelengths (FWs). All models performed extremely well, but LDA, PLS-DA, and SVM models achieved the highest performance, with kappa values and mean sensitivities exceeding 99%, while RF and KNN achieved 92-97%. Models constructed using selected FWs achieved comparable accuracy to full-spectrum models. Further analysis confirms that models constructed using NIR-derived FWs were approximately 20% more accurate than those constructed from FW deriving from the visible spectral range only, supporting the superior accuracy of spectral approaches compared to RGB for automated carcass characterisation. These findings support the robustness of Vis-NIR sensors with ML for automated tissue characterisation, highlighting their potential for industrial deployment to achieve improved accuracy and yield in automated meat deboning.