Identification of roasting degree and interpretability analysis of Yunnan arabica coffee beans based on multi-dimensional visual features and CNNs-SHAP.

This study extracted Yunnan Arabica coffee bean features through weighted fusion of CIE L*a*b* color histograms, Gray-level Co-occurrence Matrix-Local Binary Pattern composite textures, and morphological parameters, aiming to achieve accurate roasting degree identification and transparent decision-making. Convolutional neural networks outperformed all other models with the highest accuracy. Subsequent SHapley Additive exPlanations analysis demonstrated that key features exhibited a significant strong negative linear correlation with prediction outputs; negative contributions dominated light roasting samples, whereas robust positive contributions dominated dark roasting samples. External verification results clarified that the accuracy of dark-roasted samples reached 100.0%, that of light-roasted samples was 91.5%, and that of medium-roasted samples was 93.8%. The proposed intelligent system enabled automatic image acquisition, feature extraction, and inference, and supported connection to the Enterprise Resource Planning system for data traceability. This study broke deep learning's black-box limitation, providing a precise interpretable technique for coffee roasting standardization.