NIRS features and multi-model optimization fusion enabled comprehensive method for quantitative and qualitative...

To address the limitations in near-infrared spectroscopy (NIRS) for lamb meat quality (LMQ) assessment, specifically incomplete index coverage, constrained model accuracy, and the lack of a multi-index-based evaluation method, this study presents a comprehensive method for quantitative and qualitative assessment of LMQ. By combining NIRS features with multi-model optimization and fusion, the proposed method enables precise LMQ control, supports targeted marketing, and promotes maximized economic value. Quantitative models for key quality indices (KQIs) were constructed using a multi-model optimization fusion algorithm based on a convolutional neural network, the Rime-ice-based optimizer, and least squares support vector machine (CNN-RIME-LSSVM). Building on this, this study constructed a KQIs-based LMQ grading method using a multi-method approach integrating Analytic Hierarchy Process (AHP), Entropy Weighting Method (EWM), membership function, Ward's hierarchical clustering, and K-means. The grading information obtained from this grading method is associated with NIRS features to develop a qualitative model for quality grades. This qualitative model was developed using a composite algorithm, namely the Generalized Regression Neural Network optimized by the Caterpillar Fungus Optimizer (CFO-GRNN), thereby achieving rapid and intelligent grading of LMQ. Both the quantitative model (with R2P = 0.9792-0.9932 and RMSEP = 0.0155-0.0283) and the qualitative model (with Accuracy = 99.75% and Macro F1-score = 99.84%) demonstrated excellent performance. SHapley Additive exPlanations (SHAP) explains the prediction behavior of the quantitative and qualitative models, enhancing transparency and interpretability of the decision-making mechanism. This study provides a new approach for on-line quality detection and intelligent grading of LMQ.