Answers are grounded in this paper and cite their sources.
Pork is among the most widely consumed meats worldwide, rendering objective, high-throughput quality assessment essential for food safety, supply chain efficiency, and consumer trust. To address this need, we introduce a rapid, non-destructive analytical framework that integrates visible–near-infrared hyperspectral imaging (HSI, 400–1100 nm) with an interpretable hybrid deep learning architecture for the quantitative prediction of five critical pork quality indicators: lightness ( L* ), pH, protein content, moisture content, and fat content. At its core lies BO-1D-CNN-XGB-Reg—a novel, purpose-built model that synergistically combines a one-dimensional convolutional neural network (1D-CNN) for hierarchical spectral feature extraction, an XGBoost regressor for robust, nonlinear prediction, and Bayesian optimization for principled, global hyperparameter tuning. Evaluated on a rigorously curated dataset of 400 pork samples exhibiting substantial heterogeneity in breed, cut, and post-mortem aging, the model consistently outperformed established benchmark algorithms—including PLSR, SVR, and standalone CNN—achieving high predictive performance, as quantified by coefficient of determination (R²p): L* (0.94), pH (0.93), protein (0.94), moisture (0.97), and fat (0.94). Crucially, to ensure scientific interpretability beyond predictive accuracy, we applied SHAP (SHapley Additive exPlanations) to quantify feature attribution across the full spectral range; this analysis robustly linked spectrally influential wavelengths to well-documented biochemical and biophysical mechanisms—including myoglobin redox transitions, protein secondary structure changes, water–protein interactions, and lipid oxidation products. Collectively, this work delivers a validated, end-to-end proof-of-concept that bridges deep learning and explainable AI (XAI), establishing a transparent, high-fidelity framework for next-generation meat quality analysis.
No takes yet. Share an insight, caveat, or question.
Wang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: