Scholarlink

<p>This study utilizes a handheld VIS-NIR spectrometer equipped with Machine Learning, under LED and Bulb illumination, to rapidly predict the freshness stage of minced mutton, veal, and their mixtures over 7 days (the first day and the subsequent six days of storage at 4&#xa0;°C) using SVM &amp; MLP classification models. A total of 84 samples were analyzed, with 5 distinct points per sample. Among all configurations, the highest performance for mutton samples was achieved using SVM under LED illumination with stratified sampling, yielding a prediction accuracy and F1-score of 96.4%. The corresponding 5-fold cross-validation accuracy was 90.3%, confirming strong generalization. For veal samples, the optimal result was obtained with an MLP under LED illumination using stratified sampling, yielding a prediction accuracy and F1-score of 96.4% in prediction, with a cross-validation accuracy of 86.7%. In mixed meat samples (40% mutton, 60% veal), the best classification was achieved using an MLP under Bulb illumination with stratified sampling, yielding a prediction accuracy of 85.7% and an F1-score of 85.2%. The corresponding cross-validation accuracy was 72.1%, reflecting the increased spectral complexity of heterogeneous tissues. Across all models, stratified sampling consistently outperformed the Kennard-Stone method, and LED illumination generally enhanced classification performance, except in mixed samples. These findings confirm the feasibility of VIS-NIR spectroscopy combined with machine learning, providing a means for real-time, fine-grained, day-level classification of freshness in minced meat, and demonstrating its practical suitability for handheld spectrometers and on-site quality control applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Machine learning-based rapid freshness evaluation of minced red meat (mutton and veal) using miniaturized portable VIS-NIR spectroscopy

  • 期刊论文

  • 发表日期：

    2026年5月25日

查看全文
  • Ali Mohammad Kazempour,
  • Sajad Kiani

摘要

This study utilizes a handheld VIS-NIR spectrometer equipped with Machine Learning, under LED and Bulb illumination, to rapidly predict the freshness stage of minced mutton, veal, and their mixtures over 7 days (the first day and the subsequent six days of storage at 4 °C) using SVM & MLP classification models. A total of 84 samples were analyzed, with 5 distinct points per sample. Among all configurations, the highest performance for mutton samples was achieved using SVM under LED illumination with stratified sampling, yielding a prediction accuracy and F1-score of 96.4%. The corresponding 5-fold cross-validation accuracy was 90.3%, confirming strong generalization. For veal samples, the optimal result was obtained with an MLP under LED illumination using stratified sampling, yielding a prediction accuracy and F1-score of 96.4% in prediction, with a cross-validation accuracy of 86.7%. In mixed meat samples (40% mutton, 60% veal), the best classification was achieved using an MLP under Bulb illumination with stratified sampling, yielding a prediction accuracy of 85.7% and an F1-score of 85.2%. The corresponding cross-validation accuracy was 72.1%, reflecting the increased spectral complexity of heterogeneous tissues. Across all models, stratified sampling consistently outperformed the Kennard-Stone method, and LED illumination generally enhanced classification performance, except in mixed samples. These findings confirm the feasibility of VIS-NIR spectroscopy combined with machine learning, providing a means for real-time, fine-grained, day-level classification of freshness in minced meat, and demonstrating its practical suitability for handheld spectrometers and on-site quality control applications.