欧亚国际

   本研究采用了多光谱法（VideometerLab 多光谱成像系统）来检测牛肉掺杂马肉，研究采用了多种化学计量法和机器学习技术。样品储藏在4 °C，会显著影响了模型效果。利用了2步SVM来检测纯样品以及刚出来的新鲜样品。此模型独立验证总正确区分率达到95.31%。

   近年来，食品工业和检测机构主要重点聚焦于检测制假和掺假行为。本研究目的是调查多光谱成像法（与数据分析相结合）在检测牛肉馅掺假（掺杂马肉）应用领域的潜力，并探讨冷藏储存条件下模型的效果。

   基于此原因，我们获得了来自3个不同批次的掺杂有马肉的牛肉馅的110个样品的18个波段的光谱图像。将样品在4 °C储存6、24和48小时后，再次拍摄图像。分类模型（偏较小二乘判别分析，随机森林，支持向量机）基于前两个批次创建，第3个批次放在一旁用于额外独立验证。

   结果显示，很容易区分新鲜和储存样品，而掺假样品的检测分类模型性能则显著受到储存期间肉色变化影响。利用2步SVM模型，可正确区分所有纯肉和新鲜肉，独立批次验证总正确分类可达95.31%。

Highlights

Multispectral imaging was used in the detection of beef adulteration with horsemeat.Various chemometric and machine learning techniques are applied. Samples were stored at 4 °C, significantly affecting model performance.A two-step SVM was applied for the detection of pure and freshly-ground samples.The model yielded 95.31% overall correct classification for independent validation.

Abstract

In recent years, detection of fraudulent and deceptive practices has become a major priority in the food industry and inspection authorities. The aim of this study was to investigate the potential of multispectral imaging coupled with data analysis methods for the detection of minced beef adulteration with horsemeat, as well as to explore model performance during storage in refrigerated conditions.

For this reason, multispectral images of 110 samples from three different batches of minced beef and horsemeat in 18 wavelengths were acquired. Images were taken again after samples were stored at 4 °C for 6, 24 and 48 h. Classification models (partial least squares discriminant analysis, random forest, support vector machines) based on the first two batches were developed while the third batch was set aside for external/independent validation.

Results showed that freshly-ground and stored samples were clearly distinguishable, whereas classification model performance for detection of adulterated samples was significantly affected by changes in meat color during storage. Using a two-step SVM model however, all pure and freshly-groundsamples were classified correctly and the overall correct classification was equal to 95.31% for independent batch validation.