高光谱图像结合特征变量筛选定量检测羊肉中狐狸肉掺假

为了探讨快速无损检测羊肉糜中狐狸肉掺假含量的可行性，该研究利用高光谱技术结合特征变量筛选方法开展了其定量检测研究。利用遗传算法、竞争性自适应重加权算法和二维相关光谱分析（Two-Dimensional Correlation Spectroscopy，2D-COS）3种方法分别对代表性样品全部846个波长进行特征波长筛选，得到207、34和14个特征波长；基于全部波长和特征波长建立羊肉糜中狐狸肉掺假含量的偏最小二乘回归（Partial Least Squares Regression，PLSR）和支持向量回归（Support Vector Regression，SVR）模型并进行比较。研究结果表明，基于全部波长和特征波长建立的SVR模型性能均优于PLSR模型。其中，利用2D-COS方法提取的14个特征波长建立的SVR模型（即2D-COS-SVR模型）性能最优，其预测集决定系数和均方根误差分别为0.928和3.00%，相对分析误差为4.85，表明高光谱结合2D-COS-SVR模型可以有效实现羊肉糜中狐狸肉掺假的定量检测。该研究结果为开发低成本肉类掺假检测系统提供技术支持和参考依据。

Quantitative detection of fox meat adulteration in mutton by hyper spectral imaging combined with characteristic variables screening

This study aims to explore the rapid nondestructive detection of fox meat adulteration in minced mutton using hyperspectral imaging technology combined with characteristic variables screening. A quantitative detection model was also established. A total of 120 adulterated mutton samples were first prepared by adding fox meat into minced mutton at different levels, including 5%, 10%, 15%, 20%, 25%, 30%, 35%, 40%, 45%, and 50%. Spectral information of samples with different adulterated contents was then obtained using visible near-infrared hyperspectral imaging. Some approaches were selected for spectral pre-processing, including First Derivative (1D), the Mean Center (MC), Multiplicative Scattering Correction (MSC), and Standard Normalized Variate (SNV). A Partial Least Squares Regression (PLSR) model was established to determine the 1D optimal pre-processing using the original and the pre-processed spectra. The prediction performance of model was significantly improved, where the determination coefficient (R2) value of calibration set increased from 0.925 to 0.940, the R2 value of cross-validation set increased from 0.894 to 0.911, the R2 value of validation set increased from 0.896 to 0.912, and the relative error increased from 2.37 to 2.73, indicating better prediction ability of model, compared with no pre-processed spectra. The pre-processed spectra effectively enhanced the difference of spectral data. There were also obvious absorption and reflection bands at specific wavelengths. Furthermore, Genetic Algorithm (GA), Competitive Adaptive Reweighted Sampling (CARS), and Two-Dimensional Correlation (2D-COS) analysis were used to screen the characteristic wavelengths after 1D pre-processing. The PLSR and Support Vector Regression (SVR) models were then established to compare with the total 846 wavelengths and characteristic ones. It was found that 207, 34, and 14 characteristic wavelengths were obtained by GA, CARS, and 2D-COS. More importantly, the performances of all SVR models using the whole wavelengths and characteristic wavelengths were better than that of PLSR model. Among them, the best performance was achieved in the SVR model with 14 characteristic wavelengths from 2D-COS, where the R2 value and root mean square error of cross-validation set were 0.928 and 3.00%, respectively, while the relative error of validation set was 4.85. Consequently, the hyperspectral imaging combined with 2D-COS-SVR model can effectively realize the quantitative detection of the fox meat adulterated in the minced mutton. The findings can also provide a strong technical support for the development of a low-cost meat adulteration detection system.