牛肉丸是一种口感独特的肉类深加工食品。 不法商贩为了谋取利益, 在牛肉中掺入猪肉、 鸡肉等廉价肉制作肉丸冒充纯牛肉丸售卖。 传统的肉品掺假检测方法费时费力, 成本高昂。 高光谱成像技术具有快速无损、 低成本等优点, 因此对牛肉丸中掺假猪肉和鸡肉进行高光谱成像检测。 首先分别制作纯牛肉丸和混有掺假肉猪肉和鸡肉的牛肉丸, 掺假肉占原料肉质量比例分别为5%, 10%, 15%, 20%, 25%。 采集所有肉丸样本的高光谱信息并提取光谱数据。 分别采用1st Der, 2nd Der, MC, MSC, SG和SNVT六种预处理方法对所提取光谱进行预处理, 建立全波段下偏最小二乘(PLS)掺假含量预测模型, 并比较模型预测效果得出最佳预处理方法。 对最佳预处理方法处理后的光谱数据进行特征波长的筛选, 筛选方法有: 连续投影法(SPA)、 竞争性自适应重加权算法(CARS)、 联合区间偏最小二乘法(siPLS), 并创新性地联用siPLS与CARS的联合区间偏最小二乘-竞争性自适应重加权算法(siPLS-CARS)。 最后比较不同波长筛选方法下的模型预测效果。 研究表明, 牛肉丸掺猪肉和鸡肉PLS预测模型最佳预处理方法分别为MSC和1st Der。 SPA, CARS和siPLS-CARS分别筛选了掺猪肉样品光谱中的13, 51和32个特征波长, siPLS将全光谱分为14个子区间, 联合第1, 3, 7, 13子区间进行建模, 其中CARS筛选波长后的PLS预测模型效果最好, RC和RP分别为0.981 4和0.972 1, RMSECV和RMSEP分别为0.016 3和0.020 3。 SPA, CARS和siPLS-CARS分别筛选了掺鸡肉光谱中的15, 61和28个特征波长, siPLS将全光谱分为15个子区间, 联合第7, 8, 11, 12子区间进行建模, 最佳波长筛选方法也是CARS, 此时PLS预测模型RC和RP分别为0.990 2和0.987 8, RMSECV和RMSEP分别为0.012 3和0.012 6。 siPLS-CARS相比于siPLS不仅缩减了特征波长数量, 且提高了模型预测的精度； 相比于CARS筛选出的波长更少, 但精度略低。 掺鸡肉样品预测模型效果整体优于掺猪肉样品。 研究结果表明高光谱成像技术可以实现牛肉丸中掺假的含量预测, 为牛肉丸掺假快速检测提供理论基础。

Beef meatball is a deep-processed meat product with a unique taste. In the market, some unscrupulous traders cashed in on mixing beef with cheap meat such as pork and chicken to make meatballs. The traditional methods of meat adulteration detection are time-consuming and costly. Hyperspectral imaging technique has the advantages of fast, non-destructive and low cost on meat test. Therefore, the detection of beef meatballs adulterated with pork and chicken was carried out by hyperspectral imaging technique in this study. Adulterated meat was added to the beef meatballs at a level of 0, 5%, 10%, 15%, 20% and 25% of the quality of raw meat respectively. All meatballs hyperspectral data were collected while their spectral data were extracted. The spectral data were pretreated by six methods, first derivative (1st Der), second derivative (2nd Der), mean centering (MC), multiplicative scatter correction (MSC), Savitzky-Golay (SG), standard normal variate transformation (SNVT), which established the Partial least squares model of adulteration content at the full-wave band and obtained the optimum pretreatment method by comparison. After the optimum pre-processing method, the characteristic wavelengths were screened by successive projections algorithm (SPA), competitive adaptive reweighted sampling (CARS), synergy interval partial least squares (siPLS), synergy interval partial least squares-competitive adaptive reweighted sampling (siPLS-CARS), for the purpose of comparing, the prediction effects of models were evaluated on different screening wavelengths methods. The results suggested that the best pre-processing methods of PLS prediction model for beef meatballs adulterated with pork and chicken were MSC and 1st Der. 13, 51 and 32 characteristic wavelengths of adulterated pork spectra were screened by SPA, CARS and siPLS-CARS, respectively. The characteristic subinterval combinations were screened by siPLS: the full-wave band was divided into 14 subintervals, which was then combined with the 1st, 3rd, 7th, and 13th subintervals to establish PLS prediction models. The prediction model of adulterated pork content by CARS wavelength screening method had the best effect, with the RC and RP at 0.981 4 and 0.972 1 respectively, while RMSECV and RMSEP at 0.016 3 and 0.020 3 respectively. 15, 61 and 28 characteristic wavelengths of adulterated chicken spectra were screened by SPA, CARS and siPLS-CARS, respectively. The full spectrum was divided into 15 subintervals by siPLS, combined with the 7th, 8th, 11th, and 12th subintervals to establish PLS prediction models. Analogously, the prediction model of adulterated chicken content by CARS wavelength screening method had the best effect as well, with RC and RP at 0.990 2 and 0.987 8 respectively, and RMSECV and RMSEP at 0.012 3 and 0.012 6 respectively. In this study, compared with siPLS, siPLS-CARS not only reduced the number of characteristic wavelengths but also improved the accuracy of the model prediction. Compared with CARS, it screened for fewer wavelengths, but with slightly lower accuracy. Compared with adulterated pork, the prediction model of adulterated chicken was better on the whole. The research results suggested that hyperspectral imaging technique can realize the content prediction of adulterated pork and chicken in beef meatballs, which provides a theoretical basis for rapid detection of beef meatball adulteration.