摘要：采用近红外光谱分析技术建立预测模型，可以实现桑叶含水率、粗蛋白和可溶性糖的实时无损检测，达到筛选优质桑叶的目的。本试验分别利用烘干法、凯氏定氮法和蒽酮-硫酸比色法测定桑叶含水率、粗蛋白和可溶性糖。选用MicroNIR1700和DLPNIRScanNano两种型号的近红外光谱仪分别采集鲜桑叶和桑叶粉的近红外光谱，光谱预处理后进行偏最小二乘回归法（PLS）建模，并采用竞争性自适应重加权算法（CARS）、无信息变量消除算法（UVE）和随机蛙跳算法（RF）筛选特征波长变量，以提高模型的预测精度。

试验结果证明，采用原始光谱数据建模得到的鲜桑叶含水率的预测模型效果最好，MicroNIR1700光谱的RMSEC、RMSECV和RMSEP分别为0.96%、1.31%和0.96%，R2、R2和R2分别为0.931、0.871和0.938，Bias为-0.01%。DLPNIRScanNano光谱的RMSEC、RMSECV和RMSEP分别为1.09%、1.32%和1.31%，C、RCV和RP分别为0.913、0.873和0.881，Bias为-0.10%。CARS筛选MicroNIR1700光谱波长变量建立的鲜桑叶粗蛋白模型预测结果最好。RMSEC、RMSECV和RMSEP分别为0.55g/100g、0.88g/100g和1.01g/100g，C、RCV和RP分别为0.945、0.863和0.841，Bias为0.32g/100g。桑叶粉MicroNIR1700光谱的UVE优选结果最佳，校正集的RMSEC和R2分别为0.58g/100g和0.933，交叉验证RMSECV和R2分别为0.82g/100g和0.867，预测集的RMSEP和R2分别为0.93g/100g和0.864，Bias为0.02g/100g。UVE筛选的鲜桑叶和桑叶粉粗蛋白DLPNIRScanNano光谱波长的建模效果最优，鲜桑叶DLPNIRScanNano光谱的校正集的RMSEC和R2分别为0.50g/100g和0.950，交叉验证的RMSECV2分别为0.92g/100g和0.831，预测集的RMSEP和R2分别为1.15g/100g和0.804，Bias为-0.03g/100g，桑叶粉DLPNIRScanNano光谱的RMSEC、RMSECV和RMSEP分别为0.62g/100g、0.93g/100g和1.01g/100g，R2、R2和R2分别为0.923、0.826和0.824，Bias为0.46g/100g。鲜桑叶可溶性糖MicroNIR1700和DLPNIRScanNano光谱数据的UVE筛选结果最好，MicroNIR1700光谱的RMSEC、RMSECV和RMSEP分别为1.72%、2.38%和2.39%，R2、R2和R2分别为0.823、0.670和0.742，Bias为1.36%。DLPNIRScanNano光谱的RMSEC、RMSECV和RMSEP分别为1.70%、2.50%和1.80%，R2、R2和R2分别为0.813、0.608和0.821，Bias为0.57%。Randomfrog优选的桑叶粉MicroNIR1700光谱数据建模效果最好，校正集、交叉验证和预测集的均方根误差分别为1.41%、2.34%和2.14%，相关系数分别为0.849、0.608和0.735，Bias为-0.63%。桑叶粉DLPNIRScanNano光谱的UVE优选结果建模效果最好，校正集的RMSEC和R2分别为1.14%和0.921，交叉验证的RMSECV和CV分别为2.13%和0.732，预测集的RMSEP和RP分别为2.58%和0.560，Bias为-1.18%。结果表明，基于近红外光谱技术结合波长优选可以实现对桑叶含水率、粗蛋白快速无损检测。同时，还需要继续研究如何建立可以检测桑叶可溶性糖的预测模型。

关键词：桑叶；近红外光谱分析技术；偏最小二乘回归法；波长优选

Abstract：The near-infrared spectroscopy (NIRS) technique was used to establish the prediction model, which could realize the real-time non-destructive testing of mulberry leaf water content, crude protein and soluble sugar, and to achieve the purpose of screening high quality mulberry leaves. In this study, moisture content, crude protein and soluble sugar were measured by drying method, Kjeldahl method and anthrone- sulfuric acid colorimetry method. The Near Infrared spectra of fresh mulberry leaf and mulberry leaf powder were collected by the near-infrared spectroscopy of MicroNIR1700 and DLP NIRScan Nano. The PLS model was used for spectral preprocessing and Competitive Adaptive Reweighted Sampling (CARS), Uninformative Variables Elimination (UVE) and Randomfrog (RF) to improve the prediction accuracy of the model.

The experimental results show that the prediction model of the moisture content of fresh mulberry leaves obtained by using the original spectral data is the best.The RMSEC, RMSECV and RMSEP of the MicroNIR170 spectrum were 0.96%,    1.31% and 0.96%, respectively, and R2 , R2 were 0.931, 0.871 and 0.938, respectively, and Bias was -0.01%. The RMSEC, RMSECV and RMSEP of the DLP NIRScan Nano spectrum were 1.09%, 1.32% and 1.31%, respectively, and R2 , R2     and R2  were 0.913, 0.873 and 0.881, respectively, and Bias was -0.10%. CARS screened MicroNIR1700 spectral wavelength variables established the fresh  mulberry leaf  crude  protein  model  to  predict  the  best  results. The RMSEC,RMSECV and RMSEP were 0.55g/100g, 0.88g/100g and 1.01g/100g, R2 , R2