不同预处理方法对烟草近红外光谱预测模型的影响

Influence of Data Pre-processing on NIR Prediction Model for Tobacco Chemical Composition

采用基线校正、去卷枳、一阶微分、二阶微分、主成分回归(PCR)和偏最小二乘(PLS)法对198个烟叶样品的近红外光谱和总糖、还原糖、总烟碱含量数据进行了处理,建立了相应的总糖、还原糖和总烟碱校正模型,并将这些模型的回归参数作了比较。结果表明:①二阶微分处理光谱建立的模型的相关系数比基线校正、去卷枳、一阶微分预处理法建立的模型的高,而其相对偏差比基线校正、去卷枳、一阶微分法的低;②PLS算法建立的定量分析模型比PCR算法的好。

The data of NIR spectra, total sugar, reducing sugar and total nicotine from 198 tobacco samples were processed by four different data pre-processing approaches, baseline correction, decongvolution, first differential and second differential, and two different statistic methods, principle component regression (PCR) and partial least square (PLS). The respective NIR calibration models for total sugar, reducing sugar and total nicotine in tobacco were developed and the regressive parameters of those models were compared. The results indicated that: 1) the model developed by NIR spectra processed with second differential approach had the highest correlation coefficient, and the lowest relative deviation; and 2) the quantitative model established with PLS was better than that with PCR.