基于近红外光谱的烤烟油分识别研究OA北大核心CSTPCD

[Objective]In order to scientifically predict the flue-cured tobacco leaf oil levels and achieve rapid spectral identification of flue-cured tobacco with different oil levels.[Method]A total of 299 samples of Yunyan 87 tobacco leaves with different oil levels and full leaf position from representa-tive tobacco planting counties were collected by near-infrared spectra.Five methods,including first derivative(Dl),normalization(NOR),wavelet transform(WAVE),standard normalization(SNV),and multivariate scattering correction(MSC),were used to preprocess the spectral data.We investigated the discriminative performance of two discriminant models,linear partial least squares dis-criminant analysis(PLS-DA)and nonlinear least squares support vector machine(LS-SVM).[Result]The accuracy of the PLS-DA oil level classification model training set constructed based on principal component analysis of near-infrared raw spectral data could reach 100.0％,but the test set was only 79.8％.After D1,NOR,SNV,and MSC preprocessing,the accuracy of the model's test set had been improved to 85.9％,90.0％,83.8％,and 83.8％,respectively.The training set accuracy of the LS-SVM oil classification model based on direct construction of near-infrared raw spectral data also reached 100.0％,and the test set reached 92.9％.After NOR,WAVE,SNV,and MSC prepro-cessing,the accuracy of the test set was improved to over 95.0％,with the highest accuracy of 99.0％achieved by MSC preprocessing.[Conclusion]The oil levels discrimination model constructed by com-bining multiple scattering correction preprocessing with LS-SVM method has the best effect,improving the efficiency of flue-cured tobacco leaf oil levels determination.