Determination of Calcium and Magnesium in Tobacco by Near-Infrared Spectroscopy and Least Squares-Support Vector Machine
TIAN Kuang-da1, QIU Kai-xian1, LI Zu-hong2, Lü Ya-qiong2, ZHANG Qiu-ju2, XIONG Yan-mei1, MIN Shun-geng1*
1. Department of Applied Chemistry, College of Science, China Agricultural University, Beijing 100193, China 2. Qujing Tobacco Company, Yunnan Tobacco Company, Qujing 655000, China
Abstract：The purpose of the present paper is to determine calcium and magnesium in tobacco using NIR combined with least squares-support vector machine (LS-SVM). Five hundred ground and dried tobacco samples from Qujing city, Yunnan province, China, were surveyed by a MATRIX-I spectrometer (Bruker Optics, Bremen, Germany). At the beginning of data processing, outliers of samples were eliminated for stability of the model. The rest 487 samples were divided into several calibration sets and validation sets according to a hybrid modeling strategy. Monte-Carlo cross validation was used to choose the best spectral preprocess method from multiplicative scatter correction (MSC), standard normal variate transformation (SNV), S-G smoothing, 1st derivative, etc., and their combinations. To optimize parameters of LS-SVM model, the multilayer grid search and 10-fold cross validation were applied. The final LS-SVM models with the optimizing parameters were trained by the calibration set and accessed by 287 validation samples picked by Kennard-Stone method. For the quantitative model of calcium in tobacco, Savitzky-Golay FIR smoothing with frame size 21 showed the best performance. The regularization parameter λ of LS-SVM was e16.11, while the bandwidth of the RBF kernel σ2 was e8.42. The determination coefficient for prediction (R2c) was 0.975 5 and the determination coefficient for prediction (R2p) was 0.942 2, better than the performance of PLS model (R2c=0.959 3, R2p=0.934 4). For the quantitative analysis of magnesium, SNV made the regression model more precise than other preprocess. The optimized λ was e15.25 and σ2 was e6.32. R2c and R2p were 0.996 1 and 0.930 1, respectively, better than PLS model (R2c=0.971 6, R2p=0.892 4). After modeling, the whole progress of NIR scan and data analysis for one sample was within tens of seconds. The overall results show that NIR spectroscopy combined with LS-SVM can be efficiently utilized for rapid and accurate analysis of calcium and magnesium in tobacco.
[1] CHU Xiao-li(褚小立). Molecular Spectroscopy Analytical Technology Combined with Chemometrics and Its Applications(化学计量学方法与分子光谱分析技术). Beijing: Chemical Industrial Press(北京：化学工业出版社), 2011. 23. [2] Vapnik V N. Statistical Learning Theory. New York: Wiley-Interscience, 1998. [3] Suykens J A K，Vandewalle J. Neural Processing Letters, 1999, 9(3): 293. [4] Morón A，Cozzolino D. The Journal of Agricultural Science, 2002, 139(4): 413. [5] WANG Dong, DING Yun-sheng, YUAN Xing-fen(王 冬，丁云生，袁杏芬). Modern Instruments and Medical Treatment(现代仪器), 2008, 14(5): 19. [6] Kennard R W, Stone L A. Technometrics. American Society for Quality Control, 1969, (11): 137. [7] Xu Qingsong, Liang Yizeng. Chemometrics and Intelligent Laboratory Systems, 2001, 56(1): 1.
