A new multivariate calibration model transfer method of near-infrared spectral analysis

Abstract

Near-infrared (NIR) has been successfully applied for rapid and nondestructive analyses in diverse fields, but the issue of calibration model transfer has not been clarified thus far. Typically, the calibration model developed for one instrument cannot be applied directly to other instruments in many cases. To achieve calibration model transfer between the master and slave instruments, an improved piecewise direct standardization designated Rank–Kennard–Stone-PDS (Rank-KS-PDS) is proposed in this paper. An optimal sample selection method (Rank-KS) is proposed to select the transfer samples. The process of sample selection is based simultaneously on the distribution of spectral space and of property space. Thus, the samples selected by Rank-KS have greater representativeness and wider coverage. The Rank-KS-PDS method has been used to predict the alkaloid and glycoside content in tobacco. Comparative studies of calibration model transfer with the proposed method, random selection (RS), and KS have also been performed. For the prediction model of alkaloids, the comparative experiment results have verified the reduction in the number of transfer samples required from 12 to 8 and in the root-mean-square error of prediction (RMSEP) from 0.1440 to 0.1250 with the proposed method. For the prediction model of glycosides, the accuracy of model transfer was found to improve significantly, accompanied by a reduction in RMSEP from 1.6945 to 1.5850 without any increase in the number of transfer samples. With the proposed Rank-KS-PDS method, the selected transfer samples based on one property can be directly applied to the calibration model transfer of other properties with satisfactory results.