[A local regression method for near-infrared spectral quantitative analysis of tobacco samples]

A local regression method based on distance criterion in principal component (PC) space for near-infrared (NIR) spectral quantitative analysis was proposed. In this method, principal component analysis (PCA) is firstly utilized to extract the information of the NIR spectra, and then, the calibration subsets are individually selected for each prediction sample according to the distance between the sample and calibration samples in the PCs space. Finally, the PLS local model for every prediction sample is established individually and the prediction of the sample is done with the local model. It was found that the Euclidean distance can more effectively measure the similarity of the samples than Mahalanobis distance. With an application of the local regression method to the quantitative determination of chlorine and nicotine in tobacco samples, it is proved that the prediction precision of local regression method is better than that of global regression methods, especially in the situation of predicting the low concentration components.