A weighted multiscale regression for multivariate calibration of near infrared spectra

Abstract

A weighted multiscale regression for building a combined model in multivariate calibration of near infrared spectra is proposed. In the approach, the spectra are decomposed into different scale blocks (or frequency components) by wavelet transform (WT) at first, then partial least squares (PLS) models are built with the decomposed components, and at last a combined model is built by a weighted averaging. The weight of each model is determined by the prediction residual error sum of squares (PRESS) value obtained with Monte Carlo cross validation (MCCV). The underlying philosophy of the strategy is that useful information may be embedded in all the components obtained by WT, although the higher and lower frequency components mainly represent noise and background, respectively. To validate the effectiveness and universality of the proposed method, it was applied to two different sets of near-infrared (NIR) spectra of tobacco lamina. Compared with the results obtained with commonly used PLS methods, the proposed method is proved to be a high-performance tool for multivariate calibration of complex NIR spectra.