Piecewise preprocessing of near-infrared spectra for improving prediction ability of a PLS model

In this study, a new strategy for preprocessing of near-infrared spectra, piecewise preprocessing (PP) is proposed. Unlike routine in optimization of preprocessing methods, in PP a spectrum is split into a number of intervals alone wavelength and the optimization of preprocessing method is independently implemented to each interval, that means that different intervals in the spectrum may select different preprocessing methods. And genetic algorithm (GA) is used in the optimization of preprocessing methods or their combinations on each wavelength interval. This strategy was tested with three near infrared (NIR) spectra datasets. Some common spectral preprocessing algorithms, such as Standard Normal Variate (SNV), multiplicative signal correction (MSC), Savitzky-Golay smoothing (smooth), first Savitzky–Golay derivative (1D), second Savitzky–Golay derivative (2D), and their combinations are used. The performance of PP was compared with the traditional method selection strategies. The results show that the proposed strategy PP is very effective in improving prediction ability of PLS models built with the pretreated spectra.