Improvement of Partial Least Squares (PLS) model by wavelength selection and sample selection

Abstract

Several methods to improve the performance of Partial Least Squares (PLS) regression have been proposed. These methods can be classified into two kinds, called sample selection and wavelength selection method, respectively. In the sample selection method, multi-objective Genetic Algorithms (GA) were combined with PLS to remove specific samples with systematic errors from data set. It also enables to analyze the factor of the systematic errors in detail. In the wavelength selection method, Moving Window PLS (MWPLS), Changeable Size Moving Window PLS (CSMWPLS) and Searching Combination Moving Window PLS (SCMWPLS) were proposed. These methods are based on the use of moving window which selects local wavelength region in the spectral data. By calculating PLS with the move of the window, it is possible to search the informative region or combination of several informative regions. These methods were applied to near infrared and infrared spectral data to evaluate their performances. The results showed the remarkable improvement of PLS model by these methods.