A novel wavelength interval selection based on split regularized regression for spectroscopic data

摘要

Wavelength selection has become a critical step in the analysis for near-infrared (NIR) spectroscopy with high co-linearity and large number of spectral variables. In this study, a novel wavelength interval selection method based on split regularized regression and partial least squares (SplitReg-PLS) is developed. SplitReg-PLS is a two-step approach, which combines the advantage of the SplitReg and PLS methods. SplitReg presents interesting properties, which can split the variables into groups and pool the regularized estimation of the regression coefficients together as groups. The PLS regression is one of the most popular methods for multivariate calibration, and is performed on the selected group variables by using the SplitReg. The SplitReg-PLS method can automatically select successive strongly correlated and interpretable spectral variables related to the response, which provides a flexible framework for variable selection. The performance of the proposed procedure is evaluated by three real NIR datasets. The results indicate that SplitReg-PLS is a good wavelength interval selection strategy.