A wavelength selection method based on randomization test for near-infrared spectral analysis

Partial least squares (PLS) regression has been widely used in the analysis of near-infrared (NIR) spectroscopy. The informative wavelength selection can improve the predictive ability of the PLS models by reducing the bias introduced by the uninformative wavelength. A new method based on randomization test was proposed for wavelength selection in NIR spectral analysis. In the proposed method, a regular PLS model and a number of random PLS models are constructed at first. Then, with the regression coefficients of these models, a statistic, P, which is defined as the ratio of the number of the coefficients that are bigger than the corresponding coefficient in the regular model to the total number of the random models, is calculated for each variable. Therefore, the variables with very low P values will be the important ones for building a stable model, whereas the variables whose P value is bigger than a threshold can be eliminated. To validate the performance of the proposed method, it was applied to the PLS modeling of two NIR spectral data sets. Results show that the proposed method can effectively select the informative wavelength from the measured NIR spectra, and enhance the prediction ability of the PLS model.