An Objective Wavelength Selection Method Based on Moving Window Partial Least Squares for Near-Infrared Spectroscopy

An objective wavelength selection method is proposed for near-infrared (NIR) spectroscopy mainly to overcome the possible subjectivity introduced by moving window partial least squares regression (MWPLS). This improved procedure (iMWPLS) introduced an indicator to evaluate importance of each wavelength and then all wavelengths were ranked by these indicators. On the basis of the indicator ranking, a series of PLS models were constructed by starting with one wavelength and incorporating a new one until all wavelengths were involved. Finally, according to root mean square error of cross-validation (RMSECV) obtained by each model, wavelengths that constructed the optimal model were selected as informative ones while the others were discarded. Subsequently, this new objective procedure was applied to two real standard NIR datasets and the prediction performance was compared with full-spectrum PLS and the original MWPLS. Results demonstrated that iMWPLS could achieve an effective wavelength selection and improve predictive accuracy in near-infrared spectroscopy.