Improvement of Partial Least Squares Modelling for Determination of Soil Nitrogen by Fourier Transform Near-Infrared Spectrometry

The concentrations of nitrogen in soil were analyzed by the Fourier transform near-infrared (FT-NIR) spectrometry. Using partial least squares (PLS) regression, wavelength selection is a vital task for improving the modeling ability because the predictive results partially depend on the signal-to-noise ratio of the modeling wavelengths, The discrete single-and-favorite combination linear regression (DSFCLR) method was proposed for selecting the informative wavelength combination. And the discrete combination partial least squares (DCPLS) models were established on the informative wavelengths. Compared to moving window partial least squares modeling and full-range partial least squares modeling, DCPLS modeling observe improved predictive results and appreciate validation effects. Considering the optimal selected discrete combination contained only 32 wavelengths, the computational complexity was substantially reduced. discrete combination partial least squares models with appropriate discrete wavelengths corresponded to the characteristic absorption of nitrogen can effectively overcome collinearity interruption for the linear regressions. Therefore, DSFCLR method has physical and chemical significance while retaining the simplicity of linear regression. discrete single-and-favorite combination linear regression combined with DCPLS modeling is expected to be a new chemometric technique in spectroscopic analysis for selecting discrete wavelength combination.