Uninformation Variable Elimination and Successive Projections Algorithm in Mid-Infrared Spectral Wavenumber Selection

A hybrid effective wavenumber selection method, uninformation variable elimination (UVE) combined with successive projections algorithm (SPA), were applied in mid-infrared (MIR) spectra analysis. As 3727 wavenumbers in full MIR spectra were too complicated to be trained directly in the chemomtric models, UVE was executed to eliminate noisy, useless or irrelevant wavenumbers, and 1241 wavenumbers were obtained. SPA was then operated to move collinearly and redundant wavenumbers. Finally, 15 effective MIR wavenumbers were obtained, and inputted into least-squarer support vector machine (LS-SVM) for calcium content prediction in powdered milk. A good prediction result was obtained with coefficient of determination (r2) =0.930, residual predictive deviation (RPD) =3.773 and root mean square error of prediction set (RMSEP) =30.162. The results of UVE-SPA are little worse than those of UVE, but were better than those of full MIR spectra. As the numbers of wavenumbers are much smaller than those selected by UVE, the performance of UVE-SPA was still acceptable. The good performance showed a potential application using UVE-SPA to select MIR effective wavenumbers.