Hybrid variable selection in visible and near-infrared spectral analysis for non-invasive quality determination of grape juice

Several wavelength variable selection algorithms were compared to analyze visible and near-infrared (Vis–NIR) spectra for the non-invasive quantitative determination of soluble solids content (SSC) and pH in grape juice. In order to eliminate useless variables and improve the signal/noise ratio, the pretreated full spectra were firstly calculated by different informative variable selection methods. Uninformation variable elimination (UVE) did better than interval partial least squares (iPLS), synergy interval partial least squares (siPLS) and backward interval partial least squares (biPLS). Successive projections algorithm (SPA) was further operated to select variables. Finally, nine and eleven variables were obtained for respectively SSC and pH analyses. The better results of UVE–SPA–PLS models compared to those of SPA–PLS models in both SSC and pH analyses show that it is necessary to execute UVE before SPA, which can both reduce the calculation time and increase the model's performance. Furthermore, two common used calibration methods, PLS and multiple linear regression (MLR), were compared. UVE–SPA–MLR obtained better results than UVE–SPA–PLS in both SSC and pH analyses. The coefficients of determination for prediction set (rp2) and residual predictive deviation (RPD) obtained by UVE–SPA–MLR are 0.979 and 6.971 for SSC, and 0.951 and 5.432 for pH. The overall results demonstrate that it is feasible to non-invasively determine SSC and pH of grape juice using Vis–NIR spectroscopy, UVE–SPA is a powerful tool to select the efficient variables, and UVE–SPA–MLR is simple and excellent for the spectral calibration.