Exploring Near and Midinfrared Spectroscopy to Predict Trace Iron and Zinc Contents in Powdered Milk

Abstract

Near infrared (NIR) and mid-infrared (MIR) spectroscopy were investigated to predict iron and zinc contents in powdered milk. A hybrid variable selection method, namely, uninformative variable elimination (UVE) combined with successive projections algorithm (SPA), was applied to select the most effective wavenumber variables from full 2756 NIR and 3727 MIR variables, respectively. Finally, 18 NIR and 18 MIR variables were selected for iron content prediction, and 17 NIR and 12 MIR variables for zinc content prediction. The obtained effective wavenumber variables were input into partial least-squares (PLS) and least-squares-support vector machines (LS-SVM), respectively. The selected MIR variables obtained much better results than NIR to predict both iron and zinc contents in both the PLS and LS-SVM models. The iron content prediction results based on LS-SVM with 18 MIR spectra were as follows: coefficient of determination (r(2)) was 0.920, residual predictive deviation (RPD) was 3.321, and root-mean-square error of prediction (RMSEP) was 1.444. The zinc content prediction results based on LS-SVM with 12 selected MIR spectra were as follows:r(2) was 0.946, RPD was 4.361, and RMSEP was 0.321. The good performance shows that UVE-SPA is a powerful variable selection tool. The overall results indicate that MIR spectroscopy incorporated to UVE-SPA-LS-SVM could be applied as an alternative fast and accurate method to determine trace mineral content in powdered milk, such as iron and zinc.