Assessment of protein content and insect infestation of maize seeds based on on-line near-infrared spectroscopy and machine learning

Abstract

Maize is one of the most important crops in the world. The protein content and insect infestation significantly influence seed vigor and growth. Therefore, it is vital to detect the protein content and insect infestation seeds rapidly and non-destructively. In this study, a near-infrared (NIR) spectra acquisition device (900–1700 nm) was designed and employed for on-line seed quality detection. Support vector machine (SVM), logistic regression (LR), and partial least square regression discrimination analysis (PLS-DA) were used for insect infestation seeds classification. PLS and least squares-support vector machine (LS-SVM) were adopted for protein detection. Competitive adaptive reweighted sampling (CARS) and successive projections algorithm (SPA) were applied to select the feature wavelengths to reduce the redundant data and identify important information. As for insect infestation seeds, CARS-SPA-LR model achieved the classification using only 7 features with an accuracy of 0.83. In terms of protein prediction, the LS-SVM models obtained the best results for grain protein content (GPC, %) and absolute GPC (Ab_GPC, mg/kernel), respectively. The optimal models only used 22 and 21 feature wavelengths selected by CARS-SPA, with the RMSEP of 3.38 % and 2.38 mg/kernel, and RPD was 2.08 and 2.11, respectively. The results indicated that the NIR on-line acquisition system could be applied for qualitative and quantitative analysis of maize seeds.