Application of hyperspectral imaging and chemometrics for variety classification of maize seeds

Abstract

Seed variety classification is important for assessing variety purity and increasing crop yield. A hyperspectral imaging system covering the spectral range of 874–1734 nm was applied for variety classification of maize seeds. A total of 12 900 maize seeds including 3 different varieties were evaluated. Spectral data of 975.01–1645.82 nm were extracted and preprocessed. Discriminant models were developed using a radial basis function neural network (RBFNN). The influence of calibration sample size on classification accuracy was studied. Results showed that with the expansion of calibration sample size, calibration accuracy varied slightly, but prediction accuracy changed from the increasing form to the stable form. Accordingly, the optimal size of the calibration set was determined. Optimal wavelength selection was conducted by loading of principal components (PCs). The RBFNN model developed on optimal wavelengths with the optimal size of the calibration set obtained satisfactory results, with calibration accuracy of 93.85% and prediction accuracy of 91.00%. Visualization of classification map of seed varieties was achieved by applying this RBFNN model on the average spectra of each sample. Besides, the procedure to determine the optimal sample quantity proposed in this study was verified by support vector machine (SVM). The overall results indicated that hyperspectral imaging was a potential technique for variety classification of maize seeds, and would help to develop a real-time detection system for maize seeds as well as other crop seeds.