Application of Hyperspectral Imaging to Discriminate the Variety of Maize Seeds

Hyperspectral imaging technique was utilized as a rapid and nondestructive tool to classify different varieties of maize seeds in the current study. The feasibility of combining spectral data with texture features to improve the classification accuracy for maize seeds was analyzed. Average and pretreated spectra with detrending were extracted from the region of interest of hyperspectral images over the wavelength region of 400-1000 nm, and six optimal spectral wavelengths were selected by successive projection algorithm. Meanwhile, five textural feature variables were extracted by gray-level run-length matrix analysis. Least-square support vector machine was developed to classify different varieties of maize seeds based on spectral, textural, or fusion data. The least-square support vector machine model based on full pretreated spectral data (91.667 %) achieved better results than those based on full spectral data (90.741 %) and optimal spectral data (87.037 %). On the other hand, an accuracy of 88.889 % based on data fusion was achieved, which was superior to the results based on spectra (87.037 %) or texture (85.185 %) alone. At last, the resulting classification maps were developed to visualize different varieties of maize seeds. The current study indicated that combining spectral data with textural features was an effective method to improve the classification accuracy for maize varieties.