Identification of Moldy Corn Kernels Using Visible/near-Infrared Hyperspectral Images

Visible/near-infrared (VIS/NIR) hyperspectral imaging was used to identify moldy corn kernels in this work. A total of 200 corn kernels in the same variety collected from field were divided into moldy group and health group. Hyperspectral images of moldy and healthy kernels were acquired with a spectral range of 400-1000 nm. Band math and principle component analysis (PCA) were employed to remove background, bad pixels and noise. After that, PCA was performed again on the cleaned hyperspectral images. Score images were used to initially judge whether the corn kernels were moldy or not. A support vector machine (SVM) model based on the first three PCs was established, and the classification accuracies were 98.00%, 96.00% and 97.00% for calibration set, validation set and cross validation set. Further, four wavelengths (664, 515, 970 and 440nm) were selected by Successive projections algorithm (SPA). A new SVM model was built, the classification accuracies of calibration, validation and cross validation set were 98.67%, 98.00% and 98.00% respectively. based on the characteristic wavelengths SVM model, a prediction map of moldy kernels was established. The map showed the position of the moldy corn kernels. All the results illustrated the VIS/NIR hyperspectral imaging has the potential to identify and separate moldy corn kernels from healthy ones.