Variety classification of maize kernels using near infrared (NIR) hyperspectral imaging

Abstract. Variety classification of maize kernels was evaluated using near infrared (NIR) hyperspectral imaging in this work. Firstly, NIR hyperspectral images of kernels of four widely used maize varieties were acquired within effective spectral range of 1000-2500 nm. Spectral math was used to compensate for minor lighting differences, and band math combined with threshold method was used to remove the background from images. Minimum noise fraction (MNF) was adopted to reduce noise. Texture features (mean, variance, homogeneity, contrast, dissimilarity, entropy, second moment, and correlation) as appearance character of each maize kernel were calculated and extracted to establish classification model combined with spectra data. Moving average smoothing and standard normal variate were applied on the raw spectra extracted from hyperspectral images. Four optimal wavelengths (1352.20 nm, 1615.50 nm, 1733.10 nm, and 2478.20 nm) were selected by competitive adaptive reweighted sampling (CARS) method. Partial least squares discriminant analysis (PLSDA) was employed to build varieties classification models, based on full wavelength data, the four wavelengths data, and combination of spectral and textural features at four wavelengths, respectively. Results demonstrated that PLSDA model based on combination of spectral and textural features had the best performance with accuracies of 0.89, 0.83 for calibration and prediction set, which indicated the hyperspectral imaging technique with combination of spectral and textural features had a potential of application for variety classification.