Study on disease level classification of rice panicle blast based on visible and near infrared spectroscopy

Visible and near infrared (Vis-NIR) spectroscopy was used to fast and non-destructively classify the disease levels of rice panicle blast. Reflectance spectra between 325 and 1 075 nm were measured. Kennard-Stone algorithm was operated to separate samples into calibration and prediction sets. Different spectral pretreatment methods, including standard normal variate (SNV) and multiplicative scatter correction (MSC), were used for the spectral pretreatment before further spectral analysis. A hybrid wavelength variable selection method which is combined with uninformative variable elimination (UVE) and successive projections algorithm (SPA) was operated to select effective wavelength variables from original spectra, SNV pretreated spectra and MSC pretreated spectra, respectively. UVE was firstly operated to remove uninformative wavelength variables from the full-spectrum. Then SPA selected the effective wavelength variables with less colinearity after UVE. Least square-support vector machine (LS-SVM) was used as the calibration method for the spectral analysis in this study. The selected effective wavelengths were set as input variables of LS-SVM model. The LS-SVM model established based on SNV-UVE-SPA obtained the best results. Only six effective wavelengths (459, 546, 569, 590, 775 and 981 nm) were selected from the full-spectrum which has 600 wavelength variables by UVE-SPA, and their LS-SVM model's performance was further improved. For SNV-UVE-SPA-LS-SVM model, coefficient of determination for prediction set (R2(p)), root mean square error for prediction (RMSEP) and residual predictive deviation (RPD) were 0.979, 0.507 and 6.580, respectively. The overall results indicate that Vis-NIR spectroscopy is a feasible way to classify disease levels of rice panicle blast fast and non-destructively. UVE-SPA is an efficient variable selection method for the spectral analysis, and their selected effective wavelengths can represent the useful information of the full-spectrum and have higher signal/noise ratio and less colinearity.