Modeling research on wheat grain in different quality categories using multilayer support vector machine

Near Infrared Reflectance (NIR) spectroscopy is a 'green' nondestructive testing technology and it has been widely used in grain crop analysis. The experimental data were collected using 161 wheat samples from the major wheat-producing area in China. The original spectral data was represented by four characteristic variables extracted by Partial Least Squares based Dimension Reduction (PLSDR). Besides, Mahalanobis distance method, second derivative and SNV were used to preprocess spectra. A two-tier classification model based on SVM algorithm was used to achieve the classification of wheat quality. The experimental results indicated that the two-tier SVM classification model was effective in identifying the quality of wheat grain with the recognition rates of common, strong-gluten, middle-gluten and weak-gluten wheat samples being 93.3%, 87.5%, 72.7% and 92.3%, respectively, and the rejection rates of them being 90.0%, 97.4%, 100.0% and 95.2%, respectively. The model realized rapid and accurate classification of wheat, besides it could be applied to the detection system of wheat quality.