Modeling for mung bean variety classification using visible and near-infrared hyperspectral imaging

This study was carried out to investigate the feasibility of using visible and near infrared hyperspectral imaging for the variety classification of mung beans. Raw hyperspectral images of mung beans were acquired in the wavelengths of 380-1023 nm, and all images were calibrated by the white and dark reference images. The spectral reflectance values were extracted from the region of interest (ROI) of each calibrated hyperspectral image, and then they were treated as the independent variables. The dependent variables of four varieties of mung beans were set as 1, 2, 3 and 4, respectively. The extreme learning machine (ELM) model was established using full spectral wavelengths for classification. Modified gram-schmidt (MGS) method was used to identify effective wavelengths. Based on the selected wavelengths, the ELM and linear discriminant analysis (LDA) models were built. All models performed excellently with the correct classification rates (CCRs) covering 99.17%-99.58% in the training sets and 99.17%-100% in the testing sets. Fifteen wavelengths (432 nm, 455 nm, 468 nm, 560 nm, 705 nm, 736 nm, 760 nm, 841 nm, 861 nm, 921 nm, 930 nm, 937 nm, 938 nm, 959 nm and 965 nm) were recommended by MGS. The results demonstrated that hyperspectral imaging could be used as a non-destructive method to classify mung bean varieties, and MGS was an effective wavelength selection method.