Nondestructive Identification of Millet Varieties Using Hyperspectral Imaging Technology

In this study, we used eight millet varieties and took visible-near-infrared hyperspectral images of 480 millet samples. Spectral and image characteristics, including texture and color features, of the millet samples were extracted from the hyperspectral images. Support vector machine (SVM) models for millet variety identification were established using the extracted spectral and image characteristics. An attention-convolutional recurrent neural network (attention- CRNN) model with attention mechanism was introduced for the identification of millet varieties, and the SVM and attention- CRNN models for millet variety identification were established using an image and spectral features fusion method. We found that the highest mathematical transformation method was the reciprocal logarithmic method. The identification accuracy of the SVM cultivar classification model with the reciprocal logarithmic spectral characteristics curve was 73.13%. The overall identification accuracy of the SVM model for the eight millet varieties using the image features was only 61.25%. The identification accuracy of the SVM model using the image and spectral information fusion method greatly improved the overall accuracy rate to 77.5%, and the minimum discrimination accuracy of the millet varieties increased from 50 to 65%. The overall identification accuracy of the attention-CRNN model was 87.50%, which is 10% higher than that of the SVM model, and the minimum discrimination accuracy of the millet varieties increased from 65 to 90%. The results show that the attention-CRNN model improved the overall identification accuracy of the eight millet varieties and greatly improved the minimum identification accuracy. The attention-CRNN model shows great potential for the nondestructive identification of millet and possibly other small grain varieties.