Discrimination of small sample tea varieties based on convolutional neural network and deep convolutional generative adversarial network enhanced near-infrared diffuse reflectance spectral dataset

Abstract Near-infrared diffuse reflectance spectroscopy is widely recognized as a rapid, non-destructive, and environmentally friendly detection technology. However, in order to ensure the accuracy and stability of the detection model, a large number of sample data is necessary. This paper proposed the rapid and non-destructive detection of small sample tea variety recognition based on the near-infrared diffuse reflectance spectrum data extended by convolutional neural network (CNN) and deep convolutional generative adversarial network (DCGAN). The near-infrared diffuse reflectance spectra of 240 tea samples were obtained by Lambda 950 spectrometer using eight of the most popular tea varieties on the Chinese market. Firstly, the spectral data was enhanced using translation, linear superposition, noise addition, and DCGAN methods, and the quality of the generated spectra was evaluated using the support vector machine (SVM) and gradient boosting decision tree (GBDT) methods. Compared with other methods, the DCGAN has the highest accuracy of 91.75%. Secondly, the optimal number of iterations of DCGAN was confirmed to be 6000 by SVM and GBDT methods. To further augment the precision of identifying small samples of tea, two additional classification models of the Extreme Gradient Boosting (Xgboost) and CNN were applied to the DCGAN. Finally, the results demonstrated that the CNN achieved the highest identification accuracy of 98.68% compared with SVM (90.46%), GBDT (90.42%), and Xgboost (88.83%) with an additional 100 samples and 6000 iterations. Therefore, the combination of deep convolutional generative adversarial network enhanced near-infrared diffuse reflectance spectral dataset and the CNN successfully realizes the identification of small sample tea varieties. The experimental results strongly indicate that this method holds significant potential for practical implementation in the field of small sample tea varieties identification.