A tea classification method based on near infrared spectroscopy (NIRS) and transfer learning

Abstract

Tea is one of the most popular and widely consumed beverages worldwide, and accurately identifying its type is important for consumers. NIRS, a technology that uses near-infrared light for material analysis, is often employed for this purpose. Traditionally, automatic identification of NIRS has relied on classical machine learning methods. However, these conventional algorithms tend to lack accuracy when dealing with complex spectra. This article proposes a tea classification method based on a 1-dimensional residual network(1DResNet) model combined with transfer learning. The method is implemented in several steps. First, the 1DResNet model is pre-trained using a pre-training dataset. Then, the parameters of the feature extraction layers are frozen, and the model is fine-tuned using a fine-tuning dataset. Finally, the fine-tuned 1DResNet model is tested on a separate test dataset. Compared to traditional machine learning algorithms like Partial Least Squares Discriminant Analysis (PLS-DA), K-Nearest Neighbor (KNN), and Multilayer Perceptron (MLP), the fine-tuned 1DResNet model demonstrates significantly improved classification accuracy (by more than 4.32%). Furthermore, compared to a 1DResNet model without fine-tuning, accuracy improves by 4.96%. When compared to a fine-tuned 1-dimensional Convolutional Neural Network (1DCNN), the accuracy increases by 4%.This notable improvement highlights the potential of the fine-tuned 1DResNet model in handling complex spectral data. The method also performs well in transfer learning tasks; both black tea and green tea classification results demonstrate that the 1DResNet model with fine-tuning has strong potential for migration tasks. Overall, this classification method offers broader application prospects.