NIR Spectroscopy with AI Proves to be a Powerful Combination for Tea Classification

A variety of loose leaf teas in wooden bowls © Konstantin Sigaev - stock.adobe.com

Tea, one of the world's most popular beverages, is not just a daily ritual for millions; it is also a significant economic commodity. However, the tea industry faces challenges in ensuring accurate classification of its diverse varieties, with counterfeit or misrepresented teas often entering the market (1,2). Traditional classification methods, which rely on non-precise human sensory evaluation or costly chemical analysis, have proven inadequate for this task (1,2). But a method combining near-infrared spectroscopy (NIRS) and artificial intelligence (AI) may soon offer an alternative (3)

In a recent study, Long Liu, Bin Wang, Xiaoxuan Xu, and Jing Xu from Nankai University have demonstrated a novel tea classification method using a fine-tuned 1DResNet model. The team's findings, published in the journal Infrared Physics & Technology, show that their approach outperforms traditional machine learning (ML) techniques and offers promising applications for the tea industry (3).

The Need for Accurate Tea Classification

Accurate classification of tea types is essential for both consumer protection and market fairness. Tea sellers often mislabel or arbitrarily label tea varieties, which can harm consumers and disrupt the market. Given that sensory-based tea classification is subjective and prone to human error, there has been a growing need for more reliable, scientific methods. Traditional chemical analysis techniques, while precise, involve complex sample preparations, expensive equipment, and skilled professionals. Furthermore, these methods generate significant waste, which can harm the environment (1–3).

To address these issues, researchers have turned to near-infrared (NIR or NIRS) spectroscopy, which is widely used for material analysis due to its ability to detect molecular fingerprints without damaging the sample. NIR works by analyzing the absorption of near-infrared light by samples, allowing scientists to identify the chemical composition of the material. However, traditional ML algorithms have struggled to handle the complexity of the spectra obtained from NIR data (3).

A Novel Solution: 1DResNet and Transfer Learning

To overcome these challenges, the research team proposed a hybrid approach that integrates the 1DResNet model with transfer learning. The 1DResNet, a variant of the convolutional neural network (CNN), is specifically designed for one-dimensional data like the NIR spectral data. The model includes "residual connections" that allow it to handle deeper networks without suffering from the common issue of gradient vanishing, resulting in more accurate predictions (3).

The process begins with pre-training the 1DResNet model using a dataset of known tea types. The model’s feature extraction layers are then frozen, and the remaining layers are fine-tuned using a smaller, fine-tuning dataset. This approach improves the model's ability to handle complex spectral data. Finally, the fine-tuned model is tested using a separate test dataset to evaluate its classification performance (3).

Reported Results

The researchers compared their fine-tuned 1DResNet model to traditional ML algorithms like partial least squares discriminant analysis (PLS-DA), K-nearest neighbor (KNN), and multilayer perceptron (MLP). The results were striking: the fine-tuned 1DResNet model demonstrated a significant improvement in classification accuracy—by more than 4.32% over the traditional methods. The fine-tuned model even outperformed the standard 1DResNet model by 4.96%, showcasing the power of transfer learning in enhancing the model's performance (3).

In addition, when compared to another deep learning model, the fine-tuned 1DResNet achieved a 4% higher accuracy than the fine-tuned 1-dimensional convolutional neural network (1DCNN). The team also demonstrated the model's ability to handle "migration tasks," where it successfully classified black and green tea types, further proving the flexibility and broader applicability of the method (3).

The Role of NIRS in Tea Classification

NIR has long been used in food and agriculture for its ability to analyze samples without destroying them. The method works by measuring the absorption of light at various wavelengths, providing information on the chemical composition of the material. For tea, NIR spectra reveal information about key components such as polyphenols, proteins, and carbohydrates, which vary between tea types. The detailed peaks and troughs in the NIR spectra allow the model to distinguish between different teas based on these molecular signatures (3).

Preprocessing steps, such as standard normal variate (SNV) transformation, are applied to the NIRS data to enhance its quality, minimizing noise and spectral variations. These preprocessing methods, coupled with the AI-driven 1DResNet model, allow for high-accuracy classification even with complex spectra (3).

This new method of tea classification, based on the fine-tuned 1DResNet model and NIR, not only addresses the limitations of traditional classification techniques but also provides a potentially useful tool for the tea industry. The approach offers a non-destructive, cost-effective, and efficient way to ensure that consumers receive the tea varieties they are promised, helping to maintain market integrity and consumer trust (3).

References

(1) ISO 20715:2023—Classification Of Tea Types Home Page. https://blog.ansi.org/iso-20715-2023-classification-of-tea-types/ (accessed 2024 01-21).

(2) A New Look at Tea Classification Page. https://teaepicure.com/tea-classification/ (accessed 2024 01-21).

(3) Liu, L., Wang, B., Xu, X. and Xu, J., 2025. A tea classification method based on near infrared spectroscopy (NIRS) and transfer learning. Infrared Physics & Technology, p.105713. 10.1016/j.infrared.2025.105713