Robust Classification of Tea Based on Multi-Channel LED-Induced Fluorescence and a Convolutional Neural Network

Hongze Lin, Zejian Li, Huajin Lu, Shujuan Sun, Fengnong Chen, Kaihua Wei, Dazhou Ming

TL;DR

This paper introduces a novel system for classifying tea varieties using multi-channel LED-induced fluorescence and a Convolutional Neural Network (CNN). By combining spectral data from seven different LEDs into a 2D matrix, their CNN approach achieves significantly higher classification accuracy than traditional methods, offering a fast, compact, and robust solution for tea authentication.

What Problem Does This Paper Address?

The classification of tea is crucial for determining its quality, geographical origin, and market value. Current methods rely heavily on subjective human tasters, which can be inconsistent and easily influenced by external factors. This subjectivity creates a demand for objective, reliable, and non-destructive analytical techniques. While optical methods like UV-Vis, NIR, and FTIR spectroscopy, along with chemometrics, have been explored, they often face challenges such as high equipment costs, complex sample preparation, or limitations in capturing the full spectral information needed for accurate classification. Specifically, single-source excitation in fluorescence spectroscopy can lead to highly variable spectra, hindering robust classification.

Methods

The core of this research lies in a two-pronged approach: a novel fluorescence excitation system and an advanced machine learning model.

  • Multi-Channel LED-Induced Fluorescence System: Instead of a single light source, the researchers developed a system using seven LEDs with wavelengths ranging from UV to blue. These LEDs are sequentially activated to excite fluorescence from tea leaves. This multi-channel approach aims to capture a broader range of fluorescent components within the tea. The emitted fluorescence is collected by a lens and fiber optic cable and sent to a spectrometer.
  • Convolutional Neural Network (CNN): The spectral data from each of the seven LEDs are not treated as independent 1D signals. Instead, they are merged to form a 2D matrix. This "figure-like" representation is then fed into a custom-designed CNN. The CNN architecture includes a convolution layer, a flattening layer, and two fully connected layers. Notably, a pooling layer was omitted as the spectral data's inherent positional information (wavelength) was deemed important. The CNN uses Rectified Linear Unit (ReLU) activation and Softmax for output, trained with stochastic gradient descent.
  • Baseline Comparison: For validation, the researchers also implemented a Principal Component Analysis (PCA) combined with k-Nearest Neighbors (kNN) classification method, a common technique in spectral analysis.

Datasets

The study utilized nine different types of dried and fermented tea samples:

  • Six grades of green tea: Wuniuzao (G-WNZ), Huangshan Maofeng (G-HM), Vietnam Green tea (G-V), and three grades of Xihu Longjing (G-LJ1, G-LJ2, G-LJ3).
  • Two types of black tea: Keemun black tea (B-KM) and Lipton yellow label tea beverage (B-L).
  • One type of white tea: Mancheng white tea (W-MC).

The samples were divided into training (50%) and test (50%) sets for model evaluation. No new benchmark datasets were introduced; the focus was on applying the developed system to existing tea varieties.

Key Insights and Ideas

  • 2D Representation is Key: The most significant insight is the successful transformation of 1D fluorescence spectra from multiple excitation sources into a 2D matrix. This allows CNNs, which excel at image-like pattern recognition, to leverage the interdependencies between different excitation wavelengths, revealing richer information than individual spectra alone.
  • LEDs as a Viable Alternative: The use of multiple LEDs as excitation sources offers a more compact, cost-effective, and user-friendly alternative to traditional laser-induced fluorescence systems, making advanced spectral analysis more accessible.
  • Chlorophyll as a Discriminator: The study highlights that the fluorescence peaks around 680 nm and 740 nm, primarily attributed to chlorophyll, show distinct patterns between green and black teas, providing a strong basis for classification.
  • CNN's Superior Feature Extraction: The CNN's ability to automatically learn complex features from the 2D spectral data significantly outperforms traditional PCA-kNN methods, demonstrating its power in handling intricate spectral relationships.

Key Results

  • CNN Achieves High Accuracy: The proposed CNN method achieved an impressive test set accuracy of 0.993.
  • PCA-kNN Performance: The baseline PCA-kNN method, when using all seven LEDs, achieved accuracies of around 0.826-0.839, with the best performance (0.839) achieved using only LED4. This indicates that while PCA-kNN can classify, it doesn't fully exploit the multi-channel data as effectively as the CNN.
  • Improved Separation: PCA analysis of the CNN's intermediate layer outputs showed that previously overlapping tea varieties became well-separated, suggesting the CNN effectively learned discriminative features.
  • LED1's Broad Excitation: LED1 (around 371 nm) was found to be particularly effective in exciting fluorescence from not only chlorophyll but also other compounds, contributing significantly to the classification.

Limitations and Future Work

  • Limitations:
    • The study acknowledges that while the multi-channel LED system is compact and cost-effective compared to conventional fluorospectrophotometers, it still relies on a spectrometer.
  • The paper notes that the origin of fluorescence in the 500-600 nm range is not fully understood, though it's found to be useful for discrimination.
  • Future Work:
    • The authors suggest exploring the relationship between fluorescence spectra and the fermentation stage of tea during production.
  • Further studies could investigate the optimization of LED selection and combination for even greater classification efficiency.

Summary updated

Abstract

A multi-channel light emitting diode (LED)-induced fluorescence system combined with a convolutional neural network (CNN) analytical method was proposed to classify the varieties of tea leaves. The fluorescence system was developed employing seven LEDs with spectra ranging from ultra-violet (UV) to blue as excitation light sources. The LEDs were lit up sequentially to induce a respective fluorescence spectrum, and their ability to excite fluorescence from components in tea leaves were investigated. All the spectral data were merged together to form a two-dimensional matrix and processed by a CNN model, which is famous for its strong ability in pattern recognition. Principal component analysis combined with k-nearest-neighbor classification was also employed as a baseline for comparison. Six grades of green tea, two types of black tea and one kind of white tea were verified. The result proved a significant improvement in accuracy and showed that the proposed system and methodology provides a fast, compact and robust approach for tea classification.
BibTeX
@inproceedings{lin2019robust,
  author = {Hongze Lin and Zejian Li and Huajin Lu and Shujuan Sun and Fengnong Chen and Kaihua Wei and Dazhou Ming},
  title = {{Robust Classification of Tea Based on Multi-Channel LED-Induced Fluorescence and a Convolutional Neural Network}},
  booktitle = {Sensors},
  year = {2019},
  url = {https://openreview.net/forum?id=Z4xD5c3zQN},
}
Figure