An improved deep convolutional generative adversarial network for quantification of catechins in fermented black tea.

Area of Science:

• Analytical Chemistry
• Food Science
• Machine Learning

Background:

• Accurate quantification of catechins in fermented black tea is vital for quality assessment.
• Hyperspectral imaging combined with chemometrics offers non-destructive analysis but is limited by small datasets.
• Insufficient sample size hinders the performance of regression models for catechin quantification.

Purpose of the Study:

• To address the challenge of limited sample sizes in hyperspectral analysis of black tea catechins.
• To propose an improved deep convolutional generative adversarial network with a labeling module (DCGAN-L) for hyperspectral data augmentation.
• To enhance the rapid and non-destructive quantification of catechins in fermented black tea.

Main Methods:

• Developed DCGAN-L with spectral and label generating modules for hyperspectral data augmentation.
• Generated synthetic spectra and proposed an indicator for quality evaluation.
• Developed a novel label generation method using Euclidean distances and weighted allocation.
• Augmented the training dataset with synthetically generated data.

Main Results:

• The DCGAN-L model successfully augmented the hyperspectral dataset for catechin quantification.
• Evaluated data augmentation using Random Forest (RF) and Broad Learning System (BLS) regression models.
• Average R² for RF and BLS models increased by 0.044 and 0.164, respectively, after data augmentation.
• Demonstrated improved performance of regression models with the augmented dataset.

Conclusions:

• The proposed DCGAN-L model effectively addresses the issue of limited sample size in hyperspectral analysis.
• DCGAN-L enables rapid, non-destructive quantification of catechins in black tea.
• The method significantly improves the accuracy of catechin quantification models.