Improvement of near-infrared spectroscopic assessment methods for the quality of Keemun black tea: Utilizing transfer learning.
Journal:
Food research international (Ottawa, Ont.)
PMID:
40253124
Abstract
Keemun black tea, a renowned Chinese black tea, presents challenges in quality assessment due to variability in data across different years. To address this, we developed transfer learning algorithms using near-infrared spectral data. The qualitative algorithm, BLACK TEA-GRADING, utilizes stochastic Fourier features of the limit learning machine, improving F1 scores from 0.7035 to 0.8138 in grading tea across different years. The quantitative algorithm, BLACK TEA-SUBSTANCE ANALYSIS, which is based on the inverse Gram matrix, was developed to predict the main flavor substances. Without the algorithm, the MSE exceeded 0.04 and R values were below 0.7; with it, MSE dropped below 0.015 and R values exceeded 0.8. Therefore, our method leverages transfer learning and near-infrared spectroscopy to enhance the accuracy of Keemun black tea quality assessment across different years, promoting the use of near-infrared spectroscopy in tea quality evaluation.