Near-infrared spectroscopy coupled with Gramian angular field two-dimensional convolutional neural network for white tea adulteration detection.

Abstract

Background: The flavor profile and product quality of white tea, heavily dependent on its place of origin, significantly influence consumers' purchasing decisions. Quantitative adulteration testing for tea origin has encountered challenges due to the poor performance in random external validation, which has severely hindered the practical application of near-infrared (NIR) technology.

Results: This study employs a two-dimensional convolutional neural network (2D-CNN) deep learning model combined with Gramian angular field (GAF) image coding technology (GAF-2D-CNN) to quantitatively detect geographical origin adulteration of white tea using near-infrared spectral (NIRS) data. The results demonstrate that the GAF-2D-CNN model can effectively process raw spectral data and predict the untrained random adulteration ratio data with high accuracy. The average R² and root mean square error in the external verification of the original data reach 0.9754 and 0.0349, respectively, which meet practical production needs. Moreover, the GAF-2D-CNN significantly outperforms traditional regression models and 1D-CNN models.

Conclusion: This study introduces the application of the NIR spectral image coding method in tea regression and highlights the advantages of deep learning image processing in the tea industry. © 2025 Society of Chemical Industry.