Integrating Hyperspectral Data and Deep Learning for Non-Destructive Prediction of Tea Quality Parameters Across...

Achieving rapid and non-destructive assessment of tea quality is essential for intelligent tea production and quality control. In this study, an integrated hyperspectral and deep learning framework was developed to estimate tea quality constituents across seasons and physical states. Samples included field fresh leaves, dried tea leaves, and tea powder, were collected in spring, summer, and autumn. Tea polyphenols and catechins were predicted using original reflectance, harmonic features, and wavelet features fused into multi-domain indices. Extreme gradient boosting, Gaussian process regression, and convolutional neural networks (CNN) were systematically compared to construct the quality estimation models. The result showed that three-feature indices consistently outperformed two-feature indices, yielding R² from 0.48 to 0.71. CNN achieved the best overall performance among the three modeling approaches, with its optimal accuracy obtained for tea powder samples in autumn, yielding R² values of 0.81 and 0.76 for tea polyphenols and catechins, respectively. This framework provides an accurate, non-destructive tool for tea quality evaluation and traceability, offering technical support for intelligent agriculture and quality control across the tea industry chain.