Time-resolved monitoring of key tea fermentation indicators via Raman-NIR spectroscopic fusion and deep learning-based predictive modeling.

Routine monitoring of black tea fermentation relies on destructive and time-consuming wet chemistry methods. Therefore, a non-destructive analytical strategy combining Raman and near-infrared (NIR) spectroscopy with deep learning was developed for simultaneously monitoring key fermentation indicators. Fresh Camellia sinensis leaves were fermented for 4 h under controlled temperature and humidity, with samples collected at nine time points. Concentrations of total tea polyphenols, theabrownins, and theaflavins were determined using standard reference methods, while corresponding Raman spectra (400-2500 cm-1) and NIR spectra (970-1700 nm) were acquired. The spectral data were integrated via low-level fusion and subsequently modeled using convolutional neural networks (CNN) and long short-term memory (LSTM) networks. The fused LSTM model yielded the best predictive performance, with Rp values of 0.9395, 0.9345, and 0.8596 for total tea polyphenols, theabrownins, and theaflavins, respectively. These results demonstrate the potential of Raman-NIR spectral fusion combined with deep learning for non-destructive testing of key quality indicators during black tea fermentation.