Comparison of machine learning and deep learning models for detecting quality components of vine tea using smartphone-based portable near-infrared device

Abstract

Tea polyphenols (TPs) and dihydromyricetin (DMY) are critical quality attributes of vine tea. This study developed a smartphone-based portable near infrared (NIR) device integrated with machine learning (ML) and deep learning (DL) approaches for rapid prediction of TPs and DMY in vine tea. NIR spectra of the vine tea samples were acquired using the developed portable device and smartphone software, while the contents of TPs and DMY were determined using UV–Vis spectrophotometer and high-performance liquid chromatography (HPLC). To accurately analyze the spectral data, various ML and DL models were evaluated and compared. Results indicate that DL models, including convolutional neural networks (CNN), long-short-term memory (LSTM) and CNN-LSTM, demonstrated superior predictive performance compared to traditional ML approaches in large sample environments. Thereinto, CNN-LSTM exhibited the optimal predictive performance for TPs (Rp = 0.9816; RPD = 5.24) and DMY (Rp = 0.9900; RPD = 7.11). Additionally, the optimal model 's validation performance are commendable, with a maximum coefficient of determination (0.9483 for TPs and 0.9625 for DMY). This demonstrates that the developed intelligent portable NIR instrument coupled with DL tools enables rapid on-site detection of vine tea quality components. Furthermore, it provides a potential strategy for real-time quality monitoring during vine tea online processing.