Deep learning-based fine-tuning transfer improves the generalizability of tea component prediction using miniature...

Model accuracies for predicting tea components using near-infrared spectroscopy are often constrained by sample type. Herein, spectral data from four tea types (green, black, oolong, and yellow) were collected using a low-cost, miniature near-infrared spectrometer. Various strategies were explored to develop generalized predictive models for catechins and caffeine across different tea types. Local models trained on single tea types were unable to accurately predict the samples from other types. Although global models that incorporated diverse tea samples demonstrated improved performance, their accuracies remained unacceptable. To address this limitation, a deep learning-based fine-tuning strategy was implemented, enabling the transfer of accuracy from local models to other tea types. Fine-tuning the convolutional and fully connected layers of pre-trained models resulted in accurate predictions of catechin and caffeine content in all tea samples, outperforming traditional transfer component analysis. This study presents a generalizable and cost-effective method for the determination of tea components.