Transfer learning-assisted snapshot multispectral imaging for robust prediction of chemical components during black...

Rapid and accurate monitoring of key chemical constituents during black tea fermentation is crucial for quality control and process optimization. However, chemical and spectral heterogeneity caused by different processing conditions compromises the robustness of conventional spectroscopic calibration models. This study developed a transfer learning (TL)-assisted snapshot multispectral imaging framework for robust prediction of catechins, theaflavins, theabrownins, and thearubigins under different processing conditions. Three progressive fine-tuning strategies based on pretrained convolutional neural networks (CNNs) were designed to improve feature transferability. These models were evaluated across two withering conditions and an independent batch for external validation. Results showed that TL models improved prediction performance, with R_p values increased by 0.47%-3.15% over retrained CNN models on the target domain, and maintained robust prediction on the external validation set. This study provides a transferable strategy to improve the robustness of spectroscopic analysis, enabling reliable chemical quantification across varying processing conditions during black tea fermentation.