Multimodal Sensor Fusion for Non-Destructive Tea Quality Evaluation: Deep Learning-Enabled Methods, Applications, and Challenges

Tea quality evaluation is increasingly moving from subjective sensory assessment and destructive laboratory analysis toward rapid, non-destructive, and data-driven approaches. This review summarizes recent advances in multimodal sensing integrated with deep learning for tea quality evaluation, with emphasis on sensor complementarity, data-fusion strategies, representative applications, and deployment-related limitations. Major sensing modalities, including machine vision, near- and mid-infrared spectroscopy, Raman and fluorescence spectroscopy, hyperspectral imaging, and electronic nose/electronic tongue systems, are discussed in relation to their ability to characterize appearance, chemical composition, aroma, flavor, processing status, and safety-related attributes. Applications are examined for quality grading, chemical composition prediction, aroma and flavor characterization, fermentation monitoring, and safety-related extensions across representative tea products, including green tea, black tea, dark tea, matcha, and jasmine tea. Overall, multimodal approaches can outperform single-sensor systems only when the selected modalities provide complementary, rather than redundant, information layers. However, practical translation remains constrained by small and weakly standardized datasets, insufficient external validation, sensor instability, limited model transferability, high computational cost, and insufficient interpretability. Future research should prioritize standardized datasets, leakage-free validation protocols, interpretable multimodal modeling, truly independent external validation, interoperable multi-sensor platforms, and lightweight deployable models.