Artificial Intelligence-Driven Tea Quality Evaluation: A State-of-the-Art Review

Traditional tea quality evaluation relies largely on subjective human sensory assessment and destructive chemical laboratory assays, suffering from poor objectivity, low throughput and limited scalability. In the past five years, artificial intelligence-driven deep learning has enabled rapid, non-destructive, data-oriented quality inspection for various teas. This review summarizes representative advances covering computer vision, electronic nose, near-infrared spectroscopy, hyperspectral imaging, electronic tongue and multi-sensor multimodal fusion. CNNs, YOLO-series detectors, Transformer and CLIP achieve 85–99% classification accuracy in tea grading, aroma analysis, fermentation monitoring and geographical tracing, alongside accurate quantitative prediction of individual chemical constituents (e.g., R2 = 0.92–0.97 for four catechins in black tea). In most reviewed studies, multimodal frameworks achieved higher performance than single-sensor systems by exploiting complementary sensing information. However, industrial deployment is hindered by small non-standard datasets, low model interpretability, sensor drift, poor cross-domain transferability and high costs. Future directions include standardized benchmark datasets, explainable multimodal models, lightweight networks and compatible multi-sensor hardware platforms.