Multispectral imaging for tea grading
Research into using reflected light to classify Oriental Beauty tea samples.
Research team: High-Speed Networks Laboratory, National Tsing Hua University. First half of 2025.
The assessment challenge
Oriental Beauty tea is still assessed by people. Licensed tasters judge appearance, aroma, and taste, and the report describes that process as inconsistent and difficult to scale.
This research asks whether photographs taken under different lighting conditions can support a more consistent classification of sample categories. The images record how the tea reflected light. Those measurements became inputs for models trained to predict the labels assigned to the samples.
The models do not taste the tea. They classify labeled sample groups from reflected-light measurements, as a research step toward more objective quality assessment.
Building the classification pipeline
The study combined controlled sample preparation with industrial multispectral imaging. Tea samples were presented under consistent conditions and photographed repeatedly.
Image processing separated the tea from the background and extracted measurements from the different imaging channels. These included average intensity and measures of variation within the tea regions.
The team used the resulting features to train Random Forest and XGBoost models. Their task was to predict the three labeled sample categories used in the study. Those categories are 1 flower, 2 flower, and not selected (no grade).
This approach connects a physical measurement with a classification decision: the imaging system captures the signal, and the model learns how that signal relates to the sample labels.
Reported results
Both models made most of their classification errors between the one-flower and two-flower sample categories. This identifies a specific area for further development: distinguishing the sample groups that were harder to separate in the reported evaluation.
Results reported for the first half of 2025. The Random Forest and XGBoost classification reports each cover 390 evaluation records across three labeled sample categories. The study includes repeated sample imaging; these records should not be interpreted as 390 independent tea batches.
Exploratory model work: The report also describes a preliminary ResNet-18 experiment that reached a peak validation accuracy of 71%.
What the research shows
The study shows that reflected-light measurements can become useful inputs for classifying Oriental Beauty samples. Sample presentation and image processing decide what reaches the model. The sample labels decide the judgment it is learning to make.
For Plexys, this is another research setting for physical intelligence: using observations of a material to support assessment. The interim report treats the work as an initial prototype. Further development would widen sample and batch coverage and test a grading workflow across more operating conditions.