A Lab Can Sort Real Fuding From a Lookalike. No Shop Does.

Near-infrared light sorted bai mudan grown in Fuding from the same grade grown elsewhere in Fujian correctly 97.96 percent of the time. The model that did the sorting needed one extracted feature from the spectrum and no taster at all. The work was published in the journal Foods in January 2023¹ by a group at Fujian Agriculture and Forestry University.

Bai mudan, white peony, is white tea's second grade: one bud with one or two leaves, withered and dried, not rolled and not fired. Fuding, a coastal county in Fujian, is the name that grade sells best under.

The researchers give their reason for building the test in their own opening. "It has been reported that Fuding white tea sold on the market is far more than the actual production."

More Is Sold Than Is Grown

Fuding's white tea holds national geographical-indication protection, which reserves the name for leaf grown and made inside a defined boundary. That protection is a legal instrument attached to the name. It does not travel with the leaf.

Chinese reporting on the district's own brand-protection work states the consequence plainly: counterfeit marks and falsified origins occur from time to time².

Set two tins on a counter, both labeled Fuding, and a buyer has the label, the seal, and the seller's word. Traditionally, the claim rested on an expert's taste, a method the paper finds vulnerable to subjective sway.

What the Instrument Reads

The group collected 579 bai mudan samples from seven provinces. Fujian supplied 389 of them, then Guizhou 55, Hunan 40, Sichuan 40, Yunnan 34, Zhejiang 12, Guangxi 9. Within the Fujian set, 192 were from Fuding and 197 from other Fujian districts.

Every sample was scanned across the near-infrared band, 4,000 to 10,000 cm-1, counted in wavenumbers (wave cycles per centimeter). After preprocessing, the models were built on the 4,000 to 7,200 stretch. Near-infrared light sits just past the red end of the visible range, and the leaf returns it in a pattern set by the chemical bonds inside it.

Three classifiers were run against those spectra: k-nearest neighbors, random forest, and support vector machine.

The accuracy rose as the question narrowed. Province against province, 88.97 percent. District against district inside Fujian, 93.88 percent. Fuding against the other Fujian districts in the set, 97.96 percent.

The best model on that last question was the plainest of the three. A k-nearest-neighbors classifier with k=3, run on the original spectra with no extra preprocessing and using linear discriminant analysis to extract features, placed a sample on a single variable. It was trained on 291 samples and validated on 98, and four-fold cross-validation held the result.

The Method It Replaces

Origin claims have rested on sensory evaluation, which is to say on a trained taster's judgment. The paper is blunt about the cost of that. The results are "easily influenced by experts' physical conditions and mental factors, leading to the results being more subjective and lacking repeatability."

That is a register that cannot be kept. The same taster on a different day may enter a different verdict, and nothing in the method says which entry was right.

The Limit on the Figure

The 97.96 percent belongs to that one grade and to this dataset. One grade was sampled, and Fuding was tested against a bounded set of other Fujian districts, in one collection of 579 teas.

It is not a finding about white tea as a class. Silver needle (bai hao yin zhen), gong mei and shou mei were not in the sample, and each of those grades would need its own. Neither is it a certificate that any given tin is genuine.

What it establishes is narrower and holds: origin leaves a mark in the leaf that survives withering and drying, and an instrument can read that mark in one pass.

What Runs at the Counter Instead

No tea meets a spectrometer between the field and the shelf. What it meets is a QR code.

Fuding has built a big-data traceability platform² holding records for 3,137 tea companies, 76,043 registered tea farmers and 1,666 leaf brokers, covering 362,000 mu of tea gardens (about 60,000 acres), with more than 20 million traceability labels issued. Scan one and the phone returns the product name, the claimed origin, the processing date, and access to production records and test reports.

Use of the geographical-indication mark itself has risen alongside it, from 347 companies in 2021 to 513 in 2024, an increase of 47.8 percent, which the same reporting puts at 85 percent of the district's production enterprises.

A Code Checks the Seller, Not the Leaf

The two systems ask different questions. The spectrometer asks the leaf where it grew. The code asks the seller what was filed, and returns the filing.

A label can be printed on a bag that never held Fuding tea, and enforcement is what catches that case afterward. Replying to a 2025 city proposal on online infringement³, Fuding's Market Supervision and Management Bureau reported that since the start of 2024 it had filed 12 cases against e-commerce sellers on platforms including Douyin and Taobao, and asked Douyin, Pinduoduo, Taobao and WeChat to close 31 online shops. It filed 14 trademark-violation cases, with fines and confiscations totaling 31,350 yuan (about US$4,400). It ran 225 batch quality inspections in the district, which came back 100 percent compliant.

Those are trademark and platform actions, and the inspections test the product against quality standards. None of them reads a sample and says where it grew.

What Is Still Not Asked of the Leaf

A buyer holding the tin can check the name on the label, the code on the seal, and the record the code returns. Each of those checks reads something a seller wrote down.

The one check that reads the leaf came back 97.96 percent correct in a paper published in 2023, on one grade, in one dataset, in a university laboratory in the same province the tea comes from. It has not been asked of a tin on a shelf.