This guideline (open access) provides practical recommendations for the preparation and analysis of solid and liquid food samples using stable isotope techniques.

Emphasis is placed on harmonized protocols for sample preparation, including drying, encapsulation, and homogenization, and on the use of internationally recognized isotope reference scales: Vienna Pee Dee belemnite (VPDB) for carbon, atmospheric air (AIR) for nitrogen, Vienna standard mean ocean water (VSMOW) for hydrogen and oxygen, and Vienna Canyon Diablo troilite (V‐CDT) for sulfur. Best practices for selecting and applying certified reference materials, multipoint normalization, correcting for instrument linearity, and uncertainty propagation using regression approaches are discussed. Quality control measures—such as blanks, replicates, and matrix‐matched standards—are essential to ensure reproducibility and interlaboratory comparability.

These guidelines aim to standardize stable isotope methodology in food science and support users in producing accurate, defensible, and globally comparable isotopic data.

The UK National Food Crime Unit provides a free periodic e-mail newsletter.  You can subscribe, and see back-issues, here.

The January 2026 issue has just been circulated.  It includes:

• impact of goat pox on authentic feta production and steps businesses can take
• horizon scanning
• NFCU Global Alliance update
• NFCU annual report
• report of NFCU operation against illegal bushmeat

The horizon scanning section highlights macro-economic pressures on supply of olive oil, cardomon, salmon and  wild caught white fish.

A key advantage of the Direct Analysis in Real Time (DART) mass spectrometry (MS) ion source is its ability to ionise the sample without the need for extraction.  In this study (open access), the authors compared DART with a previously-published extraction-based MS method to analyse key components in olive oil.

Having optimised and validated DART-MS, they then used it to build a discrimination model between different classes of edible oils.  They analysed a reference set of 80 samples from different regions of Greece (Crete, Peloponnese, Central Greece, and the North Aegean) to discriminate authentic extra virgin olive oil (EVOO).  These were from 10 oil categories including 35 EVOOs, 15 lower-quality olive oils (five of each category: refined, olive pomace, and ordinary), and 30 vegetable oils (five of each type: sunflower, corn, soybean, canola, sesame, and linseed).

They report that multivariate statistical analysis revealed clear discrimination of EVOO from other oils and enabled detection of EVOO adulteration down to 1 % with vegetable oils and 5 % with lower-quality olive oils. Key authenticity markers, including phenols, squalene, and triacylglycerols (TAGs), were identified.

They conclude that the proposed method demonstrates high potential for rapid, reliable EVOO authentication in routine quality control.

The EU forced labour regulations 2024/3015 have an implementation date on December 2027.  As well as prohibiting “'all work or service which is exacted from any person under the threat of a penalty and for which the person has not offered himself or herself voluntarily.' within EU member states, they also put a due diligence requirement on companies to check that such practices do not occur in their supply chain.  The way that Ireland have implemented the enforcement of these regulations (Irish Statutory Instrument 623 of December 2025) is typical.  The Irish law applies to businesses of all sizes, in all sectors, and includes requirements to

• Update compliance and training programmes: Governance and training programmes should be updated to educate procurement teams, in-house counsel, risk managers, and senior leadership on their obligations.
• Supply-chain mapping: Conduct supply-chain mapping to identify any products or parts that may fall under the forced-labour import ban.
• Due-diligence: Incorporate due-diligence measures into procurement, contracting and supplier oversight.

A practical example within the food industry would by the business-to-business supply of canned tomatoes.  At the moment, most companies’ supply chain mapping and VACCP assessments would take these back to their country of origin (e.g. “China”, which accounts for many of the tomatoes on the EU/UK b-2-b market).  However, these new regulations imply a due-diligence requirement to go further down the chain, to give assurance that the production is not linked to previous reports of Uyghur forced labour in the Xinjiang region.

Photo by Marwan Ahmed on Unsplash, (with no implication that this image shows forced labour)

In this study (open access) the researchers used a panel of three different DNA test protocols to verify the labelled species in dairy products sampled from Greek supermarkets over the winter of 2024.  They tested 74 samples in total encompassing cow, sheep and goat products. This included 15 different commercial brands of goat yoghurt, 7 brands of sheep yoghurts, 3 brands of goat kefir, and, samples of feta cheese and sheep-goat cheese from 16 to 11 different geographical origins and 7 brands of goat cheese were analysed.  The brands are anonymised within the publication.

They report widespread adulteration, particularly in goat yogurts (40 %) and cheeses (40 %), as well as in three kefirs and several mixed and whey-based cheeses. Notably, only 7 out of 17 - nominally goat - feta samples contained detectable goat DNA..

Image from the publication

While pH is known to vary between different variety of whiskies it lacks a statistically rigorous exclusionary standard to be used as an authentication marker for Scotch whisky.

This study (open access) addressed this gap by performing statistical distribution fitting analysis on the pH of 32 authentic single malt and 33 authentic blended Scotch whiskies on the market in Taiwan, utilizing the three-parameter lognormal distribution to establish the 99.7% authentic pH ranges for the first time: 3.47–4.46 for single malt and 3.73–4.67 for blended whisky.

 Validation using seized counterfeit samples confirmed that an abnormally elevated pH is a strong indicator of adulteration.

Consequently, this authors propose using a pH threshold as a rapid, non-destructive, and cost-effective forensic exclusionary criterion. Although the pH value feature alone is insufficient to confirm authenticity, it is ideal as a first screening test.

Photo by Ambitious Studio* | Rick Barrett on Unsplash

This study (purchase required) used, high-performance liquid chromatography, combined with chemometric analysis, to classify buffalo vs cows ricotta based on the profile of water soluble peptides.  The authors then identified specific peptides that could be used as species markers . Both mid-infrared spectroscopy and electrophoresis were also investigated as peptide measurement methods by were found to give insufficient discrimination, with IR overly affected by storage time of the extracts.

The authors created  11 experimental cheese formulations by increasing the proportions of cow whey mixed with buffalo whey. Water-soluble peptides were analysed using mid-infrared spectroscopy, high-performance liquid chromatography and electrophoresis. The data obtained from mid-infrared spectroscopy and high-performance liquid chromatography were statistically processed using principal component analysis, analysis of covariance and multiple linear regression..

High-performance liquid chromatography identified 14 peptide peaks, with three recognized as specific markers for cow whey in adulterated samples. PCA explained 77% of the variance, distinguishing pure and adulterated ricotta. Multiple linear regression modelling of high-performance liquid chromatography data predicted cow whey concentration with a correlation of R = 0.87. High-performance liquid chromatography with chemometrics was effective for detecting buffalo ricotta adulteration.

When applied to 14 commercial samples, the model suggested that nine contained adulteration ranging from 10% to 100% cow whey.

Photo by Conor Brown on Unsplash

In this study (open access) the authors report the development of a simple point-of-use extraction system and spectral imaging protocol.  It has potential as a point-of-use food inspection tool to check for species adulteration in processed meat.

Samples were macerated (approximately 200 mg) in 500 μL of sterile saline (0.9%). The extract was centrifuged. 2 μL of the supernatant from each sample was used to obtain the UV-Vis spectrum (180–800 nm; 0.5 nm intervals).

30 samples of each of four species (beef, pork, chicken and pacu fish) were used to develop and validate a classification model. Spectral data were preprocessed using standard normal variate transformation and analyzed using principal component analysis.  The authors report that this revealed distinct clustering, particularly for beef. Support vector machine algorithms were trained, achieving an overall accuracy of 89.3% in leave-one-out cross-validation and 86.1% in external validation.

Photo by amirali mirhashemian on Unsplash

This study (open access) compared four different test approaches (DNA barcoding rbcL, DNA barcoding matK , ITS2 barcoding vs the NCBI database, ITS2 barcoding vs the BOLD database) in an authenticity survey of 100 herbal infusions on the Portuguese market.  Samples included 94 single-species products and six polyherbal formulations.

The authors report that DNA extraction was successful for 94 samples, while six single-species products failed to amplify any of the tested barcodes. Among the 88 remaining single-species samples, ITS2 showed the highest amplification success (100 %), outperforming the barcodes rbcL (94 %) and matK (84 %).

Sanger sequencing confirmed the labelled species in 69.3 % of cases with rbcL and 48.9 % with matK. While 63 samples would be considered authentic solely based on barcoding (i.e., if either rbcL or matK matched the label), ITS2 metabarcoding revealed that many of these contained additional undeclared species, indicating that barcoding alone overestimated product authenticity. Of the 85 samples successfully analysed by ITS2 metabarcoding, only 27 (32 %) fully matched their label, while 58 (68 %) contained either additional undeclared species or complete substitutions. Several products contained undeclared species in significant proportions, indicating potential economic adulteration.

The authors conclude that their results revealed (i) the importance of curated and comprehensive databases, with a higher number of species being identified by NCBI database, (ii) the superior sensitivity of ITS2 metabarcoding, and (iii) the widespread mislabelling in commercial herbal products.

Photo by Alice Pasqual on Unsplash

This study (purchase required) shows the potential for stable isotope ratio analysis (SIRA) to be used to verify whether the cocoa in bean-to-bar chocolate originated from the Amazon region.  Higher-sugar chocolate is too complex a product to test in this way.  Such tests will be important for checking compliance with future EU Deforestation Regulations. The proposed method can also be used to estimate the cocoa content of chocolate.

Chocolate is a complex product in terms of its carbon isotope distribution.  Cocoa, from a C₃-photosynthetic plant, is its main raw material, while sugar from sugarcane (C₄-metabolism) is also commonly included. The authors analyzed the δ¹³C and δ¹⁵N composition of Brazilian chocolates, including Conventional, Bean-to-bar, and Imported brands across White, Milk, Semisweet, Dark <70 %, and Dark ≥70 % versions. They report that conventional White, Milk, and Semisweet chocolates showed no significant isotopic differences, with average δ¹³C values around −22 ‰, indicating high C₄-derived ingredient content. Bean-to-bar chocolates presented δ¹³C values 2–3 ‰ lower, and those made with Amazon cocoa were ∼1.5 ‰ lighter than those from Atlantic Forest, enabling accurate prediction of cocoa origin. Imported chocolates showed even lighter δ¹³C values, suggesting greater use of C₃-based ingredients. δ¹³C and δ¹⁵N values also enabled reliable estimation of cocoa content.

Photo by Boudhayan Bardhan on Unsplash

In this survey a total of 107 tuna cans were collected between 2020 and 2022 from 7 different Portuguese commercial brands.  Samples were purchased from main supermarket chains, large-scale distributors and local stores.  Samples were categorised by the quarter of the year when the tuna was canned (inferred from the expiry date) with each sample was tested in triplicate.  Extracted DNA was purified and iteratively tested using molecular metabarcoding methods.

The researchers report (open access) that the occurrence of different species was observed only in products canned in brine or water (i.e. all products canned in oil were Skipjack Tuna). Skipjack tuna was predominant across all canning liquids and brands analysed. Nonetheless, other species like Thunnus obesus and T. albacares, or Auxis spp. (not considered true tuna) were also detected. The use of different species was limited to cans produced during the second quarter of the year, which could reflect differences in seasonal availability of different tuna species or in sourcing strategies/market preferences of each company. For four brands, multiple species were detected inside the same can.  This violates current European legislation.

The researchers conclude that these results provide the first broad assessment of species used in the Portuguese tuna canning industry and showed the inclusion of vulnerable species is limited.

Photo by Grooveland Designs on Unsplash

The EU Joint Research Centre (JRC) have published today their monthly collation of fraud media reports for October and November 2025.  The full index of reports can be found here

The JRC collation underpins a searchable front-end for media reports of food fraud incidents.  It allows filtering by commodity, country, fraud type and other key criteria.

This is just one of the incident databases available from different organisations.  Different databases collect different information, in different ways, and therefore show a different angle on the true picture.  All of these sources are signposted on FAN.  Best practice is to use a combination of all sources, but the final critical question is “how vulnerable is my own supplier”.

• JRC – These are solely media reports.  They exclude cases not in the public domain, and can be biased by shocking but highly localised incidents in local food supply within poorly regulated countries.  For the past few years, FAN member Bruno Sechet has produced a useful infographic based on each month's data
• EU Agri-Food Suspicions – These are solely EU Official Reports, and only suspicions.  The root cause of each incident is unknown.  The data include cases less likely to be deliberate fraud such as pesticide residues above their MRLs or unpermitted (but labelled) additives.  FAN produce our own infographic on a rolling 3-month basis.
• Food Industry Intelligence Network Fiin SME Hub – These are aggregated anonymised results from the testing programmes of large (mainly UK) food companies.  The testing programmes are targeted and risk-based, not randomised, and the fraud risks within the suppliers of large BRC-certified retailers and manufacturers may be different than the companies supplying small manufacturing businesses or hospitality firms.

Many testing laboratories also supply their own customers with incident collations, and there are many commercial software systems that scrape reports from the internet.  All collect and treat the data slightly differently.  FAN produce a free annual aggregate of "most adulterated foods" from three of the largest commercial providers (Fera Horizonscan, Meriux SafetyHud, FoodChainID), which gives very high level smoothed data.

This research (purchase required) presents a nanozyme-based colorimetric sensor array for effective coffee variety discrimination by analyzing characteristic components.  Nanozymes are nano-scale molecules (for example, some metal oxides) that can catalyse biological reactions in an analogous way to enzymes.  

The authors report the development of an optimized sensor array containing 13 nanozymes.  No details of the reference samples or Machine Learning training are reported in the public abstract.  They report that their sensor enabled the discrimination of coffee key compounds (no details given in the abstract) within the concentration range of 0.8 μmol/L to 100 μmol/L. The sensing mechanism involves coffee components modulating nanozyme peroxidase-like activity through electron transfer and hydrophobic interactions.

They report that this technology successfully distinguished two main coffee categories and their subtypes with high accuracy.

They also report the development of a mobile phone app based on their sensor.  It achieved identification of coffee varieties and the detection of semi-quantitative adulteration levels. The conclude that this portable platform demonstrates significant potential for commercial coffee quality monitoring, providing a reliable tool for authenticity verification in the coffee industry.

Photo by Art Rachen on Unsplash

Combining data from multiple analytical tests into a single classification model can give greater discrimination than a model based on a single analytical technique.  There are three levels of data fusion: low-, mid- and high-level. Low-level fusion is the concatenation of analytical data obtained from several different sources.  Mid-level strategies imply that only the most significant features are fused after conducting a feature extraction step. High-level fusion combines the classification or regression results after they have been extracted separately from each type of data source.

This study (open access) aimed to build a model, testing different low- and mid- level fusion approaches of Raman and ATR FT-IR spectroscopy data, that can discriminate honey from different botanical origins (acacia, linden, colza, and raspberry) and different harvest years (2020 and 2021).  The models were trained using honeys exclusively collected from Romania but was then validated using honeys (of both the classes sought and of unrelated classes) from other countries of origin.

The authors tried different data fusion approaches.  They concluded that data fusion provided more accurate classification results than those resulting from a single input data type (i.e. Raman or IR) in most of the cases. Nevertheless, the simple increase of the number of input variables through the concatenation of the experimental data will not automatically generate an improvement in the models' prediction rate.  In order to obtain the best results using the data fusion approach, it is essential to find the best way of reducing the input space to those variables that have the highest discrimination capacity.

They found that the best performances were obtained when the low-level data fusion approach was used for both botanical and harvesting year recognition models. The differentiation potential of the classifiers was proven using an external validation set, leading to test accuracies between 84% and 100% for the two investigated classification criteria. The single difference was that for the botanical differentiation, the use of the fingerprint spectral regions proved to be more effective, while for the harvesting year classification, the involvement of the entire spectral ranges led to a better performance.

Mid-level data fusion provided similar differentiation accuracies in cross-validation, either for the use of the fingerprint regions or of the entire measured spectral ranges.

Photo by Roberta Sorge on Unsplash

Analytical tests that are underpinned by statistical classification models are often based on reference sets of relatively few (in statistical terms) “authentic samples”.  There is a risk that these may not reflect the entire scope of variability within an authentic food population.  Laboratories building these models therefore need to take great care in how they process the reference data (e.g. dimension reduction, feature selection) to avoid the problem of over-fitting.  Over-fitting results in a statistical model too tightly tailored to the reference set which then fails when applied to a sample that differs in some way.  There is best-practice guidance available for laboratories – see signposts on FAN.

This latest research (open access) develops specific recommendations and a workflow for laboratories to deal with dimension reduction.  It comes from a statistical, rather than analytical, scientific journal. The authors evaluated different statistical approaches, using a model dataset of ICP-MS data from 28 apples of 4 origin classes.  They compared Linear Discriminant Analysis (LDA) and Partial Least Squares Discriminant Analysis (PLS-DA) algorithms. Their workflow integrated Principal Component Analysis (PCA) for feature extraction, followed by supervised classification using LDA and PLS-DA. Model performance and stability were systematically assessed. The dataset was processed with normalization, scaling, and transformation prior to modeling. Each model was validated via leave-one-out cross-validation and evaluated using accuracy, sensitivity, specificity, balanced accuracy, detection prevalence, p-value, and Cohen’s Kappa.

They report that, as a linear projection-based classifier, LDA provided higher robustness and interpretability in small and unbalanced datasets. In contrast, PLS-DA, which is optimized for covariance maximization, exhibits higher apparent sensitivity but lower reproducibility under similar conditions. They also emphasise the importance of dimensionality reduction strategies, such as PCA-based variable selection versus latent space extraction in PLS-DA, in controlling overfitting and improving model generalisability.

They conclude that their proposed algorithmic workflow provides a reproducible and statistically sound approach for evaluating discriminant methods in chemometric classification.

This study (open access) builds on a previously-published proof of concept.  The authors are working towards producing a hand-held multi-mode scanner (combining fluorescence, visible, NIR, and short-wave IR spectroscopy) to support species verification of white fish fillets in business-to-business supply (currently reliant, largely, on visual recognition by experienced traders).

The explain that one of the key challenges in using machine learning for fish species identification is managing the large number of classes, as the variety of fish species is extensive. In their previous research, they introduced a novel multi-mode, highly multi-class machine learning framework based on a hierarchy of dispute models. This approach involved training a global model, and then recognizing groups of classes that have feature subspaces too similar for effective single-stage classification. By partitioning the overall space into smaller, distinct subspaces, they trained specialized models that are more tailored to these specific subsets of the dataset. In practice, the global model initially classified a sample to determine the appropriate subspace, while the dispute model then identified the precise species within that subspace.

The objective of this latest study was to apply this approach data acquired with the multi-mode handheld spectroscopy device. Tissue spectra were acquired at 25 positions on 68 fillets from 11 species, in both frozen and thawed states.

They report that feature-level fusion across the four spectroscopy modes enabled higher classification accuracy than any single mode alone. A global machine-learning model classified all species with 85 ± 2.8 %, while specialized dispute models for commonly misclassified species improved performance to 90 % ± 6.1 %. Individual models for thawed and frozen fillets achieved 90 ± 6.0 % and 90 ± 5.4 %, respectively, with dispute models in the thawed dataset increasing accuracy to 93 ± 4.3 %.

They conclude that their results demonstrate that portable multi-mode spectroscopy, combined with machine learning, can provide a fast, non-destructive and reliable tool for on-site fish species identification.

Authentication of Extra Virgin Olive Oil (EVOO) sometimes requires a panel of different tests and – with more sophisticated adulteration – a weight of evidence interpretation.  For more crude adulterations a single test is often enough.

One of the available tests is for fatty acids ethyl esters (FAEE).  These are more concentrated in lower quality oils (e.g.improperly stored or overripe), formed from ethanol which is a result of fermentation. EU legislation specifies a maximum 35 mg per kg FAEE concentration in EVOO.

FAEE concentration is officially measured using gas chromatography (GC) after recovery by silica gel column chromatography. While highly accurate, this method is complex, time-consuming, and relatively expensive.

This paper (purchase required) reports an alternative approach to FAEE measurement by using infra-red spectroscopy (FT-IR) with machine learning. A dataset of 170 olive oil samples with FAEE concentrations ranging from 1.81 mg/kg to 109.00 mg/kg were analysed using FTIR. Spectral data were preprocessed and used to train various regression models.

The authors report that the best performance was obtained with an XGBoost model. Explainable AI techniques (SHAP) enabled interpretation of the model and identification of spectral regions mostly associated with FAEE content.

They conclude that combining FT-IR spectroscopy with advanced ML models—particularly XGBoost—can effectively predict the concentration of FAEE.

Photo by Massimo Adami on Unsplash

Here is our latest monthly graphic from the EC Reports of Agri-Food Fraud Suspicions, showing a rolling 3-month trend. 

Our interpretation of the reports is subjective. In order to show consistent trends we have excluded cases which appear to be unauthorised sale but with no intent to mislead consumers (e.g. unapproved food additives, novel foods which are declared on pack), we have excluded unauthorised health claims on supplements, and we have excluded residues and contaminants above legal limits.  We have grouped the remaining incidents into crude categories.  Our analysis is intended only to give a high-level overview. 

It is notable how consistent is the relative frequency of different types of fraud.  The highest proportion always relate to falsified or unlicenced trade in high risk food (illegal operators, missing or falsified health certificates, attempts at illegal import) and relating to falsified or missing traceability documentation.  Particularly prominent over the past 3 months were:

• Documentation forgeries (eg invoices)  to falsify traceability
• Smuggling (deliberate avoidance of import checks)
• Excess water or low net weight of frozen seafood
• Non-prosecco “Prosecco”
• Adulterants in Dubai chocolate

Some incidents relate to absence of expected “premium” ingredients in manufactured food.  This is a reminder that compliance is judged not only against the ingredient declaration but also against the artwork (including any pictures alongside online sales) which give the impression that the product contains a particular ingredient.

These Agri-Food suspicions are just one of the incident databases available.  Different databases collect different information, in different ways, and therefore show a different angle on the true picture.  All of these sources are signposted on FAN.  Best practice is to use a combination of all sources, but the final critical question is “how vulnerable is my own supplier”.

• JRC – These are solely media reports.  They exclude cases not in the public domain, and can be biased by shocking but highly localised incidents in local food supply within poorly regulated countries.  They now incorporate a search and trending tool to produce graphs and charts
• EU Agri-Food Suspicions – These are solely EU Official Reports, and only suspicions.  The root cause of each incident is unknown.  The data include pesticide residues above their MRLs. unapproved supplements and novel foods, and unapproved health claims.
• Food Industry Intelligence Network Fiin SME Hub – These are aggregated anonymised results from the testing programmes of large (mainly UK) food companies.  The testing programmes are targeted and risk-based, not randomised, and the fraud risks within the suppliers of large BRC-certified retailers and manufacturers may be different than the companies supplying small manufacturing businesses or hospitality firms.  The Fiin dataset has just (November 2025) been updated.

Many testing laboratories also supply their own customers with incident collations, and there are many commercial software systems that scrape reports from the internet.  All collect and treat the data slightly differently.  FAN produce a free annual aggregate of "most adulterated foods" from three of the commercial providers, which gives very high level smoothed data.

The use of toxic testile dyes, such as the Sudan Red group, to adulterate food has been a high risk alert since the early 2000’s.  Over the past few years there have been persistent reported incidents with no apparent decline.  Original watch-outs were red spices and sauces, but more recently the focus has been on the adulteration of cheaper vegetable oils with red dye to pass them off as palm oil.  Palm oil from West Africa has been particularly implicated.

A recent media report from Ghana suggests that – far from improving – the problem is increasing in the case of palm oil on the local market.

Sudan Red dyes are classified as a Group 3 human carcinogen by the IARC and their widespread use in food is an obvious health concern for the local population.  For companies importing palm oil from countries where adulteration is endemic within the local market then traceability becomes key; being sure that your own stock comes from plantations and refineries with good and trusted oversight and has not been substituted for cheaper (adulterated) oil.  It is relatively easy to test for Sudan dyes, and periodic analysis is always a good way to check that assumptions about strong traceability are correct.

Photo by IKRAM ULLAH on Unsplash

This paper (open access) reports the development of a classification model for the geographic origin on black tea based upon measuring a panel of 15 trace elements by X-ray fluorescence (XRF).  XRF is a non-destructive technique.  The only sample preparation required is grinding the tea leaves into a fine powder.

The model could discriminate between 10 major tea-producing regions.  It was built using reference samples obtained, via tea industry contacts, directly from plantations or primary processing facilities.   791 black tea samples were collected in total: Assam (272 samples), Burundi (40 samples), Darjeeling (145 samples), Ethiopia (40 samples), Keemun (115 samples), Kenya region 1 (41 samples), Kenya region 2 (40 samples), Malawi (40 samples), Rwanda (10 samples), and Sri Lanka (48 samples).

Two unsupervised analysis techniques were used to visualize high-dimensional data, and six supervised models were employed to discriminate the ten GI regions.

The authors report that machine learning models, including random forest, support vector machine, k-nearest neighbours, linear discriminate analysis, and the deep learning multilayer perceptron (MLP) model, demonstrated superior predictive capabilities compared to the traditional partial least squares discriminant analysis model. The MLP model achieved the highest performance, with a 97.7 % overall F1 score in predicting the geographical origins of 532 authentic samples across ten GI regions.

The authors also Identified Rb, Sr, Mn, Si, and Cl as geographical markers for African region discrimination.

The conclude that their work could form the basis and foundation for an international database of tea Geographic Origin, enabling cheap and quick authenticity verification testing.

Photo by Oleg Guijinsky on Unsplash