News

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.

Spain has a legal limit of 3% for undeclared vegetable proteins in meat patties.  The aim of this open-access study was to evaluate the feasibility of point-based near infrared spectroscopy (NIRS) and hyperspectral imaging (HSI) to verify compliance.

The model was trained on patties prepared in-house.  They were all prepared from the same cut of beef, so the robustness of the model has not been verified.  A total of 240 patties were fabricated, of which 60 contained pea (PP), 60 contained soybean (SP), and 60 chickpea protein (CP) at levels from 1 up to 6 % (w/w). 60 pure beef patties were included.

The authors report that they could clearly discriminate the type of protein added, using either partial least squares-discriminant analysis (PLS-DA) or linear discriminant analysis (LDA), with >90 % of the samples in the test set correctly classified. Based on protein inclusion, LDA discriminated 100 % of the PP, SP and CP samples with both NIR and HSI. PLS-DA classified 100 % of the PP and CP burgers using the NIR instrument. To manage double classification tasks, a hierarchical model classifier (HMC) was proposed for both NIR and HSI spectra, achieving classification rates of at least 83% by combining LDA and PLS-DA models at the nodes.

The authors conclude that NIR spectroscopy is suitable for detecting low levels (1 %) of vegetable protein flours added to beef burgers.

In this paper (open access) the authors built a classification model to discriminate premium from non-premium grades of vacuum packed sliced Iberian ham.  They used a Near Infrared Scanner reading directly through the packaging.

The model was constructed from a database of 312 purchased from retail on a weekly basis over a two-year period (2023–2024). The samples were obtained as vacuum-packed slices from a range of commercial brands and industrial producers, in a manner analogous to typical consumer purchasing behaviour in supermarkets, encompassing the four official commercial categories: black seal, red seal, green seal, and white seal. These samples were preliminarily grouped into premium (201 samples) and non-premium (111 samples) categories based on their commercial labelling.  The researchers further verified the label categorisation by free fatty acid analysis.

The classification was based on the quality and sensory differences that appear in products derived from animals fed with natural resources (acorn and grass) in extensive systems (premium category), as opposed to those from animals fed with compound feeds (non-premium category}

The authors report 100 % sensitivity, specificity, and non-error rate (NER) for both of two different NIR sensors tested during external validation.  They report that a lower cost miniaturised model performed less well, with 100 % sensitivity but 85.71 % specificity and 94.74 % NER, limiting its applicability for samples near the classification threshold.

They conclude that their results confirm the suitability of NIRS technology for rapid and non-destructive in situ classification of high-value foods, including pre-sliced Iberian ham.

This study (purchase required) evaluated two portable NIR spectrometers (900–1700 nm and 1450–2450 nm) and a benchtop FTIR device (4000–550 cm⁻¹) for authenticating edible insect flours. The reference data were constructed from flours produced in-house from insects or larvae purchased online: mealworm (Tenebrio molitor) larvae (23 samples), buffalo worm (Alphitobius diaperinus) larvae (28 samples) and crickets (Acheta domesticus) (28 samples).  Data-Driven Soft Independent Modelling Class Analogy (DD-SIMCA) and soft Partial Least Squares Discriminant Analysis (sPLS-DA), were used on the spectral data.

Principal Component Analysis (PCA) showed that spectral data of pure insect flours were clustered in the scores plot. DD-SIMCA achieved 100 % sensitivity (SNS) in the test set using FTIR for all insects. NIR Spectrometer in the range of 1450–2450 nm reached 100 % SNS and 100 % specificity (SPS) for buffalo worm and mealworm flour. sPLS-DA showed class sensitivity (CSNS) between 75 % and 100 %, for all three devices tested, with spectrometer in the range of 1450–2450 nm reaching class efficiency rate (CEFF) and total efficiency (TEFF) values ranging from 93 % to 100 %. Also, PLSR achieved RMSEP values as low as 0.44 %, demonstrating its robustness as a tool.

The authors conclude that IR spectroscopy with soft modelling is a non-destructive solution for authenticating insect flours, filling the current gap in rapid and reliable analytical tools for this emerging industry.

Photo by Olga Kudriavtseva on Unsplash

Seed fraud, particularly the misrepresentation of rice paddy (unhusked rice grain) as rice seed, is a growing concern that threatens sustainability efforts.

This study (open access) proves the concept of using a portable NIR spectroscopic device, combined with chemometric analysis, for rapid onsite identification of rice seed and paddy varieties for real-time verification of seed authenticity.

A total of 280 rice samples, representing four varieties (Agra, Amankwatia, Legon 1, and Jasmine 85) across two categories (seeds and paddy), were analyzed.

After applying various pre-processing techniques and principal component analysis (PCA), the authors report that linear discriminant functions 1 and 2 revealed distinct clustering patterns for both the varieties and categories (rice seed and paddy). Among the classification algorithms used, Random Forest (RF) achieved 100 % accuracy for rice seed identification and 97.38 % for paddy identification in the test sets. Support Vector Machine (SVM) demonstrated 98.15 % accuracy in distinguishing between rice seed and paddy for detecting seed fraud.

The authors conclude that such a portable NIR device can reliably perform varietal identification and seed authenticity checks, including use by seed inspectors, farmers, and regulatory officers.

Photo by Prahlad Inala on Unsplash

In this study (open access) researchers developed and piloted a single in-line sensor to classify yoghurt as either sheep, goat or milk origin and simultaneously check viscosity and pH Quality Attribute Specifications.  Their goal is a rapid in-line sensor that incorporates automated decision making, for routine use in the dairy industry.

Their reference dataset was sourced directly from two reputable Spanish companies and included both pasteurised and UHT yoghurts.

They found that the animal origin of milk could be predicted by building models based on the spectral data between 400 and 600 nm whilst viscosity and pH could be predicted by building models based on the spectral data between 800 and 1800 nm. To identify the animal origin of milk, they used Partial Least Squares-Discriminant Analysis (PLS-DA), achieving 100 % accuracy (95 % confidence interval). The model used to predict pH and viscosity was built with Partial Least Squares Regression (PLSR). The predictive power was generally very good (MSE=0.04–0.06; R²=0.94–0.96; MAE=0.16–0.17).

They conclude that their study demonstrates that the proposed spectroscopic method offers a more efficient approach for the simultaneous prediction of pH, viscosity, and milk origin in yogurt compared to existing methods, that require separate and slower analyses. Further work still needs to be carried out to optimize the model and achieve real-time monitoring that enables automated decision-making.

[picture – from the publication]

In this study (purchase required) the authors propose and develop a strategy for a field-based screening test for crude honey adulteration (adulteration with inverted sugars) using Near InfraRed Spectroscopy (NIR) hand-held scanners.  They developed a single class classification model that was sufficient to either give an “unadulterated” verdict or to refer the sample for confirmatory (IRMS) analysis.

The authors developed their SIMCA model using “genuine” adulterated honeys that had been previously seized in a Brazilian police operation that had cracked down on industrial-scale addition of invert sugar to honeys over a three year period, along with unadulterated honeys collected by police during the same operation.  The traditional SIMCA was improved by optimizing the class boundaries based on receiver operating characteristic (ROC) curves and the estimate of an uncertainty region, thus optimising the model for a screening application.

Photo by Roberta Sorge on Unsplash

Cocoa is high on many companies’ current risk radar for authenticity threats, due to recent supply pressures and price increases. Carob has legitimate uses as a cocoa replacement, and carob flour has been cited as a potential cocoa adulterant.

 A number of chemometric classification methods to differentiate cocoa from carob recently have been proposed, including one featured in our blogs in January based on DART-MS.  In a more recent publication (purchase required) the authors use the alternate method of near and mid-infrared spectroscopy before applying various chemometric approaches.

Spectral data were collected using four different infrared spectrometers: a benchtop FT-NIR system, two portable NIR instruments, and a benchtop FT-MIR-ATR. Reference samples included pure cocoa, pure carob, and their mixtures with carob concentrations ranging from 0 % to 60 %. Both classification and regression models were developed to detect and quantify the presence of carobs in cocoa powder. Classification models, including Random Forest (RF), Support Vector Machine (SVM), Multi-Layer Perceptron (MLP), k-Nearest Neighbors (kNN), Linear Discriminant Analysis (LDA), and Soft Voting Classifiers, demonstrated superior performance for discriminating between cocoa powder, carob powder, and cocoa-carob mixtures, particularly using the benchtop FT-NIR. Similarly, regression models - RF, SVM, MLP, kNN, Partial Least Squares Regression, and Voting Regressor- exhibited robust predictive capabilities, particularly, FT-MIR and portable NIR.

Overall, these findings highlight and prove the potential of NIR and MIR spectroscopy as rapid, robust, and non-destructive tools for screening and quality control in food authentication.

Photo by lindsay Cotter on Unsplash

A2 cows’ milk is from selective breeds that produce a higher ratio of the A2/A1 form of β-casein in the milk.  It is sold mostly in Australia, New Zealand, China, and the United States and commands a price premium over conventional milk.  Authenticity testing has been difficult, typically requiring genetic techniques.

In this study (purchase required) the researchers piloted portable NIR to differntiate A2 milk versusnon-A2 milk and their mixtures using a portable NIR spectrometer.  They built a 1-class classification model.  63 samples of whole A2 milk were selected (authentic set), and 40 samples (fraudulent set) composed of non-A2 milk and mixtures in 3 different proportions (10, 25, and 50% v/v) of non-A2 milk in A2 milk. The abstract gives no further details of the reference samples in terms of production systems or seasonality.  For spectra collection, a MicroNIR was used.

Full data were pre-processed using different methods, but they found the most effective approach was the combination of the first derivative with Savitzky-Golay smoothing and Standard Normal Variate (SNV). A Data-driven Soft Independent Modeling of Class Analogy (DD-SIMCA) was applied. Using the Kennard-Stone algorithm, the authentic samples were split into two sets (45 for calibration and 20 for external validation). The non-A2 and fraudulent samples were added to the external validation set, and the model’s performance was evaluated using the metrics of sensitivity, specificity, accuracy, and precision.

They report that the DD-SIMCA model, utilizing 2 PCs, showed 100% results in all metrics, indicating no errors in the recognition of authentic samples.

They conclude that the model is suitable for use with portable equipment. Additionally, this fast and non-invasive technique can be optimized for applications in industrial management, food control, and A2 product authentication.

In this paper (open access) the authors trained a Machine Learning model to differentiate between Top, Bottom and Spontaneous fermented bottled beers.  Data were collected using a non-invasive hand held NIR scanner pointed directly through the unopened bottle using a customised foam attachment.  The model was trained on 25 samples of major brands purchased online, rather than reference samples of verified traceability, but the training samples covered a wide range of beer types from stouts to light ales, and a wide range of bottle types and colours.

The authors report good classification based on fermentation method.  They consider that evidence of a wrong fermentation method could be one quick and easy check that could flag counterfeits.  They also correlated the NIR data with sensory panel assessments and SPME-GC-MS data and concluded that non-invasive NIR has the potential to classify beers based on their aroma profiles.

Image from the paper

In this paper (open access) two optical spectroscopic techniques,  Laser-Induced Breakdown Spectroscopy (LIBS) and UV-Vis-NIR absorption spectroscopy, are assessed for EVOO adulteration detection, using the same reference database of olive oil samples. In total, 184 samples were studied, including 40 EVOOs and 144 binary mixtures with pomace, soybean, corn, and sunflower oils, at various concentrations (ranging from 10 to 90% w/w). The reference class of “pure” EVOOs were limited to oils from a specific geographic region (either Crete, Lesvos, Kalamata or Achaia, with a different model built for each case).

The emission data from LIBS, related to the elemental composition of the samples, and the UV-Vis-NIR absorption spectra, related to the organic ingredients content, were analyzed, both separately and combined (i.e., fused), by Linear Discriminant Analysis (LDA), Support Vector Machines (SVMs), and Logistic Regression (LR). In all cases, very highly predictive accuracies were achieved, attaining, in some cases, 100%.

The authors conclude that both techniques have the potential for efficient and accurate olive oil verification test protocols, with the LIBS technique being better suited as it can operate much faster.