Dairy products have been the target of some high-profile incidents of adulteration, resulting in sickness and deaths by the addition of melamine to milk and infant formula. Melamine detection is difficult because it mimics high protein content in routine quality tests. While it can be detected by more advanced tests, those tests are expensive and difficult to implement on a large scale. Other adulterants of milk include non-fat solids and inorganic salts in milk powder as well as representation of a lower quality milk as a higher quality product. Some cheeses are strictly regulated by manufacturing standards and designation. Products that do not meet standards are considered to be adulterated. Butter and yogurt can also be targets for adulteration, often by impurities containing fat and protein that can cheat routine quality tests. There is a need for fast, non-invasive testing for adulteration of dairy products that can replace current methods. NIR spectroscopy has been examined for this purpose and the results of some studies are summarized below.



  • Milk
  • Milk Powder
  • Cheese
  • Butter
  • Yogurt


  • Melamine
  • Cow Milk in Camel Milk
  • Dicyandiamide
  • Aminotriazole
  • Biuret
  • Soy Protein Isolate
  • Pea Protein Isolate
  • Calcium Carbonate
  • Maltodextrin, Sucrose
  • Authentic Origin Designation
  • Tallow
  • Edible Gelatin
  • Industrial Gelatin
  • Soy Protein

Scientific References and Statistics


Melamine Detection by Mid- and Near-Infrared (MIR/NIR) Spectroscopy: A Quick and Sensitive Method for Dairy Products Analysis Including Liquid Milk, Infant Formula, and Milk Powder – Balabin, Smirnov, Talanta 85 (2011) 562-568

Melamine (2,4,6-triamino-1,3,5-triazine) is a nitrogen-rich chemical most frequently used in making plastics. In routine quality tests like the Kjeldahl and Dumas methods, the high nitrogen content increases the apparent protein content, making it a chemical that can be used for food adulteration mimicking high protein content. Melamine contamination has been reported in liquid and powdered milk, infant formula, frozen yogurt, pet food, biscuits, candy, and coffee drinks. Two high profile incidents resulted in recalls of pet and human food in 2007 and infant formula in 2008, creating a widespread global food safety scare. Ingestion of melamine may lead to reproductive damage, bladder or kidney stones, and bladder cancer. The current FDA method for detecting melamine in infant formula is liquid chromatography-triple-quadrupole tandem mass spectroscopy (LC-MS/MS). While effective with a limit of detection as low as 0.25 ppm, it requires extensive sample preparation and cleanup, skilled labor, and is time-consuming and expensive, making it ill-suited for testing large numbers of samples. Vibrational spectroscopy offers a cost-effective and fast alternative to current methods and both NIR and Mid-IR spectroscopy were examined for detecting melamine adulteration in infant formula, milk powder, and liquid milk.

The initial sample set consisted of sixty infant formula samples and seventy-two each of milk powder and liquid milk. All samples were first checked for the absence of melamine using HPLC. After verifying the samples to be absent of contamination, they were mixed in random proportions to create six hundred ninety infant formula samples and six hundred sixty milk powder and liquid milk samples. Four separate melamine brands from three different producers were used as the adulterant. The range of melamine concentration was set to be very low (0.11 ppm) to very high (2000 ppm). Between one gram to five grams were prepared for each sample and all samples were homogenized before spectra collection. Samples were scanned right after preparation to minimize experimental errors. NIR spectra were collected from 9000 cm-1 to 4500 cm-1 using 8 cm-1 resolution. Sixty-four scans were collected per reading and averaged into one spectrum. An 8 mm in diameter cylindrical glass cell was used and five spectra were collected per sample, rotating the cell between each collection. These five spectra were then averaged into a single spectrum per sample. Mid-IR spectra were collected from 4000 cm-1 to 500 cm-1 using 2 cm-1 resolution. Thirty-two scans were averaged per spectrum and an ATR background was used. This collection process was repeated between five to seven times for each sample and all spectra from each sample were averaged into a single spectrum per sample. Before calibration modeling, nine different preprocessing methods were applied to both sets of spectra. Fairly poor results were obtained using Partial Least Squares (PLS) and Orthogonal Projections to Latent Structures (OPLS), indicating the possibility that a non-linear relationship existed between both sets of spectra and melamine concentration. The non-linear regression methods Polynomial-PLS (Poly-PLS), Artificial Neural Network (ANN), Support Vector Regression (SVR), and Least Squares Support Vector Machine (LS-SVM) were analyzed. In order to keep the model unbiased towards the accurate prediction of samples with high melamine content, the data sets were split into a low set and high set. The low data set used samples with a melamine concentration of 17.3 ppm or lower and the high set used samples with a melamine concentration of 17.3 ppm to the highest concentration of 2000 ppm. The results listed below are averaged results for both NIR and Mid-IR.

Low Melamine Concentration (Infant Formula, Milk Powder, Liquid Milk):

Average Error (PLS & OPLS)RMSEP = 1.31 +/- 0.07 ppm
Average Error (Poly-PLS, ANN, LS-SVM, SVR)RMSEP = 0.28 +/- 0.05 ppm

High Melamine Concentration (Infant Formula, Milk Powder, Liquid Milk):

Average Error (PLS, OPLS, Poly-PLS)RMSEP = 15.0 +/- 6.0 ppm (Estimated)
Average Error (ANN, LS-SVM, SVR)RMSEP = 6.1 +/- 0.9 ppm

The desired detection threshold for melamine adulteration is 1.0 ppm or less for lower concentrations. Analysis of the non-linear calibration models showed that the threshold of detection was 0.76 +/- 0.11 ppm, making both NIR and Mid-IR acceptable methods in practice for determining melamine concentration in all three types of milk products. In the case of the high melamine concentration, prediction error was much higher for Mid-IR than NIR. The results of both sets of models were verified by an independent validation set chosen from the samples. Overall, statistics from the calibration models showed an ability to measure infant formula, milk powder, and liquid milk with equal efficiency. The results here appear good enough to use NIR spectroscopy as a screening tool to detect adulterated samples that can be passed on for more advanced tests if melamine is detected. However, these models would require more validation before being used in a real setting. NIR spectroscopy rarely has a threshold of sensitivity low enough to measure parameters at a ppm level, even in the case of water which is known to be a very strong absorber in the NIR. It is possible that the change in melamine is colinear with other changes in the dairy product, thus creating an indirect correlation in the calibration models. However, while an indirect correlation is acceptable in NIR spectroscopy, such models require careful validation and the wavelength ranges used for the correlation must be carefully examined. Such analysis was not presented in this study. With such low concentrations and a non-linear relationship between NIR spectra and melamine concentration, careful calibration work must be done to use NIR spectroscopy for melamine detection in other products. The potential was demonstrated in this study to use NIR spectroscopy and calibration models to measure melamine adulteration in milk products but more careful examination of the results is required. If properly validated, NIR spectroscopy offers a much quicker and less expensive alternative to traditional reference methods for monitoring melamine adulteration.

Detection of Melamine Adulteration in Milk by Near-Infrared Spectroscopy and One-Class Partial Least Squares – Chen, Tan, Lin, Wu, Spectroschimica Acta Part A: Molecular and Biomolecular Spectroscopy 173 (2017) 832-836

Melamine is a nitrogen containing compound that has been implicated in global food scares involving milk products. It contains 66.7% nitrogen by mass and is used as an adulterant to increase the apparent protein content. The traditional Kjeldahl test for protein does not measure protein directly but determines protein from the nitrogen content without considering the source. High doses of melamine in dairy products can result in kidney stones, renal failure, and has resulted in deaths of babies after consuming melamine-adulterated infant formula. The publicity from incidents of adulterated dairy products has resulted in the development of a number of testing methods. However, these tests are often expensive, time-consuming, require skilled labor and the use of toxic solvents, and are ill-suited to use as a large-scale quality assurance tool. An ideal analytical method to verify the quality and authenticity of food products requires speed with little sample preparation and low cost. NIR spectroscopy was examined as a method for determining melamine adulteration in milk. Milk powder was procured from a local supermarket for the study and was confirmed to be free of melamine. Sixty-two 100 ml samples of milk liquor were prepared over two days with a week interval in between. Forty-two of these samples were set aside as pure samples and the remaining twenty-two were prepared as adulterated samples. 99% pure melamine was procured from a vendor and different concentrations of melamine were dissolved in the remaining twenty samples of milk liquor. Melamine concentration ranged from 0.001 g/100 ml to 0.29 g/100 ml, which is the upper limit of solubility of melamine in water. NIR spectra were collected using an FT-NIR spectrometer from 10000 cm-1 to 4000 cm-1 at 3.856 cm-1 intervals. Thirty-two scans were collected per reading and averaged into one spectrum. One spectrum was collected for each pure milk sample. Three spectra were collected for each adulterated sample, making a total of one hundred two NIR spectra. A One-Class Partial Least Squares (OC-PLS) classification model was created by assigning a value of 0 to all pure milk spectra and 1 to all adulterated samples. A Variable Importance (VI) index was used to select the forty most important input variables for the classification. Samples were split into a training set to create the model and a test set for model validation.


Accuracy 89%Sensitivity 90%Specificity 88%

The results shown above were determined from predictions using the test set spectra and prove the feasibility of using NIR spectroscopy and a classification model as a screening tool to determine the presence of melamine adulterant in milk. Future work should include more types of milk and different concentrations of melamine in order to increase the robustness of the model. Implementing NIR spectroscopy as a method for detecting melamine adulteration offers a less expensive and time-consuming alternative to current methods and can be used as a screening tool to find adulterated samples that can be sent for more advanced testing if melamine is detected.

FT-NIRS Coupled With Chemometric Methods As A Rapid Alternative Tool for the Detection & Quantification of Cow Milk Adulteration in Camel Milk Samples – Mabood, Jabeen, Hussain, et al., Vibrational Spectroscopy 92 (2017) 245-250

Camel milk is considered to have high nutritional value in comparison to milk produced by cows. It is a rich source of Vitamin A and C, has a high content of potent immunoglobins, and does not contain lactoglobulin and A1 casein, making it suitable for consumption by people with bovine dairy allergies. Thus, it sells for a higher market price and is subject to adulteration using cheaper forms of milk. FT-NIR spectroscopy was examined as a method for determining cow milk adulteration in samples of camel milk. Three samples of camel milk were procured for the study and prepared in triplicate form. Each separate camel milk sample was adulterated with different percent levels of cow milk adulterant: 2%, 5%, 10%, 15%, and 20%. Including the three pure samples, a total of fifty-four samples were created. 70% of the samples were used as a calibration set to create the model and the remaining 30% was used as a validation set. All samples were scanned from 700 nm to 2500 nm at 2 cm-1 spectral resolution using a 0.2 mm pathlength sealed cell. Two calibration models were created: Partial Least Squares Discriminant Analysis (PLS-DA) to determine the presence of the cow milk adulterant and Partial Least Squares (PLS) to quantify the amount of adulterant present.


Presence of AdulterantR2= 0.973RMSEP= 0.0801


Amount of AdulterantR2= 0.926RMSEP= 1.32%

Both calibration models showed good results and proved the feasibility of the measurement. In the case of PLS-DA, an arbitrary value of 0 was assigned to the pure camel milk samples and 1 to samples spiked with 10% cow milk adulterant. The model predicts a number and a threshold of 0.5 was chosen to determine the presence of adulterant. A predicted value less than 0.5 indicates no adulterant and a predicted value greater than 0.5 indicates an adulterant is present in the sample. The high correlation and low RMSEP show that this model can be used to determine the presence of cow milk adulterant in camel milk. In the case of PLS, the results show that the model can predict the amount of adulterant present within an accuracy of 1.32%. It must be noted that while the models showed good results, using them in a practical setting for different kinds of milk and adulterants would require a much larger sample set. Natural products often show variability in NIR spectra due to many factors, such as region of origin, different types of food fed to animals, different soil for plant growing, and so forth. Incorporating different samples encompassing any potential variability is important when building calibration models. Predictions performed on the validation set proved that the models could work in a real-time setting, using the PLS-DA model as a detection tool and the PLS model as a quantification tool, providing information that would be very difficult to find using conventional methods for adulteration detection.

Non-Targeted NIR Spectroscopy and SIMCA Classification for Commercial Milk Powder Authentication: A Study Using Eleven Potential Adulterants – Karunathilaka, Yakes, He, et al., Heliyon 4 (2018) e00806.

NIR spectroscopy was evaluated as a method for rapid screening of commercial milk powder products as authentic or being mixed with known and unknown adulterants. Milk powder is a known target of adulteration, ranking only second to olive oil according to the USP database on food fraud and economic adulteration. Milk powder adulteration can cause adverse health effects, as highlighted by two incidents of milk and wheat gluten adulteration with melamine. One incident was linked to renal failure in cats and dogs due to pet food adulteration in the United States. The other was linked to infant formula in China, causing thousands of cases of renal complications in children as well as at least six confirmed deaths. A number of commercially available milk powder samples were procured for the study, representing manufactured products in sixteen different states and twenty-four different companies and brands. Eleven different milk powder adulterants were selected for the study based on their history or potential use as adulterants. These can be divided into four categories: 1. Low Molecular Weight, Nitrogen-Rich Compounds, 2. Plant Proteins, 3. Inorganic Salts, 4. Non-Fat Solids. Some of the adulterants were blended as well to increase the scope of the potential adulterants that could be used. Different levels of all adulterants were mixed with the pure milk powder samples before NIR spectra were collected. Three different spectrometers were used for the study: Two benchtop FT-NIR instruments and a handheld NIR device. Principle Component Analysis (PCA) was performed on all three sets of data and SIMCA classification models were created for determining the presence of an adulterant in the milk powder. After model creation, separate test sets of both pure and adulterated samples were scanned for model validation.

FT-NIR Spectrometer #1

Instrument Parameters:

12500 cm-1 to 4000 cm-1, sixty-four scans per average, 16 cm-1 spectral resolution

100% Correct Adulterant Classification and Concentration (%w):

Melamine 0.6% to 2.0%
Dicyandiamide 2%
Aminotriazole 0.4% to 2.0%
Biuret 0.2% to 2.0%
Soy Protein Isolate5% to 20%
Pea Protein Isolate5% to 20%
Calcium Carbonate2%
Maltodextrin 2% to 20%
Sucrose 7% to 50%

FT-NIR Spectrometer #2

Instrument Parameters:

10000 cm-1 to 4000 cm-1, thirty-two scans per average, 16 cm-1 spectral resolution

100% Correct Adulterant Classification and Concentration (%w):

Melamine 0.4% to 2.0%
Aminotriazole 0.4% to 2.0%
Biuret 0.2% to 2.0%
Cyanuric Acid2%
Soy Protein Isolate5% to 20%
Pea Protein Isolate2% to 20%
Maltodextrin 5% to 20%
Sucrose 10% to 50%

Handheld Spectrometer

Instrument Parameters:

6266 cm-1 to 4167 cm-1, 10 scans per average, 11 nm optical resolution

100% Correct Adulterant Classification and Concentration (%w):

Biuret 0.4% to 2.0%

Results for all three spectrometers are shown above. Both FT-NIR benchtop spectrometers showed 100% specificity and accuracy for determining the presence of an adulterant in milk powder if the adulterant was at a sufficiently high percentage, which varied based on the type of adulterant present. In the case of the handheld NIR device, results were much worse. This is most likely due to a narrower wavelength range and lower resolution than the benchtop FT-NIR instruments. The results here prove the feasibility of using FT-NIR spectrometers as a tool for determining the presence of adulterant in milk powder and show that FT-NIR spectrometers are much better suited for such analysis than handheld NIR spectrometers.


Screening of Grated Cheese Authenticity by NIR Spectroscopy – Cevoli, Fabbri, Gori, et al., Journal of Agricultural Engineering 2013; volume XLIV(s2):e53

Parmigiano-Reggiano (PR) cheese is one of the oldest traditional cheeses produced in Europe and has a Protected Designation of Origin in Italy. It is manufactured exclusively from whole PR wheels that correspond to the production standard. Grated PR cheese must be matured for a period of twelve months and characterized by a rind content of less than 18%. NIR spectroscopy was examined as a method for determining the authenticity of PR grated cheese. Four hundred samples were procured for the study with the following classifications: Compliance PR, Non-Compliance PR, PR with Rind Content > 18%, and Competitors (various commercial brands of grated cheeses obtained from local markets). NIR spectra were collected using an FT-NIR spectrometer in diffuse reflectance mode from 12500 cm-1 to 4000 cm-1. Thirty-two scans were averaged per spectrum and 8 cm-1 spectral resolution was used. Three replicate spectra were collected per sample. Various pre-processing treatments were performed on the NIR spectra. Principle Component Analysis (PCA) was first performed as an exploratory tool to define discrimination between Compliance and Non-Compliance PR samples or Competitors. Artificial Neural Network (ANN) models were created using software to test the feasibility of predicting each specific class from the spectral data. Reference values of Rind % and Months of Ripening were used with the NIR spectra to create Partial Least Squares (PLS) regression models for predicting these values from the spectra.


Compliance PR Classification100% for Training Set100% for Validation Set
Competitors Classification100% for Training Set95.5% for Validation Set
Non-Compliance PR Classification100% for Training Set100% for Validation Set
Rind Content > 18% Classification100% for Training Set100% for Validation Set


Rind %R2= 0.982RMSEP= 1.473%
Months of RipeningR2= 0.986RMSEP= 0.805

The results obtained in this study for both types of models were excellent and confirmed the ability of NIR spectroscopy to be used as a screening tool for determining grated cheese authenticity. The ANN model was able to 100% predict compliance or non-compliance in PR samples and detect competitor grated cheese at an accuracy above 95%. More competitor samples in the model will likely improve these results. In the case of PLS, the results were especially good considering there was some question in the reference values for rind %. A regression model can only predict within the error of the reference method and these results should improve as well with more accurate reference testing. Ripening can be predicted to within an accuracy of less than one month. NIR spectroscopy can be used as a fast, non-destructive screening tool for determining the authenticity of grated cheese.


Robust New NIRS Coupled With Multivariate Methods for the Detection and Quantification of Tallow Adulteration in Clarified Butter Samples – Mabood, Abbas, Jabeen, et al., Food Additives & Contaminants: Part A, 35:3, 404-411

Food adulteration has become a big problem on a global scale during recent years. Increased population, higher supply and demand for food, and less detectable methods of adulteration have all contributed to the problem. Food authenticity and detection of adulteration have become a priority for both food producers and consumers, as adulteration results in reduced profits, bad publicity, and in some cases, presents a health risk to the public. Dairy products are no exception to the adulteration issue and one potential adulterant in butter is tallow, an animal fat material which causes increased serum cholesterol and triglycerides levels when consumed. Tallow is even used to make candles and soap and is obviously an unsuitable substitute for butter at any concentration. Visual examination is difficult for determining the presence of adulterants and wet chemistry methods are time-consuming and expensive. NIR spectroscopy was examined as a fast method requiring little sample preparation for determining the presence of tallow adulterant in butter. Nine portions of pure butter samples with no tallow adulteration were set aside. Tallow was prepared by melting animal fat and collecting the oil portion poured out from the solid residue. Nine samples of each of the following tallow concentrations by weight in butter were prepared: 1%, 3%, 5%, 7%, 9%, 11%, 13%, 15%, 17%, and 20%. Including the nine samples with no tallow adulteration, a total of ninety-nine samples were used for the study. NIR spectra were collected in reflectance mode from 10000 cm-1 to 4000 cm-1 using 2 cm-1 resolution. A transflectance sample accessory with a total pathlength of 0.5 mm was used for collection. Various pre-treatments were performed on the NIR spectra, after which Principle Component Analysis (PCA), Partial Least Squares-Discriminant Analysis (PLS-DA), and Partial Least Squares (PLS) chemometric methods were used to build the models.


redict Value (0 = no tallow, 1 = presence of tallow)R2= 0.95RMSEP= 0.062


Tallow %R2= 0.973RMSEP= 1.537%

The results presented were excellent and proved the feasibility of using NIR spectroscopy to determine the presence of tallow adulterant in butter as well as quantitatively measure the amount of tallow with reasonable prediction error. After various pre-treatments were performed, the wavenumber range from 7500 cm-1 to 4000 cm-1 using first derivative with fifteen smoothing points was used for the calibration models. Examination of the scores plot after PCA showed clear classification and separation between each group of samples, proving that changes to the NIR spectra occur as more tallow is added to the samples. PLS-DA uses the arbitrary values of 0 and 1 for classification purposes between two groups. The model generates a number based on NIR spectra. A number less than 0.5 classifies as the group assigned to 0 and a number greater than 0.5 classifies as the group assigned to 1. In this case, the RMSEP of 0.062 is more than accurate enough to classify the sample as non-adulterated or adulterated. PLS predicts a quantitative value from NIR spectra and a calibration model. The RMSEP shows a prediction accuracy with error slightly greater than 1.5% tallow, which is accurate enough for real-time use. In order to properly validate the model, 30% of the samples were removed from the PLS model, a new model was created without those samples, and the NIR spectra of those samples were used with the new model to predict the percentage of tallow adulteration. These results proved the validity of the model and all predictions showed an error of less than 2% tallow. More samples over the range of values encompassing different types of butter will improve the results and make the model robust enough for universal application for tallow adulteration screening of butter.


The Feasibility of Using Near-Infrared Spectroscopy and Chemometrics for Untargeted Detection of Protein Adulteration in Yogurt: Removing Unwanted Variations in Pure Yogurt – Xu, Yan, Cai, et al., Journal of Analytical Methods in Chemistry, Volume 2013, Article ID 201873

In recent years, scandals involving dairy product adulteration have led to the development of new targeted analytical methods to detect the presence of adulterants. Melamine has especially been examined as a milk adulterant but new types of adulterants are being reported all the time. One such adulterant is different types of non-milk proteins in yogurt. Because of the evolving nature of adulterants, new types of untargeted analyses are needed to determine whether or not a product is pure and unadulterated, with the focus being on not necessarily identifying the specific adulterant but rather determining the presence of one. If a product is determined to be impure, further testing can be done to identify the adulterant. NIR spectroscopy was examined as a method for protein adulteration identification in yogurt. Yogurt was manufactured from milk and bacteria cultures using the standard method specifically for the study. The yogurt was divided into nineteen portions. Three portions were kept pure with no adulterant. Six portions were adulterated with edible gelatin ranging from 1% to 8% by weight. Five portions were adulterated with industrial gelatin ranging from 0.5% to 5% by weight. Five portions were adulterated with soy protein powder ranging from 0.5% to 5% by weight. In order to keep the thickness uniform in all the samples, pure water was added which is the common practice in protein adulteration. NIR spectra of both the pure and adulterated samples were collected from 12000 cm-1 to 4000 cm-1 in diffuse reflectance mode. Spectral resolution was 8 cm-1, scanning interval was 3.857 cm-1, and sixty-four scans were collected per reading and averaged into one spectrum. In total after dividing the portions, sixty spectra of pure samples and one hundred ninety-seven spectra of adulterated samples were collected. Spectrum of pure water was collected by averaging five measurements of water film on the reflectance background. In order to remove the influence of water variation, all NIR spectra of the pure and adulterated yogurt samples were orthogonally projected (OP) on the complement space of water spectrum using an algorithm, minimizing the influence of the water difference on the classification. Standard Normal Variate (SNV) processing was used as well to reduce scattering effects and correct interference caused by variations. The groups were divided into a training set and test set and the OCPLS class modeling algorithm was used to classify samples using the following sets of NIR spectra: raw spectra, OP, and SNV.

OCPLS Classification:

Raw Spectra Test Set17/20
Raw Spectra Training Set163/197
SNV Spectra Test Set18/20
SNV Spectra Training Set181/197
OP Spectra Test Set18/20
OP Spectra Training Set181/197

The results showed that both pre-processing methods had a positive effect on the models, with the OP spectra classification showing slightly better results than the SNV spectra classification. Because water has such a strong absorbance in the NIR wavelength range, it is possible that without pre-processing, the results for the raw spectra and SNV may be classifying based on the differences in water and not the presence of an adulterant. Careful analysis of the wavelength ranges used to determine the classification will show this, but such analysis was not performed in this study. One purpose of analyzing the classification in this manner was to determine the minimum threshold for each adulterant that could be detected from the NIR spectra. None of the non-adulterated samples were incorrectly classified as having an adulterant present. For the samples that were incorrectly classified as being pure while having an adulterant present, the concentration was 0.5% for edible gelatin, 1% for industrial gelatin, and 1% for soy protein powder. All samples with an adulterant concentration of 1% edible gelatin, 2% industrial gelatin, and 2% soy protein powder (or higher) were correctly classified. These can be considered the safe thresholds of detection that the models can accurately use to detect protein adulteration in yogurt. The potential was demonstrated to use NIR spectroscopy as a method for protein adulteration in screening of yogurt and further work with a larger sample set at lower concentrations of protein adulterants should improve the results.