Research report: Detection and quantification of water adulteration in goat milk using FT-NIR spectroscopy combined with partial least squares and logistic regression

Goat milk commands a premium price on the dairy market, and that premium makes it a target. Diluting goat milk with water, or stretching it with cheaper cow milk, is one of the simplest ways to commit dairy fraud, and one of the hardest to catch quickly. A research team in Brazil now reports that a benchtop Fourier-transform near-infrared spectrometer, paired with two workhorse statistical models, can detect and quantify exactly this kind of adulteration in ternary mixtures of goat milk, cow milk and water, with prediction errors of only a few percent.

Their approach is deliberately practical: rather than relying on expensive laboratory assays for foreign DNA or proteins, they used near-infrared light, which interacts with the C–H, N–H and O–H chemical bonds that dominate the composition of milk, and let multivariate statistics do the interpretation.

To build the test set, the researchers prepared ternary mixtures following a simplex-centroid experimental design, a scheme that systematically samples the full compositional triangle formed by water, cow milk and goat milk. Each mixture was made up to 10.0 milliliters in test tubes and kept refrigerated until analysis. In total, 95 FT-NIR spectra were collected from 48 distinct mixtures, each measured in duplicate, on a Bruker MPA spectrometer operating in diffuse reflectance mode between 12000 and 4000 cm−1, with 32 scans per sample at a spectral resolution of 8 cm−1. Before modeling, spectra flagged as outliers by a Hotelling T2 test at the 95 percent confidence level were removed, and the remaining data were split into a calibration set of 67 spectra, about 70 percent, and an independent external validation set of 29 spectra, about 30 percent.

The raw materials were characterized thoroughly before any spectroscopy was done. Raw cow milk came from the Milk and Derivatives Laboratory at the Federal University of Lavras, pooled from different milkings and animals of various breeds, while raw goat milk arrived frozen from the Capril Eldorado farm in Perdões, Minas Gerais. Physicochemical analysis showed apparent pH values of 6.72 for cow milk and 6.68 for goat milk, titratable acidity of 0.16 and 0.15 percent lactic acid respectively, and identical relative densities of 1.033. Electrical conductivity was higher in goat milk, 4.64 against 4.25 mS/cm, reflecting a greater concentration of ions such as calcium, potassium and sodium, and total fat was also higher in goat milk at 4.3 versus 3.8 g/100 g. The authors note that the potential effect of freezing on goat milk composition was deliberately disregarded, citing earlier reports of no significant changes in the main constituents of Saanen goat milk after freezing.

Exploratory analysis of the spectra identified three wavelength ranges that carried the strongest compositional information: 10800 to 9800, 8800 to 7200 and 6400 to 5600 cm−1. Correlating spectral intensity at each wavelength with the percentages of goat milk, cow milk and water showed that bands above 7000 cm−1 had the highest correlation values, with the water correlation negative throughout that range. A variance analysis of the normalized spectral intensities, deconvoluted into nine Gaussian components with an excellent fit, pinpointed key peaks at 10208.3, 8636.6, 8277.2, 7677.7, 7465.6, 7343.1, 7269.7 and 5948.9 cm−1, corresponding to overtone and combination vibrations of C–H, N–H and O–H bonds. A preference map built from the same spectra captured more than 95 percent of the spectral variance in its first two factors and placed pure water, goat milk and cow milk at the vertices of a triangle, with mixtures falling along the edges and interior in positions that closely reconstructed the original experimental design.

The quantitative workhorse was partial least squares regression, run in SAS OnDemand after mean-centering and scaling the spectral matrices. Separate predictor models were built for the percentage of goat milk, cow milk and water in each mixture. The models performed well across a wide range of compositions, with root-mean-square errors of 1.2 to 1.5 percent on training data and 1.8 to 2.0 percent on the external validation set. In the conclusion the authors report calibration performance of R2 = 0.995 with an RMSEC of 1.3 percent, and external validation performance of R2 = 0.986 with an RMSEP of 1.9 percent, figures they compare favorably against the 2 to 3 percent errors reported in earlier goat–cow milk quantification studies.

For classification, the team turned to logistic regression, implemented in Orange Data Mining software. Samples containing 90 percent or less goat milk, meaning at least 10 percent added water or cow milk, were labeled as fraud, a threshold chosen on the basis of typical adulteration levels reported in industry studies; samples above 90 percent goat milk were classed as not suspected of fraud. The classifier performed almost perfectly on training data, with R2 = 0.999, and on validation data it misclassified only a single sample, giving R2 = 0.947, or 94.7 percent accuracy. The authors note this is on par with the 93 to 95 percent accuracy achieved in other studies using support vector machines on Raman spectra for milk adulteration detection. Adulteration levels below 10 percent were not evaluated, since fraudsters generally prefer to dilute with higher percentages to increase profits; one cited industry case reported 35.3 percent added water and starch in raw milk.

The spectroscopy itself is fast and non-destructive, which is precisely its appeal for routine quality control. Near-infrared spectra arise from the vibration of molecules at specific frequencies in the 780 to 2500 nanometer range, allowing chemical bonds to be identified by comparison with reference spectra. In this study, samples with high water content, particularly pure water, produced spectra clearly distinct from those containing milk, and principal component analysis separated water-rich samples from milk-rich ones, which clustered toward the center of the score plot. The authors also describe a very low-cost accessory built for under ten dollars, an electronic circuit of infrared LEDs, photodiodes and diodes connected to a PC sound card, that measures milk flow time through a burette to calculate relative viscosity, with a timing error below 0.45 seconds, illustrating the frugal engineering ethos behind the project.

Important caveats remain. The models were calibrated and validated exclusively on laboratory-prepared mixtures, which ensures strict experimental control but says nothing yet about performance on commercial goat milk, where processing and formulation variations could challenge the models. The authors state explicitly that future studies will assess application to commercial samples and that real-world deployment will require further validation, along with extension to other adulterants such as bovine milk proteins and starches. They also flag that the classification result depended partly on the design of the mixture set, which contained far fewer samples not suspected of fraud, and on random sampling by the algorithm, although repeated draws gave essentially the same result.

Even with those limitations, the study offers a promising template for milk authenticity testing. Established methods for detecting adulteration, such as polymerase chain reaction, enzyme-linked immunosorbent assay and high-performance liquid chromatography, are accurate but slower and more demanding, while the FT-NIR plus chemometrics workflow could serve as a rapid screening tool at the dairy gate or in the laboratory. The authors suggest that integration with inline or handheld NIR devices may eventually enable on-site monitoring, though they caution that development and validation of portable instrumentation are still needed. For an industry where a splash of water can quietly erode both value and consumer trust, a technique that quantifies dilution to within roughly two percent, and flags fraudulent samples with nearly 95 percent accuracy, is a meaningful step toward faster, more accessible fraud detection.

Subject of Research: Detection and quantification of water and cow milk adulteration in goat milk using FT-NIR spectroscopy with partial least squares and logistic regression models

Article Title: Detection and quantification of water adulteration in goat milk using FT-NIR spectroscopy combined with partial least squares and logistic regression

Article References: Silveira, P. G., Lima, C. M. G., Shurmasti, D. K., de Paula Silva, C. R., Contado, J. L., Carvalho dos Anjos, V. D., Bell, M. J. V., Verruck, S., & da Rocha, R. A. (2025). Detection and quantification of water adulteration in goat milk using FT-NIR spectroscopy combined with partial least squares and logistic regression. Dairy Science and Management, 2(1), Article 10. https://doi.org/10.1186/s44363-025-00008-8

Image Credits: AI Generated

DOI: 10.1186/s44363-025-00008-8

Keywords: goat milk, adulteration, FT-NIR spectroscopy, partial least squares, logistic regression, food fraud, dairy quality control, chemometrics, ternary mixtures, milk authentication, near-infrared spectroscopy, Brazil