Rapid detection of milk adulteration using AI-driven portable colorimetric spectroscopy

This study presents a non-destructive, artificial intelligence (AI)-assisted spectro-analytical system for rapid milk adulteration detection using a laboratory-developed portable reflectance spectrophotometer operating across 410–940 nm. Six adulterants—urea, formalin, hydrogen peroxide, baking soda, sucrose, and cornstarch—were evaluated individually to establish compositional–spectral relationships within a single-adulterant modelling framework. Physicochemical characterization confirmed adulterant-dependent shifts in solids-not-fat (SNF; 3.29–9.18%), fat (1.5–5.4%), and protein (2.01–3.40%), supporting class-level discrimination. Analysis of variance (ANOVA) revealed active spectral windows at 410–510 nm and 680–940 nm, corresponding to chromophore absorption and near-infrared (NIR) overtone regions. Strong wavelength–property associations were identified, including negative correlations at lower wavelengths (r = − 0.81 to − 0.93) for protein and SNF, and positive correlations at higher wavelengths (r = 0.74–0.89) for carbohydrate-based adulterants. Principal component analysis (PCA) explained 91.6% variance, confirming distinct adulterant-specific clustering. ANOVA-based feature selection isolated three discriminatory spectral zones per adulterant. Classification models—Decision Tree, Logistic Regression, support vector machine (SVM), and ensemble methods—achieved ≥ 99% accuracy with inference times below 0.15 ms, with Decision Tree and Logistic Regression offering optimal edge efficiency. The study demonstrates compact, low-cost, AI-enabled multispectral milk monitoring for real-world deployment.