Enhancing authentication specialty milks by near-infrared spectroscopy combined with a convolutional neural network (CNN)

Specialty milks (camel, donkey, and horse milk) have high nutritional value; however, they face issues with adulteration by budget-friendly cow milk and mislabeling their proportions. In this study, Near-infrared (NIR) spectroscopy, combined with machine learning, was utilized to differentiate between raw cow milk and specialty milks, as well as to quantify the levels of adulteration in mixtures of specialty milk and cow milk. The variable selection methods integrated with classifiers, including partial least squares-discriminant analysis (PLS-DA), support vector machine (SVM), k-nearest neighbors (KNN), and decision tree (DT), achieved perfect (100 %) classification accuracy. Moreover, the Convolutional Neural Network (CNN) model not only delivered equally high performance but also provided intuitive visualizations, via Gradient-weighted Class Activation Mapping (Grad-CAM), of the distinct spectral regions driving its decisions. For adulteration analysis, compared with the partial least squares regression (PLSR) models, the CNN model demonstrated superior performance in predicting the levels of specialty milks adulteration with cow milk across a concentration range of 2–80 % ( R²pre ≥ 0.9995 and RPD ≥ 47.91). These findings suggest that NIR spectroscopy combined with CNN machine learning provides an efficient solution for addressing classification and adulteration issues in raw specialty milk, thereby facilitating its quality control in the market. • Discrimination of specialty milks by NIR spectra combined with CNN models. • Grad-CAM revealed important spectral variables in different milk types. • The CNN model quantified cow milk adulteration (2–80 %) in specialty milks.