Non-destructive classification of milk types using visible-near infrared spectroscopy and machine learning

초록

Milk authentication is important for ensuring product quality, consumer confidence, and regulatory compliance in the dairy industry. This study developed a non-destructive milk classification framework by integrating Visible-Near Infrared (Vis-NIR) spectroscopy with machine learning to differentiate cow milk, goat milk, and reconstituted milk. A total of 3000 spectra (1000 per class) were collected and divided into training (n = 2400) and independent test (n = 600) datasets using stratified sampling. Four preprocessing pipelines and ten machine learning models were systematically evaluated. Spectral analysis revealed clear class-dependent differences, particularly within the 600-1000 nm region, corresponding to absorption features associated with water, lipids, and proteins. Principal Component Analysis (PCA) demonstrated distinct clustering among milk classes, indicating strong spectral separability. Cross-validation performed on the training dataset showed excellent performance, with XGBoost achieving the highest mean accuracy (99.35%), followed by LightGBM (99.20%) and Gradient Boosting (99.00%). Independent test evaluation further confirmed strong generalization capability, where XGBoost achieved the highest accuracy (99.17%). Overall, the results demonstrate the potential of Vis-NIR spectroscopy combined with machine learning for rapid and reliable milk authentication.

키워드