Noninvasive multiclass milk contaminants detection using hyperspectral imaging and hybrid ensemble learning.

Abstract

Food contamination remains a serious global concern due to its health risks, with milk being one of the most commonly adulterated foods in developing countries such as Pakistan, India, and Bangladesh. Accurate detection of milk contamination is essential for ensuring consumer safety and maintaining food industry standards. This study explores both invasive and noninvasive approaches for contamination analysis. The invasive method uses the Lactoscan system to assess parameters such as fat, conductivity, protein, density, solids, lactose, temperature, pH, and SNF across varying contamination levels. The noninvasive method employs hyperspectral imaging using the Specim FX-10 system (400-1,000 nm) to detect contamination through spectral and spatial analysis. Preprocessing involved image resizing and region of interest selection for feature extraction, as well as radiometric correction using the empirical line method. Postprocessing included noise reduction and spectral smoothing using the Savitzky-Golay filter. The resulting clean spectral data were classified using a hybrid ensemble learning (HEL) framework, which combines voting and stacking ensembles of gradient boosting, XGBoost, LightGBM, and multilayer perceptron models. Comparative results show the HEL approach significantly outperforms existing methods, achieving 100% training and 96% validation accuracy-demonstrating its potential for real-time, noninvasive milk quality assurance.