Hyperspectral imaging-driven detection of neutraliser adulteration in milk using spatial-spectral features

The excessive use of neutralisers such as sodium carbonate and bicarbonate in milk poses serious health risks, violates food safety regulations, and remains a growing concern in the dairy industry. This highlights the need for non-destructive and reliable methods to ensure milk quality and consumer safety. This study aims to detect and classify neutralisers in milk using Hyperspectral Imaging (HSI) integrated with Machine Learning (ML). A hybrid approach combining novel spectral indices and spatial derivative features is proposed to enhance detection accuracy. Fresh and adulterated milk samples were prepared by adding 0.025 g, 0.05 g, and 0.075 g of neutralisers per 100 ml of milk. Spectral data were acquired using a Resonon Pika-L hyperspectral camera operating in the 400-1000 nm range with 300 spectral bands. Novel spectral indices were developed from the reflectance profiles of pure neutralisers (650-975 nm). These indices, along with second-order spatial derivative features, were analysed using ML models including Extreme Gradient Boosting (XGBoost) and Random Forest (RF). The XGBoost and RF models achieved accuracies of 99.6% and 99.0%, respectively, outperforming both the conventional method and deep learning models. The ML based hybrid feature approach effectively discriminated between adulterated and pure samples across varying concentrations.