Lightweight residual deep learning framework for urea and formalin adulteration detection in milk using hyperspectral imaging

Milk adulteration is a major concern for dairy quality and food safety, especially when harmful adulterants such as urea and formalin are present at closely related concentration levels that are difficult to distinguish. This study proposes a lightweight Squeeze-and-Excitation Residual 1D Network (SE-ResNet1D) combined with Hyperspectral Imaging (HSI) to detect the presence of urea and formalin in milk at a finer level. A custom hyperspectral milk dataset was developed using a Resonon Pika L system from the 400 nm to 1000 nm wavelength range over 300 bands, including pure milk and closely spaced adulteration levels of urea (0.02–0.125 g/50 mL) and formalin (0.1–0.6 mL/50 mL). The raw spectral data was pre-processed by multiplicative scattering correction and Savitzky–Golay smoothing, followed by Principal Component Analysis (PCA) with PCA-30 as the optimum input data representation. The model achieved accuracies of 98.8% for urea and 98.1% for formalin, outperforming compared models while maintaining computational efficiency with only 17,419 trainable parameters. The framework was evaluated against machine learning and deep learning baseline models, with ablation analysis to assess architectural contributions. The results obtained show that the proposed model could be used as a fast, non-destructive, and feasible solution for intelligent dairy quality monitoring.