An interpretable CNN-grad-CAM framework for detection and quantification of cow milk adulteration in goat milk using...

Cow milk adulteration in goat milk is difficult to detect because of subtle compositional differences. A rapid, non-destructive method combining excitation-emission matrix (EEM) fluorescence spectroscopy with a convolutional neural network (CNN) was developed for the detection and quantification of cow milk adulteration. The CNN directly processed raw EEM matrices as image-like inputs without prior fluorescence decomposition. For binary classification, it achieved 98.65% accuracy, with precision, recall, specificity, and F1-score exceeding 98%, whereas it outperformed SVM in seven-class classification. For quantitative prediction, CNN yielded marginally better results than SVR, with an RMSEP of 4.84% and an R_P² of 0.91, compared with 4.88% and 0.90 for SVR, respectively. Grad-CAM and PARAFAC showed partial overlap between CNN-discriminative and protein-related regions, indicating that protein signals contributed to, but did not solely determine, the model decisions. This EEM-CNN framework enables rapid, interpretable screening of cow milk adulteration in goat milk under controlled conditions.