Rapid and nondestructive quantitative detection of milk adulteration using hyperspectral imaging and optimized LSTM modeling

Milk adulteration poses persistent challenges to food authenticity, quality assurance, and consumer safety, while conventional analytical methods are often destructive, labor-intensive, and unsuitable for rapid routine screening. This study aimed to develop a rapid and nondestructive framework for the quantitative detection of milk adulteration using visible–near-infrared hyperspectral imaging (400–1000 nm) combined with an optimized long short-term memory model. Four common adulterants, namely starch, sugar, soymilk powder, and salt, were added to pure milk at concentrations ranging from 0.5 to 16 g/100 mL. A total of 440 samples from one commercial brand were used for model development, whereas 880 samples from two independent brands were reserved for external validation to assess inter-brand generalization. After spectral preprocessing and wavelength selection, long short-term memory models optimized by population-based algorithms were established for concentration prediction. Among them, the gray wolf optimization-based model showed the best performance, achieving a coefficient of determination of 0.9904 and a root mean square error of prediction of 0.0369 on the internal test set. In external validation, the model maintained satisfactory predictive accuracy, with coefficients of determination of 0.952 and 0.945 for the two unseen brands, respectively. These findings indicate that hyperspectral imaging combined with optimized deep learning has strong potential for robust quantitative detection of milk adulteration and may provide a useful basis for future dairy quality control. However, further validation under practical industrial conditions is still required before real-time online implementation.