A novel microscopic hyperspectral imaging detection method for quantitative powdered food adulteration assessment: A comparative study of CNN-RNN attention models.

Microscopic hyperspectral imaging (MHSI) was employed to acquire particle-resolved hyperspectral information for the quantitative analysis of adulteration in powdered foods. Lotus root powder adulterated with maize starch and cassava starch was used as a model system. To characterize the complex spectral patterns associated with particle scattering, three hybrid deep sequence models, namely CNN-LSTM-Attention, CNN-GRU-Attention, and CNN-BiLSTM-Attention, were developed and systematically compared with conventional models. An optimal analytical workflow was further established by evaluating six spectral preprocessing techniques and four wavelength selection methods. Among the evaluated models, CNN-BiLSTM-Attention exhibited the best predictive performance, achieving an R2 value of 0.9457 and an RMSE value of 1.3871 on the prediction set. When combined with pseudo-color visualisation, the optimized model enabled pixel-wise mapping of predicted adulteration levels across the sample. Within regions exhibiting stable imaging quality, the pseudo-color distribution progressively shifted toward colors corresponding to higher predicted concentrations as the adulteration level increased, allowing the spatial variation in predicted adulterant concentration to be intuitively visualized. These results demonstrate the potential of MHSI coupled with deep learning for quantitative adulteration analysis and provide a promising analytical strategy for the quality control of lotus root powder and other high-value powdered food products.