Quantitative analysis of corn adulteration in sweet potato starch using a CNN-LSTM hybrid model

Starch, as a primary storage carbohydrate, serves as a key energy source in the human diet. It is often difficult to distinguish between different starch types visually due to their similar appearance. Traditional starch detection methods suffer from several limitations like low accuracy, complex pretreatment processes, and sample damage. Therefore, to achieve a rapid and accurate quantification of starch adulteration, terahertz time-domain spectroscopy (THz-TDS) was utilized. First, time-domain spectral data were obtained from a series of adulterated samples. Then, four pretreatment methods were applied to process the original spectra. Both the full spectra and SPA-selected spectral features were used as inputs to develop traditional regression models, and a deep learning model based on a hybrid convolutional neural network and long short-term memory network (CNN-LSTM) model. The results demonstrate that the CNN-LSTM model achieved the best performance (Rp = 0.9765, RMSEP = 4.85 %), indicating its potential for rapid and non-destructive monitoring of starch quality.