Abstract: This study proposed an analytical method based on near-infrared spectroscopy (NIR) and one-dimensional convolutional neural network (1D-CNN) for the identification and quantification of adulterated sweet potato starch. For the qualitative and quantitative analysis of adulterants in sweet potato starch, NIR spectra of pure and adulterated sweet potato starch with different levels of corn, potato and cassava starch at 10% intervals were preprocessed by different methods including first-order derivative (1st), continuous wavelet transform (CWT), multiplicative scatter correction (MSC), and standard normal variate transformation (SNV). Both raw and preprocessed spectra were utilized as input signals for 1D-CNN to establish classification models for starch types and prediction models for sweet potato content. The performance of the 1D-CNN models was systematically compared with that of traditional partial least squares (PLS) models. The results demonstrated that these spectral preprocessing methods improved the accuracy of classification and quantitative models to varying degrees, 1st and CWT being more effective than MSC and SNV. For classification models, the prediction accuracy of 1D-CNN was higher than that of PLS. For the prediction set, the prediction accuracy of 1D-CNN based on preprocessed spectra for starch types was 100%. Both PLS and 1D-CNN quantitative models could accurately predict the content of sweet potato starch in samples with single adulterants with similar coefficient of determination for prediction set (Rp2) and root mean squared error of prediction (RMSEP). Compared with the PLS method, the 1D-CNN method was more effective in classification than in quantitation. This study shows that the combination of NIR, 1D-CNN and PLS allows identification and quantification of adulterated sweet potato starch, having practical significance for quality and safety screening of starch adulteration.

Key words: near-infrared spectroscopy; sweet potato starch; adulterated starch; convolutional neural network; partial least squares