Nondestructive classification of native starches by using machine learning-assisted hyperspectral imaging: effects of particle size and moisture content

In this study, a framework has been presented for classifying native starches by using hyperspectral imaging in combination with machine learning, to improve the classification performance of machine learning models under controlled preprocessing conditions such as particle size and moisture content. Hyperspectral analysis revealed a substantial overlap among the different types of native starches, indicating limitations in the classification on the basis of raw hyperspectral data. The starch classification performance of the machine learning model varied depending on the particle size and moisture content of the starch. The highest classification accuracy (~97%) was achieved at a particle size of 100 mesh and moisture content of 50%, and particle aggregation was effectively suppressed to maintain a uniform dispersion state. These results were confirmed using microstructural analysis of the starches on the basis of their scanning microscopic images and particle size distributions. Overall, the proposed preprocessing approach demonstrates the strong potential of machine learning-combined hyperspectral technology as a rapid and nondestructive analytical method for identifying native starches in the food industry.