Combining comparative spectroscopic analysis with machine learning for origin identification and wheat flour blending ratio prediction in buckwheat flour

Identifying the origin of buckwheat flour, detecting wheat flour contamination, and intentionally adulterating it are important for ensuring food authenticity and quality assurance. In this study, excitation–emission matrix (EEM) fluorescence spectra, near-infrared (NIR) spectra, and Fourier transform infrared (FT-IR) spectra combined with machine learning were used within the same experimental system. Machine learning techniques were applied to identify the origin of buckwheat flour, perform product identification, and predict wheat flour mixing ratios in mixed buckwheat and soft wheat flour samples. Samples with wheat flour mixing ratios of 0, 5, 10, and 20% were prepared using Hokkaido-produced buckwheat and soft wheat flours. A k-nearest neighbor (KNN) model using NIR spectra as input demonstrated a test accuracy of 0.9917 for origin identification. For product identification, a convolutional neural network (CNN) model using FT-IR spectra as input showed the highest accuracy. In predicting the flour mixing ratio, models using FT-IR spectra achieved high classification accuracy and strong regression performance. The absorption band around 1650–1700 nm, associated with C–H overtone vibrations in organic components such as lipids and carbohydrates, contributed to identification and quantification in NIR. In contrast, the wavenumber band around 1000 cm⁻¹, related to carbohydrate-associated vibrations such as C–O and C–C stretching, strongly contributed to identification and quantification in FT-IR. These findings suggest that the fusion of spectroscopic analysis and machine learning has potential as a rapid and nondestructive quality assurance method for buckwheat flour products.