Research on rapid and non-destructive detection of adulterated wheat flour using hyperspectral imaging and machine learning

Detection of adulterated wheat flour is crucial for public health. This study aims to establish a detection model based on hyperspectral imaging and machine learning for identifying wheat flour adulteration. Hyperspectral images of wheat flour with different degrees of mold and germination were collected by a hyperspectral camera. The raw spectral data were preprocessed with various methods to enhance quality. Uninformative Variable Elimination (UVE), Competitive Adaptive Reweighted Sampling (CARS), and Successive Projections Algorithm (SPA) were utilized to select feature spectral bands. Random Forest Regression (RFR), Gradient Boosting Regression (XGBR), and other three models were used to establish optimized quantitative analysis models. The results showed that, for moldy wheat flour, all models achieved R² values above 0.95. In particular, the combination of RFR and XGBR yielded R² values exceeding 0.99 and root mean square error (RMSE) values below 0.02. The performance for germinated wheat flour detection was slightly lower, but some models still achieved R² values above 0.95. Furthermore, spectral visualization analysis was performed to generate intuitive heat maps, illustrating the spatial distribution of inferior components in the samples. This study confirms the feasibility of using hyperspectral imaging and machine learning for rapid, non-destructive detection of adulterated wheat flour.