Machine learning-combined hyperspectral imaging analysis for the non-destructive identification of wheat flours with...

This study aimed to non-destructively identify wheat flours with different gluten strengths through the application of machine learning-combined hyperspectral imaging analysis. The performance of this approach was compared to conventional instrumental methods, specifically Mixolab, Fourier Transform Infrared (FTIR) spectroscopy, Rapid Visco-analyzer (RVA). The Mixolab measurement demonstrated that high values of mixing properties (water absorption, dough stability/development time) were observed in the order of the strong, medium, and weak flour samples. FTIR revealed the strong wheat flour with elevated peak intensities in the ranges of 1500-1700 and 2800-3500 cm^-1, while the weak wheat had dominant peaks in the range of 800-1500 cm^-1. The weak flour exhibited lower hyperspectral signal intensity, while the strong flour showed slightly higher intensity beyond 1395 nm. Linear discriminant analysis captured a significant dataset portion with two discriminants (≥99 %). Three machine learning models (decision tree, random forest, and k-nearest neighbor) achieved superb wheat flour classification with Mixolab features. For FTIR, RVA, and hyperspectral results, dimension-reduced datasets were generally more effective in classification than the original results. Furthermore, these prediction capabilities were verified by validating the models with independent datasets. This investigation demonstrates the potential of machine learning-combined hyperspectral analysis for the non-destructive identification of wheat flours, determining their suitability for specific food applications.