Discrimination of wheat gluten quality utilizing terahertz time-domain spectroscopy (THz-TDS)

[Display omitted] •Terahertz time-domain spectroscopy combined with machine learning algorithms for wheat gluten identification, the comprehensive discrimination accuracy of wheat glunten by the SSA-SVM model reached 100%.Present an efficient, stable, and accurate approach and exemplary method for identifying wheat gluten.•Adopting terahertz spectroscopy technology for characteristic identification of crops such as wheat holds practical significance.•Competitive Adaptive Reweighted Sampling (CARS) was used to select the frequency variables of terahertz spectrum, which reduces the amount of spectral data.•Analysis and compare the results of the detection models of the terahertz spectral data after processing. Wheat is an important food crop in the world, and wheat gluten quality is one of the important standards for judging the use of wheat. In this study, a combination of chemometric and machine learning methods based on THz-TDS were used to identify three different gluten wheats (high gluten, medium gluten, and low gluten). After collecting the time-domain spectral information of the samples, the frequency-domain spectra, refractive index spectra and absorption coefficient spectra of the samples were obtained by calculating the optical parameters. The experimental results showed that there were differences in the refractive indices and absorption coefficients of wheat with different gluten levels. More importantly the differences in refractive index spectra were more significant. The Competitive Adaptive Reweighted Sampling (CARS) method was applied to select characteristic frequencies from the refractive index spectra within the frequency range of 0.1 to 1.5 THz, to establish a discrimination model for wheat gluten strength. We analysed and compared four discriminative models of Support Vector Machines (SVM), Back Propagation Neural Networks (BPNN), Improved Convolutional Neural Networks (Improved CNN) and Sparrow Algorithm Optimised Support Vector Machines (SSA-SVM). The final results indicated that the SSA-SVM model demonstrated the optimal discrimination performance, achieving an accuracy rate of 100% as reflected in the confusion matrix. In summary, this study provides an efficient, accurate, and non-destructive discrimination method for wheat gluten strength, offering a theoretical basis for differentiating wheat with varying gluten strengths in production processes. It holds practical significance for industrial production reference.