Abstract: In order to select the optimal near infrared wavelengths for the determination of wheat flour protein, a novel attenuation elimination-binary dragonfly algorithm (AE-BDA) was proposed based on a combination and modification of single-binary dragonfly algorithm (single-BDA) with exponential and linear attenuation functions. Single-BDA and AE-BDA were separately employed to select the characteristic wavelengths for proteins from the near infrared spectra of 160 wheat flour samples. A quantitative prediction model for wheat flour protein content was established by partial least square regression (PLSR) and used to evaluate the results of wavelength selection. The results indicated that compared with single-BDA, fewer but more stable wavelengths were selected using AE-BDA and the established model had better prediction performance with a determination coefficient of 0.972 7 and a root mean square error of prediction (RMSEP) of 0.281 1. The average number of characteristic wavelengths selected from 8 experiments using AE-BDA was 15.8, accounting for 12.6% of the original wavelengths, of which 3 wavelengths were selected in each experiment. According to the analysis of the near-infrared spectra, the selected wavelengths were contained in the major absorption bands of wheat flour proteins and background components. In conclusion, AE-BDA can select the few characteristic wavelengths from near-infrared spectra of wheat flour with high computational efficiency, giving a predictive model with higher accuracy and stability. The proposed method can provide a simpler and more effective wavelength optimization strategy for near-infrared modelling.

Key words: attenuation elimination-binary dragonfly algorithm; protein; near-infrared spectroscopy; wavelength selection