Abstract: The wavelength variable selection method for infrared spectroscopy primarily employed multiple iterations to screen wavelength variables with significant model contribution from the full-spectrum data. This approach exhibited poor resistance to redundant information interference, making it difficult to accurately reflect the true characteristics of the sample. To address this, a wavelength variable selection method based on the SPA algorithm was proposed for near-infrared spectroscopy of vine tea-flavored beer. Multiscattering correction was applied to the near-infrared spectral data to enhance absorption information correlated with component content and reduce noise. The continuous projection algorithm was used for preliminary wavelength screening. Combined with correlation and collinearity penalties for the target components in wisteria tea flavor detection, this approach selected a wavelength subset with low information redundancy and strong predictive capability. Combined entropy and mutual information were used to assess wavelength redundancy, constructing a scoring function incorporating information content and redundancy penalties to select the subset with the highest overall score. The results showed that after wavelength variable optimization using this method, spectral absorption coverage remained stable at approximately 95%, indicating highly satisfactory optimization outcomes.

Key words: SPA algorithm, vine tea-flavored beer, near-infrared spectroscopy, wavelength variables, feature selection