Near-Infrared Spectroscopy-Based Maturity Classification of Wine Grapes Using the SG-SFLA-Transformer Model

Abstract

A novel approach combining multiple preprocessing techniques, feature extraction algorithms, and a Transformer-based classification model was proposed to improve the classification accuracy of wine grape maturity. A total of 360 Cabernet Sauvignon samples were collected at four ripening stages. Spectral data covering characteristic NIR regions of 1100 to 1400 nm, 1770 to 1857 nm and 1910 to 2050 nm were analyzed along with physicochemical data. Eight preprocessing methods were applied to enhance full-spectrum data quality and reduce redundancy through wavelength selection. A spectral clustering algorithm was used to classify maturity stages, and spectral errors were optimized with functional group correction. The optimal classification method, combining Savitzky-Golay smoothing and the shuffled frog-leaping algorithm (SG-SFLA), followed by a Transformer model, achieved an accuracy of 94.44%. This method offers a reliable solution for determining grape ripeness.