ATR-FTIR Spectroscopy and Deep Learning for Chemometric Analysis and Geographical Classification of Red Wines

This study leverages attenuated total reflection Fourier-transform infrared (ATR-FTIR) spectroscopy combined with a novel residual composite convolutional neural network (Res-MCNN) to address the critical challenge of geographical origin authentication in red wines. By analyzing 200 samples from distinct Chinese regions (Xinjiang and Hebei), we demonstrate the efficacy of ATR-FTIR in capturing region-specific spectral fingerprints, particularly highlighting the role of titratable acids and phenolic compounds as discriminative markers. The proposed Res-MCNN model integrates advanced preprocessing techniques, including competitive adaptive reweighted sampling (CARS) for wavelength selection and decision tree pooling for feature optimization, achieving an exceptional classification accuracy of 97.975%. Our findings reveal that while alcohol and sugar content variations introduce sensitivity in model performance, the robust extraction of high-dimensional nonlinear spectral features ensures reliable origin traceability when key chemical parameters are controlled. The proposed Res-MCNN model achieved a classification accuracy of 97.975%, a sensitivity of 98.2%, a specificity of 97.6%, and an AUC of 0.989, demonstrating highly reliable geographical discrimination performance. Furthermore, controlled experiments showed that variations in alcohol (10%-15%) and sugar content (1-20 g/L) altered accuracy by no more than 0.6%, confirming that the model maintains stable traceability performance when key chemical parameters remain within typical ranges.