Wine variety traceability by data fusion of near-infrared(NIR) spectroscopy and mid-infrared (MIR) spectroscopy combined with GAF and ResNet

Abstract

Wine varietal authentication is critical for ensuring quality and preventing market fraud. Spectroscopy method is commonly used for rapid and nondestructive analysis of wine varieties, but the method typically rely on one-dimensional(1D) spectrum inputs and conventional machine learning algorithms, limiting complex compositional feature capture and reducing classification accuracy. To address these challenges, this study proposes a novel wine varietal tracing approach integrating NIR and MIR spectroscopy, Gramian Angular Field (GAF) image encoding, and ResNet-based deep learning. The NIR and MIR spectra of 172 wine samples from three grape varieties: Cabernet Sauvignon, Merlot, and Cabernet Gernischt were collected. Spectral fusion and data augmentation techniques constructed fused spectra and expanded the dataset to 1720 samples. Then the 1D NIR, MIR, and fused spectra were transformed into two-dimensional(2D) images using GAF encoding, creating Gramian Angular Difference Field (GADF) and Gramian Angular Summation Field (GASF) representations. These images were fed into a deep residual network (ResNet) for varietal classification respectively. The designed ResNet architecture incorporates residual blocks and attention mechanisms, significantly enhancing feature extraction and classification performance. Experimental results show that the proposed model achieves 100 % classification accuracy on the test set, outperforming traditional machine learning methods and 1D-CNN. The result indicates that the integrating of infrared spectroscopy, GAF image encoding, and ResNet is a feasible approach for wine varietal tracing, providing a novel solution for food quality tracing analysis.