SAGT-Net: A multi-scale Transformer embedded with GCN based on FT-NIR for Vitis 'Shine Muscat' maturity...

Vitis 'Shine Muscat' remains green when fully ripe, making visual maturity identification unreliable. Traditional judgment methods depend on growers' experience, which is subjective and error-prone. This study proposes the Spectral Adaptive Graph Transformer Network (SAGT-Net), a multi-scale Transformer combined with Graph Convolutional Network (GCN), to achieve rapid and non-destructive maturity classification of Vitis 'Shine Muscat' using Fourier Transform Near-Infrared (FT-NIR) spectroscopy. First-derivative Savitzky-Golay (SG-1st) smoothing was selected as the optimal spectral preprocessing method. After dataset partitioning and data augmentation, the model integrates one-dimensional convolutional neural network (1D-CNN), spectral attention and adaptive multi-scale GCN. It achieves a test accuracy of 95.16% and F1-score of 95.15%. Key spectral bands were further identified via t-SNE and SHAP analyses. The results show that the SAGT-Net model achieves high-precision maturity classification of Vitis 'Shine Muscat', providing a methodological framework and technical support for near-infrared nondestructive testing of this cultivar.