Near-infrared spectroscopy combined with multi-source feature fusion and transformer for identifying the extent of sulfur fumigation in dried ginger.

The widespread use of sulfur fumigation in dried ginger processing poses significant quality and safety concerns, creating a pressing need for rapid, non-destructive discrimination techniques. While Near-Infrared (NIR) spectroscopy holds promise, effectively extracting subtle, fumigation-induced spectral features remains a challenge. The primary contribution of this study is the development of a novel multi-branch, multi-scale feature extraction framework that synergistically combines 1D-Convolutional Neural Networks (1D-CNN) and a Transformer model. This architecture processes raw and preprocessed NIR spectra in parallel, employing a hybrid strategy that fuses statistically significant features (identified by t-test filtering) with automatically learned deep features (from 1D-CNN). The integrated feature set is then modeled by the Transformer to capture global dependencies for highly accurate classification. Our model achieved exceptional performance, with 95.24% identification accuracy and average precision, recall, and F1-scores all exceeding 95%, significantly outperforming conventional methods. This work not only provides a robust and efficient solution for quality control of dried ginger but also establishes a new, transferable paradigm for feature extraction from spectroscopic data, with broad potential applications in pharmaceuticals, food, and agriculture.