A method for food origin identification and ingredient content prediction based on S-transform and multi-task deep learning

Lei Shi, Zetong Zhang, Yu Yang, Dandan Zhai, Peng Li,
A method for food origin identification and ingredient content prediction based on S-transform and multi-task deep learning,
Journal of Food Composition and Analysis,
Volume 147,
2025,
108123,
ISSN 0889-1575,
https://doi.org/10.1016/j.jfca.2025.108123.
(https://www.sciencedirect.com/science/article/pii/S088915752500938X)
Abstract: As the demand for efficient, non-destructive food quality assessment grows, near-infrared (NIR) spectroscopy has shown promise in evaluating food origin and ingredient content, though challenges remain in performing both tasks simultaneously. This study proposes FFMNet, a multi-task deep learning model, utilizing the S-transform to convert NIR data into spectral time-frequency diagrams as input. Through its three core networks—Feature Capture, Feature Interaction, and Multitasking Header Networks—FFMNet achieves outstanding performance. On soybean and wheat flour datasets, it attains 99.09 % and 98.71 % accuracy for origin identification, with R² values of 0.9382 and 0.9885 for water-soluble protein content prediction and protein content prediction, respectively, outperforming traditional models. Grad-CAM visualization confirms its focus on key spectral bands. These results demonstrate FFMNet’s potential to enhance food origin and quality analysis, offering valuable applications in the food industry.
Keywords: NIR spectroscopy; S-transform; Multi-task; Deep learning; Grad-CAM