NIRDiffusion: A diffusion-model-based framework for enhanced quality assessment of industrial plant materials
Yang Yu, Siqi Wang, Jinlong Duan, Wei Zhang, Qifu Wang, Dandan Zhai, Qin Yao, Zhiqing Yang, Peng Li
Abstract
Soybean and American ginseng, as industrial plant-derived materials, require precise quantification of soluble proteins and ginsenosides due to their importance in industrial applications such as bio-based materials and pharmaceutical ingredient production. Near-infrared (NIR) spectroscopy, combined with advanced AI models, offers a rapid, non-invasive, and cost-effective approach for evaluating the quality of industrial plant-derived materials. However, its application is limited by challenges such as high spectral dimensionality, environmental noise, and small labeled datasets. To address these issues, this study proposes a diffusion-model-based representation learning framework, NIRDiffusion, which utilizes a probabilistic, iterative process of perturbation and reversal to capture multi-scale non-linear spectral latent features while effectively suppressing noise. The proposed NIRDiffusion combined with a 1D-CNN regression model achieves R 2 values of 0.973 and 0.949 in predicting water-soluble protein (WSP) in soybean seeds and total ginsenosides (TG) in American ginseng, respectively. Comparative analyses reveal significant prediction error reductions of at least 17.89 % (WSP) and 11.21 % (TG) relative to popular unsupervised feature extraction and three self-supervised learning approaches. These results demonstrate that NIRDiffusion effectively captures discriminative features in NIR spectral data, enabling accurate and scalable quality assessment of plant-derived materials through integration with predictive modeling.