Adaptive conditional diffusion-based data augmentation for NIR spectroscopy quantitative analysis under limited...

Accurate quantification of key chemical constituents is critical for quality control in food and pharmaceutical products, yet conventional analytical methods are often destructive, time-consuming, and costly. Combining near-infrared (NIR) spectroscopy with deep learning shows strong potential for quantitative regression. However, labeled spectra typically require expensive, labor-intensive reference assays, resulting in limited training data that constrain model training, generalization, and real-world deployment. Although data augmentation can mitigate small-sample issues, existing generative approaches often struggle with stable synthesis and precise conditional control under continuous regression labels, which may cause spectrum-label inconsistency and limit downstream gains. To address these challenges, we propose an Adaptive Conditional Diffusion Model (ACDM) for NIR spectral generation with continuous regression conditioning, designed for regression-oriented augmentation. ACDM integrates robust conditioning via classifier-free guidance and Feature-wise Linear Modulation (FiLM), label-guided adaptive noising through a condition-guided noise injection network, spectral feature recalibration using Squeeze-and-Excitation modules to preserve fine-grained absorption structures, and an improved one-dimensional U-Net tailored to NIR signals. Experiments on three representative NIR regression datasets (American ginseng, corn kernel, and tablet samples) show that ACDM generates spectra with high distributional consistency and spectral-shape fidelity. More importantly, ACDM-based augmentation consistently improves regression performance. Under the highest-performing predefined augmentation setting, ResNet1D achieves R² values of 0.9462, 0.9370, and 0.9311 on the three datasets, with low prediction errors and stable residual behavior. These results demonstrate that ACDM effectively enhances NIR regression under limited labeled data in the three evaluated sample systems and supports its potential for regression-oriented spectral augmentation.