Interpretable Spectral Evidence Learning from Vis/NIR Imaging for Non-Destructive Authentication of Herbal Medicines

Area of Science:

• Agricultural Science
• Analytical Chemistry
• Computer Science

Background:

• Ensuring the quality and authenticity of herbal medicines is crucial for public health and regulatory compliance.
• Visible and near-infrared (Vis/NIR) imaging offers a rapid, non-destructive method for analyzing herbal materials.
• Existing authentication methods may lack interpretability or require extensive spectral data.

Purpose of the Study:

• To establish an interpretable spectral evidence learning framework for Vis/NIR imaging-based authentication of Codonopsis Radix (CR) and Aurantii Fructus (AF).
• To evaluate the effectiveness of different spectral preprocessing techniques and machine/deep learning classifiers.
• To introduce and assess a lightweight diffusion (LD) module for spectral augmentation and denoising.

Main Methods:

• Collected compact 31-band Vis/NIR spectra at region of interest (ROI) and sample levels for CR and AF.
• Compared spectral preprocessing methods (smoothing, MSC, SNV) with Linear SVM, and deep learning classifiers.
• Implemented a fold-contained lightweight diffusion (LD) module for class-conditioned spectral augmentation and denoising.

Main Results:

• Linear SVM models without LD achieved high performance (e.g., macro-F1 of 0.9238 for CR, 0.9018 for AF) under grouped cross-validation.
• LD augmentation significantly improved model performance, with LD-augmented SVMs reaching macro-F1 values of 0.9427 (CR) and 0.9197 (AF).
• LD increased the overall mean macro-F1 from 0.7302 to 0.8189 across various model-dataset combinations, and identified key spectral bands for authentication.

Conclusions:

• The developed interpretable spectral evidence learning framework is feasible for compact Vis/NIR imaging-based authentication of herbal materials.
• The lightweight diffusion (LD) module enhances authentication accuracy and provides spectral augmentation and denoising capabilities.
• Selected spectral bands within the Vis/NIR range contain sufficient discriminative information for effective herbal material authentication using this imaging approach.