Explainable hyperspectral authentication of Pinellia ternata against Pinellia pedatisecta across tuber, slice, and powder forms with GC–TOF–MS corroboration

Technical Summary: Explainable Hyperspectral Authentication of Pinellia ternata against Pinellia pedatisecta

Problem Statement
The authentication of Pinellia ternata (the source of the medicinal material Pinelliae Rhizoma or Banxia) is critical due to the prevalence of adulteration by the morphologically similar Pinellia pedatisecta. The latter contains higher levels of toxic lectins and differs in therapeutic indications, posing safety risks. Current authentication methods face three primary limitations:

1. Preprocessing Dependence: Conventional hyperspectral imaging (HSI) workflows rely on fixed, manually selected preprocessing steps (e.g., Savitzky–Golay derivatives), which are not jointly optimized with the classification model and require significant domain expertise.
2. Single-Form Validation: Most studies evaluate models on a single physical form (e.g., only tubers), failing to account for the variability introduced when medicinal materials are processed into slices or powders, which alters morphological cues and light scattering properties.
3. Lack of Chemical Corroboration: Explainable AI (XAI) methods often identify discriminative wavelengths without validating them against independent chemical evidence, leaving the biological plausibility of model decisions unverified.

Methodology
The study proposes HyGD-Net (Hyperspectral Gated Derivative Network), an end-to-end deep learning framework designed to address these limitations. The workflow involves:

• Data Acquisition: Hyperspectral images were acquired in Visible and Near-Infrared (VNIR; 400–1000 nm) and Short-Wave Infrared (SWIR; 900–1700 nm) ranges for P. ternata and P. pedatisecta in three forms: intact tubers, slices, and powders.
• HyGD-Net Architecture:
  • Learnable Filter Module: Instead of fixed preprocessing, the network incorporates a learnable Savitzky–Golay (SG) derivative filter initialized with classical SG coefficients. This layer adaptively optimizes first- and second-order derivative transformations during training.
  • Triple-Branch CNN: Three parallel convolutional branches process the raw spectrum, the first-order derivative, and the second-order derivative independently to extract representation-specific features.
  • Position-Wise Gated Fusion: A gating mechanism adaptively modulates the contribution of each branch at every spectral position, allowing the model to emphasize informative representations dynamically.
• Experimental Design: Seven scenarios were tested: three "in-form" (single form), three "mixed-form" (two forms combined), and one "all-form" (all three forms combined) to evaluate robustness against physical form variation.
• Explainability and Validation: Integrated Gradients (IG) were used to identify consensus discriminative wavelengths across models and scenarios. These wavelengths were cross-validated against volatile organic compound (VOC) profiles obtained via Gas Chromatography–Time-of-Flight Mass Spectrometry (GC–TOF–MS).

Key Results

• Classification Performance:
  • SWIR Range: HyGD-Net achieved 100.00% accuracy, sensitivity, and specificity across all seven experimental scenarios. ResNet also achieved 100.00%, while other models (CNN, EfficientNetV2-S, Swin Transformer, SVM) performed near-ceiling levels (≥98.88%).
  • VNIR Range: HyGD-Net achieved the highest mean accuracy (94.46%) with the lowest cross-scenario variability (standard deviation of 2.81 percentage points). It outperformed all benchmarks, including CNN (91.19% mean) and ResNet (90.51% mean). Notably, in the challenging "all-form" scenario, HyGD-Net maintained 92.76% accuracy, whereas ResNet dropped to 90.51% and EfficientNetV2-S to 89.87%.
• Ablation Analysis:
  • Removing the learnable SG filter (replacing it with fixed weights) reduced VNIR accuracy by up to 3.16 percentage points in specific scenarios.
  • Replacing the triple-branch architecture with a single branch caused the largest performance drop, with a mean decrease of 6.30 percentage points in VNIR, highlighting the necessity of processing raw and derivative spectra independently.
  • Replacing gated fusion with simple concatenation resulted in accuracy reductions of 0.23–4.09 percentage points in VNIR.
• Chemical Corroboration: IG analysis identified recurrent consensus wavelengths in the 400–500 nm, 900–1000 nm, and 1100–1200 nm ranges. GC–TOF–MS analysis revealed distinct volatile profiles between the species, including species-specific aldehydes and acids. The identified wavelengths corresponded plausibly to functional group absorptions (e.g., O–H overtones near 950–1000 nm and C–H overtones near 1157 nm) associated with these chemical differences.

Significance and Claims
The paper claims that HyGD-Net provides a robust, nondestructive authentication framework that eliminates the need for manual preprocessing selection by internalizing derivative transformations. The study demonstrates that:

1. Form Robustness: The model maintains high accuracy across diverse physical forms (tuber, slice, powder) and mixed-form scenarios, addressing a gap in existing literature where models are often validated on single forms.
2. Joint Optimization: By learning spectral transformations end-to-end, the model adapts to task-specific data while retaining the mathematical grounding of spectroscopic priors.
3. Explainability with Chemical Context: The integration of XAI with independent chemical profiling (GC–TOF–MS) moves beyond simple band identification, offering a scientifically defensible basis for the model's decisions by linking spectral features to functional-group-level chemical differences.

The authors conclude that this framework offers a practical quality-control tool for Pinelliae Rhizoma authentication, capable of handling the variability inherent in post-harvest processing while providing chemically corroborated explanations for its predictions.