Abstract
<title>Abstract</title> <p> <bold>Background</bold> <italic>Pinellia ternata</italic> is the botanical source of <italic>Pinelliae Rhizoma</italic> (Banxia), a medicinal material widely used in Chinese medicine. The morphologically similar adulterant <italic>P. pedatisecta</italic> reportedly contains higher levels of toxic lectins and differs in therapeutic indications and processing requirements, raising safety and efficacy concerns. We developed an explainable hyperspectral framework for nondestructive authentication across physical forms. <bold>Methods</bold> Visible and near-infrared (VNIR; 400–1000 nm) and short-wave infrared (SWIR; 900–1700 nm) hyperspectral images were acquired from both species in tuber, slice, and powder forms. HyGD-Net integrates Savitzky–Golay-initialized learnable derivative filters, three convolutional branches, and position-wise gated fusion. It was compared with six classifiers across seven multi-form scenarios. Integrated Gradients identified recurrent wavebands, and GC–TOF–MS profiling assessed their chemical plausibility. <bold>Results</bold> HyGD-Net achieved VNIR accuracies of 91.55–100.00%, with a mean of 94.46% and the lowest cross-scenario variability. In SWIR, HyGD-Net and ResNet achieved 100.00% accuracy in all scenarios, while other models also approached ceiling performance. Ablation analysis showed a mean VNIR gain of 6.30 percentage points from the triple-branch architecture and gains of 0.23–4.09 percentage points from gated fusion. Integrated Gradients identified recurrent regions within 400–500, 900–1000, and 1100–1200 nm. GC–TOF–MS revealed distinct volatile profiles and provided functional-group-level context for these regions. <bold>Conclusions</bold> HyGD-Net enables nondestructive authentication across physical forms without manually selected derivative preprocessing. By combining learnable spectral transformation with chemically corroborated attribution, it provides accurate and explainable discrimination of <italic>P. ternata</italic> from <italic>P. pedatisecta</italic> . These findings demonstrate its potential as a practical quality-control tool for <italic>Pinelliae Rhizoma</italic> authentication. </p>