Spectral-Structural Decoupled Hypergraph Neural Network for Fritillaria Species Identification Using Hyperspectral...

Fritillaria thunbergii Miq. (ZBM) and Fritillaria hupehensis Hsiao et K.C. Hsia (HBBM) are two common medicinal and edible species of the genus Fritillaria, both possessing high economic and pharmacological value. Given their differences in pharmacological effects and economic value, rapid, nondestructive, and accurate identification of these two species is of practical significance. In this study, hyperspectral imaging (HSI) was used for the species identification of sliced Fritillaria samples. Based on the extracted one-dimensional spectra, conventional machine learning models, including Logistic Regression (LR), Support Vector Classification (SVC), and Extreme Gradient Boosting (XGBoost), as well as deep learning models, including one-dimensional Convolutional Neural Network (1D-CNN), one-dimensional Residual Network (1D-ResNet), Transformer, and CNN-Transformer, were first constructed. The one-dimensional spectra were further transformed into two-dimensional Gramian Angular Difference Field (GADF) images, followed by the construction of two-dimensional Convolutional Neural Network (2D-CNN) and two-dimensional Residual Network (2D-ResNet) models. To further explore structural relationships among spectral samples, graph-structured and hypergraph-structured data were constructed from both one-dimensional spectra and GADF images, and the corresponding Graph Convolutional Network (GCN) and Hypergraph Neural Network (HGNN) models were established. On this basis, a Spectral-Structural Decoupled Hypergraph Neural Network (SSD-HGNN) was proposed, in which deep spectral features extracted by 1D-ResNet were used as node attributes, while deep GADF image features extracted by 2D-ResNet were used to construct hyperedges. This design enabled decoupled fusion of one-dimensional spectral information and two-dimensional structural information within a hypergraph learning framework. The results showed that most models achieved satisfactory identification performance, demonstrating the feasibility of HSI for distinguishing ZBM and HBBM slices. SSD-HGNN achieved competitive overall performance, with accuracies of 0.9848, 0.9533, and 0.9441 on the training, validation, and test sets, respectively. Although SSD-HGNN did not achieve the highest test accuracy among all evaluated models, it achieved the highest validation accuracy and effectively integrated discriminative one-dimensional spectral features with two-dimensional GADF-based structural relationships. These results demonstrate that the proposed spectral-structural decoupled hypergraph learning strategy provides a competitive multimodal approach for rapid and nondestructive identification of Fritillaria species.