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

Area of Science:

• * Botany and Pharmacognosy: Focuses on the identification of medicinal plant species.
• * Spectroscopy and Imaging: Utilizes hyperspectral imaging for material analysis.
• * Machine Learning and Artificial Intelligence: Applies advanced algorithms for classification and identification tasks.

Background:

• * Two Fritillaria species, Fritillaria thunbergii (ZBM) and Fritillaria hupehensis (HBBM), are valuable medicinal and edible plants.
• * Accurate and rapid identification is crucial due to differences in pharmacological effects and economic value.
• * Existing identification methods may be destructive or lack efficiency.

Purpose of the Study:

• * To develop a rapid, nondestructive method for distinguishing between ZBM and HBBM slices.
• * To evaluate the effectiveness of hyperspectral imaging (HSI) combined with various machine learning models for species identification.
• * To propose and validate a novel Spectral-Structural Decoupled Hypergraph Neural Network (SSD-HGNN) for enhanced identification accuracy.

Main Methods:

• * Hyperspectral imaging (HSI) was employed to acquire spectral data from sliced Fritillaria samples.
• * One-dimensional (1D) spectral data were analyzed using conventional (LR, SVC, XGBoost) and deep learning (1D-CNN, 1D-ResNet, Transformer) models.
• * Spectra were transformed into 2D Gramian Angular Difference Field (GADF) images, analyzed with 2D-CNN and 2D-ResNet.
• * Graph (GCN) and hypergraph (HGNN) models were constructed using spectral and GADF image data.
• * A novel SSD-HGNN model was developed, integrating 1D-ResNet spectral features and 2D-ResNet GADF image features within a hypergraph framework.

Main Results:

• * Most evaluated models demonstrated satisfactory performance in distinguishing ZBM and HBBM slices, confirming HSI feasibility.
• * The proposed SSD-HGNN achieved high accuracy: 0.9848 (training), 0.9533 (validation), and 0.9441 (test).
• * SSD-HGNN achieved the highest validation accuracy and effectively fused spectral and structural information, outperforming other models in integrated performance.

Conclusions:

• * Hyperspectral imaging is a viable technique for the rapid and nondestructive identification of Fritillaria species.
• * The developed Spectral-Structural Decoupled Hypergraph Neural Network (SSD-HGNN) offers a competitive multimodal approach for species identification.
• * The SSD-HGNN effectively combines spectral and structural information, paving the way for advanced plant identification systems.