Acta Photonica Sinica, Volume. 54, Issue 4, 0410002(2025)

Hyperspectral Image Classification Method Based on Dynamic Graph-spectral Feature Extraction

Chenjie XU1...2, Dan LI1,2,* and Fanqiang KONG2 |Show fewer author(s)
Author Affiliations
  • 1Key Laboratory of Space Photoelectric Detection and Perception Ministry of Industry and Information Technology,Nanjing University of Aeronautics and Astronautics,Nanjing 210016,China
  • 2College of Astronautics,Nanjing University of Aeronautics and Astronautics,Nanjing 210016,China
  • show less

    Hyperspectral remote sensing integrates spatial and spectral imaging technologies to capture continuous spectral image data containing spatial and spectral information. Hyperspectral image (HSI) classification is one of the key studies in hyperspectral interpretation. HSI classification has been applied in military target recognition, land resources and other fields. Despite significant advancements in machine learning and deep learning methods, challenges persist due to the high dimensionality of HSI data, limited training samples, and the inherent complexity of spatial and spectral features.Aiming at the problem of low classification accuracy caused by limited samples in HSI classification, this study proposes a novel HSI classification framework based on Dynamic Graph-Spectral Feature Extraction (DGSFEC). The proposed method enhances the classification accuracy by integrating dynamic graph construction, dynamic graph convolution spatial feature extraction, and hierarchical spectral feature modeling within a unified architecture. The contributions of this work are threefold: (1) the introduction of a novel dynamic graph construction method that adaptively captures spatial relationships, (2) the development of a multi-resolution dynamic graph feature extraction network that enhances the perception of cross-domain spatial features, and (3) the implementation of multi-level spectral feature extraction network that combines shallow local and deep global features to achieve refined spectral information processing. The Dynamic Graph Construction (DGC) addresses the limitations of static graph-based methods that fail to capture the spatial variability within HSI. By utilizing a sliding window mechanism, the dynamic graph adjusts adaptively to the spatial relationships of pixels, incorporating spatial similarity and distance information. This method effectively captures both local and global spatial dependencies, resulting in a more accurate and representative graph structure for HSI. Unlike static graphs, the dynamic approach offers flexibility and computational efficiency, ensuring the graph better reflects real spatial relationships. The DGSFEC model is structured into two synergistic branches. The Dynamic Graph Feature Extraction Network (DGCFN) is designed to effectively extract spatial information at different resolutions through dynamic graph convolutions, and combine with dynamic spatial convolutions and conditional position coding to enhance the perception of cross-domain spatial features. This module integrates fine-grained local features and global spatial information to provide richer spatial feature descriptions for HSI classification. Simultaneously, the Region-Global Spectral Feature Network (RGSFN) focuses on both local and global spectral feature modeling. RGSFN employs a multi-level convolutional structure, which is then combined with 3D Transformer encoders cross-layer feature fusion mechanism to integrate shallow local features with deep global features. This branch enhances the model's capacity to identify subtle spectral differences between classes. To further improve classification accuracy, the Cross Attention Feature Fusion (CAFF) network is proposed, which combines spatial and spectral features extracted from the two branches. The CAFF module employs a cross-attention mechanism to dynamically align spatial and spectral information, capturing the complementary relationships between the two feature types. This fusion enhances the model’s robustness in distinguishing complex patterns and improves generalizability across datasets with varying spatial and spectral characteristics.Extensive experiments were conducted on three benchmark hyperspectral datasets Indian Pines, University of Pavia, and Salinas. Results demonstrate that the DGSFEC framework outperforms state-of-the-art methods, achieving higher Overall Accuracy (OA), Average Accuracy (AA), and Kappa coefficients across all datasets. Notably, the proposed method achieves smoother and accurate classification maps in complex scenarios, highlighting its robustness and generalizability. This is attributed to integrating cross-domain spatial features and global spectral similarity features from DGCFN and RGSFN, ensuring classification performance on complex scenes. In addition, the DGFCN decreases computational complexity, attributed to its dynamic graph construction and feature optimization strategies. Comparative analyses reveal that the DGSFEC balances computational demands and accuracy more effectively than existing graph-based and transformer-based approaches.In conclusion, the DGSFEC method offers a classification performance advancement in HSI classification by effectively utilizing limited training samples and extracting both spatial and spectral features. The future work will focus on optimizing the training cost and exploring limited sample, low-parameter classification models to further enhance the comprehensive performance of HSI classification.

    Keywords
    Tools

    Get Citation

    Copy Citation Text

    Chenjie XU, Dan LI, Fanqiang KONG. Hyperspectral Image Classification Method Based on Dynamic Graph-spectral Feature Extraction[J]. Acta Photonica Sinica, 2025, 54(4): 0410002

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Category:

    Received: Oct. 16, 2024

    Accepted: Dec. 16, 2024

    Published Online: May. 15, 2025

    The Author Email: LI Dan (danli@nuaa.edu.cn)

    DOI:10.3788/gzxb20255404.0410002

    Topics