Structural-Spectral Graph Convolution With Evidential Edge Learning for Hyperspectral Image Clustering

Jianhan Qi, Yuheng Jia, Hui Liu, Junhui Hou

Abstract

Hyperspectral image (HSI) clustering is a critical unsupervised task for material identification. Traditional methods often struggle with complex spectral patterns and noisy graph topologies. This paper introduces a novel framework, SSGCO-EGAEL, which integrates a Structural-Spectral Graph Convolutional Operator (SSGCO) with Evidence-Guided Adaptive Edge Learning (EGAEL). By leveraging 1D convolutions for spectral continuity and dynamically refining graph edges based on clustering evidence, the proposed method achieves state-of-the-art performance across multiple benchmarks, improving accuracy by up to 6% while remaining computationally efficient.

Introduction

Hyperspectral imaging captures hundreds of narrow spectral bands, providing a unique "fingerprint" for every pixel. However, the high cost of manual labeling makes unsupervised clustering—grouping pixels without prior labels—essential. Current state-of-the-art methods typically group pixels into superpixels to reduce computational load and then apply Graph Neural Networks (GNNs). Despite their success, these methods face two primary hurdles: a feature extraction gap where simple linear layers fail to capture complex spectral patterns, and noisy graphs where spatial adjacency creates incorrect edges between different material classes.

Methodology

The proposed SSGCO-EGAEL framework addresses these challenges through two main innovations integrated into a contrastive learning pipeline.

Structural-Spectral Graph Convolutional Operator (SSGCO)

Unlike standard graph layers, SSGCO sequentially applies 1D convolutions to capture high-order spectral patterns followed by graph convolutions to aggregate spatial information from neighbors. This dual approach ensures the network captures both the chemical signature and the physical context of each segment.

Evidence-Guided Adaptive Edge Learning (EGAEL)

EGAEL treats the graph as a dynamic entity. It employs a small neural network to predict connection strengths between superpixels. By using "evidence"—derived from clustering confidence and node similarity—it supervises edge weights, effectively pruning links between segments likely belonging to different classes.

Overview of the SSGCO and EGAEL modules.

Contrastive Framework

The model utilizes a teacher-student setup (online and target networks). It encourages neighborhood alignment by producing similar representations for a superpixel and its neighbors, while prototype contrast ensures that different clusters remain distinct.

Experimental Results

The model was evaluated on four standard HSI benchmarks: Indian Pines, Pavia University, Botswana, and Trento.

Performance Gains

SSGCO-EGAEL set new state-of-the-art records, with accuracy (ACC) gains of 2.61% on Indian Pines, 6.06% on Pavia University, 4.96% on Botswana, and 3.15% on Trento compared to previous best methods. Statistical analysis via pairwise t-tests confirmed that the method significantly outperformed competitors in over 91% of cases.

Clustering accuracy comparison across representation networks and the impact of edge refinement.

Key Insights

Ablation studies highlight the importance of spectral continuity; 1D convolutions are significantly more effective than linear projections. Furthermore, the research demonstrates that as clustering confidence improves, the model can retroactively fix topological errors. Deleting incorrect edges directly leads to a steady climb in clustering accuracy, as evidenced by the correlation between deletion ratio and performance metrics.

Conclusion

SSGCO-EGAEL offers a robust and efficient solution for HSI clustering by better integrating spectral data and refining graph topology. While highly effective, the current reliance on superpixel averaging discards internal spatial structures. Future research will explore the use of 2D convolutions on padded superpixels to capture finer spatial details.

Summary updated

BibTeX
@article{qi2025structural,
  author = {Jianhan Qi and Yuheng Jia and Hui Liu and Junhui Hou},
  title = {{Structural-Spectral Graph Convolution With Evidential Edge Learning for Hyperspectral Image Clustering}},
  journal = {IEEE Trans. Image Process.},
  year = {2025},
  doi = {10.1109/tip.2025.3636098},
}
Figure