Multi-level Graph Subspace Contrastive Learning for Hyperspectral Image Clustering

Jingxin Wang, Renxiang Guan, Kainan Gao, Zihao Li, Hao Li, Xianju Li, Chang Tang

Introduction

Hyperspectral imaging (HSI) captures hundreds of narrow spectral bands, enabling precise material identification. However, clustering these high-dimensional data points without labels remains a significant challenge. Traditional subspace clustering often fails to capture nonlinear structures, while standard Graph Convolutional Networks (GCNs) tend to focus on local neighborhoods, suffering from "myopia" that ignores global context. Furthermore, existing methods often underutilize the synergy between different data views (e.g., spectral vs. texture) and the relationship between local pixels and the global scene.

The MLGSC Framework

To address these limitations, we present the Multi-level Graph Subspace Contrastive Learning (MLGSC) framework. MLGSC integrates multi-view graph construction with a novel multi-level contrastive learning strategy and attention-based pooling.

The detailed architecture of MLGSC, highlighting the multi-level contrastive learning and attention pooling mechanisms.

The framework operates through five core stages:

  1. Multi-view Graph Construction: Two distinct graph views are generated from spectral-spatial features (via PCA) and texture features (via morphological operations).
  2. Attention Pooling: Instead of simple averaging, an attention mechanism weighs node importance to distill a global graph representation, capturing the overall structure of the HSI scene.
  3. Multi-level Contrastive Learning: Consistency is enforced at two scales: node-level (local regions across views) and graph-level (global representations across views).
  4. Weighted Feature Fusion: An adaptive module merges multi-view features into a discriminative representation.
  5. Self-Expression Module: A robust affinity matrix is generated through data reconstruction, followed by spectral clustering.

Experimental Results

The MLGSC model was evaluated on four benchmark datasets: Indian Pines, Pavia University, Houston-2013, and Xu Zhou. The quantitative results demonstrate that MLGSC significantly outperforms state-of-the-art (SOTA) methods, achieving near-perfect accuracy on several benchmarks.

Dataset MLGSC Overall Accuracy (OA) Improvement over Sub-optimal SOTA
Indian Pines 97.75% +4.55%
Pavia University 99.96% +2.21%
Houston 92.28% +2.98%
Xu Zhou 95.73% +3.80%

Qualitatively, the clustering maps produced by MLGSC exhibit superior spatial consistency and cleaner boundaries compared to traditional methods.

Qualitative comparison of clustering maps across different methods. MLGSC produces significantly cleaner results with less isolated noise.

The visual evidence confirms that MLGSC effectively reduces "pretzel noise"—isolated misclassified pixels—by leveraging global perception through attention pooling and robust feature extraction via contrastive learning.

Discussion and Conclusion

The success of MLGSC stems from three key insights:

  • Global Perception: Attention pooling allows the model to distinguish land cover types that appear similar locally but differ in global distribution.
  • Contrastive Regularization: Multi-level contrastive learning forces the model to identify invariant features across texture and spectral views, enhancing robustness to noise.
  • Subspace-GCN Synergy: Combining the structural modeling of GCNs with the mathematical rigor of subspace self-expression creates a powerful complementary effect.

While MLGSC achieves impressive results, future work will focus on optimizing computational complexity and hyperparameter sensitivity to scale the model for full-scale satellite imagery.

Summary updated

Abstract

Hyperspectral image (HSI) clustering is a challenging task due to its high complexity. Despite subspace clustering shows impressive performance for HSI, traditional methods tend to ignore the global-local interaction in HSI data. In this study, we proposed a multi-level graph subspace contrastive learning (MLGSC) for HSI clustering. The model is divided into the following main parts. Graph convolution subspace construction: utilizing HSI’s spectral and texture feautures to construct two graph convolution views. Local-global graph representation: local graph representations were obtained by step-by-step convolutions and a more representative global graph representation was obtained using an attention-based pooling strategy. Multi-level graph subspace contrastive learning: multi-level contrastive learning was conducted to obtain local-global joint graph representations, to improve the consistency of the positive samples between views, and to obtain more robust graph embeddings. Specifically, graph-level contrastive learning is used to better learn global representations of HSI data. Node-level intra-view and inter-view contrastive learning is designed to learn joint representations of local regions of HSI. The proposed model is evaluated on four popular HSI datasets: Indian Pines, Pavia University, Houston, and Xu Zhou. The overall accuracies are 97.75%, 99.96%, 92.28%, and 95.73%, which significantly outperforms the current state-of-the-art clustering methods.
BibTeX
@inproceedings{wang2024multi,
  author = {Jingxin Wang and Renxiang Guan and Kainan Gao and Zihao Li and Hao Li and Xianju Li and Chang Tang},
  title = {{Multi-level Graph Subspace Contrastive Learning for Hyperspectral Image Clustering}},
  booktitle = {IJCNN},
  year = {2024},
  url = {https://openreview.net/forum?id=rf75oFJnw9},
}
Figure