SLCGC: A lightweight Self-supervised Low-Pass Contrastive Graph Clustering Network for Hyperspectral Images

Yao Ding, Zhili Zhang, Aitao Yang, Yaoming Cai, Xiongwu Xiao, Danfeng Hong, Junsong YuanIEEE Trans. Multimedia 2025

TL;DR

This paper introduces SLCGC, a novel self-supervised graph clustering network for hyperspectral images (HSIs). It tackles the challenges of noise and computational complexity in HSI clustering by employing a low-pass graph denoising mechanism and a contrastive learning approach, achieving high accuracy with reduced computational cost.

What Problem Does This Paper Address?

Hyperspectral Image (HSI) clustering, the task of grouping pixels with similar spectral characteristics, is crucial for applications like remote sensing and environmental monitoring. However, it's a challenging problem due to several factors:

  • Lack of Labeled Data: HSIs are often acquired without ground truth labels, making supervised learning difficult.
  • Spatial-Spectral Complexity: HSIs contain rich spectral information across many bands, but also complex spatial relationships between pixels, making it hard to capture both effectively.
  • Noise Sensitivity: Real-world HSI data is prone to noise, which can significantly degrade clustering performance.
  • Computational Demands: Existing deep learning methods, especially those using graph convolutional networks (GCNs), can be computationally expensive, limiting their practical deployment.

Current methods often struggle to balance clustering accuracy, feature discriminability, computational efficiency, and robustness to noise. This paper aims to address these limitations by proposing a self-supervised, lightweight, and robust graph clustering network.

Methods

The proposed SLCGC network integrates several novel components:

  • Homogeneous Region Generation: Instead of processing individual pixels, the method first groups pixels into spectrally consistent "superpixels." This reduces the graph complexity significantly while preserving local spatial-spectral coherence.
  • Low-Pass Graph Denoising: A graph denoising mechanism is introduced to suppress high-frequency noise present in the graph topology. This is achieved using a low-pass graph filter, which helps in stable feature propagation across the graph.
  • Dual-Branch Graph Contrastive Learning: The core of the self-supervised learning lies here. The network uses two separate branches (MLPs) to generate augmented views of the graph features. Gaussian noise is added to one branch to create variations. A novel "cross-view contrastive loss" is then employed to enforce structural consistency between these augmented views, learning noise-invariant representations.
  • Feature Fusion and K-means Clustering: The learned latent embeddings from the contrastive learning are fused linearly. Finally, a standard K-means algorithm is used to obtain the final cluster assignments.

The key innovation is the combination of graph denoising with a contrastive learning framework that doesn't rely on GCNs, making it more lightweight.

Datasets

The proposed SLCGC method was evaluated on three widely used hyperspectral image datasets:

  • Salinas: A dataset known for its diverse land cover types.
  • Pavia University (PU): Another benchmark dataset for HSI analysis.
  • Trento: A dataset often used to test the robustness of algorithms.

The paper does not introduce new benchmark datasets.

Key Insights and Ideas

  • Superpixels as Graph Nodes: Aggregating pixels into superpixels is an effective strategy to reduce graph complexity and computational burden without losing crucial spatial-spectral information.
  • Low-Pass Filtering for Robustness: Explicitly denoising the graph structure using a low-pass filter significantly improves robustness to noise, a common issue in HSI data.
  • Contrastive Learning Without GCNs: The paper demonstrates that effective self-supervised feature learning for HSI clustering can be achieved using MLPs and contrastive loss, bypassing the computational overhead of GCNs.
  • Cross-View Contrastive Loss: This novel loss function is crucial for learning noise-invariant representations by ensuring structural consistency between different augmented views of the data.

Key Results

The SLCGC method demonstrated strong performance across the evaluated datasets:

  • High Clustering Accuracy: On the Salinas dataset, SLCGC achieved an Overall Accuracy (OA) of 85.48%, outperforming existing methods by a significant margin (e.g., 6.67% improvement over the second-best NCSC). Similar improvements were observed on the PU (OA=68.23%) and Trento (OA=88.98%) datasets.
  • Computational Efficiency: The paper highlights that SLCGC has significantly lower training and testing times compared to other deep learning methods, making it more practical for real-world applications. For instance, its training time on the Salinas dataset was 44.43s, considerably less than methods like NCSC (437.58s) or AdaGAE (144.86s).
  • Robustness: The low-pass graph denoising mechanism contributes to the method's robustness against noise, as evidenced by its consistent performance across different datasets.

Limitations and Future Work

  • Hyperparameter Sensitivity: While the paper provides optimal hyperparameter values, the performance can be sensitive to these choices, as shown in the sensitivity analysis figures.
  • Future Work: The authors suggest exploring even more lightweight methods for graph contrastive learning. Additionally, they propose investigating reinforcement learning for HSI preprocessing to further enhance feature extraction capabilities of clustering methods.

Summary updated

Abstract

Self-supervised hyperspectral image (HSI) clustering remains a fundamental yet challenging task due to the absence of labeled data and the inherent complexity of spatial-spectral interactions. While recent advancements have explored innovative approaches, existing methods face critical limitations in clustering accuracy, feature discriminability, computational efficiency, and robustness to noise, hindering their practical deployment. In this paper, a self-supervised efficient low-pass contrastive graph clustering (SLCGC) is introduced for HSIs. Our approach begins with homogeneous region generation, which aggregates pixels into spectrally consistent regions to preserve local spatial-spectral coherence while drastically reducing graph complexity. We then construct a structural graph using an adjacency matrix A and introduce a low-pass graph denoising mechanism to suppress high-frequency noise in the graph topology, ensuring stable feature propagation. A dual-branch graph contrastive learning module is developed, where Gaussian noise perturbations generate augmented views through two multilayer perceptrons (MLPs), and a cross-view contrastive loss enforces structural consistency between views to learn noise-invariant representations. Finally, latent embeddings optimized by this process are clustered via K-means. Extensive experiments and repeated comparative analysis have verified that our SLCGC contains high clustering accuracy, low computational complexity, and strong robustness. The code source will be available at https://github.com/DY-HYX.
BibTeX
@article{ding2025slcgc,
  author = {Yao Ding and Zhili Zhang and Aitao Yang and Yaoming Cai and Xiongwu Xiao and Danfeng Hong and Junsong Yuan},
  title = {{SLCGC: A lightweight Self-supervised Low-Pass Contrastive Graph Clustering Network for Hyperspectral Images}},
  journal = {IEEE Trans. Multimedia},
  year = {2025},
}
Figure