Research map / Graph learning / Graph self-supervised learning
Structure-Adaptive Multi-View Graph Clustering for Remote Sensing Data
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SAMVGC, a structure-adaptive multi-view graph clustering framework for remote sensing data that refines superpixel anchor graphs in embedding space and treats central and confusing samples differently via contrastive and consistency losses.
Problem
Multi-view clustering of remote sensing data with GNN methods relies on a fixed predefined graph that contains noisy background connections, and confusing boundary samples compromise cluster compactness.
Approach
Adaptively refine the anchor adjacency matrices using similarity in the learned embedding space, fused with previous graphs, and split samples into central and confusing groups: prototype contrastive learning on central samples and KL-based intra/inter-view consistency learning on confusing samples.
Key results
- SAMVGC reaches 95.04% ACC on Trento, outperforming MDFL, TMPCC and MDC by 9.03, 6.47 and 11.65 points respectively.
- On Augsburg, SAMVGC achieves 67.80% ACC and 55.21% NMI, versus 52.70% ACC for the next best baseline TMPCC.
- Removing the graph refinement module (w/o GE) lowers ACC by 2.01-3.92 points across the four datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SAMVGC instead of the strongest reported baseline | Remote-sensing multi-view clustering · Trento | ACC: 93.90±1.97 -> 95.04±0.36 | Helped |
| use SAMVGC instead of the strongest reported baseline | Remote-sensing multi-view clustering · MUUFL | ACC: 48.23±4.21 -> 50.50±1.50 | Helped |
| use SAMVGC instead of the strongest reported baseline | Remote-sensing multi-view clustering · Augsburg | ACC: 52.70±2.61 -> 67.80±0.20 | Helped |
| use SAMVGC instead of the strongest reported baseline | Remote-sensing multi-view clustering · MDSA | ACC: 40.56±3.84 -> 47.59±0.70 | Helped |
| add graph spatial correlation adaptive refinement (ablation) | SAMVGC · Trento; MUUFL; Augsburg; MDSA | ACC: 92.89 -> 95.04; 46.58 -> 50.50; 65.11 -> 67.80; 45.58 -> 47.59 | Helped |
| add central-sample contrastive learning loss (ablation) | SAMVGC · Trento; MUUFL; Augsburg; MDSA | ACC: 87.07 -> 95.04; 42.59 -> 50.50; 64.34 -> 67.80; 44.98 -> 47.59 | Helped |
| add confusing-sample consistency learning loss (ablation) | SAMVGC · Trento; MUUFL; Augsburg; MDSA | ACC: 93.16 -> 95.04; 46.17 -> 50.50; 66.89 -> 67.80; 46.24 -> 47.59 | Helped |
Limitations
- Relies on a fixed number of superpixels M and k-nearest neighbors, both requiring tuning
- Large central-sample rate alpha incurs excessive time and space costs
- Evaluation limited to four remote sensing datasets
Experimental setup
Datasets: Trento MUUFL Augsburg MDAS
Benchmarks: Trento MUUFL Augsburg MDAS
Metrics: ACC Kappa NMI ARI PUR
Baselines: FMVACC AWMVC AMKSC MFLVC MDC CVCL SDMVC GCFAgg TMPCC MDFL CMSCGC
Builds on: SLIC superpixels graph convolutional networks prototype contrastive learning K-means
Closest papers by meaning
- SURER: Structure-Adaptive Unified Graph Neural Network for Multi-View Clustering (AAAI 2024 · similarity 0.88)
- A Non-parametric Graph Clustering Framework for Multi-View Data (AAAI 2024 · similarity 0.86)
- Medusa: A Multi-Scale High-order Contrastive Dual-Diffusion Approach for Multi-View Clustering (CVPR 2025 · similarity 0.85)
- HetSSNet: Spatial-Spectral Heterogeneous Graph Learning Network for Panchromatic and Multispectral Images Fusion (ICML 2025 · similarity 0.85)
- Embedded Feature Selection on Graph-Based Multi-View Clustering (AAAI 2024 · similarity 0.85)
- From Spectrum-free towards Baseline-view-free: Double-track Proximity Driven Multi-view Clustering (ICML 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.