Research map / Graph learning / Knowledge and temporal graphs

HetSSNet: Spatial-Spectral Heterogeneous Graph Learning Network for Panchromatic and Multispectral Images Fusion

ICML 2025 · Graph learning › Knowledge and temporal graphs · approach: Heterogeneous graph learning · 1 citations (OpenAlex)

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 HetSSNet, a heterogeneous graph learning network that models PAN/LR-MS pansharpening relationships with a spatial-spectral heterogeneous graph, relationship pattern generation, and local/global aggregation with contrastive learning.

Problem

Pansharpening fuses PAN and LR-MS images to produce HR-MS images; CNN and Transformer models treat pixels on rigid Euclidean grids, which is ill-suited to irregular ground objects in remote sensing imagery.

Approach

Builds a heterogeneous graph whose nodes are PAN patches and LR-MS spectral-band features, with edges encoding spatial, intra-spectral, and cross-spectral relations; generates all basic relationship patterns from edge-type combinations, then aggregates them with learnable weights via local GCN and global pattern-similarity GCN, aligned by contrastive loss.

Key results

• On GaoFen-2 reduced-resolution, HetSSNet reports PSNR 49.541, SSIM 0.994, SAM 0.015, ERGAS 0.955, SCC 0.990, outperforming the best baseline MDCUN (PSNR 48.668).
• On WorldView-3 reduced-resolution, HetSSNet achieves PSNR 40.623 and SAM 0.054, exceeding LGTEUN (40.124 PSNR).
• On full-resolution real-world scenes, HetSSNet achieves best D_lambda, D_s and QNR across the three datasets in the reported table (e.g., GaoFen-2 QNR 0.860).

Measured effects reported in the paper

Surprising findings

• Ablation shows a 2-layer aggregation performs best and deeper layers degrade performance, while the authors state the goal is not to address over-smoothing.

Limitations

• Graph construction is O(n^2) in node count, limiting scalability to large images (complexity analysis).
• Evaluation limited to three satellite datasets and pansharpening; the authors note future need for more challenging scenarios such as fast-moving objects.
• Full-resolution evaluation uses non-reference metrics only, with no GT available.

Experimental setup

Datasets: WorldView-3 QuickBird GaoFen-2

Benchmarks: WorldView-3 reduced-resolution QuickBird reduced-resolution GaoFen-2 reduced-resolution full-resolution real-world scenes (20 images per dataset)

Metrics: PSNR SSIM SAM ERGAS SCC D_lambda D_s QNR

Baselines: SRPPNN DCFNet CTINN Hyperformer SFIINet BiMPan MDCUN MSDDN LGTEUN FAMENet GPCNet SFIM

Builds on: MHGCN GPCNet

Closest papers by meaning

• SSDiff: Spatial-spectral Integrated Diffusion Model for Remote Sensing Pansharpening (NeurIPS 2024 · similarity 0.88)
• S2MAE: A Spatial-Spectral Pretraining Foundation Model for Spectral Remote Sensing Data (CVPR 2024 · similarity 0.85)
• A Selective Re-learning Mechanism for Hyperspectral Fusion Imaging (CVPR 2025 · similarity 0.85)
• Kumaraswamy Wavelet for Heterophilic Scene Graph Generation (AAAI 2024 · similarity 0.85)
• FMPM-DNet: Hyperspectral Pansharpening Dynamic Network Based on Feature Modulation and Probability Mask (AAAI 2025 · similarity 0.85)
• Structure-Adaptive Multi-View Graph Clustering for Remote Sensing Data (AAAI 2025 · similarity 0.85)