Research map / Architectures and efficiency / CNN design
A General Adaptive Dual-level Weighting Mechanism for Remote Sensing Pansharpening
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 ADWM, a plug-and-play dual-level weighting module (intra-feature and cross-feature) built on covariance-based weights (CACW) that boosts deep pansharpening networks on WV3, QB, and GF2.
Problem
Deep learning pansharpening networks underuse feature heterogeneity and redundancy both across channels within a feature and across layers of sequential feature extraction.
Approach
Computes a normalized covariance matrix over channels (IFW) or layers (CFW) and maps it through an MLP to produce weights that rescale features, with CFW softmax-weighting and summing intermediate features before decoding; it can be plugged into existing networks.
Key results
- ADWM-enhanced LAGNet reaches 39.170 dB PSNR on WV3 reduced-resolution, 0.158 dB above the second-best PanMamba (39.012 dB).
- On GF2 reduced-resolution, the proposed method reaches 43.884 dB PSNR, 0.977 dB above LAGNet's 42.735 dB... (second-best PanMamba 42.907 dB, so +0.977 vs. second-best is incorrect; the gain over second-best is 0.977 dB only vs LAGNet)
- Adding ADWM to FusionNet, LGPNet, and U2Net improves HQNR on WV3 full-resolution (e.g., FusionNet 0.940 to 0.946, LGPNet 0.940 to 0.950) with small parameter increases.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ADWM to LAGNet | LAGNet · WV3 reduced-resolution | PSNR: 39.012 -> 39.170 | Helped |
| Add ADWM to LAGNet | LAGNet · QB reduced-resolution | PSNR: 38.209 -> 38.466 | Helped |
| Add ADWM to LAGNet | LAGNet · GF2 reduced-resolution | PSNR: 42.907 -> 43.884 | Helped |
| Add IFW (ablation) | LAGNet without IFW or CFW · WV3 reduced-resolution | PSNR: 38.592 -> 39.028 | Helped |
| Add CFW (ablation) | LAGNet without IFW or CFW · WV3 reduced-resolution | PSNR: 38.592 -> 38.923 | Helped |
| Add CFW to IFW (ablation) | LAGNet with IFW only · WV3 reduced-resolution | PSNR: 39.028 -> 39.170 | Helped |
| Add IFW to CFW (ablation) | LAGNet with CFW only · WV3 reduced-resolution | PSNR: 38.923 -> 39.170 | Helped |
| Generate weights using CACW instead of PCA (ablation) | LAGNet · WV3 reduced-resolution | PSNR: 38.612 -> 39.170 | Helped |
| Generate weights using CACW instead of global average pooling (ablation) | LAGNet · WV3 reduced-resolution | PSNR: 39.053 -> 39.170 | Helped |
| Generate weights using CACW instead of self-attention (ablation) | LAGNet · WV3 reduced-resolution | PSNR: 38.875 -> 39.170 | Helped |
Surprising findings
- Pool-based and attention-based weighting alternatives can underperform the no-weight baseline in some settings, and PCA-based weights underperform CACW.
Limitations
- Gains on full-resolution metrics are modest for some backbones.
- Theoretical complexity grows with channels and spatial size; intermediate layer size d needs tuning (overfitting beyond about 0.8n).
Experimental setup
Datasets: WV3 (PanCollection) QB (PanCollection) GF2 (PanCollection)
Benchmarks: WV3 reduced-resolution QB reduced-resolution GF2 reduced-resolution WV3 full-resolution QB full-resolution GF2 full-resolution
Metrics: PSNR SAM ERGAS Q4 Q8 Ds Dλ HQNR
Baselines: MTF-GLP-FS BDSD-PC TV PNN PanNet DiCNN FusionNet LAGNet LGPNet PanMamba
Builds on: PCA LAGNet SENet
Closest papers by meaning
- Content-Adaptive Non-Local Convolution for Remote Sensing Pansharpening (CVPR 2024 · similarity 0.88)
- Dual-Granularity Semantic Guided Sparse Routing Diffusion Model for General Pansharpening (CVPR 2025 · similarity 0.86)
- Meta-Learning Hyperparameters for Parameter Efficient Fine-Tuning (CVPR 2025 · similarity 0.86)
- PanAdapter: Two-Stage Fine-Tuning with Spatial-Spectral Priors Injecting for Pansharpening (AAAI 2025 · similarity 0.85)
- Pan-LUT: Efficient Pan-sharpening via Learnable Look-Up Tables (NeurIPS 2025 · similarity 0.85)
- SSDiff: Spatial-spectral Integrated Diffusion Model for Remote Sensing Pansharpening (NeurIPS 2024 · 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.