Research map / Architectures and efficiency / CNN design
Adaptive Rectangular Convolution 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 Adaptive Rectangular Convolution (ARConv), which learns per-pixel kernel height and width and adapts the number of sampling points, and embeds it in a U-Net-based pansharpening network (ARNet).
Problem
Standard convolutions in CNN-based remote sensing pansharpening use fixed square windows and a fixed number of sampling points, which handles the wide range of object scales in remote sensing imagery poorly.
Approach
Treats kernel height and width as learnable per-pixel parameters, selects the number of sampling points from the mean learned height and width, samples on a rectangular deformable grid with bilinear interpolation, and adds an affine modulation; ARConv replaces standard convolutions in a U-Net (ARNet).
Key results
- On WV3 reduced-resolution, ARNet achieves SAM 2.885, ERGAS 2.139, Q8 0.921, outperforming CANNet (2.930 / 2.158 / 0.920).
- On QB, ARNet reaches SAM 4.430, ERGAS 3.633, Q4 0.939, the best among compared methods.
- Replacing convolutions with ARConv improves FusionNet (ERGAS 2.467 to 2.395) and CANNet (SAM 2.930 to 2.885) on WV3; ARConv also beats DCNv2 (ERGAS 2.149 vs 2.425 on WV3).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ARConv in place of standard convolution in ARNet | ARNet · WV3 reduced-resolution | SAM: 2.930 ± 0.593 -> 2.885 ± 0.590 | Helped |
| Use ARConv in place of standard convolution in ARNet | ARNet · WV3 reduced-resolution | ERGAS: 2.158 ± 0.515 -> 2.139 ± 0.528 | Helped |
| Use ARConv in place of standard convolution in ARNet | ARNet · QB reduced-resolution | SAM: 4.507±0.835 -> 4.430±0.811 | Helped |
| Use ARConv in place of standard convolution in ARNet | ARNet · GF2 reduced-resolution | SAM: 0.707±0.148 -> 0.698±0.149 | Helped |
| Add height and width adaptation (ablation) | ARNet · WV3 reduced-resolution | SAM: 2.925±0.593 -> 2.885±0.590 | Helped |
| Add sampling points adaptation (ablation) | ARNet · WV3 reduced-resolution | SAM: 2.911±0.603 -> 2.885±0.590 | Helped |
| Add affine transformation (ablation) | ARNet · WV3 reduced-resolution | SAM: 3.020±0.614 -> 2.885±0.590 | Helped |
| Set height and width learning range to 1-3 (ablation) | ARNet · WV3 reduced-resolution | SAM: 2.885±0.590 -> 2.923±0.600 | Hurt |
| Set height and width learning range to 1-9 (ablation) | ARNet · WV3 reduced-resolution | SAM: 2.885±0.590 -> 2.896±0.588 | Hurt |
| Set height and width learning range to 1-36 (ablation) | ARNet · WV3 reduced-resolution | SAM: 2.885±0.590 -> 3.044±0.646 | Hurt |
| Set height and width learning range to 1-63 (ablation) | ARNet · WV3 reduced-resolution | SAM: 2.885±0.590 -> 3.066±0.593 | Hurt |
| Replace FusionNet convolution layers with ARConv | FusionNet · WV3 reduced-resolution | SAM: 3.325±0.698 -> 3.171±0.650 | Helped |
Surprising findings
- Ablation shows that removing the number-of-sampling-points adaptation yields only a small change (ERGAS 2.152 vs 2.139), while removing affine transformation causes the largest drop.
Limitations
- Training requires a 100-epoch exploratory phase followed by fixing a randomly selected kernel configuration, which complicates convergence.
- Intermediate-layer kernel heatmaps appear disordered, so the size correlation is only partially validated visually.
- Kernel height/width range is a hyperparameter; performance peaks at a specific range (1-18) and degrades outside it.
Experimental setup
Datasets: WorldView-3 (WV3) QuickBird (QB) GaoFen-2 (GF2)
Benchmarks: WV3 QB GF2
Metrics: SAM ERGAS Q8 Q4 Dλ Ds HQNR
Baselines: EXP MTF-GLP-FS TV BSDS-PC CVPR2019 LRTCFPan PNN PanNet DiCNN FusionNet DCFNet LAGConv
Builds on: Deformable Convolution (DCNv2) LAGConv CANConv U-Net Dynamic Snake Convolution
Closest papers by meaning
- Content-Adaptive Non-Local Convolution for Remote Sensing Pansharpening (CVPR 2024 · similarity 0.89)
- SSDiff: Spatial-spectral Integrated Diffusion Model for Remote Sensing Pansharpening (NeurIPS 2024 · similarity 0.85)
- A General Adaptive Dual-level Weighting Mechanism for Remote Sensing Pansharpening (CVPR 2025 · similarity 0.84)
- Pan-LUT: Efficient Pan-sharpening via Learnable Look-Up Tables (NeurIPS 2025 · similarity 0.83)
- PanAdapter: Two-Stage Fine-Tuning with Spatial-Spectral Priors Injecting for Pansharpening (AAAI 2025 · similarity 0.83)
- Dual-Granularity Semantic Guided Sparse Routing Diffusion Model for General Pansharpening (CVPR 2025 · similarity 0.82)
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.