You Only Encode Once: Sub-Pixel-Shifted Frames for Arbitrary-Scale Remote Sensing Super-Resolution with 2D Gaussian Splatting
Organizations: Zhengzhou University · Beijing University of Technology · Anhui University
Abstract
Arbitrary-scale super-resolution (ASSR) of remote sensing images remains challenging due to the limited high-frequency information in a single low-resolution (LR) observation. While multi-frame SR can exploit complementary sub-pixel information, existing approaches often require real multi-frame training sequences and repeated feature encoding, increasing data acquisition demands and computational costs. To address these limitations, we propose SPS-GS (Sub-Pixel-Shifted Gaussian Splatting), an encode-once framework that learns from multi-phase LR observations synthesized from individual high-resolution images, without requiring real multi-frame training sequences. Specifically, Phase-Coherent Aliasing Demodulation (PCAD) exploits shift-dependent aliasing through Fourier-based alignment, phase-aware residual fusion, and least-squares aliasing estimation, enabling multi-frame information to be integrated before a single deep encoding pass. Gaussian Projection Alignment (GPA) then iteratively refines Gaussian centers and colors using multi-observation back-projection to promote observation consistency. The refined Gaussians support continuous rendering at arbitrary scales. Extensive experiments on five remote sensing datasets demonstrate strong reconstruction performance against single-image and multi-frame SR baselines, while maintaining competitive inference efficiency at common scales. The code will be publicly released.
Figures & tables
| Methods | AID [ 54 ] | UCMerced [ 60 ] | WHU-RS19 [ 10 ] | RSSCN7 [ 67 ] | NWPU-RESISC45 [ 8 ] | |||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Bicubic | 32.90 | 27.58 | 24.56 | 22.60 | 31.09 | 25.70 | 22.48 | 20.42 | 33.67 | 27.73 | 24.45 | 22.42 | 30.24 | 26.36 | 24.10 | 22.55 | 30.98 | 26.36 | 23.66 | 21.77 |
| Single-image methods | ||||||||||||||||||||
| Meta-SR [ 23 ] | 35.75 | 29.60 | 25.84 | 23.44 | 35.10 | 28.60 | 24.17 | 21.31 | 36.65 | 29.83 | 25.78 | 23.23 | 32.53 | 27.80 | 25.03 | 23.15 | 33.51 | 28.15 | 24.76 | 22.43 |
| LIIF [ 7 ] | 35.75 | 29.66 | 25.91 | 23.51 | 35.16 | 28.74 | 24.33 | 21.42 | 36.72 | 29.92 | 25.88 | 23.35 | 32.53 | 27.85 | 25.10 | 23.25 | 33.55 | 28.19 | 24.84 | 22.54 |
| LTE [ 31 ] | 35.77 | 29.68 | 25.05 | 21.54 | 35.17 | 28.74 | 23.80 | 19.98 | 36.74 | 29.90 | 24.96 | 21.54 | 32.54 | 27.86 | 24.11 | 20.49 | 33.58 | 28.22 | 24.30 | 21.06 |
| Methods | SSIM | LPIPS | DISTS | FID | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Bicubic | 0.9083 | 0.7405 | 0.6075 | 0.5538 | 0.2133 | 0.4912 | 0.7125 | 0.8514 | 0.1081 | 0.2185 | 0.3400 | 0.4839 | 14.92 | 39.94 | 98.80 | 208.60 |
| Single-image methods | ||||||||||||||||
| Meta-SR [ 23 ] | 0.9466 | 0.8169 | 0.6670 | 0.5752 | 0.1059 | 0.3046 | 0.5101 | 0.7064 | 0.0757 | 0.1680 | 0.2656 | 0.3725 | 9.72 | 43.04 | 94.87 | 168.22 |
| LIIF [ 7 ] | 0.9467 | 0.8171 | 0.6702 | 0.5799 | 0.1043 | 0.3186 | 0.5360 | 0.7277 | 0.0752 | 0.1736 | 0.2745 | 0.3850 | 9.73 | 42.27 | 97.22 | 166.94 |
| LTE [ 31 ] | 0.9469 | 0.8177 | 0.6668 | 0.5741 | 0.1049 | 0.3149 | 0.5324 | 0.7275 | 0.0759 | 0.1739 | 0.2722 | 0.3855 | 9.65 | 43.30 | 86.86 | 151.63 |
| Methods | #Params | Inference time (ms) | Memory | ||
|---|---|---|---|---|---|
| (M) | (GiB) | ||||
| Meta-SR | 13.68 | 142.5 | 96.9 | 82.5 | 3.81 |
| LIIF | 15.92 | 442.9 | 400.7 | 381.6 | 1.43 |
| LTE | 13.34 | 131.3 | 84.0 | 71.2 | 0.89 |
| LMF | 12.61 | 85.3 | 38.2 | 27.4 | 1.30 |
| HIIF | 14.08 | 244.6 | 200.1 | 182.7 | 6.56 |
| Datasets | Methods | ||||
|---|---|---|---|---|---|
| PROBA-V [ 34 ] | Bicubic | 47.31 | 45.38 | 43.65 | 42.27 |
| BIPNet | 47.76 | 45.86 | 43.07 | 39.91 | |
| HighRes-net | 48.71 | 46.57 | 43.68 | 41.28 | |
| TR-MISR | 48.56 | 46.44 | 44.10 | 42.36 | |
| PIUnet | 49.04 | 46.96 | 44.58 | 42.53 | |
| SPS-GS (Ours) | 49.73 | 47.18 | 44.53 | 42.45 |
| PCAD | GPA | PSNR | Avg. | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Res. | LS | Haar | |||||||||||
| 0 | 35.71 | 31.85 | 29.66 | 27.25 | 25.91 | 30.08 | |||||||
| 7 | ✓ | ✓ | 39.56 | 35.16 | 31.81 | 28.72 | 26.91 | 32.43 | |||||
| ✓ | 7 | ✓ | ✓ | 39.98 | 35.31 | 32.14 | 29.00 | 27.13 | 32.71 | ||||
| ✓ | ✓ | 7 | ✓ | ✓ | 40.02 | 35.33 | 32.18 | 29.02 | 27.17 | 32.74 | |||
| ✓ | ✓ | ✓ | 7 | ✓ | ✓ | 40.48 | 35.77 | 32.60 | 29.36 | 27.46 | 33.13 | ||
| Representation | Avg. | Time | |||||
|---|---|---|---|---|---|---|---|
| 2DGS decoder (Ours) | 40.48 | 35.81 | 32.65 | 29.40 | 27.48 | 33.17 | 69.6 |
| LIIF decoder [ 7 ] | 40.51 | 35.82 | 32.70 | 29.42 | 27.55 | 33.20 | 488.1 |
| LTE decoder [ 31 ] | 40.50 | 35.84 | 32.73 | 29.46 | 27.52 | 33.21 | 705.9 |
Image Super-Resolution
Image super-resolution reconstructs high-resolution images from lower-resolution observations, balancing detail, fidelity, and computational cost.
See all 271 papers