.png){kind=icon}
.png){kind=icon}
Return
DGPIR: Degradation-Guided Prompting for All-in-One Radiometric Restoration of Level-1 Optical Satellite Imagery
.png){kind=icon}
PRE
AI
K
L
H
X
L
L
DOI：10.1109/tgrs.2026.3716720.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Commercial optical satellite constellations have greatly increased the volume and diversity of Level-1 (L1) optical products acquired across different sensors, orbit groups, and land-cover scenes. Although standard radiometric calibration and relative radiometric correction remove most systematic detector- or channel-level response differences, weak postcorrection residuals may remain because of time-varying on-orbit response, residual sensor-dependent variations, and complex scene radiance distributions. These residuals, including cross-track low-frequency color distortion, residual striping, and random noise, are often subtle, spatially nonuniform, sensor- and band-dependent, and tightly coupled with genuine land-cover textures, making reliable postcorrection restoration difficult without introducing over-correction, spectral bias, or texture degradation. To address this problem, we propose DGPIR, a degradation-guided prompt-based all-in-one framework for radiometric restoration of L1 optical satellite imagery. DGPIR uses the radiometric residual category identified by scene-level quality inspection, together with image-derived degradation statistics, to generate scale-specific dynamic prompts. These prompts are injected into a Transformer backbone through prompt-aware attention (PAA) and prompt-gated feed-forward (PGF) modulation, enabling a shared model to adapt its restoration behavior to physically different residual categories. A dual-branch inference strategy is further designed to match the spatial scale of L1 residuals: a global branch estimates full-swath low-frequency color distortion, while a patch-based branch handles stripe and noise residuals under practical memory constraints. Experiments on representative L1 restoration tasks show that DGPIR achieves consistent improvements over state-of-the-art all-in-one restoration baselines. For low-frequency color distortion, DGPIR improves PSNR by up to 0.60 dB and achieves strong cross-track radiometric consistency, with profile mean absolute error (Profile MAE) reduced to 0.0031 and profile correlation (Profile Corr.) reaching 0.8145 under the high-intensity setting. For destriping, DGPIR improves PSNR by 2.09 dB over the strongest all-in-one baseline in the mild-stripe regime and maintains the lowest RMSE under severe stripe degradation. It also shows stable denoising behavior, improves mixed-degradation restoration, and produces physically meaningful low-frequency correction profiles on additional real L1 scenes from different orbit groups and satellite platforms. The model maintains practical inference efficiency, requiring 95.84 GFLOPs and 424.18-MB peak GPU memory. The project repository is available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/zhaxia191812/DGPIR</uri>
Keywords:
All-in-one image restoration (AiOIR)
low-frequency color distortion correction
prompt learning
remote sensing image restoration (RSIR)
Journal
IF:
8.6
Papers:
2.1W
Citations:
10.7W
Organization
Cited Papers