.png){kind=icon}
.png){kind=icon}
Return
Dual adversarial network-driven combined multiple-temporal Sentinel-1/2 images for fine cloud removal in urban regions
.png){kind=icon}
PRE
AI
Z
G
W
P
Z
Z
X
DOI：10.1080/01431161.2026.2697508.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Severe cloud occlusion severely limits the practical application of optical remote sensing data. Dense cloud cover over highly urbanized regions leads to extensive loss of fine ground features, including buildings and roads, while existing cloud removal algorithms struggle to simultaneously preserve spectral fidelity and recover subtle details, frequently suffering from texture distortion and blurred boundaries in reconstruction outcomes. Although SAR can penetrate cloud cover, distinct imaging mechanisms make single-SAR-assisted methods prone to semantic distortion during fine-scale urban reconstruction. To tackle the problem of missing detailed terrain induced by heavy cloud occlusion in cities, this paper proposes a dual-adversarial-network-based cloud removal approach via multi-source and multi-temporal fusion of Sentinel-1 and Sentinel-2 imagery. A multiple-city coregistered multi-temporal Sentinel-1/2 dataset named UCR (Urban Cloud Removal) is established to alleviate the shortage of training samples for urban cloud removal. The presented DGAN is optimized through two-stage collaborative adversarial training: the first stage achieves accurate cross-domain mapping from SAR to optical space to provide structural priors, and the second stage refines cloud-region reconstruction with multi-temporal optical constraints. Benefiting from complementary multi-source information, the proposed method realizes high-precision and high-fidelity cloud elimination as well as detailed terrain restoration for urban areas. Comparative experiments on the UCR dataset demonstrate that DGAN outperforms state-of-the-art algorithms in quantitative metrics, such as PSNR and SSIM, which validates its effectiveness in recovering intricate ground objects over complex urban landscapes.
Keywords:
Deep learning
Sentinel-1/2
GAN
urban region
cloud removal
Journal
IF:
2.6
Papers:
1.2W
Citations:
2.7W
Organization
Cited Papers