遥感
计算机科学
云计算
环境科学
过程(计算)
地质学
计算机视觉
自动化
实时计算
领域(数学)
全球定位系统
作者
Junjie Ning,Zhenhui Sun,Yufan Wang,Kaixiang Wei,Lang Zhang,Chen Zhang
标识
DOI:10.1080/01431161.2026.2717440
摘要
Cloud cover limits the use of optical remote-sensing imagery. We present CGD-Net, a deterministic single-pass conditional reconstruction network for Sentinel-2 cloud removal that combines multi-temporal optical observations, Sentinel-1 SAR guidance, and a refined cloud-mask prior. The network separates absolute reflectance prediction from prior-guided local refinement and fuses the two paths with an asymmetric cloud gate, reducing the transfer of cloudy-input brightness bias into thick-cloud reconstruction. Temporal reference selection, mask-reliability gating, multi-scale SAR–optical modulation, and cloud-region residual refinement provide complementary spatial and spectral cues. Under the new ROI/sub-ROI-isolated split, CGD-Net obtains 30.95 dB PSNR, 0.884 SSIM, and 5.35° SAM on SEN12MS-CR-TS, and 33.75 dB PSNR, 0.919 SSIM, and 5.05° SAM on AllClear. Controlled tests of mask perturbation, registration error, temporal mismatch, extreme cloud coverage, regional variation, SAR speckle, and cross-dataset transfer delimit both its robustness and its remaining uncertainty.
科研通智能强力驱动
Strongly Powered by AbleSci AI