首页|期刊导航|北京测绘|GF-1宽幅影像单图像超分辨率重建方法研究

GF-1宽幅影像单图像超分辨率重建方法研究OA

Single-image super-resolution reconstruction for GF-1 wide-swath imagery

中文摘要英文摘要

为进一步提升高分一号(GF-1)宽幅影像的空间分辨率,拓宽其在地理国情监测,特别是水文水利领域的应用场景,本文对GF-1宽幅影像单图像超分辨率开展研究.研究从插值法、重建法等算法门类中选取6种算法完成复现,构建分辨率16 m至2 m的GF-1影像测试集进行实验,验证6种算法在低、中、高倍率下的超分辨率性能与效果,利用均方误差、视觉保真度等指标构建评价体系.实验表明:针对GF-1宽幅影像中低倍超分需求,各类算法均可以满足,其中深度学习类算法效果最优,插值法运算效率优势突出.在重点考虑效率时,可选用插值法开展中低倍超分任务;算力充足时,可采用深度学习算法获得最佳效果.高倍超分宜采用深度学习算法,但需关注伪影问题.

To further enhance the spatial resolution of Gaofen-1(GF-1)wide-swath imagery and expand its applicability in national condition monitoring—particularly hydrology and water resources—this study investigated the single-image super-resolution(SISR)problem for GF-1 wide-swath data.Six representative algorithms were selected from two broad categories—interpolation-based methods and adversarial neural-network approaches—and implemented.A test set was constructed that spanned resolutions from 16 m down to 2 m,to validate the super-resolution performance and effectiveness of the six algorithms under low,medium,and high scaling(up-sampling)factors.Assessment metrics included mean-square error(MSE)and visual-fidelity indices.The experiments show that,for the medium-and low-factor SISR requirements of GF-1 wide-swath imagery,all six methods can meet the task;deep-learning-based algorithms achieve the highest reconstruction quality,whereas interpolation-based techniques provide a clear advantage in computational efficiency.Consequently,when efficiency is a prior-ity,interpolation methods are recommended for medium-and low-factor super-resolution tasks;when sufficient computing resources are available,deep-learning algorithms should be adopted to achieve optimal results.For high-factor super-resolution,deep-learning methods remain the preferred choice,but attention should be paid to the artifacts they may introduce.

李珏;沈鹏;高露雄

长江水利委员会水文局,湖北 武汉 43001095791部队,甘肃 酒泉 735018长江水利委员会水文局,湖北 武汉 430010

天文与地球科学

高分一号(GF-1)超分辨率机器学习深度学习

Gaofen-1(GF-1)super-resolutionmachine learningdeep learning

《北京测绘》 2026 (9)

1241-1246,6

10.19580/j.cnki.1007-3000.2025080063

评论