Remote Sensing Technology and Application, Volume. 40, Issue 4, 969(2025)
Super-Resolution Reconstruction Methods for Remote Sensing Images: A Review and Experiments
Figures & Tables(19)
Fig. 1. Development of super resolution methods for remote sensing images
Fig. 2. Schematic diagram of interpolated image super-resolution reconstruction
Fig. 3. Schematic diagram of super-resolution reconstruction of reconstructed images
Fig. 4. Schematic diagram of image super-resolution reconstruction using shallow learning
Fig. 11. Histogram of the result evaluation scores of each super-resolution reconstruction method
Fig. 12. Qualitative comparison of ×4 super-resolution reconstruction of NWPU-RESISC45 Dataset （wetland-028） images by different SR methods
Fig. 13. Qualitative comparison of ×4 super-resolution reconstruction of NWPU-RESISC45 Dataset （wetland-160） images by different SR methods
Fig. 14. Qualitative comparison of ×4 super-resolution reconstruction of NWPU-RESISC45 Dataset （wetland-360） images by different SR methods
Fig. 15. Qualitative comparison of ×4 super-resolution reconstruction of NWPU-RESISC45 Dataset （wetland-169） images by different SR methods
Table 1.
Commonly used data sets based on deep learning super-resolution models View tableView in ArticleTable 1.
Commonly used data sets based on deep learning super-resolution models 数据集名称 分辨率 样本
数量
格式 内容 DIV2K 2 560✕1 440 1 000 PNG 人物、动物、风景等 Flickr2K 2 560✕1 440 2 650 PNG 人物、动物、风景等 OST 553✕440 10 624 PNG 户外场景 FFHQ 1 024✕1 024 70 000 PNG 人脸 CelebA-HQ 1 024✕1 024 30 000 JPG 人脸 Set5 512✕512 5 PNG 人物、自然场景等 Set14 500✕480 14 PNG 人物、自然场景等 BSD100 481✕321 100 JPG 自然场景 Urban100 984✕797 100 PNG 城市街景 AID 600✕600 10 000 JPG 自然场景 UC Merced 256✕256 2 100 PNG 自然场景 WHU-RS19 600✕600 1 005 TIF 自然场景 NWPU-RESISC45 256✕256 31 500 PNG 自然场景
Table 2.
Training parameter configuration View tableView in ArticleTable 2.
Training parameter configuration 实验环境/参数 软/硬件型号及参数 操作系统 Windows11 CPU 英特尔® Xeon(R) Gold 6248 2.99GHz GPU NVIDIA GTX A5000 24GB 内存 64GB CUDA V11.7 torch 2.0.1-gpu python 3.9 迭代次数 10 000 批次 16 初始学习率lr0 0.000 1 学习率lr 0.01 优化器 Adam 线程数 8
Table 3.
Quantitative comparison of results of various super-resolution reconstruction methods View tableView in ArticleTable 3.
Quantitative comparison of results of various super-resolution reconstruction methods 类别 方法 PSNR↑ SSIM↑ LPIPS↓ NIQE↓ PI↓ 插值 Bicubic 29.32 0.67 0.4427 7.1095 7.7862 CNN
SRCNN 25.55 0.53 0.3993 9.8029 8.3432 VDSR 26.10 0.55 0.2609 8.1731 6.9092 DRRN 33.89 0.87 0.1841 7.1886 6.1021 GAN SRGAN 26.05 0.54 0.2048 4.0019 3.8272 ESRGAN 26.58 0.57 0.1530 5.4197 4.1732 DDPM EDiffSR 27.68 0.61 0.1306 4.2200 3.8162
Table 4.
Parameters and average inference time of different models View tableView in ArticleTable 4.
Parameters and average inference time of different models 方法 网络结构 参数量（M） 平均推理时间/ms 是否支持实时部署 Bicubic - - < 1 √ SRCNN CNN 57K 3.20 √ VDSR ResNet 0.67 8.50 × SRGAN GAN 16.7 36.10 × ESRGAN GAN 48.6 23.54 × EDiffSR Diffusion 23.5 50.30 ×
Tools
Get Citation
Copy Citation Text
Chengquan ZOU, Jiangcheng HUANG, Zhengbao SUN, Yutong YANG. Super-Resolution Reconstruction Methods for Remote Sensing Images: A Review and Experiments[J]. Remote Sensing Technology and Application, 2025, 40(4): 969
Paper Information
Category:
Received: Sep. 16, 2024
Accepted: --
Published Online: Aug. 26, 2025
The Author Email: Zhengbao SUN (zbsun@ynu.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20