Research on Cross-Sensor Remote Sensing Image Super-Resolution Method Based on Diffusion Models

In remote sensing image (RSI) super-resolution (SR), traditional deep learning methods have made remarkable progress. However, these methods struggle to handle the complex mapping between cross-sensor images. Although generative adversarial networks can reconstruct fine details, their training proce...

Full description

Saved in:
Bibliographic Details
Main Authors: Ru Miao, Kai Yang, Ke Zhou, Jia Song, Shihao Fu, Cong Liu, Yuanxing Wang
Format: Article
Language:English
Published: IEEE 2025-01-01
Series:IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Subjects:
Online Access:https://ieeexplore.ieee.org/document/11085125/
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1849391075047768064
author Ru Miao
Kai Yang
Ke Zhou
Jia Song
Shihao Fu
Cong Liu
Yuanxing Wang
author_facet Ru Miao
Kai Yang
Ke Zhou
Jia Song
Shihao Fu
Cong Liu
Yuanxing Wang
author_sort Ru Miao
collection DOAJ
description In remote sensing image (RSI) super-resolution (SR), traditional deep learning methods have made remarkable progress. However, these methods struggle to handle the complex mapping between cross-sensor images. Although generative adversarial networks can reconstruct fine details, their training process is relatively challenging. Recently, diffusion-based generative models have effectively improved the visual quality of reconstructed images through the iterative reverse diffusion mechanism. Therefore, this study proposes a cross-sensor RSI SR method based on diffusion models, which aims to effectively reconstruct low-resolution (LR) RSI into high-resolution (HR) images, addressing sensor differences and information loss. To extract rich prior information from LR images and improve image reconstruction, we designed a prior feature extraction module, which aids in recovering HR features. In addition, we designed an efficient sampling strategy that starts with LR images and progressively adds noise, replacing the traditional method that begins with pure noise. This approach leverages the LR image as an approximation of the intermediate state in the Markov chain, reducing the number of steps required by the diffusion model and improving generation efficiency. Finally, we used RSI from the Sentinel-2 (10 m resolution) and Gaofen-2 (0.8 m resolution) sensors to validate the reconstruction performance of the diffusion model. Experimental results confirm that the proposed method effectively utilizes LR priors for HR reconstruction while reducing computational costs.
format Article
id doaj-art-d8c91feb509045aa9a96d361675ffc82
institution Kabale University
issn 1939-1404
2151-1535
language English
publishDate 2025-01-01
publisher IEEE
record_format Article
series IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
spelling doaj-art-d8c91feb509045aa9a96d361675ffc822025-08-20T03:41:11ZengIEEEIEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing1939-14042151-15352025-01-0118185281854210.1109/JSTARS.2025.359068711085125Research on Cross-Sensor Remote Sensing Image Super-Resolution Method Based on Diffusion ModelsRu Miao0https://orcid.org/0000-0003-0349-2971Kai Yang1https://orcid.org/0009-0008-9656-8392Ke Zhou2https://orcid.org/0009-0006-6739-4448Jia Song3https://orcid.org/0000-0002-9051-1925Shihao Fu4https://orcid.org/0009-0003-4194-7997Cong Liu5Yuanxing Wang6School of Computer and Information Engineering, Henan Engineering Research Center of Spatial Information Processing, Henan University, Kaifeng, ChinaSchool of Computer and Information Engineering, Henan Engineering Research Center of Spatial Information Processing, Henan Technology Innovation Center of Spatio-Temporal Big Data, Henan University, Kaifeng, ChinaSchool of Computer and Information Engineering, Henan Engineering Research Center of Spatial Information Processing, Henan Technology Innovation Center of Spatio-Temporal Big Data, Henan University, Kaifeng, ChinaState Key Laboratory of Resources and Environmental Information System, Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences, Beijing, ChinaState Key Laboratory of Resources and Environmental Information System, Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences, Beijing, ChinaSchool of Computer and Information Engineering, Henan Engineering Research Center of Spatial Information Processing, Henan Technology Innovation Center of Spatio-Temporal Big Data, Henan University, Kaifeng, ChinaSchool of Computer and Information Engineering, Henan Engineering Research Center of Spatial Information Processing, Henan Technology Innovation Center of Spatio-Temporal Big Data, Henan University, Kaifeng, ChinaIn remote sensing image (RSI) super-resolution (SR), traditional deep learning methods have made remarkable progress. However, these methods struggle to handle the complex mapping between cross-sensor images. Although generative adversarial networks can reconstruct fine details, their training process is relatively challenging. Recently, diffusion-based generative models have effectively improved the visual quality of reconstructed images through the iterative reverse diffusion mechanism. Therefore, this study proposes a cross-sensor RSI SR method based on diffusion models, which aims to effectively reconstruct low-resolution (LR) RSI into high-resolution (HR) images, addressing sensor differences and information loss. To extract rich prior information from LR images and improve image reconstruction, we designed a prior feature extraction module, which aids in recovering HR features. In addition, we designed an efficient sampling strategy that starts with LR images and progressively adds noise, replacing the traditional method that begins with pure noise. This approach leverages the LR image as an approximation of the intermediate state in the Markov chain, reducing the number of steps required by the diffusion model and improving generation efficiency. Finally, we used RSI from the Sentinel-2 (10 m resolution) and Gaofen-2 (0.8 m resolution) sensors to validate the reconstruction performance of the diffusion model. Experimental results confirm that the proposed method effectively utilizes LR priors for HR reconstruction while reducing computational costs.https://ieeexplore.ieee.org/document/11085125/Deep learningdiffusion modelsgenerative modelsremote sensingsuper-resolution (SR)
spellingShingle Ru Miao
Kai Yang
Ke Zhou
Jia Song
Shihao Fu
Cong Liu
Yuanxing Wang
Research on Cross-Sensor Remote Sensing Image Super-Resolution Method Based on Diffusion Models
IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Deep learning
diffusion models
generative models
remote sensing
super-resolution (SR)
title Research on Cross-Sensor Remote Sensing Image Super-Resolution Method Based on Diffusion Models
title_full Research on Cross-Sensor Remote Sensing Image Super-Resolution Method Based on Diffusion Models
title_fullStr Research on Cross-Sensor Remote Sensing Image Super-Resolution Method Based on Diffusion Models
title_full_unstemmed Research on Cross-Sensor Remote Sensing Image Super-Resolution Method Based on Diffusion Models
title_short Research on Cross-Sensor Remote Sensing Image Super-Resolution Method Based on Diffusion Models
title_sort research on cross sensor remote sensing image super resolution method based on diffusion models
topic Deep learning
diffusion models
generative models
remote sensing
super-resolution (SR)
url https://ieeexplore.ieee.org/document/11085125/
work_keys_str_mv AT rumiao researchoncrosssensorremotesensingimagesuperresolutionmethodbasedondiffusionmodels
AT kaiyang researchoncrosssensorremotesensingimagesuperresolutionmethodbasedondiffusionmodels
AT kezhou researchoncrosssensorremotesensingimagesuperresolutionmethodbasedondiffusionmodels
AT jiasong researchoncrosssensorremotesensingimagesuperresolutionmethodbasedondiffusionmodels
AT shihaofu researchoncrosssensorremotesensingimagesuperresolutionmethodbasedondiffusionmodels
AT congliu researchoncrosssensorremotesensingimagesuperresolutionmethodbasedondiffusionmodels
AT yuanxingwang researchoncrosssensorremotesensingimagesuperresolutionmethodbasedondiffusionmodels