Abstract
This dissertation addresses the persistent challenge of missing or cloud-obscured optical satellite images, which hinder timely and complete Earth observation. Cloud cover disrupts remote sensing applications such as change detection and environmental monitoring, where full-scene visibility is essential. Since consistently cloud-free imagery is often unattainable, this research focuses on enhancing the effectiveness and reliability of reconstructed images. Effectiveness refers to the ability to restore cloud-covered regions for practical use, while reliability ensures the reconstructed data accurately reflects the Earth’s surface.
To achieve this, the thesis proposes multi-source data fusion methods, integrating SAR and optical imagery to improve image quality in real-world applications like crop mapping and post-flood reconstruction. Four key objectives guide the research:
1. Feature Enhancement Network (FENet) is introduced to overcome limitations in existing SAR-optical fusion methods, which often struggle with accuracy and scalability. FENet learns residual maps to preserve clear pixels and reconstruct obscured areas. It incorporates a Linear Attention Mechanism (LAM) to capture global context, enhancing reconstruction accuracy. Experiments show FENet delivers superior visual consistency and faster processing, making it suitable for large-scale deployment.
2. To address the need for efficient processing of large datasets, the Feature Pyramid Network (FPNet) is developed. FPNet combines multi-scale feature extraction with a streamlined fusion design, significantly reducing computational load. It achieves inference speeds up to 96 FPS and shortens training time, demonstrating strong practical efficiency.
3. The third objective evaluates how reconstructed cloud-free images impact downstream tasks, specifically crop type mapping. Results indicate that improved image quality modestly enhances classification accuracy and robustness, validating the utility of SAR-optical fusion in agricultural analysis.
4. Finally, the Cloud-Resilient Generation Network (CRGenNet) is proposed to enable full-scene image recovery under heavy cloud conditions. Using dual-temporal and dual-modal inputs, CRGenNet integrates DownUpBlock and FusionAttention modules to effectively fuse SAR and optical data. It consistently performs well across diverse land cover types and environmental challenges.
In summary, this thesis presents deep learning-based methods that significantly improve optical satellite image quality under cloud contamination, supporting reliable Earth observation when clear imagery is limited.
To achieve this, the thesis proposes multi-source data fusion methods, integrating SAR and optical imagery to improve image quality in real-world applications like crop mapping and post-flood reconstruction. Four key objectives guide the research:
1. Feature Enhancement Network (FENet) is introduced to overcome limitations in existing SAR-optical fusion methods, which often struggle with accuracy and scalability. FENet learns residual maps to preserve clear pixels and reconstruct obscured areas. It incorporates a Linear Attention Mechanism (LAM) to capture global context, enhancing reconstruction accuracy. Experiments show FENet delivers superior visual consistency and faster processing, making it suitable for large-scale deployment.
2. To address the need for efficient processing of large datasets, the Feature Pyramid Network (FPNet) is developed. FPNet combines multi-scale feature extraction with a streamlined fusion design, significantly reducing computational load. It achieves inference speeds up to 96 FPS and shortens training time, demonstrating strong practical efficiency.
3. The third objective evaluates how reconstructed cloud-free images impact downstream tasks, specifically crop type mapping. Results indicate that improved image quality modestly enhances classification accuracy and robustness, validating the utility of SAR-optical fusion in agricultural analysis.
4. Finally, the Cloud-Resilient Generation Network (CRGenNet) is proposed to enable full-scene image recovery under heavy cloud conditions. Using dual-temporal and dual-modal inputs, CRGenNet integrates DownUpBlock and FusionAttention modules to effectively fuse SAR and optical data. It consistently performs well across diverse land cover types and environmental challenges.
In summary, this thesis presents deep learning-based methods that significantly improve optical satellite image quality under cloud contamination, supporting reliable Earth observation when clear imagery is limited.
| Original language | English |
|---|---|
| Qualification | Doctor of Philosophy |
| Awarding Institution |
|
| Supervisors/Advisors |
|
| Award date | 18 Nov 2025 |
| Place of Publication | Enschede |
| Publisher | |
| Print ISBNs | 978-90-365-6939-2 |
| Electronic ISBNs | 978-90-365-6940-8 |
| DOIs | |
| Publication status | Published - 18 Nov 2025 |
Fingerprint
Dive into the research topics of 'Satellite SAR-optical image fusion for cloud-free image generation'. Together they form a unique fingerprint.Cite this
- APA
- Author
- BIBTEX
- Harvard
- Standard
- RIS
- Vancouver