Li et al. (2025) CloudRuler: Rule-based transformer for cloud removal in Landsat images

Identification

• Journal: Remote Sensing of Environment
• Year: 2025
• Date: 2025-07-16
• Authors: Jun Li, Qinghong Sheng, Zhaocong Wu, Bo Wang, Xiao Ling, Xiang Liu, Yang Du, Fan Gao, Gustau Camps‐Valls, Matthieu Molinier
• DOI: 10.1016/j.rse.2025.114913

Research Groups

• College of Astronautics, Nanjing University of Aeronautics and Astronautics, China
• School of Remote Sensing and Information Engineering, Wuhan University, China
• Computer and Control Engineering College, Northeast Forestry University, China
• Institute of Communication Engineering, Army Engineering University of PLA, China
• Image Processing Laboratory, University of Valencia, Spain
• VTT Technical Research Centre of Finland Ltd, Finland

Short Summary

The study proposes CloudRuler, a rule-based transformer network that integrates a cloud physical model and domain-specific rules to effectively remove thin clouds from Landsat 8 and 9 imagery.

Objective

• To develop a cloud removal method that overcomes the limitations of existing deep learning models (which often ignore position information and physical models in thermal bands) and traditional physical models (which often overlook down-transmittance in optical bands).

Study Configuration

• Spatial Scale: Pixel-level and image-level analysis of Landsat satellite imagery.
• Temporal Scale: Multi-temporal approach utilizing paired images from Landsat 8 and Landsat 9.

Methodology and Data

• Models used: CloudRuler (a transformer network incorporating a Half-Spherical Coordinate System, remote sensing mosaicking, and a cloud physical model).
• Data sources: 20 paired Landsat 8 and 9 images.

Main Results

• CloudRuler outperformed seven baseline methods based on Generative Adversarial Networks (GAN), Convolutional Neural Networks (CNN), and transformers, both quantitatively and visually.
• Ablation studies confirmed that the rule-based modules (positional information, mosaicking, and the physical model) significantly improved thin cloud removal performance.
• The joint use of Landsat 8 and 9 images proved more effective and reliable for downstream applications than using a single satellite with a longer revisit period.

Contributions

• Integration of domain-specific rules (Half-Spherical Coordinate System and remote sensing mosaicking) into a transformer architecture to better distinguish semantic meanings of features.
• Implementation of a cloud physical model that solves parameters without the limitations found in previous models.
• Demonstration of the synergy between Landsat 8 and 9 data for enhancing the reliability of cloud-free imagery.

Funding

Not specified in the provided text.

Citation

@article{Li2025CloudRuler,
  author = {Li, Jun and Wang, Yihui and Sheng, Qinghong and Wu, Zhaocong and Wang, Bo and Ling, Xiao and Liu, Xiang and Du, Yang and Gao, Fan and Camps‐Valls, Gustau and Molinier, Matthieu},
  title = {CloudRuler: Rule-based transformer for cloud removal in Landsat images},
  journal = {Remote Sensing of Environment},
  year = {2025},
  doi = {10.1016/j.rse.2025.114913},
  url = {https://doi.org/10.1016/j.rse.2025.114913}
}