Lu et al. (2026) Relaxing the clear-sky assumption: Cloud-tolerant spatiotemporal fusion via mask-guided feature modulation and temporal-memory collaboration

Identification

• Journal: International Journal of Applied Earth Observation and Geoinformation
• Year: 2026
• Date: 2026-09-21
• Authors: Sichen Lu, Juanjuan Jing, Junhua Yu, Lei Yang, Boyang Nie, Jinsong Zhou
• DOI: 10.1016/j.jag.2026.105571

Research Groups

• Aerospace Information Research Institute, Chinese Academy of Sciences
• School of Optoelectronics, University of Chinese Academy of Sciences
• School of Artificial Intelligence, China University of Geosciences Beijing
• State Key Laboratory of Internet of Things for Smart City and Department of Ocean Science and Technology, University of Macau
• School of Remote Sensing and Information Engineering, Wuhan University

Short Summary

This paper proposes a cloud-tolerant spatiotemporal fusion framework called CloudSTF, which enables the reconstruction of high-resolution images from cloud-contaminated observations. The framework integrates a Mask-guided Multi-scale Swin Transformer (M2ST) encoder and a Cross-temporal Memory-guided Fusion (CTMF) module to effectively exploit partially occluded observations.

Objective

• To develop a robust framework for reconstructing land-surface states from cloud-contaminated observations.
• To evaluate the performance of CloudSTF under realistic atmospheric conditions using the Global Cloud-shrouded Regions (GCR-STF) benchmark.

Study Configuration

• Spatial Scale: 30 m to 500 m
• Temporal Scale: Daily to weekly

Methodology and Data

• Models used: Mask-guided Multi-scale Swin Transformer (M2ST) encoder, Cross-temporal Memory-guided Fusion (CTMF) module
• Data sources: Landsat-8/9 OLI, MODIS Version 6.1 products

Main Results

• CloudSTF consistently outperforms state-of-the-art methods on the DX and LGC datasets.
• The framework achieves a significant reduction in Root Mean Square Error (RMSE) and an increase in Structural Similarity Index (SSIM) compared to traditional STF methods.

Contributions

• This study proposes a novel cloud-tolerant spatiotemporal fusion framework that can effectively reconstruct high-resolution images from cloud-contaminated observations.
• The GCR-STF benchmark provides a comprehensive evaluation of CloudSTF under realistic atmospheric conditions, demonstrating its robustness and applicability in various landscapes.

Funding

• This research was funded by the National Natural Science Foundation of China (Grant No. 62172145) and the Chinese Academy of Sciences (CAS) Strategic Priority Research Program (Grant No. XDA19040201).

Citation

@article{Lu2026Relaxing,
  author = {Lu, Sichen and Jing, Juanjuan and Yu, Junhua and Yang, Lei and Nie, Boyang and Zhou, Jinsong},
  title = {Relaxing the clear-sky assumption: Cloud-tolerant spatiotemporal fusion via mask-guided feature modulation and temporal-memory collaboration},
  journal = {International Journal of Applied Earth Observation and Geoinformation},
  year = {2026},
  doi = {10.1016/j.jag.2026.105571},
  url = {https://doi.org/10.1016/j.jag.2026.105571}
}