Luo et al. (2026) A two-stage transformer framework for reconstructing hourly 1 km all-weather land surface temperature using cross-scale transfer learning

Identification

• Journal: International Journal of Applied Earth Observation and Geoinformation
• Year: 2026
• Date: 2026-09-19
• Authors: Yuanyuan Luo, Sha Zhang, Kun Qiao, Yun Bai
• DOI: 10.1016/j.jag.2026.105602

Research Groups

• School of Geographical Sciences, Hebei Normal University
• Hebei Technology Innovation Center for Remote Sensing Identification of Environmental Change
• Hebei Key Laboratory of Environmental Change and Ecological Construction
• Hebei Key Research Institute of Humanities and Social Sciences at Universities “GeoComputation and Planning Center of Hebei Normal University”

Short Summary

This study proposes a two-stage Transformer framework, ST-LSTrans, for reconstructing hourly 1 km all-weather land surface temperature (LST) using cross-scale transfer learning. The global pre-training stage achieves an average RMSE of 1.13 K, while the local fine-tuning stage achieves an average RMSE of 1.06 K.

Objective

• To develop a high-spatiotemporal-resolution LST reconstruction framework that can capture regional thermodynamic evolution and local fine-scale thermal heterogeneity.
• To address the limitations of existing multi-source fusion frameworks in preserving regional thermal continuity and local spatial heterogeneity.

Study Configuration

• Spatial Scale: 1 km resolution, covering an area of approximately 1.46 million km2 centered on the North China Plain.
• Temporal Scale: Hourly temporal resolution for LST reconstruction.

Methodology and Data

• Models used: Two-stage Transformer framework (ST-LSTrans) with global pre-training and local fine-tuning stages.
• Data sources:
  • MODIS LST products (MOD11A1 and MYD11A1)
  • CLDAS V2.0 land surface assimilation and reanalysis data
  • ERA5 meteorological forcing variables
  • SMAP soil moisture
  • MODIS NDVI
  • SRTM DEM-derived topographic features
  • CLCD land cover

Main Results

• The global pre-training stage achieves an average RMSE of 1.13 K.
• The local fine-tuning stage achieves an average RMSE of 1.06 K, with a minimum RMSE of 0.87 K.
• Independent Landsat validation shows that the reconstructed product achieves a scene-averaged RMSE of 2.98 K and improves upon the original CLDAS background field.

Contributions

• The proposed ST-LSTrans framework provides a unified and interpretable modelling framework for high-spatiotemporal-resolution LST reconstruction.
• The study demonstrates the effectiveness of staged transfer learning in addressing the limitations of existing multi-source fusion frameworks.

Funding

• This research was funded by the National Natural Science Foundation of China (Grant No. 41821003).
• The authors acknowledge support from the Hebei Provincial Key Research and Development Program (Grant No. 2021003001).

Citation

@article{Luo2026twostage,
  author = {Luo, Yuanyuan and Zhang, Sha and Qiao, Kun and Bai, Yun},
  title = {A two-stage transformer framework for reconstructing hourly 1 km all-weather land surface temperature using cross-scale transfer learning},
  journal = {International Journal of Applied Earth Observation and Geoinformation},
  year = {2026},
  doi = {10.1016/j.jag.2026.105602},
  url = {https://doi.org/10.1016/j.jag.2026.105602}
}