A two-stage transformer framework for reconstructing hourly 1 km all-weather land surface temperature using cross-scale transfer learning

High spatiotemporal resolution land surface temperature (LST) reconstruction is essential for analyzing land-atmosphere interactions and thermal environmental dynamics. Thermal infrared observations, the primary source of regional and global LST, are affected by cloud contamination, scale discrepancies, and distribution shifts. Meanwhile, reconstruction methods struggle to provide both hourly continuity and kilometer-scale detail, and cross-scale fusion often causes spatial over-smoothing. To address these limitations, we propose ST-LSTrans, a two-stage Transformer framework for reconstructing regional hourly 1 km all-weather LST. It uses cross-scale transfer learning with global pre-training on hourly ∼6.25 km CLDAS LST to capture regional thermodynamic evolution and local fine-tuning on four-times-daily 1000 m MODIS LST to recover kilometer-scale thermal heterogeneity. A KL divergence constraint is introduced to reduce distribution shifts and spatial over-smoothing during cross-scale fusion. Validation results based on multi-source datasets show that the global pre-training stage achieved an average RMSE of 1.13 K, while the local fine-tuning stage achieved an average RMSE of 1.06 K, with the minimum RMSE reaching 0.87 K. Independent Landsat validation showed that the reconstructed product achieved a scene-averaged RMSE of 2.98 K and improved upon the original CLDAS background field. Additional evaluations against VIIRS VNP21, which is generated using an independent temperature-emissivity separation algorithm, confirmed the cross-sensor consistency of the reconstructed LST. Ground-based observations further demonstrated that ST-LSTrans preserved realistic diurnal temperature cycle characteristics under both clear-sky and persistent cloudy conditions. SHAP analysis further revealed scale-dependent transitions in LST driving mechanisms from coarse-scale atmospheric forcing to kilometer-scale terrain and land surface controls. This study provides a unified and interpretable modelling framework for high spatiotemporal resolution LST reconstruction and driving mechanism analysis.