Temporally dense 100-m land surface temperature retrieval via attention-based super-resolution deep learning

Land surface temperature (LST) is a key variable for governing surface water and energy exchanges and supports applications in hydrology, water management, climate monitoring, and agriculture. The ECOsystem Spaceborne Thermal Radiometer Experiment on Space Station (ECOSTRESS) offers LST observations at fine spatial resolution (70 m) but suffers from limited temporal coverage, cloud interference, and mission duration constraints. To overcome these limitations, we develop an Attention-based Super-Resolution deep Residual network (ASRRN) that generates 100-m LST estimates with dense temporal coverage using coarse-resolution MODIS LST and multiple auxiliary datasets. ASRRN integrates convolutional layers, residual learning, and attention mechanisms to enhance spatial feature extraction, while the model uncertainty is quantified using Monte Carlo dropout. We examine the contribution of auxiliary inputs, including Harmonized Landsat Sentinel (HLS) surface reflectance, Sentinel-1 SAR, and digital elevation model (DEM), and find that the configuration using MODIS LST with HLS (SWIR1, NDVI) and DEM yields performance comparable to ECOSTRESS. Cross-validation over 12 heterogeneous regions across the contiguous United States and two sites in Australia shows that ASRRN outperforms the Enhanced Spatial and Temporal Adaptive Reflectance Fusion Model (ESTARFM), achieving higher correlation (r = 0.984 vs. 0.898), lower RMSE (2.267 K vs. 5.764 K), and lower MAE (1.626 K vs. 4.193 K) against observations. It also outperforms the Super-Resolution Convolutional Neural Network (SRCNN) and the Enhanced Deep Super-Resolution network (EDSR). Overall, ASRRN enables generation of temporally dense, 100-m LST image time series without using fine-resolution thermal observations at estimation time, advancing high-resolution LST monitoring for various applications.