Superresolution of Land Surface Temperature Through Satellite Data Fusion

High-resolution land surface temperature (LST) is required for field-scale agriculture, heat-risk services, and land–atmosphere process studies, but existing products show a persistent spatial–temporal trade-off and strong cloud-induced gaps. We develop a hybrid superresolution framework that couples hourly ICON-EU LST with sporadic Landsat 8/9 thermal observations. A U-Net convolutional neural network is trained on 256×256-pixel tiles over central Europe using year-2023 pairs of ICON-EU inputs and five-step Landsat history, and validated on the independent year 2024. The fusion model reconstructs Landsat-scale LST with MAE of 2.55 °C and RMSE of 3.43 °C, improving on bilinear ICON-EU upscaling (MAE 3.24 °C; RMSE 4.40 °C). Qualitative examples show recovery of field and land-cover boundary thermal texture while preserving ICON-EU large-scale temperature level. The framework enables daily 100 m LST estimates independent of current satellite visibility and provides an open pipeline for reproducible NWP–satellite LST fusion.