A Two-Stage Spatiotemporal Fusion Framework for Land Surface Temperature With Environmental Variable Correction

Land surface temperature (LST) is a key parameter in Earth system science, yet a trade-off between spatial and temporal resolution limits the quality of satellite products. Spatiotemporal fusion offers a viable solution, but existing methods often suffer from limited feature utilization and insufficient correction by environmental variables. To address these challenges, we propose a novel two-stage spatiotemporal fusion framework (TSFF). The framework first employs the BMSTF-GAN (Bidirectional Multi-feature Spatiotemporal Fusion Network with Generative Adversarial Network) to generate an initial high-resolution LST estimate. This network leverages a bidirectional structure to enhance temporal constraints and a multi-scale attention mechanism to capture complex spatial dependencies. Subsequently, a Land Surface Temperature Correction Network (LST-CorrNet) refines this estimate by performing a residual correction aided by environmental variables. LST-CorrNet incorporates heterogeneous auxiliary data—digital elevation model, total column water vapor, and longwave radiation—to account for thermodynamic influences from topography, the atmosphere, and surface radiative processes. TSFF performance was evaluated through testing using the European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis v.5 (ERA5)-Land and Moderate Resolution Imaging Spectroradiometer (MODIS) LST products. The final hourly, MODIS-resolution LST products show excellent agreement with ground truth data across diverse land cover types (mean absolute error: 2.29-3.50 K; root-mean-square error: 3.20-4.75 K; correlation coefficient: 0.96-0.98). The fused LST not only preserves fine spatial details but also accurately captures temporal dynamics. This study highlights the significant benefits of a two-stage approach that uses explicit environmental variables to correct and enhance deep learning-based spatiotemporal fusion models. The code will be available athttps://github.com/zarkyoung/TSFF.