Generating spatiotemporally continuous land surface temperature (LST) data is in great demand for hydrology, meteorology, ecology, environmental studies, etc. However, the thermal infrared (TIR)-based LST measurements are prone to cloud contamination with missing pixels. To repair the missing pixels, a new XGBoost-based linking approach for reconstructing daytime and nighttime Moderate Resolution Imaging Spectroradiometer (MODIS) LST measurements was introduced. The instantaneous solar radiation and two soil-related predictors from China Data Assimilation System (CLDAS) 0.0625°/1-h data were selected as the linking variables to depict the relationship with instantaneous MODIS LST data. Other land surface properties, including two vegetation indices, the water index, the surface albedo, and topographic parameters, were also used as the predictor variables. The XGBoost method was used to fit an LST linking model by the training datasets from clear-sky pixels and was then applied to the MODIS Aqua-Terra LSTs during summer time (June to August) in 2017 and 2018 across China. The recovered LST data was further rectified with the Savitzky–Golay (SG) filtering method. The results showed the distribution of the reconstructed LSTs present a reasonable pattern for different land-cover types and topography. The evaluation results using in situ longwave radiation measurements showed the RMSE varies from 3.91 K to 5.53 K for the cloud-free pixels and from 4.42 K to 4.97 K for the cloud-covered pixels. In addition, the reconstructed LST products correlated well with CLDAS LST data with similar LST spatial patterns. The variable importance analysis revealed that the two soil-related predictors and the elevation variable are key parameters due to their great contribution to the XGBoost model performance.

Surface heat fluxes play a crucial role in the surface energy and water balance. In situ measurements are costly and difficult, and large‐scale flux mapping is hindered by surface heterogeneity. Previous studies have demonstrated that surface heat fluxes can be estimated by assimilating land surface temperature (LST) and soil moisture to determine two key parameters: a neutral bulk heat transfer coefficient (C_HN) and an evaporative fraction (EF). Here a methodology is proposed to estimate surface heat fluxes by assimilating Soil Moisture Active Passive (SMAP) soil moisture data and Geostationary Operational Environmental Satellite (GOES) LST data into a dual‐source (DS) model using a hybrid particle assimilation strategy. SMAP soil moisture data are assimilated using a particle filter (PF), and GOES LST data are assimilated using an adaptive particle batch smoother (APBS) to account for the large gap in the spatial and temporal resolution. The methodology is implemented in an area in the U.S. Southern Great Plains. Assessment against in situ observations suggests that soil moisture and LST estimates are in better agreement with observations after assimilation. The RMSD for 30 min (daytime) flux estimates is reduced by 6.3% (8.7%) and 31.6% (37%) for H and LE on average. Comparison against a LST‐only and a soil moisture‐only assimilation case suggests that despite the coarse resolution, assimilating SMAP soil moisture data is not only beneficial but also crucial for successful and robust flux estimation, particularly when the uncertainties in the model estimates are large.