Reconstruction of all-weather land surface temperature using XGBoost and spatio-temporal models with multi-source data fusion

Land surface temperature (LST) retrieved from thermal infrared (TIR) remote sensing is frequently hampered by cloud contamination, resulting in significant spatiotemporal data gaps. Existing all-weather reconstruction methods rely on a single strategy: machine learning requires sufficient clear-sky samples and degrades under dense cloud, whereas spatiotemporal interpolation cannot capture the non-linear relationships between LST and its controlling factors. These limitations are most severe in heterogeneous regions. To address them, we propose a robust all-weather LST reconstruction framework that integrates the XGBoost machine learning algorithm with a dynamic spatiotemporal adjacency model. XGBoost is trained separately for each day with clear-sky MODIS LST as the target and GLDAS LST, DEM, latitude and NDVI as predictors, and then predicts the missing LST of cloud-contaminated pixels; for days with insufficient clear-sky samples, an adjacency weighting model based on the nearest cloud-free days is used instead. This yields a seamless 1-km all-weather LST dataset for the Beijing-Tianjin-Hebei (BTH) region of China, with accuracy maintained across the full range of cloud conditions. Validated against MODIS LST, the model achieves a mean absolute error (MAE) of 1.33 K and a root mean square error (RMSE) of 1.88 K over all pixels with a valid MODIS reference, and 1.32 K and 1.90 K over the 30% of clear-sky pixels withheld from training as pseudo-cloud pixels. An independent comparison with 2-m air temperature from 31 meteorological stations (ρ = 0.79) supports the spatiotemporal consistency of the reconstruction rather than its absolute accuracy. Accuracy is highest in winter (MAE = 0.99 K) and lowest in summer (MAE = 1.74 K). This framework thus overcomes TIR observational limitations, and provides high-resolution all-weather LST data for urban heat island monitoring, agricultural drought assessment, and regional climate change studies.