Acquiring high-spatiotemporal-resolution Land Surface Temperature (LST) data in cloud-prone complex terrain faces dual challenges: cloud-induced data gaps and topography-induced texture blurring. To address these issues, this study proposes a novel framework coupling Noah-MP physical simulation with a Terrain-driven Error Decoupling and Multi-scale Adaptive Weighting (TED-MSAW) engine to generate a 1-km hourly all-weather LST dataset. First, Noah-MP is employed to establish spatiotemporally continuous ”physical anchors”, ensuring that gap-filling complies with surface energy balance laws. Second, the TED-MSAW engine, utilizing a Generalized Bayesian Triple Collocation scheme, achieves ”implicit perception” of surface complexity. It dynamically decouples terrain-induced systematic errors and adaptively adjusts fusion weights, effectively suppressing noise while preserving fine thermal textures. Validations show favorable agreement with in-situ observations, yielding an RMSE of 1.74 K at the validation sites. Furthermore, a clear-sky intercomparison against Fengyun-4B (FY-4B) products yields an RMSE of 2.65 K, which serves as additional evidence of the framework’s consistency. SHAP-based attribution analysis reveals that land cover type and elevation are dominant error drivers, exceeding slope. The model demonstrates the ability to mitigate geometric biases through its adaptive compensation mechanism. This study presents a generalizable solution integrating physical mechanisms with statistical learning to attempt to resolve the spatiotemporal resolution trade-off in complex environments.
Zhou et al. (Fri,) studied this question.