This paper proposes a coupled mechanism model-machine learning (MM-ML) framework that integrates physical constraints and data-driven learning to address the challenges of accurately estimating land surface temperature (LST) under heterogeneous land cover and extreme atmospheric conditions. To address the bias of existing separating window (SW) algorithms in wet environments and the poor interpretability and generalization performance of pure machine learning (ML) methods due to data limitations, we integrate radiative transfer modeling and data components, utilize MODTRAN simulations and global atmospheric profiles, and apply physically constrained optimization. Validated on 4,450 observations from 29 global observation sites, MM-ML achieves a mean error of 1.84K, a root mean square error (RMSE) of 2.55K, and an R-squared of 0.966, outperforming existing methods and reducing error by more than 50%, particularly under extreme conditions. Sensitivity analysis results showed that LST estimates were most sensitive to sensor irradiance, but also sensitive to water vapor, but less sensitive to emissivity, and MM-ML demonstrated excellent stability. In conclusion, this study demonstrates the effectiveness of the MM-ML framework, which combines physical interpretability and nonlinear modeling capabilities to support reliable LST calculations in complex environments and support climate monitoring and ecosystem research.
