Physics-constrained machine learning for satellite-derived evapotranspiration in China

Evapotranspiration estimates for China derived from empirical and physical models exhibit limited agreement in magnitude and trends, due to the complexity of surface characteristics and climatic conditions. Advances in knowledge-guided data-driven methods allow hybrid models that integrate remote sensing-based ET algorithms with machine learning (ML) to mitigate these discrepancies effectively. We developed two hybrid models, ML_PM and ML_SEB, by coupling the Penman Monteith (PM) and surface energy balance (SEB) algorithm with three ML methods: Artificial Neural Networks (ANN), Random Forest (RF), and Light Gradient Boosting Machine (LGBM), and then simulated China’s ET from 2001 to 2022. The hybrid models ML_PM, ML_SEB, replaced empirical formulas of the critical but uncertain parameters: surface resistance (rs) and aerodynamic resistance (ra) with ML, respectively, considering energy balance and turbulent diffusion processes. Hybrid modeling enhanced the representation of physical processes while preserving high estimation accuracy. We evaluated the hybrid models using 63 eddy covariance (EC) sites across China and compared them with process-based physical models and pure ML. Results indicated that the hybrid models ML_PM and ML_SEB substantially improved the performance of the PM (NSE: 0.49) and SEB (NSE: −0.53), achieving NSE values of over 0.8 and 0.6, respectively. ML_PM, in particular, performed on par with pure ML overall and outperformed them for grassland and forest ecosystems, owing to its stronger physical foundation. Moreover, the hybrid models demonstrated greater robustness and generalization under sparse sampling and extreme events compared to pure ML. The optimal hybrid model (RF_PM) produced an annual mean ET of 580.33 ± 9.01 mm yr⁻¹ for China over 2001–2022, with a statistically insignificant increasing trend (0.52 mm yr⁻², p > 0.05). Overall, hybrid modeling achieved an optimal balance between physical mechanism and estimation accuracy, offering a new perspective for ET estimation and enhancing our understanding of hydrological processes at regional and global scales under climate change.