Evaluation and Application of the PT-JPL Physical Model Optimized with XGBoost Algorithm in Latent Heat Flux Estimation

The static parameterization scheme in the Priestley-Taylor Jet Propulsion Laboratory (PT-JPL) model limits the dynamic capture of latent heat flux (LE) in different plant functional types (PFTs). Therefore, this study employs the Extreme Gradient Boosting (XGBoost) algorithm to optimize the constraint factors and sub-models in the PT-JPL model that are influenced by sensitive prior parameters, thereby constructing five hybrid models under the PT-JPL physical constraint framework to achieve the dynamic response of fixed prior parameters and to integrate ensemble learning (EML) with the process-based framework, ensuring that physical mechanism and high precision coexist. Comparative analysis and validation across five PFTs in the Heihe River Basin of China reveal that the XGB-LEc-PT-JPL model, optimized for vegetation transpiration, exhibits the best comprehensive performance and outperforms the pure data-driven model in several aspects. Regarding overall accuracy, the MAE and RMSE are 15.47 W/m2 and 23.85 W/m2, respectively. Although hybrid models optimized for deeper constraint factors sometimes exceed the simulation accuracy of the XGB-LEc-PT-JPL model, they often exhibit reduced parameter generalization, increasing model uncertainty. Finally, the regional scale comparison of different models reveals a consistent spatial pattern, and the XGB-LEc-PT-JPL model can still achieve good simulation accuracy. This study combines EML with physical model, providing scientific insights for understanding hydrological processes under regional climate change, as well as for ecological water resource conservation and optimal water resource allocation.