Wang et al. (2025) Estimation and mechanism analysis of global evapotranspiration based on a physics-informed deep-learning model

Identification

• Journal: Journal of Hydrology
• Year: 2025
• Date: 2025-10-01
• Authors: Jiancheng Wang, Tongren Xu, Shaomin Liu, Dongkyun Kim, Changhyun Jun, Sayed M. Bateni, Xiaoyan Li, Xin Li, Xiaofan Yang, Ziwei XU, Gangqiang Zhang, Wenting Ming
• DOI: 10.1016/j.jhydrol.2025.134351

Research Groups

• State Key Laboratory of Earth Surface Processes and Disaster Risk Reduction, Faculty of Geographical Science, Beijing Normal University, Beijing, China
• Department of Civil Engineering, Hongik University, Mapo-Gu, Seoul, the Republic of Korea
• School of Civil, Environmental and Architectural Engineering, College of Engineering, Korea University, Seoul, the Republic of Korea
• Department of Civil, Environmental and Construction Engineering & Water Resources Research Center, University of Hawaii at Manoa, Honolulu, HI, USA
• UNESCO-UNISA Africa Chair in Nanoscience and Nanotechnology College of Graduates Studies, University of South Africa, Muckleneuk Ridge, Pretoria, South Africa
• National Tibetan Plateau Data Center, State Key Laboratory of Tibetan Plateau Earth System, Environment and Resources, Institute of Tibetan Plateau Research, Chinese Academy of Sciences, Beijing, China

Short Summary

This paper introduces a physics-informed deep-learning model, Self-attention Influence (SAI), for global evapotranspiration (ET) estimation, demonstrating superior spatial extrapolation and robustness, especially in data-poor regions, and providing explainable insights into ET mechanisms and climate impacts.

Objective

• To develop a physics-informed deep-learning model (SAI) that enhances the accuracy, spatial extrapolation, and environmental adaptability of global evapotranspiration (ET) estimation, particularly in data-poor regions, by overcoming data imbalance issues.
• To utilize the explainable features of the SAI model to analyze the mechanisms of global ET, including the effects of El Niño-Southern Oscillation and temperature.

Study Configuration

• Spatial Scale: Global, site-scale, and basin-scale, with a specific focus on data-poor regions such as South America and Africa.
• Temporal Scale: Analysis of global ET from 2000 to 2021.

Methodology and Data

• Models used: Self-attention Influence (SAI) method, a physics-informed deep-learning model that integrates environment cognition and parameter calibration learning.
• Data sources: In-situ observations (for model training and evaluation), and implied global datasets for ET estimation and analysis of climate phenomena.

Main Results

• At site-scale, the SAI model reduced the Root Mean Square Error (RMSE) by 25.15 % and increased the Coefficient of Determination (R²) by 38.35 % in data-poor regions (e.g., South America and Africa) compared to pure machine learning methods.
• At basin-scale, the SAI model exhibited the lowest estimation uncertainty of 2.79 mm/month in data-poor basins.
• Based on the water balance method, the SAI model achieved the highest estimation accuracy with a centered RMSE of 20.97 mm/month.
• The model realized a collaborative breakthrough in accuracy, adaptability, and explainability, effectively addressing data imbalance problems globally.
• The explainable method confirmed the physical consistency of the physics-informed components within the SAI model, validating its spatial extrapolation capabilities.
• The study revealed the global effects of El Niño-Southern Oscillation and temperature on ET during the period 2000 to 2021.

Contributions

• Introduction of the novel Self-attention Influence (SAI) model, a physics-informed deep-learning approach that significantly improves global evapotranspiration (ET) estimation.
• Demonstrated enhanced spatial extrapolation, environmental adaptability, and estimation robustness, particularly in data-poor regions, addressing a critical limitation of existing machine learning models.
• Achieved a collaborative breakthrough in accuracy, adaptability, and explainability for ET estimation, effectively overcoming data imbalance challenges.
• Provided an explainable framework to analyze ET mechanisms, confirming the physical consistency of the model and revealing the global impacts of climate drivers like El Niño-Southern Oscillation and temperature.

Funding

[No funding information was provided in the paper text.]

Citation

@article{Wang2025Estimation,
  author = {Wang, Jiancheng and Xu, Tongren and Liu, Shaomin and Kim, Dongkyun and Jun, Changhyun and Bateni, Sayed M. and Li, Xiaoyan and Li, Xin and Yang, Xiaofan and XU, Ziwei and Zhang, Gangqiang and Ming, Wenting},
  title = {Estimation and mechanism analysis of global evapotranspiration based on a physics-informed deep-learning model},
  journal = {Journal of Hydrology},
  year = {2025},
  doi = {10.1016/j.jhydrol.2025.134351},
  url = {https://doi.org/10.1016/j.jhydrol.2025.134351}
}