Chen et al. (2026) A hybrid Penman-Monteith and machine learning model for simulating evapotranspiration and its components

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-01-21
• Authors: Han Chen, Stephen P. Good, Kelly Caylor, Richard P. Fiorella, Lixin Wang
• DOI: 10.1016/j.jhydrol.2026.134985

Research Groups

• Biological & Ecological Engineering, Oregon State University (OSU), Corvallis, OR, USA
• Water Resources Graduate Program, Oregon State University, Corvallis, OR, USA
• Institute of Surface-Earth System Science, School of Earth System Science, Tianjin University, Tianjin, China
• Bren School of Environmental Science and Management, University of California, Santa Barbara, Santa Barbara, CA, USA
• Earth and Environmental Sciences Division, Los Alamos National Laboratory, Los Alamos, NM, USA
• Department of Earth and Environmental Sciences, Indiana University Indianapolis, Indianapolis, IN, USA

Short Summary

This study develops Residual Neural Network–Penman–Monteith (RNN-PM), a novel hybrid model that integrates physical processes with machine learning to accurately simulate and partition evapotranspiration (ET) into soil evaporation and vegetation transpiration. Validated at NEON flux sites, RNN-PM reliably reproduces ET and its components, demonstrating superior performance and generalization compared to existing models.

Objective

• To develop and validate a novel hybrid dual-source evapotranspiration (ET) model, Residual Neural Network–Penman–Monteith (RNN-PM), capable of accurately simulating and explicitly partitioning total ET into its components (soil evaporation and vegetation transpiration), addressing limitations of existing hybrid models.

Study Configuration

• Spatial Scale: Multiple National Ecological Observatory Network (NEON) flux sites across various ecosystems.
• Temporal Scale: High-frequency measurements of partitioned soil evaporation and vegetation transpiration over an unspecified period.

Methodology and Data

• Models used:
  • RNN-PM (Residual Neural Network–Penman–Monteith): A hybrid dual-source ET model combining the Penman–Monteith framework with three specialized residual neural networks for estimating canopy, soil surface, and aerodynamic conductances.
  • Seven established dual-source ET models (for comparison).
  • Conventional machine learning models (for comparison).
  • Purely physical process-based models (for comparison).
• Data sources:
  • High-frequency partitioned soil evaporation (E) and vegetation transpiration (T) data from National Ecological Observatory Network (NEON) flux sites.

Main Results

• The RNN-PM model reliably reproduces total evapotranspiration (ET) and the transpiration fraction (T/ET).
• For ET simulation, RNN-PM achieved an average Kling–Gupta efficiency (KGE) of 0.89 and a root-mean-square error (RMSE) of 0.55 mm/day.
• For T/ET simulation, RNN-PM achieved a KGE of 0.87 with an RMSE of 0.06.
• RNN-PM demonstrated robust generalization capabilities, accurately simulating ET and its components beyond the initial training dataset, even under extreme climatic conditions.
• The RNN-PM model outperformed both conventional machine learning models and purely physical process-based models in simulating ET components in most cases.
• Among purely physical process-based dual-source models, those based on surface temperature decomposition showed improved performance with decreasing leaf area index (LAI), while conductance-based models declined with decreasing LAI.

Contributions

• Introduces RNN-PM, a novel hybrid dual-source ET model that explicitly partitions total ET into soil evaporation and vegetation transpiration, overcoming a significant limitation of most existing hybrid models.
• Achieves high accuracy in simulating both total ET and its components (T/ET), as evidenced by high KGE and low RMSE values.
• Demonstrates robust generalization capability, accurately simulating ET components under diverse and extreme climatic conditions, which is a common limitation for purely machine learning models.
• Provides a scalable approach for improving the representation of land–atmosphere interactions in Earth system models.
• Offers a comprehensive comparison with established physical and machine learning models, highlighting the superior performance of the hybrid RNN-PM approach.

Funding

Not specified in the provided text.

Citation

@article{Chen2026hybrid,
  author = {Chen, Han and Good, Stephen P. and Caylor, Kelly and Fiorella, Richard P. and Wang, Lixin},
  title = {A hybrid Penman-Monteith and machine learning model for simulating evapotranspiration and its components},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.134985},
  url = {https://doi.org/10.1016/j.jhydrol.2026.134985}
}