Xie et al. (2026) Improving evapotranspiration estimation by integrating process-based biophysical variables into a deep learning approach

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2026
• Date: 2026-01-08
• Authors: Mingming Xie, Jianyun Zhang, Zhenxin Bao, Linus Zhang, Zheng Duan, Guoqing Wang, Cuishan Liu, Feifei Yuan, Xiaoxiang Guan
• DOI: 10.1016/j.ejrh.2026.103114

Research Groups

• College of Hydrology and Water Resources, Hohai University, Nanjing, China
• The National Key Laboratory of Water Disaster Prevention, Nanjing Hydraulic Research Institute, Nanjing, China
• Research Center for Climate Change, Ministry of Water Resources, Nanjing, China
• Yangtze Institute for Conservation and Development, Nanjing, China
• Division of Water Resources Engineering, LTH, Lund University, Lund, Sweden
• Department of Physical Geography and Ecosystem Science, Lund University, Lund, Sweden
• GFZ Helmholtz Centre for Geosciences, Section Hydrology, Potsdam, Germany

Short Summary

This study developed a hybrid Penman-Monteith-Leuning (PML) and Long Short-Term Memory (LSTM) model (PML-LSTM) to improve evapotranspiration (ET) estimation by integrating process-based biophysical variables into a deep learning framework. The PML-LSTM model demonstrated superior accuracy and generalization ability across diverse vegetation types and extreme weather conditions compared to standalone PML and LSTM models.

Objective

• To assess the ET estimation performance of the PML, LSTM, and PML-LSTM models and compare their applicability across different vegetation types and weather conditions.
• To analyze the temporal and spatial generalization capabilities of the models at unseen sites and under extreme weather conditions.
• To explore the effectiveness of incorporating biophysical process variables in improving the ET estimation performance of LSTM models.

Study Configuration

• Spatial Scale: 103 FLUXNET2015 flux tower sites globally, representing 10 major vegetation types across diverse climatic and ecological regions.
• Temporal Scale: Daily, multi-year records for each site, with data pre-processed and resampled to daily resolution.

Methodology and Data

• Models used:
  • Penman-Monteith-Leuning (PML) model (specifically PML-V2)
  • Long Short-Term Memory (LSTM) network
  • Hybrid PML-LSTM model (integrating PML-simulated biophysical variables into LSTM)
• Data sources:
  • FLUXNET2015 global flux observation network: Daily meteorological observations (daylight air temperature, precipitation, wind speed, vapor pressure deficit, air pressure, incoming shortwave and longwave radiation) and water flux observations (sensible heat, latent heat flux/ET) from 103 eddy covariance flux tower sites.
  • MODIS satellite products: Leaf Area Index (LAI) (MCD15A3H.006, 500 m, 4-day), Surface albedo (MCD43A3.006, 500 m, daily), Surface emissivity (MOD11A2.006, 500 m, 8-day).
  • NOAA Global Monitoring Laboratory: Monthly atmospheric CO2 concentration.
  • PML-simulated biophysical variables (for PML-LSTM input): Stomatal conductance (gs), canopy conductance (Gc), canopy available energy (Ac), and equilibrium soil evaporation (Es_eq).

Main Results

• The hybrid PML-LSTM model achieved superior overall performance, with median Nash-Sutcliffe Efficiency (NSE) values of 0.851 (local level), 0.913 (vegetation type level), and 0.933 (group level) during validation, significantly outperforming the standalone PML (0.843, 0.788, 0.766) and LSTM (0.818, 0.879, 0.873) models.
• Integrating process-based biophysical variables into the PML-LSTM model enhanced ET estimation accuracy and model generalization, leading to more robust spatiotemporal performance under leave-one-out cross-validation and extreme weather extrapolation experiments (Root Mean Square Error generally below 0.5 mm/day).
• Model behaviors varied with data availability: the PML model showed greater robustness under data-scarce conditions at the local level, while LSTM and PML-LSTM models benefited from larger training datasets at the type and group levels.
• Under extreme high-temperature and high-vapor pressure deficit (VPD) conditions, the PML-LSTM model consistently outperformed other models, with NSE values for certain vegetation types increasing from near-zero to above 0.8.

Contributions

• Developed a novel hybrid PML-LSTM model that effectively integrates key process-based biophysical variables (stomatal conductance, canopy conductance, canopy available energy, equilibrium soil evaporation) from the PML model into a deep learning framework.
• Provided a systematic and comprehensive evaluation of ET estimation performance and spatiotemporal generalization abilities of process-based, data-driven, and hybrid models across three distinct modeling levels (local, vegetation type, and forest/non-forest groups) and under various validation strategies, including extreme weather extrapolation.
• Demonstrated that incorporating physically meaningful biophysical variables significantly enhances the accuracy, robustness, and generalization capacity of deep learning models for ET estimation, particularly in challenging scenarios like data scarcity or extreme weather events.
• Offered practical guidance for selecting and optimizing ET modeling strategies in ecohydrological research by highlighting the trade-offs and complementary strengths of different approaches.

Funding

• National Key R&D Program of China (grant No. 2022YFC3205200)
• National Natural Science Foundation of China (grant No. 52121006, 52479020)
• China Yangtze Power Co., Ltd (grant No. Z242302052)
• West Light Foundation of the Chinese Academy of Sciences (xbzg-zdsys-202215)
• China Scholarship Council (202406710092)

Citation

@article{Xie2026Improving,
  author = {Xie, Mingming and Zhang, Jianyun and Bao, Zhenxin and Zhang, Linus and Duan, Zheng and Wang, Guoqing and Liu, Cuishan and Yuan, Feifei and Guan, Xiaoxiang},
  title = {Improving evapotranspiration estimation by integrating process-based biophysical variables into a deep learning approach},
  journal = {Journal of Hydrology Regional Studies},
  year = {2026},
  doi = {10.1016/j.ejrh.2026.103114},
  url = {https://doi.org/10.1016/j.ejrh.2026.103114}
}