Zhao et al. (2026) Learning evaporative fraction with memory

Identification

• Journal: Hydrology and earth system sciences
• Year: 2026
• Date: 2026-08-25
• Authors: Wenli Zhao, Alexander J. Winkler, Markus Reichstein, René Orth, Pierre Gentine
• DOI: 10.5194/hess-30-5373-2026

Research Groups

• Max Planck Institute for Biogeochemistry (Jena, Germany)
• Columbia University (New York, USA)
• ELLIS Jena Unit (Jena, Germany)
• Faculty of Environment and Natural Resources, University of Freiburg (Freiburg, Germany)

Short Summary

This study develops an explainable machine learning model to predict evaporative fraction (EF) with memory effects using long-term observations from 90 eddy covariance sites across diverse plant functional types. The model accurately captures EF dynamics, particularly during post-rainfall pulses and soil moisture dry-down events.

Objective

• Investigate the mechanisms underlying EF dynamics with memory effects.
• Develop an explainable machine learning framework to quantify feature importance and memory effects in EF predictions.

Study Configuration

• Spatial Scale: Global, covering 90 eddy covariance sites across diverse plant functional types (PFTs).
• Temporal Scale: Daily scale, using long-term observations (>5 years) from ICOS, AmeriFlux, and FLUXNET2015 Tier 1 datasets.

Methodology and Data

• Models used: Long Short-Term Memory (LSTM) architecture with memory effects.
• Data sources: Eddy covariance measurements, remote sensing data, soil attribute data, rooting depth information, and long-term evapotranspiration data.

Main Results

• The LSTM model accurately captures EF dynamics, particularly during post-rainfall pulses and soil moisture dry-down events.
• Expected Gradients (EG) analysis reveals precipitation and vapor pressure deficit as dominant drivers of EF in woody savanna, savanna, open shrubland, and grassland ecosystems.
• Air temperature emerges as the dominant factor in deciduous broadleaf, evergreen needleleaf, and mixed forests.

Contributions

• This study advances understanding of EF regulation and memory effect dynamics.
• The developed explainable machine learning framework provides a novel approach to quantify feature importance and memory effects in EF predictions.

Funding

• Max Planck Society (Germany)
• National Science Foundation (USA)
• European Research Council (ERC) grant "Understanding the role of memory effects in EF prediction"

Citation

@article{Zhao2026Learning,
  author = {Zhao, Wenli and Winkler, Alexander J. and Reichstein, Markus and Orth, René and Gentine, Pierre},
  title = {Learning evaporative fraction with memory},
  journal = {Hydrology and earth system sciences},
  year = {2026},
  doi = {10.5194/hess-30-5373-2026},
  url = {https://doi.org/10.5194/hess-30-5373-2026}
}