Woo (2025) Estimating actual evapotranspiration from widely available meteorological data with a hybrid CNN–LSTM

Identification

• Journal: Agricultural Water Management
• Year: 2025
• Date: 2025-12-18
• Authors: Dong Kook Woo
• DOI: 10.1016/j.agwat.2025.110078

Research Groups

• Department of Civil Engineering, Keimyung University, Daegu, Republic of Korea

Short Summary

This study develops a hybrid Convolutional Neural Network – Long Short-Term Memory (CNN–LSTM) model to estimate daily actual evapotranspiration (ETa) directly from widely available meteorological and soil data, demonstrating high accuracy (R²=0.92, RMSE=0.38 mm d⁻¹) across 167 FLUXNET sites and global applicability with ERA5 forcings.

Objective

• To develop a machine-learning framework for estimating daily actual evapotranspiration (ETa) directly from routinely measured station-type meteorological and soil data, without relying on reference evapotranspiration (ETo) as a primary proxy.
• To provide a contemporary ETa benchmark that can support hydrological model evaluation and the diagnosis of recent changes in land–atmosphere coupling under a rapidly changing climate.

Study Configuration

• Spatial Scale: Global. Model trained on 167 FLUXNET eddy-covariance sites spanning diverse biomes globally. Applied globally using ERA5 reanalysis fields.
• Temporal Scale: Daily ETa estimates. Input sequences of 36 days. Training data from 1996–2018. Global application for two benchmark years: 2012 and 2024.

Methodology and Data

• Models used: Hybrid Convolutional Neural Network – Long Short-Term Memory (CNN–LSTM) framework. 21 empirical reference evapotranspiration (ETo) formulations (temperature-based, radiation-based, mass-transfer) were calculated and one (A1) was used as an auxiliary predictor.
• Data sources:
  • FLUXNET2015 Tier-1 and FLUXNET-CH4 synthesis (eddy-covariance measurements of ETa and meteorological variables from 167 sites).
  • Meteorological forcings: Solar radiation (Rs), vapor pressure deficit (VPD), air pressure (Pa), minimum/mean/maximum air temperature (Ta), precipitation (PPT), wind speed (WS), and volumetric soil water content (SWC).
  • Static site descriptors: Percentages of sand, silt, and clay (soil texture), and a categorical vegetation index (Veg).
  • ERA5 reanalysis data (for global application and benchmarking).

Main Results

• The hybrid CNN–LSTM model achieved high fidelity in estimating daily ETa against independent flux-tower measurements, explaining 92% of observed variance (R² = 0.92) with a root-mean-square error (RMSE) of 0.38 mm d⁻¹ for both validation and independent test subsets.
• A leave-one-out sensitivity analysis revealed that solar radiation, air pressure, wind speed, and vapor pressure deficit were the most informative predictors. The categorical vegetation class degraded model performance, and ETo provided only modest marginal benefit as an auxiliary predictor.
• Global application of the trained model with ERA5 forcings for 2012 and 2024 reproduced broad spatial patterns of ETa, with pixel-wise scatterplots clustering tightly around the 1:1 line, indicating strong global correspondence and temporal consistency.
• Daily performance against ERA5 was strongest in snow-influenced continental climates (median R² consistently exceeding 0.80 and RMSE remaining well below 0.65 mm d⁻¹), while arid and high-latitude regions showed larger discrepancies.

Contributions

• Developed a robust, accurate, and generalizable hybrid CNN-LSTM model for direct daily ETa estimation using widely available meteorological and soil data, effectively bypassing the reliance on ETo as a primary proxy.
• Demonstrated that ETo, while informative as a diagnostic predictor, offers only modest marginal value when its constituent meteorological drivers are already included, reinforcing the conceptual mismatch of using ETo as a direct surrogate for ETa.
• Provided a contemporary, climate-dependent global ETa benchmark derived from station-type data and ERA5 forcings, which is valuable for hydrological model evaluation, drought monitoring, irrigation planning, and land-surface model assessment.
• Identified the most influential meteorological drivers for ETa estimation and highlighted the limited utility of coarse vegetation descriptors in this modeling context.

Funding

• National Research Foundation of Korea (NRF), South Korea grant (RS-2024-00336959) funded by the Korea government (MSIT).

Citation

@article{Woo2025Estimating,
  author = {Woo, Dong Kook},
  title = {Estimating actual evapotranspiration from widely available meteorological data with a hybrid CNN–LSTM},
  journal = {Agricultural Water Management},
  year = {2025},
  doi = {10.1016/j.agwat.2025.110078},
  url = {https://doi.org/10.1016/j.agwat.2025.110078}
}