Gyamfi et al. (2026) Physics-informed spatio-temporal graph neural networks for evapotranspiration prediction: Case of the Korean Peninsula

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2026
• Date: 2026-03-10
• Authors: Kwame Adutwum Gyamfi, Eun-Sung Chung, Young Hoon Song, Shamsuddin Shahid
• DOI: 10.1016/j.ejrh.2026.103314

Research Groups

• Civil Engineering Department, Seoul National University of Science and Technology, Seoul, South Korea
• Department of Applied Artificial Intelligence, Seoul National University of Science and Technology, Seoul, South Korea
• Regional Climate Change Center, National Center for Meteorology, Jeddah, Saudi Arabia

Short Summary

This study develops a physics-informed spatio-temporal graph neural network for evapotranspiration prediction across the Korean Peninsula, integrating climate variables, soil moisture, and a surface energy-balance constraint. The model demonstrates strong skill, particularly under dry conditions, and projects substantial increases in evapotranspiration under future climate scenarios, highlighting increasing evaporative demand and water stress.

Objective

• To evaluate the effectiveness of a physics-informed graph neural network in improving spatio-temporal learning of evapotranspiration.
• To determine the individual and combined contribution of climate variables in predicting evapotranspiration.
• To analyze how modeling choices influence projected evapotranspiration responses under future climate scenarios.

Study Configuration

• Spatial Scale: Korean Peninsula (33° N to 43° N, 124° E to 131° E), using observations from 372 stations.
• Temporal Scale: Historical observations from 1950 to 2014; future projections for 2021–2060 (near-future) and 2061–2100 (far-future) at monthly resolution.

Methodology and Data

• Models used:
  • Physics-informed Spatio-Temporal Graph Neural Network (ST-GNN)
  • GraphSAGE (spatial encoder) combined with Gated Recurrent Unit (GRU) (temporal unit)
  • Physics-informed loss function based on the surface energy balance equation
  • Uncertainty quantification: Monte Carlo Dropout and Isotonic Regression
  • Hyperparameter optimization: Optuna framework (Tree-structured Parzen Estimator sampler)
  • Benchmarking models: Penman–Monteith (P-M), MODIS16, CNN+LSTM, Temporal Transformer
• Data sources:
  • Ground truth evapotranspiration: GLEAM-based ET dataset
  • Meteorological variables: University of East Anglia’s Climate Research Unit (CRU) version 4.09 (maximum temperature (tasmax), minimum temperature (tasmin), relative humidity (RH), surface downward shortwave radiation (rsds), near-surface wind speed (sfcwind))
  • Soil moisture (sm): Climate Prediction Center (CPC)
  • Future climate projections: 10 Global Climate Models (GCMs) from NASA Earth Exchange Global Daily Downscaled Projections (NEX-GDDP) and Earth System Grid Federation (ESGF) Coupled Model Intercomparison Project Phase 6 (CMIP6) under Shared Socioeconomic Pathway (SSP) scenarios SSP2–4.5 and SSP5–8.5.
  • Sensible heat flux (H): Derived from the same reference dataset as the evapotranspiration target.
  • Net radiation (Rn): Estimated from downward shortwave radiation (rsds) and downward longwave radiation (rlds).

Main Results

• The GraphSAGE+GRU architecture with the Adam optimizer achieved the best performance (MAE: 0.10 mm, RMSE: 0.12 mm, R2: 0.99, Nash–Sutcliffe efficiency (NSE): 0.91).
• Isotonic regression calibration significantly improved uncertainty estimates, reducing the Continuous Ranked Probability Score (CRPS) from 0.14 to 0.09 and aligning the 95% Prediction Interval Coverage Probability (PICP) from 0.92 to 0.96.
• The physics-informed GraphSAGE+GRU model accurately reproduced observed seasonal cycles and spatial patterns of evapotranspiration across the Korean Peninsula.
• Variable importance analysis identified maximum temperature, relative humidity, and surface downward shortwave radiation as dominant drivers of evapotranspiration.
• The explicit inclusion of soil moisture as a predictor substantially improved model performance, particularly under drought conditions, reducing RMSE and MAE by 20% during drought months.
• A hybrid graph construction strategy, combining geodesic distance and inter-station correlation, yielded the best model performance (MAE: 0.12 mm, RMSE: 0.16 mm, R2: 0.99).
• Future climate projections indicate a modest increase in average evapotranspiration under SSP2–4.5 (from approximately 36 mm/month to 40–42 mm/month) and a more pronounced, accelerated increase under SSP5–8.5 (surpassing 42 mm/month after 2060), with a substantial increase in interannual variability (coefficient of variation up to 0.07).
• Under SSP5–8.5, northern regions of the Korean Peninsula are projected to experience disproportionately greater evapotranspiration intensification, leading to a flatter evapotranspiration-latitude gradient.

Contributions

• Proposal of a physics-informed Spatio-Temporal Graph Neural Network (ST-GNN) framework for regional evapotranspiration prediction that explicitly integrates spatial interactions among meteorological stations with temporal dynamics.
• Introduction of a loss-level surface energy balance constraint within the graph-based deep learning architecture, allowing physical consistency to directly influence model optimization.
• Demonstration that incorporating soil moisture as an explicit node feature substantially improves model performance, interpretability, and robustness under both normal and drought conditions.
• Development of a comprehensive uncertainty-aware evapotranspiration prediction framework by combining Monte Carlo Dropout with isotonic regression calibration.
• Application of the proposed framework to historical observations and CMIP6 climate projections, revealing consistent increases in evapotranspiration magnitude and variability across future scenarios.

Funding

• National Research Foundation of Korea, South Korea (RS-2023-00246767_4)

Citation

@article{Gyamfi2026Physicsinformed,
  author = {Gyamfi, Kwame Adutwum and Chung, Eun-Sung and Song, Young Hoon and Shahid, Shamsuddin},
  title = {Physics-informed spatio-temporal graph neural networks for evapotranspiration prediction: Case of the Korean Peninsula},
  journal = {Journal of Hydrology Regional Studies},
  year = {2026},
  doi = {10.1016/j.ejrh.2026.103314},
  url = {https://doi.org/10.1016/j.ejrh.2026.103314}
}