Lin et al. (2025) Reanalysis-assisted AI framework for regional pan evaporation estimation in Taiwan without ground-based meteorological observations

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2025
• Date: 2025-10-18
• Authors: Hsuan‐Yu Lin, Sai Hin Lai, Yu-Ju Lin
• DOI: 10.1016/j.ejrh.2025.102863

Research Groups

Department of Civil and Water Resources Engineering, National Chiayi University, Chiayi 600355, Taiwan

Short Summary

This study develops an artificial intelligence-based framework to estimate daily pan evaporation across Taiwan without relying on ground-based meteorological station data. By integrating high-resolution reanalysis inputs with station metadata, the framework enables spatially continuous estimation of evaporation patterns, with XGBoost achieving the best performance (MAE = 0.00092 m/day, CC = 0.72, KGE = 0.58).

Objective

• Evaluate the applicability of high-resolution reanalysis-based meteorological inputs (TReAD) for spatially consistent estimation of pan evaporation (Epan) across Taiwan.
• Develop and compare artificial intelligence (AI) models for ungauged Epan estimation, validating their capability to estimate Epan at stations excluded from training.
• Investigate the generalization and robustness of these models under different spatial conditions, focusing on extending point-based station observations to regional-scale predictions.

Study Configuration

• Spatial Scale: The entire island of Taiwan, a subtropical island with complex topography. The study utilized 18 Class A Epan stations and high-resolution reanalysis data at 2 km spatial resolution for regional-scale predictions.
• Temporal Scale: Daily pan evaporation records from 2020 to 2023, with reanalysis data at 1-day temporal resolution.

Methodology and Data

• Models used: Multilayer Perceptron (MLP), Support Vector Regression (SVR), Random Forest (RF), Extreme Gradient Boosting (XGBoost), Gated Recurrent Unit (GRU). The Weather Research and Forecasting (WRF) model was used for downscaling reanalysis data.
• Data sources:
  • Daily Class A pan evaporation (Epan) observations from 18 stations operated by the Central Weather Administration (CWA) of Taiwan (2020-2023).
  • Taiwan ReAnalysis Downscaling dataset (TReAD), dynamically downscaled from ERA5 global reanalysis (ECMWF), providing continuous, high-resolution (2 km spatial, 1-day temporal) meteorological variables: Solar Radiation (J/m²), Wind Speed (m/s), Mean Temperature (K), Maximum Temperature (K), Minimum Temperature (K), Pressure (Pa), Relative Humidity (%), and Precipitation (m).
  • Station metadata: Latitude, Longitude, Elevation (m).
  • Temporal variable: Day of Year.

Main Results

• XGBoost achieved the best overall performance with the lowest Mean Absolute Error (MAE) of 0.00092 m/day and Root Mean Square Error (RMSE) of 0.00119 m/day, and the highest Nash-Sutcliffe Efficiency (CE) of 0.50, Kling-Gupta Efficiency (KGE) of 0.58, and Willmott’s Index of Agreement (WI) of 0.80, with a Correlation Coefficient (CC) of 0.72 and Percent Bias (PBIAS) of 1.30 %.
• Tree-based ensemble methods (XGBoost and Random Forest) demonstrated superior generalization capability for estimating pan evaporation in ungauged locations.
• SHapley Additive exPlanations (SHAP) analysis identified solar radiation, temperature, relative humidity, and wind speed as the most influential meteorological drivers, and revealed significant, non-linear impacts of precipitation and seasonal cycles (Day of Year) that were not apparent from conventional linear correlation analysis.
• The model successfully reproduced physically consistent spatial patterns of evaporation, showing higher rates in coastal lowlands (lower elevation, higher temperature and solar radiation, lower relative humidity, stronger winds) and lower rates in mountainous areas (higher elevation, lower solar radiation and temperature, higher relative humidity, weaker winds).
• Model performance varied spatially, with Kling-Gupta Efficiency (KGE) values generally ranging from 0.45 to 0.77 across most stations, but showing reduced performance (KGE = 0.22) at the highest elevation station due to a lack of comparable training samples.

Contributions

• Bridges sparse ground-based pan evaporation (Epan) records with high-resolution reanalysis data, enabling spatially continuous estimation across Taiwan.
• Integrates AI models with SHapley Additive exPlanations (SHAP) analysis to provide interpretable insights into both dominant and hidden meteorological drivers beyond conventional correlation methods.
• Demonstrates robust spatial generalization capability for ungauged estimation scenarios, extending point-based observations to regional-scale applications.

Funding

National Science and Technology Council, Taiwan (NSTC 113–2222-E-415–002-)

Citation

@article{Lin2025Reanalysisassisted,
  author = {Lin, Hsuan‐Yu and Lai, Sai Hin and Lin, Yu-Ju},
  title = {Reanalysis-assisted AI framework for regional pan evaporation estimation in Taiwan without ground-based meteorological observations},
  journal = {Journal of Hydrology Regional Studies},
  year = {2025},
  doi = {10.1016/j.ejrh.2025.102863},
  url = {https://doi.org/10.1016/j.ejrh.2025.102863}
}