Lee et al. (2025) An explainable AI-based approach for estimating potential evapotranspiration in ungauged areas

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2025
• Date: 2025-11-05
• Authors: Haneul Lee, Seungmin Lee, Ho-Yong Lee, Seonuk Baek, Soojun Kim
• DOI: 10.1016/j.ejrh.2025.102900

Research Groups

• Program in Smart City Engineering, Inha University, Incheon, Republic of Korea
• Department of Civil Engineering, Inha University, Incheon, Republic of Korea

Short Summary

This study develops an explainable AI-based approach using Deep Neural Networks (DNN) and Long Short-Term Memory (LSTM) models, coupled with Shapley Additive Explanations (SHAP), to accurately estimate potential evapotranspiration (PET) in ungauged areas of South Korea with limited meteorological data. The approach demonstrates that PET can be effectively estimated using only key variables like maximum temperature and average wind speed, significantly enhancing the spatial resolution of PET in data-scarce environments.

Objective

• To develop an explainable artificial intelligence model for estimating potential evapotranspiration (PET) in ungauged watersheds under limited meteorological data conditions, thereby improving the spatial resolution of PET in South Korea.
• To identify essential meteorological variables for PET estimation using SHAP analysis and develop AI models based on these selected variables for application in data-limited Automatic Weather System (AWS) stations.

Study Configuration

• Spatial Scale: Republic of Korea, focusing on 75 inland Automated Synoptic Observing System (ASOS) stations for training and 554 Automatic Weather System (AWS) stations for application.
• Temporal Scale: Daily meteorological data from 2010 to 2024.

Methodology and Data

• Models used: Deep Neural Network (DNN), Long Short-Term Memory (LSTM), FAO56 Penman–Monteith (PM) method (as reference), Shapley Additive Explanations (SHAP).
• Data sources: Daily meteorological data collected from 75 ASOS weather stations and 554 AWS weather stations operated by the Korea Meteorological Administration (KMA).

Main Results

• SHAP analysis identified maximum temperature and average wind speed as the most influential input variables for PET prediction in both DNN and LSTM models.
• AI models achieved excellent performance when trained with complete meteorological data from ASOS stations (e.g., DNN: RMSE of 0.0790 mm/day, NSE of 0.9908, PBIAS of 2.76 %, CC of 99.54 %).
• For scenarios with limited meteorological data, the DNN model in Scenario 6 (using maximum temperature, average wind speed, minimum temperature, and average temperature) showed the best performance (RMSE of 0.3082 mm/day, NSE of 0.8600, CC of 92.74 %, PBIAS of –2.0630 %).
• Application of the best-performing DNN model to AWS stations for ungauged areas yielded reasonable average performance (RMSE of 0.5326 mm/day, NSE of 0.5747, PBIAS of 25.0899 %, CC of 87.42 %) when compared to nearby ASOS stations.
• PET estimates at AWS stations tended to be underestimated, which was attributed to lower input temperature values at AWS stations compared to ASOS stations, consistent with SHAP analysis.

Contributions

• Developed an explainable AI approach (DNN, LSTM with SHAP) for PET estimation in ungauged areas with limited meteorological data, addressing the "black-box" nature of AI models.
• Quantified the contribution of individual meteorological variables to PET prediction using SHAP, identifying maximum temperature and average wind speed as key factors.
• Demonstrated the feasibility of accurate PET estimation using a minimal set of essential variables, enabling enhanced spatial resolution of PET in data-scarce regions like South Korea.
• Provided a practical alternative for water resource management and agricultural planning in areas where the full FAO56 Penman–Monteith method cannot be applied due to data limitations.

Funding

• Korea Environmental Industry & Technology Institute through the Wetland Ecosystem Value Evaluation and Carbon Absorption Value Promotion Technology Development Project funded by the Korea Ministry of Environment (MOE) (2022003630001).

Citation

@article{Lee2025explainable,
  author = {Lee, Haneul and Lee, Seungmin and Lee, Ho-Yong and Baek, Seonuk and Kim, Soojun},
  title = {An explainable AI-based approach for estimating potential evapotranspiration in ungauged areas},
  journal = {Journal of Hydrology Regional Studies},
  year = {2025},
  doi = {10.1016/j.ejrh.2025.102900},
  url = {https://doi.org/10.1016/j.ejrh.2025.102900}
}