An explainable AI-based approach for estimating potential evapotranspiration in ungauged areas

Republic of Korea Evapotranspiration is an essential component of water balance analysis for establishing long-term national water resources development and planning. Although South Korea operates 659 meteorological stations, only 75 provide the full variables required for the FAO56 Penman–Monteith (PM) method to estimate potential evapotranspiration (PET). Therefore, in this study, daily meteorological data from 2010 to 2024 were used, and Shapley Additive Explanations (SHAP) analysis was conducted to evaluate the contribution of each meteorological variable to PET prediction. Based on the SHAP results, variable selection scenarios were designed, and deep learning models, including a Deep Neural Network (DNN) and a Long Short-Term Memory (LSTM), were developed and applied to stations with limited data for evaluation. The SHAP analysis indicated that maximum temperature and average wind speed were the most influential input variables. Model performance validation at the Oksan station in Case 2 demonstrated strong predictive capability, with an RMSE of 0.3407 mm/day, NSE of 0.8007, CC of 90.01 %, and PBAIS of 9.6408 %. These results indicate that even if various meteorological factors required for PET estimation are not measured, PET can still be effectively estimated using AI with only the key variables. The proposed approach can enhance the spatial resolution of PET in ungauged regions and serve as a practical alternative in data-scarce environments. • PET was effectively estimated using explainable AI under limited data conditions. • DNN and LSTM models were developed to predict potential evapotranspiration. • SHAP analysis quantified the influence of key meteorological variables on PET. • The AI models can replace the complex PM method for PET estimation. • The proposed method can enhance the spatial resolution of PET in ungauged regions.