Interpretable machine learning workflow for estimating reference crop evapotranspiration in China's five major dry-wet regions using limited meteorological data

Accurate estimation of reference crop evapotranspiration ( ET 0 ) is crucial for improving water use efficiency and the design and operation of agricultural water management systems. Machine learning (ML) can accurately estimate ET 0 across different climatic zones in China when meteorological data are limited, but its “black box” nature restricts interpretability. This study developed an interpretable machine learning workflow to enhance ET 0 prediction transparency. It utilized four meta-heuristic algorithms and four machine learning algorithms based on meteorological data observed at 2382 stations across five climatic zones in China from 1960 to 2022. Model performance was evaluated using mean absolute error (MAE), root mean square error (RMSE), Nash-Sutcliffe efficiency coefficient (NSE), coefficient of determination ( R 2 ), and global performance index (GPI). Results indicate that the XGBoost model optimized by the Grey Wolf Optimization (GWO) algorithm (GWO-XGB) achieved the highest fitting accuracy. Its Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Nash-Sutcliffe efficiency coefficient (NSE), coefficient of determination ( R 2 ), and global prediction index (GPI) were 0.087, 0.116, 0.993, 0.993, and 1.783, respectively. Cross-validation across basins revealed that GWO-XGB maintained an R 2 above 0.96 on the independent validation dataset, indicating robust stability and generalization of the interpretable machine learning framework in ET 0 prediction. SHAP accurately captured underlying hydrological and climatic processes, identifying solar radiation and extreme temperatures as the primary predictors of ET 0 , while humidity and wind speed exerted lesser influences. This study offers a promising approach for precise ET 0 estimation in data-scarce regions, thereby supporting scientific water resource management and achieving water conservation and efficiency goals. The open-source prediction application is available at https://github.com/guangian . • Combined with EAI to obtain the relative contribution of meteorological factors. • The ET 0 in humid areas is affected by radiation and sunshine duration. • The ET 0 in arid areas is affected by radiation and temperature extremes. • The GWO-XGB model has higher ETo estimation accuracy.