Dynamic prediction and attribution of farmland evapotranspiration across year types

Accurate prediction of farmland actual evapotranspiration (ET c act ) is crucial for optimizing agricultural water resource management and food security. Using temperature and precipitation data (1984–2023) from the Luancheng station in the North China Plain, this study classified the years from 2007 to 2023 into six types: temperature year types (Average/Cold/Warm Temperature Year: ATY/CTY/WTY) and precipitation year types (High/Low/Average Precipitation Year: HPY/LPY/APY). Six machine learning models (RF, XGBoost, SVR, MLP, LightGBM, and CatBoost) were developed using daily meteorological variables and ET c act data measured by the eddy covariance method from 2007 to 2023. The results indicated significant variations in ET c act across different hydrothermal year types, with interquartile ranges (IQR) varying from 2.47 to 2.75 mm/day. The CatBoost model demonstrated superior performance for the ATY, CTY, HPY, and APY categories, and also delivered the best overall results across the entire dataset (test set R 2 = 0.80, RMSE = 0.79 mm/day). The MLP and XGBoost models exhibited specific advantages in the WTY and LPY scenarios, respectively, with R 2 values reaching 0.81 and 0.82. The SHAP framework revealed the contribution patterns of irrigated farmland ET c act driving factors. Daily solar radiation and temperature-related variables were the dominant drivers across all year types, accounting for over 27.75% and 34.34% of the contribution to ET c act , respectively. In contrast, air humidity, wind speed, and precipitation were minor contributors. Additionally, the analysis revealed that daily total solar radiation must exceed a threshold of 18.3 MJ/m 2 for a marked increase in ET c act . Applying the optimal CatBoost modeling framework to nine global FLUXNET cropland sites for independent training and validation produced test-set R 2 values ranging from 0.65 to 0.86. By integrating year-type identification with explainable machine learning to reveal the dynamic patterns of ET c act drivers under climatic stresses, this study provides theoretical support for the dynamic adaptation and optimization of models in complex contexts.