Accurate estimation of reference evapotranspiration (ETo) is key to irrigation system design and agricultural water management. Utilizing meteorological data (1960-2019) from 20 stations in China's humid and arid regions, a reference ETo value was calculated using the FAO56-Penman-Monteith (PM) method. The accuracy of 6 ensemble learning models [e.g., Adaptive boosting (AdaBoost), Gradient Boosting Decision Tree (GBDT), Categorical boosting (CatBoost), Extreme gradient boosting (XGBoost), Extra trees, and Light Gradient Boosting Method (LightGBM)] in estimating daily ETo using all available inputs was investigated. The performance of the best three models (CatBoost, GBDT and XGBoost) was then evaluated under 7 input combinations [i.e., complete and incomplete combinations of maximum and minimum temperature (T-max and T-min), relative humidity (RH), wind speed (U-2), total and extra-terrestrial solar radiation (R-s and R-a)], and 4 dataset sizes (20, 30, 40 and 60 years). CatBoost showed the highest estimation accuracy (average R-2 = 0.93), stability, and robustness. Using incomplete combinations based on temperature and other indicators to estimate daily ETo also achieved satisfactory results (R-2 > 0.91), and the key indicators contributing to a difference in ETo prediction accuracy between humid and arid regions were RH and R-a. Different models' accuracy in estimating daily ETo was not affected by dataset size (the difference of RMSE < 0.025), but its stability improves with the increase of the dataset. This study evaluated the models' performance under different data constraints and different regional applications, which provides a methodological reference for ETo simulation in global multiclimatic zones, takes into account accuracy and practicality.