Abstract: Using the carbon and water fluxes and hydrometeorological data observed by the flux tower of Yucheng Station of China Flux Observation Network （ChinaFLUX） in the lower reaches of the Yellow River Basin, the main controlling environmental factors affecting the CO₂ exchange capacity of the farmland ecosystem were determined based on the feature importance method. A machine learning model for carbon flux prediction was constructed based on all environmental factors and master environmental factors, and the mean square error （MSE）, mean absolute error （MAE） and coefficient of determination （R²） were used to evaluate the model prediction performance of the test set. The results showed that, the main environmental driving factors affecting carbon flux in Yucheng agro-ecosystems were net radiation, soil temperature, vapor pressure deficit and soil water content. Compared with single models, the ensemble models had better learning and prediction performances in the testing set. Among the single models, MLPRegressor model could better predict NEE with R² of 0.830, MSE of 3.113 and MAE of 1.283. Among the ensemble models, XGBRegressor model had better prediction performance with R² of 0.845, MSE of 2.838 and MAE of 1.149. The machine learning models using the main four environmental driving factors had the same prediction performances as the models using all environmental factors.

Key words: farmland ecosystem, carbon flux prediction, hydrometeorological factors, machine learning models