Abstract: The Net Ecosystem Exchange (NEE) serves as a critical metric for assessing ecosystem carbon budgets, offering valuable insights into carbon cycling mechanisms and informing strategies for climate change mitigation. This study leverages ground-based ChinaFLUX data collected from seven widespread forest and grassland ecosystems between 2003 and 2010. Five machine learning models-Random Forest (RF), Gradient Boosting Decision Tree (GBDT), Extreme Gradient Boosting (XGBoost), LightGBM model and linear regression, were employed in conjunction with Pearson correlation analysis and geographic detector analysis to systematically investigate the key environmental drivers influencing the interannual and seasonal variations of NEE. The primary objectives of this study are to evaluate the applicability of these models in predicting NEE variations across different temporal scales and to provide theoretical foundations for model optimization. The findings reveal a significant decline in the overall ecosystem carbon sink capacity, as indicated by NEE, from 2003 to 2010 (Slope=17.14, P<0.05). However, the carbon sink capacities at the Xishuangbanna and Haibei sites exhibited an upward trend (Slope_XSBN=-2.61 and Slope _HBGCT=-5.64). Seasonal analysis highlighted pronounced disparities in NEE during summer compared to other seasons. Notably, grassland ecosystems demonstrated a marked increase in carbon sequestration capacity during spring, with a slope of -0.74 and a P-value less than 0.05. The primary drivers of interannual variability were identified as atmospheric pressure, soil moisture, radiation, and wind speed, while seasonal fluctuations were predominantly influenced by temperature, soil moisture and soil temperature. Among the models evaluated, the RF model demonstrated the highest accuracy and precision in predicting interannual NEE (R²=0.94). For seasonal predictions, the RF model also exhibited strong performance, while the LightGBM model and XGBoost model were particularly accurate for spring and winter predictions, respectively. By integrating spatial statistics from the geographical detector with key factor identification techniques from machine learning models, this study offers a novel perspective and methodological framework for elucidating the spatiotemporal patterns of NEE and its underlying driving mechanisms.