Machine Learning-based Multi-scale Dynamics of Terrestrial Carbon Fluxes and Their Environmental Drivers Along the U.S. East Coast

With global climate change intensifying, the carbon cycle and its related processes have become a central topic in ecological research. In this study, a large-scale carbon flux estimation model for the U.S. East Coast was developed based on long-term eddy covariance observations. Through this model, carbon flux characteristics and their spatiotemporal patterns across different ecosystem types in the region were analyzed over the past two decades. By integrating correlation analysis and the Geodetector method, the roles of multiple environmental drivers in carbon flux estimation were elucidated. Subsequently, models for gross primary productivity (GPP), ecosystem respiration (ER), and net ecosystem productivity (NEP) were constructed using four machine learning algorithms: random forest (RF), artificial neural network (ANN), support vector regression (SVR), and extreme gradient boosting (XGBoost). The results indicate that: (1) The combination of eight factors, including T2M, VPD, SSRD, EVI, LSWI, LAI H, EVAVT, and DEM, exhibited the most accurate and stable performance in carbon flux estimation. Under identical input combinations, the RF model achieved the highest accuracy for GPP, NEP, and ER estimation. (2) The verification accuracy of GPP, ER, and NEP achieved R ² values of 0.88, 0.81, and 0.55, respectively. These accuracies markedly outperform those of existing carbon-flux products such as FLUXCOM, which show lower R ² values of 0.61, 0.57, and 0.28. (3) The analysis of environmental variable importance reveals that EVI is the most important variations in carbon fluxes across all ecosystems, underscoring that vegetation growth status is the most critical driver of carbon exchange processes.