Ensemble machine learning-based scaling up of wetland carbon dynamics across the Southeastern U.S.

Background Although the Southeastern United States (SE US) has been identified as a major terrestrial carbon sink, the balance of CO 2 and CH 4 fluxes from its wetlands remains poorly constrained due to strong environmental heterogeneity and limited regional-scale assessments. In this study, we used observations from 24 wetland sites located in the SE US, including freshwater and salt marshes, to train machine learning models (Random Forest, XGBoost, Support Vector Machine, Artificial Neural Network, and Partial Least Squares Regression) to predict monthly mean gross primary productivity (GPP), ecosystem respiration (RECO), net ecosystem exchange (NEE), and methane (CH 4 ) fluxes. Results Model performance measured in r 2 for GPP ranged from 0.60 to 0.70, for RECO ranged from 0.44 to 0.56, for NEE ranged from 0.45 to 0.56, and for CH 4 flux ranged from 0.33 to 0.37. Random Forest and XGBoost were the most consistent algorithms between fluxes. Feature importance analyses of these two algorithms indicated that CO 2 flux predictions were primarily associated with vegetation and climatic variables, whereas CH 4 predictions were more strongly associated with soil and hydrological properties. Using a performance-weighted multi-model ensemble (MME), we scaled up the predictions for the entire SE US from 2000 to 2020. Spatial patterns of productivity and respiration showed flux hotspots in the Everglades, the Mississippi River Valley, and the Gulf and East Coast. We also identified hotspots of CH 4 emissions in coastal North Carolina. Slightly lower CH 4 fluxes were observed between 2000–2007 in comparison to 2008–2020. Mean annual wetland carbon fluxes were 280 ± 38 Tg CO 2 -C for GPP, 233 ± 51 Tg CO 2 -C for RECO, − 20 ± 18 Tg CO 2 -C for NEE, and 5.7 ± 1.7 Tg CH 4 -C for CH 4 . Conclusions Our results provide regional estimates of CO 2 and CH 4 fluxes that improve carbon budget assessments and support future wetland management in the SE US. By identifying the spatial and temporal variability of carbon uptake and emissions across the region, this study provides a foundation for improving regional carbon accounting and informing wetland conservation and management strategies.