Spatially nonstationary drivers of aboveground biomass carbon across China’s terrestrial ecosystems

Understanding how environmental contributions to aboveground biomass carbon (AGBC) vary across space and ecosystem types is important for large-scale carbon assessment. Although machine-learning approaches can capture nonlinear relationships between AGBC and environmental variables, conventional global interpretation may obscure geographic variation in these relationships. Here, we integrated eXtreme Gradient Boosting (XGBoost), GeoShapley, conventional SHapley Additive exPlanations (SHAP), and generalized additive models (GAMs) to analyze field-based AGBC observations and multisource environmental variables across China's terrestrial ecosystems. XGBoost was used to characterize nonlinear predictor–AGBC relationships, GeoShapley to separate environmental and geographic contributions and their interactions, and GAMs to examine ecosystem-specific nonlinear attribution patterns. Spatial block cross-validation and an ablation analysis of the encoded ecosystem-type variable were further used to assess the robustness of model interpretation. Net primary productivity (NPP) remained among the environmental variables with the largest contributions, and removal of the ecosystem-type encoding produced highly consistent rankings of the remaining variables (Spearma n's r s = 0.976 , p < 0.001 ). GeoShapley–SHAP differences showed clear regional variation in the contributions of altitude, precipitation, and the normalized difference vegetation index (NDVI)-derived vegetation-cover index (TCC), whereas NPP exhibited a less distinct spatial boundary. Ecosystem-specific analyses further revealed contrasting nonlinear attribution patterns between forests and grasslands, consistent with differences in carbon retention, allocation, turnover, and environmental constraints. These results show that environmental contributions to AGBC are both geographically and ecosystem dependent, and that spatial attribution combined with ecosystem-specific interpretation provides a more informative framework than a single global variable-importance analysis for understanding national-scale AGBC variation.