Spatial heterogeneity and nonlinear climatic responses of net ecosystem productivity across China

Terrestrial ecosystems are pivotal in regulating atmospheric CO₂ concentrations, yet their carbon sequestration capacity exhibits strong spatial heterogeneity due to complex climate–ecosystem interactions. In China, where diverse climatic regimes and topographic conditions prevail, understanding the spatially explicit patterns and drivers of net ecosystem productivity (NEP) remains a critical challenge. This study applied a spatially explicit analytical framework integrating pixelwise partial correlation, XGBoost-SHAP, and ecogeographical regionalization to investigate the spatial heterogeneity, temporal trends, and climatic responses of NEP across China from 2000 to 2020. Using MODIS-derived NPP data and gridded climate datasets, we estimated NEP and analyzed its trends and interannual stability. Pixelwise partial correlation and XGBoost–SHAP were employed to disentangle linear and nonlinear climate controls. Furthermore, we conducted NEP functional zoning and centroid migration analysis to assess spatial shifts in carbon sinks and sources. Results showed pronounced spatial gradients in NEP, with higher values in southern and eastern regions. Precipitation dominated NEP variability in temperate zones, while temperature was the key control in plateau regions. Nonlinear analyses further revealed precipitation thresholds of approximately 600–800 mm in temperate regions, temperature thresholds near –10 and +5 °C in plateau regions, and precipitation saturation effects beyond about 2500 mm in humid subtropical and marginal tropical regions. The centroid of NEP-positive (carbon-sink) areas migrated northeastward, whereas that of NEP-negative (carbon-source) areas shifted southwestward, indicating a spatial redistribution of sink and source regions across China. The findings underscore the necessity of spatially differentiated assessments and demonstrate the value of integrating ecogeographical regionalization with interpretable machine learning for large-scale carbon-sink assessment and ecosystem management under climate change.