Nonlinear thresholds of landscape indicators regulating net ecosystem productivity in a subtropical region: An interpretable XGBoost–SHAP approach

Global urbanization profoundly alters regional carbon cycles, necessitating landscape pattern optimization for carbon neutrality. However, traditional linear models struggle to capture non-linear responses and threshold effects of Net Ecosystem Productivity (NEP) to landscape heterogeneity, lacking precise quantitative bases for spatial planning. Taking Hunan Province as a case study, this research integrates the XGBoost model with the SHAP framework to systematically parse the dominant driving mechanisms, ecological thresholds, and interaction paths of NEP from 2005 to 2020. Results indicate: (1) Landscape composition dictates the carbon sink baseline, but its dominance is gradually yielding to spatial configuration. Grassland_PLAND remains the primary driver of NEP spatial variability, historically outweighing forestland in marginal importance, while construction land expansion acts as the main negative constraint. Recently, the regulatory influence of landscape fragmentation and macro-connectivity has increased significantly. (2) NEP exhibits significant non-linear responses and temporal threshold shifts. The positive-effect threshold for grassland declined from 9.13% (2005) to 3.26% (2020), highlighting the critical compensatory role of remnant natural patches. Conversely, exceeding ultra-low patch density thresholds (0.30 for forest; 0.33 for cropland) triggers abrupt carbon sink declines. For construction land, a local positive response exists within the 6.0%–7.5% range but surpassing the 7.5% critical threshold intensifies its inhibitory effect. (3) The NEP regulatory mechanism shifted from single composition dominance (especially grassland and construction land proportions) to synergistic interactions between composition and spatial configuration (notably landscape fragmentation and connectivity). (4) Carbon management should pivot from area-based to threshold-led spatial refinement. Establishing quantitative safety boundaries, specifically limiting construction land below 7.5% and forestland fragmentation below 0.30, is essential to preempt abrupt carbon sink degradation. These thresholds provide robust decision support for balancing urban expansion and ecological security under tight land constraints in subtropical hilly regions.