Using interpretable machine learning to reveal carbon storage responses associated with rapid development in a “mountain-plain-hill” metropolitan area: Marginal patterns and interactions

Rapid urbanization reshapes land use and carbon storage, yet nonlinear explanatory patterns and planning implications in mountain-plain-hill metropolitan areas remain insufficiently understood. This study selected the Chengdu Metropolitan Area to construct a comprehensive “past-future” analytical framework. By coupling system dynamics (SD), the Patch-generating Land Use Simulation (PLUS) model, the Integrated Valuation of Ecosystem Services and Tradeoffs (InVEST) model, and interpretable machine learning (IML), we quantified the spatiotemporal evolution of carbon storage and explored the marginal patterns and nonlinear interactions among the main explanatory factors. The results indicate that: (1) carbon storage decreased by 6.28 Tg from 2000 to 2020, with spatial changes mainly occurring among cropland, woodland, and building land. By 2050, carbon storage is projected to increase by 52.19 Tg and 39.47 Tg under SSP126 and SSP245, respectively, but decrease by 23.16 Tg under SSP585. (2) Elevation (DEM), slope (SLOPE), and population density (POP) were consistently identified as the main explanatory factors associated with carbon storage in the model. DEM and SLOPE showed positive marginal associations with carbon storage, whereas POP showed a negative marginal association. (3) Clear SHapley Additive exPlanations (SHAP)-based nonlinear interactions were detected among the main explanatory factors. These findings suggest that carbon storage management should follow topographic gradients. Low-elevation plains should prioritize compact urban growth, cropland protection, and green network construction, whereas mid- and high-elevation areas should strengthen woodland conservation and ecological restoration.