County-level CO 2 emissions based on the urban–rural land-use structure in Hunan Province, China: evidence from human activity and SHAP value interpretation

Urban–rural land-use composition is a key planning lever for low-carbon development, yet built-up land is rarely classified into regulable functions. Using data from Hunan, China, for 2020, we integrated Sentinel-2 land-cover mapping with point-of-interest semantics to disaggregate built-up land into industrial, commercial/service, urban residential, other urban, and rural residential components and combined these shares with human-activity and physical covariates. Land-use shares were compositionally transformed before modeling. Six supervised learning algorithms were compared under five-fold spatially blocked cross-validation (CV), and the Random Forest model was retained as the primary interpretive model because it achieved the lowest mean CV-RMSE. Fold-aware SHAP results show that urban functional composition and human-activity intensity are the dominant drivers of predicted county-level total CO2 emissions. Commercial/service and industrial shares show strong nonlinear positive effects, road density exhibits a threshold-like increase, forest cover is generally associated with reduced emissions, and cropland shows a hump-shaped response. The dependence curves identify uncertainty-bounded operating ranges within which marginal structural change is associated with disproportionate changes in predicted emissions, providing spatial evidence to inform zoning, redevelopment sequencing, ecological protection, and infrastructure management for county-scale low-carbon planning.