Attribution of urban transportation carbon emissions at a 500-meter resolution: Insights from Chengdu, China

Urban transportation carbon emissions are highly uneven within cities, yet inventories often remain too coarse to support fine-grained intervention. We develop a framework for 500 m-scale spatialization, prediction, interpretation, and typological diagnosis of an urban transportation emission surface, using Chengdu as a case study. First, annual 2023 EDGAR transport-sector emissions at 0.1° resolution are conservatively clipped to the municipal boundary and downscaled to 500 m × 500 m grids using nighttime-light dasymetric allocation, preserving both citywide totals and parent-grid totals. Second, under blocked 5-fold cross-validation, a Bayesian-optimized LightGBM is used as the base learner, and a subsequent GWR residual correction yields the final LightGBM + GWR surface, with test-set performance of R2 = 0.76, RMSE = 149.18, and MAE = 63.38, while reducing residual spatial autocorrelation to Moran's I = −0.01. Third, SHAP analysis of the optimized LightGBM indicates that built-up intensity, rail-station concentration, POI density, and road-network supply are the strongest predictors, and that their effects are nonlinear, scale-dependent, and often characterized by threshold or saturation behavior. Finally, clustering of SHAP contribution vectors for all valid grids identifies three recurring spatial types: a low-emission background fabric, high-activity urban subcenters, and multimodal super-hotspots. These types support differentiated screening and comparative prioritization across background monitoring, subcenter traffic management, and the most emission-intensive multimodal cores. The resulting 500 m surface should be interpreted as a screening and comparative prioritization surface derived from a coarse inventory, rather than as validated grid-level transport carbon emission.