Structure-sensitive carbon emission mechanisms in urban morphology: Grad-CAM guided nonlinear modeling and attention-based sectoral diagnosis

Urban morphology profoundly shapes the spatial distribution of carbon emissions by regulating spatial structure, energy use, and human activity. To systematically uncover its mechanisms, this study develops a structure-sensitive analytical framework that integrates deep learning attention mechanisms with interpretable machine learning, focusing on three core tasks in the morphology–emission relationship: key factor identification, mechanism analysis, and spatial diagnosis. In an empirical case study of Nanjing, China, the framework first trains sector-specific carbon emission prediction models based on remote sensing imagery, and applies Grad-CAM to extract model attention heatmaps. These maps are used to identify morphological features most influential in emission prediction and to construct an attention-weighted morphology indicator system that reflects model-perceived spatial structures. Second, a mediation–moderation model is introduced to clarify the role of demographic and behavioral variables nested within structural pathways, and XGBoost combined with SHAP is employed to characterize the nonlinear responses and threshold behaviors of key morphological factors, revealing multi-path and sector-specific regulation mechanisms. Finally, a structural deviation index is constructed to identify high-emission zones with significant mismatches between observed morphology and model cognition. These zones exhibit threshold shifts and marginal reversals, providing critical guidance for spatial risk detection and structure-sensitive governance strategies. The findings offer new methodological support and theoretical insight for identifying morphological mechanisms of urban carbon emissions and for advancing precision-oriented mitigation.