City-Scale Assessment of Rooftop Photovoltaic Carbon Mitigation in China Using Semantic Segmentation and Interpretable Machine Learning

Abstract

Urban rooftop photovoltaics (URPV) are crucial for decarbonizing China’s building sector, yet large-scale assessment remains limited by difficulties in rooftop extraction, cross-city generalization, and linkage to forward-looking decarbonization strategies. This study developed an AI-integrated framework combining semantic segmentation, interpretable machine learning, and scenario simulation to support spatially explicit URPV planning across 690 Chinese cities. A Transformer-based model (Mask2Former) extracted rooftop areas from high-resolution imagery in 149 representative cities (mIoU = 86.6%), and a Random Forest model trained on nine socioeconomic indicators extrapolated rooftop availability to the remaining 541 cities. SHAP analysis identified total nighttime light and permanent population as the dominant predictors. The estimated national rooftop area reaches 113,264.86 km2. Under a baseline scenario (conversion factor = 0.35; PV efficiency = 0.20), the URPV carbon mitigation potential reaches 6155.80 MtCO2, equivalent to 48.9% of China’s energy-related CO2 emissions in 2023 and 1.21 times the 2021 whole-process carbon emissions of China’s building sector. Clustering identified four urban typologies—resource-rich, balanced-development, high-potential, and low-potential—supporting differentiated deployment. Accounting for urban expansion and power-mix transition, projections suggest sustained mitigation of 4700–4910 MtCO2 by 2030 under the Announced Pledges Scenario. This interpretable, data-driven framework offers scalable decision support for urban renewable energy planning and the low-carbon transformation of the built environment.