Inferring carbon emissions from human mobility data

Accurate identification, quantification, and continuous monitoring of anthropogenic carbon emissions are fundamental to effective climate mitigation. However, existing approaches often face trade-offs among temporal resolution, spatial coverage, data availability, and update frequency. Here, we propose a computational framework to estimate anthropogenic CO2 emissions by leveraging large-scale, observable proxies of human activity derived from mobility data. We find that human mobility shows a strong correlation with multi-source CO2 emissions in China (r = 0.89). By integrating mobility-derived network features with static spatial attributes such as geographic coordinates and population, machine learning models capture both the broad spatial organization and dynamic temporal variation of CO2 emissions, achieving high predictive performance. The framework performs consistently across diverse socioeconomic contexts (China, Italy, the U.S., and Mexico) and across independent CO2 emission inventories (GRACED and ODIAC), and remains robust under both stable and turbulent periods. This work proposes a scalable and generalizable approach for large-scale anthropogenic CO2 emissions estimation, providing a behavior-informed foundation for timely and adaptive climate policy.