Following the crowd: phone mobility data can track carbon emissions in near real time

Every commute, delivery, and business trip burns fuel—and much of the carbon released into the atmosphere is a byproduct of exactly that movement. That everyday link is the starting point of a study published in The Innovation, in which researchers at the National University of Defense Technology, the University of Hong Kong, and partners across three continents turn large-scale mobility data into something like a near-real-time carbon meter.

The need is practical. Conventional emission estimates rest on statistical accounts and inventories, which are accurate but slow—often lagging reality by months—and satellite inversions face their own trade-offs among resolution, coverage, and update frequency. When COVID-19 lockdowns reshaped economic activity within days in early 2020, slow instruments could not keep up. The authors asked a different question: since so many emission-generating processes involve moving people and goods, could the observable movement itself serve as the measuring stick?

To find out, the team converted anonymized cellular signaling data from China Unicom—about 318 million active users—into daily intercity mobility networks, with cities as nodes and flows as edges. From these networks they extracted structural features such as node centrality and clustering, joined them with population and geographic attributes, and trained machine-learning models against the GRACED daily emission dataset.

The window they studied could hardly have been more turbulent: January and February 2020, spanning both Spring Festival travel and the first lockdowns. Daily national mobility and CO2 emissions still moved in lockstep (r = 0.89), and city-level mobility tracked city-level emissions (r = 0.49). Across 366 cities and 60 days, the best model (LightGBM) reached R2 of 1.00, 0.99, and 0.98 for multi-source, domestic aviation, and ground transportation emissions. The framework then generalized: trained the same way, it predicted emissions in Italy, the continental US, and Mexico with R2 of 0.91, 0.95, and 0.99 under GRACED, and R2 of 0.90–0.99 against the independent ODIAC inventory. Beyond prediction, mobility networks revealed functional city clusters—groupings defined by real movement rather than administrative borders—whose emission levels relate to their position in the network, a possible template for coordinated regional climate governance.

The authors are careful with the claim's reach: mobility is a "partial but observable proxy," a complement to inventories rather than a replacement. The Chinese analysis covers a single, extraordinary two-month window; the data capture intercity rather than within-city movement; and the mobility-emission link may itself shift as work and transport change. The authors call for fusing remote sensing, energy structure, and meteorological data next. The payoff, if the framework scales, is concrete: emission estimates that update as fast as society moves, giving climate policymakers a behavior-informed reading of the present rather than a months-old accounting of the past.