Data from about 318 million phone users tracked city emissions almost in real time

Anonymized signaling records from about 318 million active mobile phone users were enough to estimate daily carbon dioxide emissions for 366 Chinese cities over 60 days in early 2020. The model, built by researchers at the National University of Defense Technology and the University of Hong Kong, matched national emissions with a correlation of 0.89 and held up when tested on three other countries.

Save more, waste less: The Cool Down’s newsletter sends smart, simple changes that help you and the planet in your everyday life.

Subscribe to The Cool Down →

We may earn a commission if you sign up.

Conventional emissions inventories are accurate but slow, often trailing reality by months. The research team behind the paper set out to shorten that wait by treating how people move between cities as a stand-in for how much fuel those cities burn. The aim, in the authors’ words as relayed by the release, is a near-real-time carbon meter.

China Unicom signaling data and a 60-day window

The study, titled “Inferring carbon emissions from human mobility data,” appeared in The Innovation, volume 7, issue 11. Its listed authors are Wang M., Zhang X. and Tan S. and colleagues, with Xin Lu of the National University of Defense Technology and Xiaoling Zhang of the University of Hong Kong listed as expert contacts. The model’s fit was scored separately for aviation and ground transportation, with the intercity network’s centrality and clustering measures doing the work of ranking cities. Their mobility data came from China Unicom cellular signaling, covering about 318 million active users, and the team declared no competing interests.

From those records the group built daily intercity networks in which cities are nodes and travel flows are the links between them. Measures such as node centrality and clustering, together with population and geographic attributes, fed a machine-learning model. The best performer was LightGBM, and its training target was the GRACED daily emission dataset, a near-real-time global gridded CO2 record described by Xinyu Dou, Zhu Liu and colleagues.

The window covered January and February 2020, a period that combined Spring Festival travel with the first COVID-19 lockdowns. That produced unusually sharp swings in movement and emissions for a model to learn from.

Correlations, R-squared scores and the three-country test

Daily national mobility and emissions tracked each other closely, with a correlation of 0.89. At the level of individual cities the link was weaker, at 0.49, a clear step down from the national picture. The full model fit better. Measured by R-squared, LightGBM scored 1.00 for multi-source emissions, 0.99 for domestic aviation and 0.98 for ground transportation, according to the EurekAlert research summary.

The group then asked whether a model trained on China could predict emissions elsewhere. Against GRACED, it reached R-squared values of 0.91 for Italy, 0.95 for the continental United States and 0.99 for Mexico. Against ODIAC, the independent Open-source Data Inventory for Anthropogenic CO2 maintained in Japan and documented in Earth System Science Data, the scores ranged from 0.90 to 0.99. The second check matters because GRACED is itself a modeled product, and agreement with a separately constructed inventory makes it harder to blame the fit on a bias shared by both. The scale is still lopsided: the Chinese sample spans 366 cities over 60 days, while the tests on Italy, the continental United States and Mexico are validations of a model that never trained on those countries. The release subtitle says up to 0.99, while the body text reports 1.00 for the best single case, a small inconsistency between the two passages.

Beyond raw prediction accuracy, the networks revealed groupings the authors call functional city clusters. These clusters follow real movement patterns rather than administrative borders, and a cluster’s emission level relates to where its cities sit in the broader network. A hub with heavy flows to many neighbors behaves differently from a peripheral city, which gives planners a way to reason about emissions at the scale at which people actually travel rather than at the scale of municipal boundaries.

The lockdown period suits that purpose. When movement collapsed over a few weeks, slow instruments such as annual inventories could not follow the change, while the phone-based signal could. The authors still treat the two as complements: an inventory remains the accounting standard, and the mobility signal offers an early estimate between releases.

Intercity travel only, and one two-month window

The authors are open about limits. Mobility is a partial but observable proxy for emissions and complements inventories rather than replacing them. The Chinese analysis covers one two-month window, so it says little about how the relationship behaves across seasons or after the pandemic distortions faded. Because signaling data capture movement between cities, trips inside a city fall outside the measurement, and the link between travel and emissions could shift as work patterns and transport technology change.

The phrase almost in real time therefore describes the method’s promise rather than a deployed service. The study demonstrated the approach retrospectively on early 2020 data, and nothing in the release says a live emissions feed is running, and the paper lists no competing interests that would suggest a commercial product behind it. The authors’ proposed next step is to combine mobility records with remote sensing, energy-mix data and weather information, which would test whether the correlation of 0.49 at city level can be improved. No release or paper detail indicates when that follow-up will appear.

This article was produced with the assistance of AI and reviewed by Morning Overview editors prior to publication.

Save more, waste less: Subscribe to The Cool Down →

We may earn a commission if you sign up.