Smartphone data and AI reveal true climate toll of urban mobility

The future of sustainable transport planning may already be sitting in people’s pockets. By transforming everyday smartphone signals into high-resolution mobility data, researchers have reconstructed how residents of Cuenca travel across the city and what those patterns mean for energy consumption and carbon emissions.

Their findings are detailed in Smart Mobility Analytics: Inferring Transport Modes and Sustainability Metrics from GPS Data and Machine Learning, published in Atmosphere, where the researchers outline a machine learning framework capable of identifying walking, cycling, tram, bus, and private vehicle trips while estimating the environmental footprint of each mode under high-altitude conditions.

Machine learning meets urban mobility

The research team moved beyond traditional mobility surveys, which are costly, infrequent, and often affected by recall bias. Instead, participants carried smartphones that continuously recorded geospatial and motion data at a frequency of one sample per second during daily trips. The dataset ultimately reached 8,997,224 samples across five transport modes: walking, cycling, tram, bus, and private vehicle.

From these raw GPS traces, the researchers derived dynamic and spatial predictors including speed, longitudinal acceleration, jerk, total longitudinal force, and minimum distance to public transport routes. Unlike many prior studies that rely only on speed profiles, this framework integrated physics-based resistance modeling and spatial proximity to transport infrastructure.

To refine the model, the team applied the Football Optimization Algorithm, a metaheuristic feature selection method inspired by team strategy dynamics. This algorithm reduced the predictor set to five core variables without sacrificing classification performance. The selected predictors captured both movement dynamics and spatial alignment with the public transport network.

Three supervised learning models were trained and tested: classification tree, support vector machine, and artificial neural network. The classification tree emerged as the most robust performer, achieving an overall accuracy of 84.2 percent. It identified pedestrian trips with over 99 percent accuracy and bicycle trips with nearly 99 percent precision. Tram detection exceeded 93 percent accuracy, while private vehicles reached more than 83 percent. Bus classification proved more challenging at roughly 64 percent accuracy, largely due to overlapping speed and stop patterns with private cars operating along shared corridors.

Despite this limitation, the model demonstrated that a compact set of physically meaningful variables can effectively distinguish multiple transport modes under real-world urban conditions. Once validated, the classifier was applied to the full dataset to infer mobility patterns and sustainability metrics at scale.

A city that walks in time, drives in distance

The results revealed a striking imbalance between time spent traveling and distance covered across transport modes.

Walking accounted for 57.93 percent of total monitored travel time, making it the dominant mode in terms of daily time allocation. Yet pedestrians covered less than one percent of total distance traveled. The average walking speed was 3.09 kilometers per hour, below commonly cited optimal walking speeds for maintaining functional health. The data suggest that pedestrian mobility in Cuenca consists largely of short, fragmented trips with frequent stops, possibly influenced by infrastructure gaps, intersection delays, or urban layout constraints.

Cycling followed a similar pattern. Although it represented 7 percent of total travel time, it contributed less than one percent of total distance. The average cycling speed of 10.63 kilometers per hour fell below expected urban cycling benchmarks, pointing to limited cycling infrastructure, topographic challenges, or low cycling efficiency in daily travel.

Motorized modes displayed the opposite pattern. Private vehicles accounted for only 10.44 percent of total travel time but represented 60.9 percent of total distance traveled. With an average speed of 29.49 kilometers per hour, private cars proved to be the most spatially efficient mode in covering longer distances.

Buses accounted for 21.86 percent of total travel time and 33.25 percent of total distance. Their average speed of 14.68 kilometers per hour positioned them between trams and private vehicles. The tram, though responsible for only 2.76 percent of travel time, covered 4.58 percent of total distance, highlighting its efficiency in medium-distance corridors.

Overall, the findings show that active modes dominate time use, while motorized modes dominate spatial reach. Residents may walk often, but they rely on vehicles for longer trips, reinforcing a structure of partial car dependency.

Age-based analysis added another dimension. Younger participants distributed their travel distances more evenly across walking, bus, and private vehicles. Middle-aged participants showed growing reliance on cars. Among those aged 39 and older, daily distance traveled by private vehicle rose sharply, reaching more than three times the distance walked. At the same time, active mobility declined with age. The data point to a gradual shift toward car dependency as residents grow older, raising concerns about long-term sustainability and public health.

Carbon burden falls on motorized transport

The study extended beyond classification to estimate energy use and CO2 emissions. For walking and cycling, energy expenditure was calculated using established metabolic models that account for body mass, speed, and terrain. For motorized modes, the researchers applied longitudinal force integration and emission factors adjusted for Cuenca’s high altitude of 2,550 meters above sea level. A locally representative air density value was incorporated to ensure physically consistent aerodynamic resistance modeling.

The results leave little ambiguity. Motorized transport generated more than 98 percent of total CO2 emissions in the sample.

Buses accounted for 78.92 percent of total emissions. Private vehicles contributed 19.27 percent. The tram represented just over 1 percent, while walking and cycling together accounted for less than 2 percent of emissions. Pedestrian activity contributed 1.58 percent, reflecting indirect metabolic emissions rather than fuel combustion, while cycling added only 0.19 percent.

The dominance of bus emissions reflects aggregate system-level activity rather than inefficiency per passenger. Because buses operate across numerous routes with extended service hours, their total emissions accumulate rapidly. The study did not include passenger occupancy data, meaning emissions were not normalized per passenger-kilometer. Even so, the overall message remains consistent: fossil-fuel-based mobility drives the city’s carbon footprint.

Private vehicles, though representing a smaller share of total emissions compared to buses, carried the highest distance burden per user and displayed lower environmental efficiency when considering limited passenger occupancy and reliance on combustion engines.

The tram emerged as a comparatively low-emission motorized alternative, largely dependent on the electricity generation mix. Active modes remained the most environmentally sustainable options, though their limited spatial contribution constrains their system-wide impact.

The authors clearly mention that the study is exploratory and based on a purposive sample of 50 participants. The monitoring period lasted 20 days, and while the dataset is extensive in terms of observations, it is not statistically representative of the entire urban population. Emission estimates are intended as system-level indicators rather than regulatory inventories.

Classification uncertainty, particularly between buses and private vehicles, may redistribute emission shares between those two categories. However, such uncertainty does not alter the broader conclusion that motorized transport overwhelmingly dominates urban CO2 output.