To understand the air we breathe at the street level, we first need to know exactly how many cars are driving on those streets every hour. However, even in a major capital like Paris, permanent traffic sensors cover less than 5% of the road network.
In this paper, researchers present a Graph Neural Network (GNN) approach to bridge this gap, reconstructing hourly traffic for every single street in Paris using only open data.
TL;DR
The authors developed a GNN-based model to estimate hourly traffic flows for the 95% of Paris streets that lack permanent sensors. By combining road attributes with a historical "climatology" baseline, the model achieves high accuracy ($R^2 = 0.89$ for interpolation) and remains robust even when extrapolating to entire districts without sensors.
What Problem Does This Paper Address?
Micro-scale air quality models are essential for urban planning and public health, but they are only as good as their inputs. These models require hourly traffic volumes for every road segment to estimate emissions of pollutants like $NO_x$ and ultrafine particles.
The challenge is that traffic counting networks are inherently sparse and biased. In Paris, sensors (inductive loops) are primarily located on major axes and motorways, leaving residential and local streets—which make up the vast majority of the network—unmonitored. Traditional traffic assignment models require heavy calibration and demand data, while commercial GPS "probe" data can be expensive and difficult to share. This paper seeks a reproducible, open-data solution for city-wide traffic reconstruction.
Methods
The researchers adapted the Graph Neural Network for Urban Interpolation (GNNUI) framework to the specific constraints of a real-world city network. Key methodological choices include:
- Edge-based Decoding: While many GNNs operate on nodes, this model decodes traffic flows directly on the edges (road segments) of a junction graph. This allows the model to retain specific segment attributes like the number of lanes, width, and regulatory speed.
- Climatological Baseline: Instead of predicting absolute flow from scratch, the model predicts a correction relative to a "climatology"—a median hourly profile calculated for different road classes.
- ZINB Likelihood: The model uses a Zero-Inflated Negative-Binomial distribution to handle the statistical nature of traffic data, which often features many zero-flow hours and high variance (over-dispersion).
- Spatial Cross-Validation: The authors employed a rigorous 8-fold spatial cross-validation, specifically distinguishing between interpolation (estimating a missing sensor in a monitored area) and extrapolation (estimating traffic in a district with no sensors at all).
Datasets
The study relies exclusively on open data to ensure the results can be reproduced:
- Road Network: The BD TOPO® database from the French national mapping agency (IGN), covering over 32,000 segments in Paris.
- Traffic Counts: Four years of hourly loop counts (2022–2026) provided by the City of Paris.
- Contextual Data: Meteorological data from Météo-France, land use from OpenStreetMap, population grids from INSEE, and public transport layers from Île-de-France Mobilités.
Key Insights and Ideas
- The "Insurance" of Climatology: The historical baseline (climatology) doesn't help much when sensors are nearby, but it acts as vital "insurance" during extrapolation, preventing the model from making wild guesses in unmonitored districts.
- Input Hierarchy: Road attributes (class, lanes, width) and time cycles are the most critical features. Interestingly, weather data had a negligible effect, likely because the "live" state of the rest of the counting network already captures the impact of weather on traffic behavior.
- Local Influence: The GNN's effective reach is quite local. Increasing the depth of the network (more "hops" in the graph) did not significantly improve performance, suggesting that a street's traffic is primarily influenced by its immediate neighbors.
- Masking Matters: The rate at which sensors are "hidden" during training was the only hyperparameter with a clear impact on performance, highlighting the importance of the masking strategy in self-supervised learning for spatial kriging.
Key Results
- High Accuracy: The model achieved an $R^2$ of 0.89 for interpolation and 0.74 for extrapolation.
- Error Margins: In interpolation, the Mean Absolute Error (MAE) was roughly a quarter of the mean flow. In extrapolation, it rose to about a third.
- Class Performance: Performance was strongest on motorways (18% relative error) but weaker on local streets (approx. 50% relative error), where sensors are rarest.
- Temporal Stability: When tested on data from 2026 (outside the training period), the frozen model's error remained consistent with previous years, proving its utility for ongoing monitoring.
Limitations and Future Work
The authors acknowledge a few key limitations:
- Bias: The model acts as a spatial interpolator pulled toward its prior, resulting in a slight underestimation of the highest flows and overestimation of the lowest.
- Local Street Validation: Because sensors are so rare on local streets (only 19 counters for over 33,000 segments), the validation in these areas is statistically thin.
- Future Data: Integrating floating car data (GPS speeds) could provide the missing observations needed to sharpen estimates for small streets. Additionally, incorporating origin-destination matrices could replace the climatology baseline with a more dynamic demand prior.
Summary updated