Source-linked AI summary
Leveraging Remote Traffic Data for Local Air Pollutant Estimation: A Scenario-Based Machine Learning Study Across London Monitoring Sites
Valeria Legaria-Santiago, Amadeo Arguelles, Magdalena Saldana-Perez, Jocelyn Richardson, Marcella Bona
TL;DR
The paper examines how much remotely acquired traffic information contributes to local ML air-pollution estimation, given inconsistent traffic effects across pollutants and locations. It evaluates interpretable tree-based models across predictor scenarios with and without neighbouring-station data, finding context-dependent benefits and comparable SHAP contributions from traffic variables in some traffic-dominated settings.
Problem
The contribution of remotely acquired traffic information to local ML air-pollution models remains insufficiently characterised across pollutants and locations.
Method
Four interpretable tree-based models are evaluated across six predictor scenarios using traffic, meteorological, temporal, and neighbouring-station data, with nested time-series cross-validation.
Results
Cross-site performance degraded by 5.76% to 61.65% in some scenarios but improved by 3.36% to 30.63% in others, while traffic variables sometimes matched or exceeded neighbouring-station pollutant contributions.
Takeaways & Limitations
Traffic data can provide valuable complementary information, particularly in traffic-dominated environments, but its predictive contribution is highly context-dependent.
Takeaways & Limitations
Cross-site findings cannot be generalised to other sites, and traffic-related benefits are not consistent across pollutants and locations.
Abstract
from arXiv · showhide
Vehicular traffic is a major source of air pollution; however, the contribution of remotely acquired traffic information to local machine-learning (ML) air-pollution models remains insufficiently characterised. This study evaluates four interpretable tree-based ML models (Random Forest, Extra Trees, LightGBM, and XGBoost) under six predictor scenarios combining progressively larger predictor sets, ranging from remotely acquired traffic, meteorological, and temporal variables alone to the inclusion of measurements from one and four neighbouring monitoring stations, to estimate NO$_2$, PM$_{10}$, PM$_{2.5}$, and O$_3$ concentrations across several sites in London. ML model performance was compared with a ridge linear regression model as a baseline, with spatial interpolation methods and with a cross-site validation experiment. When modelling without data from neighbouring stations, the RMSE for NO$_2$ ranged from 9.73 to 11.66 $μ$g/m$^3$ without traffic information, compared with 8.72 to 11.52 $μ$g/m$^3$ when traffic information was included. Additionally, for NO$_2$, SHAP analyses indicate that traffic-related variables can contribute at levels comparable to pollutant measurements from neighbouring monitoring stations in traffic-dominated~environments.
1 Introduction
The study addresses uncertain, location- and pollutant-dependent traffic–air-quality relationships by testing remotely acquired traffic information in interpretable ML models across London traffic stations. It compares traffic-inclusive and traffic-free predictor scenarios, including models with neighbouring-station observations.
- Research gap: Traffic-related predictors include vehicle counts, speed, flow, volume, density, and road congestion, often combined with meteorological variables.Vehicular traffic emits or contributes to several pollutants, including CO, PM10, PM2.5, VOCs, and NO2-related O3 formation.
- Research gap: Prior traffic-informed studies often relied on single monitoring sites, limiting generalisability and scalability because dense networks are costly to deploy and maintain.A Canadian study using RF and XGBoost achieved strong performance but covered only one roadside location and approximately eight days of traffic data.
- Research gap: Traffic effects on air pollutants vary across locations and pollutants, with NO2 responses more consistent than those of O3 and particulate matter.Lockdown studies associated reduced traffic with lower NO, NO2, and NOx, while PM10 showed no significant association and O3 and particulate-matter responses were more variable.
- Study design: This study evaluates O3, NO2, PM10, and PM2.5 estimation using traffic, temporal, and meteorological data, with comparisons to models incorporating neighbouring-station measurements.Six predictor scenarios assess whether traffic information complements monitoring infrastructure or supports estimation where stations are unavailable.
- Study design: The analysis focuses on London traffic stations, where nearby road traffic is expected to make traffic-related information especially relevant to local pollutant estimation.Traffic-station measurements represent traffic-influenced conditions along a road segment of at least 100 m.
- Study design: Four interpretable tree-based models are used so traffic-variable inclusion can be assessed through ablation and individual feature contributions can be quantified.The models are Random Forest, Extra Trees, LightGBM, and XGBoost.
2 Materials and Methods
The study combines hourly London pollutant, remotely acquired traffic, meteorological, temporal, and neighbouring-station data to estimate pollutants at traffic-type monitoring sites. It uses geographically selected auxiliary stations, preprocessing checks, and time-aware nested validation.
- Study locations and targets: The study estimates NO2, PM2.5, PM10, and O3 at three London traffic-type monitoring stations using hourly pollutant observations.The target sites are London Marylebone Road, Camden Kerbside, and Wandsworth–Putney High Street.
- Traffic data: Traffic data are collected from the road segment closest to each target station, covering distinct urban roads including Marylebone Road, Putney High Street, and Finchley Road.These roads differ in lane configuration, traffic density, and surrounding land use.
- Traffic data: Traffic predictors use current travel time and traffic level rather than direct vehicle counts, with traffic level defined relative to free-flow speed.The raw source provides current travel time, current speed, and free-flow travel time; hourly averages are aligned to pollutant timestamps.
- Meteorological and temporal data: Meteorological predictors come from OpenWeather, while temporal variables encode hour, weekday, and month using sine and cosine transformations.Temperature and wind speed are compared with monitoring-station observations; temperature agrees strongly, whereas wind-speed estimates are more uncertain.
- Neighbouring stations: For each target, the study independently selects the two nearest traffic-type and two nearest background-type stations from 13 Central London stations.Nearest stations are identified using Haversine distance, and their measured pollutants are used as additional predictors in scenarios 2 and 3.
- Preprocessing and validation: Models use 3232–3912 hourly observations per station–pollutant combination after preprocessing and are evaluated with nested chronological cross-validation.TimeSeriesSplit prevents future-to-past leakage by expanding chronologically ordered training sets; validation sites remain excluded from primary analyses until external validation.
3 Results
Across London traffic monitoring sites, model performance varied by pollutant and scenario, but traffic and neighbouring-station information generally improved predictions. Traffic was especially informative for NO2, while weather and temporal variables also contributed substantially.
- Traffic effects: Traffic-related variables reduced RMSE for NO2 across all three stations and scenarios, whereas effects for O3, PM10, and PM2.5 depended on station and scenario.Adding traffic increased RMSE in some cases, including O3 under Scen 1 and Scen 2, but improved it under Scen 3.
- Spatial information: Adding the nearest background-station pollutant measurement improved performance for every station–pollutant combination with traffic predictors.The same universal improvement was observed when four neighbouring-station measurements were added relative to Scen 1.
- Spatial information: 51.04% RMSE improvement occurred for PM2.5 at Camden Kerbside from Scen 1 to Scen 2, compared with 48.45% from Scen 1 to Scen 3.The smaller multi-station improvement suggests that additional neighbouring measurements may provide redundant information in this case.
- Spatial information: Less than 6% additional improvement from Scen 2 to Scen 3 occurred for NO2 at Putney High Street and PM10 at all sites.Performance decreased for PM2.5 and PM10 at Putney High Street, while O3 and NO2 at other stations improved by more than 15%.
- Feature contributions: Wind direction and month were among the strongest contributors, with weather generally influential and temporal variables especially important in Scenarios 1 and 2.Feature-group contributions were quantified using aggregated mean absolute SHAP values.
- Feature contributions: Traffic features ranked among the six most important predictors and contributed 26.26% at Putney High Street versus 11.88% at Marylebone Road in Scen 1.At Camden Road and Putney High Street, traffic variables ranked above temporal variables, but traffic ranked below nearest-station pollutants at all stations.
- Feature contributions: For NO2 under Scen 3, traffic information contributed at levels comparable to or greater than several neighbouring-station pollutant measurements.The comparison included PM10, PM2.5, NO, CO, and O3 predictors; NO2 was the only pollutant predictor consistently ranked higher than traffic.
- Feature contributions: Traffic contributions were generally small for PM2.5 and PM10, even when traffic produced incremental RMSE reductions.These models relied considerably more on other predictors than on traffic features.
4 Discussion
The discussion finds that traffic information can improve pollutant estimation, but its value varies by pollutant, station, scenario, and data requirements. Cross-site models may extend estimation to some locations, although fixed monitoring remains the most reliable source and deployment costs constrain large-scale use.
- Understanding the Role of Traffic and Spatial Context: O3 traffic information becomes beneficial under Scen 3, where predictors include pollutants involved in O3 chemistry, whereas simpler scenarios show weaker benefits.The authors suggest that traffic data may be more informative when combined with variables related to photochemical formation and titration, but O3 was evaluated at only one station.
- Data Requirement–Performance Trade-Offs in Monitoring Design: Nearby-station pollutant measurements generally produce lower RMSE values than traffic, temporal, and meteorological variables alone, but Scen 3 versus Scen 2 involves small gains or deterioration in some cases.Adding information from multiple monitoring stations should therefore be assessed against its additional data requirements.
- Understanding the Role of Traffic and Spatial Context: SHAP analysis found that traffic-related variables can be as important as, and sometimes more important than, pollutant measurements from auxiliary stations for NO2 and O3 prediction.The authors propose investigating whether some auxiliary-station pollutants could be excluded to reduce data requirements.
- Data Requirement–Performance Trade-Offs in Monitoring Design: For PM10, moving from one to four neighbouring stations improved RMSE by less than 5.37%, while both analysed PM2.5 cases worsened with the additional stations.These results illustrate that more monitoring-station information does not consistently justify its added data requirements.
- Cross-Site Validation and Deployment: Cross-site models may extend air-quality estimation to some locations, but the framework is intended to complement rather than replace fixed-site monitoring networks.The approach is positioned for areas with unavailable or sparse monitoring where remotely acquired traffic and meteorological data exist.
- Scalability and Spatial Deployment Considerations: Large-scale deployment requires careful query-point selection because commercial API costs depend on queried locations and sampling frequency, with traffic data particularly expensive.The authors identify spatial-query optimisation as future work for reducing operational costs.
5 Conclusions
Across five London monitoring sites, remotely acquired traffic information added predictive value for several pollutants and could support modelling where local observations are unavailable. The study also identifies transferability, seasonal coverage, data access, and London-specific conditions as important boundaries.
- Conclusions: Traffic information was systematically evaluated across six predictor scenarios and five London target locations using a scenario-based comparison of ensemble models and baselines.The design decomposed information gain as the available predictor set changed.
- Conclusions: Traffic and meteorological variables provided additional predictive value for NO2, O3, and PM10 estimation.For NO2, traffic variables could rank above auxiliary pollutant measurements in feature importance under some scenarios.
- Conclusions: Cross-site models may provide a practical alternative where historical pollution measurements are unavailable but real-time traffic, temporal, and meteorological data can be obtained.Performance changes relative to locally trained models ranged from 5.76% degradation to 30.63% improvement, although the results were not generalisable to other sites.
- Conclusions: Locally trained models remain preferable when sufficient historical data are available because successful transferability could not be determined from the limited number of sites.Geographical proximity alone may not be sufficient; urban morphology, road characteristics, nearby emission sources, and local environmental conditions may also matter.
- Conclusions: The framework can be reproduced and deployed without additional sensing infrastructure using publicly accessible or commercially available data sources, subject to provider request limits.The data sources include UK-AIR, OpenWeather, and TomTom.
- Conclusions: The study’s generalisability is constrained by its June-to-December period, ULEZ setting, restricted data redistribution, and limited cross-site validation.Future work should assess other cities and climates, extend monitoring to the full year, use broader transferability samples, and explore additional data and modelling approaches.
- Conclusions: Incorporating information directly related to pollution sources may approximate source influence without additional sensors or extensive monitoring networks.The authors also suggest considering other source-related data, including satellite imagery, traffic-flow video, and descriptors of surrounding environments.
5.3 Author contributions
The paper assigns contributions across conceptualisation, data curation, analysis, methodology, software, validation, writing, funding, and supervision, and reports no conflicts of interest.
- Author contributions: Conceptualisation, data curation, formal analysis, investigation, methodology, validation, visualisation, and writing were shared among the listed authors.The contribution statement identifies specific authors for each role.
- Author contributions: Funding acquisition, project administration, resources, software, supervision, and review and editing were assigned to designated authors.The statement distinguishes operational, technical, supervisory, and editorial responsibilities.
- Author contributions: The authors report funding from SIP-IPN, ProRed, STFC, SECIHTI, Fundación Politécnico IPN, and QMUL programmes.The listed support covers multiple institutional and doctoral-programme sources.
- Author contributions: The authors declare no conflicts of interest.This is stated explicitly in the disclosure section.
A Supplementary Material
The supplementary material documents station coverage, validation-related analyses, data quality, model optimisation, interpolation settings, performance, feature contributions, and observed-versus-predicted pollutant series.
- Supplementary Material: The ancillary files include supplementary figures for all monitoring stations, including validation stations analysed separately from the main experimental dataset.The material is listed as accompanying the submission.
- Supplementary Material: Figures S1–S5 compare OpenWeather meteorological data or show correlation matrices linking temporal, remotely acquired traffic, meteorological, and pollutant variables across London stations.The correlation matrices cover Camden–Euston Road, Marylebone Road, Putney High Street, Camden Kerbside, and Westminster–Oxford Street.
- Supplementary Material: Tables S1–S3 report data-source observations, preprocessing and missing values, and model hyperparameter search spaces.The model abbreviations include RF, ETR, LGBM, and XGB.
- Supplementary Material: Tables S4–S5 document Ordinary Kriging variogram settings and best-performing models with metrics by pollutant, station, and scenario.Table S4 focuses on NO2 interpolation, while Table S5 covers the model-performance comparison.
- Supplementary Material: Table S6 reports feature-group contributions by pollutant, station, and scenario, including neighbouring-station pollutant measurements.The N-S pollutant notation refers to pollutants measured at neighbouring stations for the relevant scenario and target site.