PaperlayerView exact revision on arXiv

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

arXiv:2608.23219v1cs.LG

TL;DR

Evidence remains limited on how traffic information contributes to local air-pollution predictions, as traffic–pollutant relationships vary by location and pollutant. This study evaluates traffic and other predictors across six scenarios at London monitoring sites, finding additional predictive value for NO_2, O_3, and PM_10, with traffic contributions comparable to or greater than neighbouring-station measurements.

  • Problem

    Evidence is limited on how traffic information contributes to local air-pollution predictions, while traffic–pollutant relationships vary across locations and pollutants.

    Source: §1 · pp. 1–2

  • Method

    The study evaluated traffic, meteorological, temporal, and neighbouring-station predictors across six scenarios and five London target locations.

    Source: §5 · p. 23

  • Results

    Traffic-related information provided additional predictive value for NO_2, O_3, and PM_10, with SHAP contributions comparable to or greater than neighbouring-station pollutant measurements.

    Source: §5 · p. 23

  • Takeaways & Limitations

    Source-related information may help approximate pollution-source influence without additional sensors or extensive monitoring networks.

    Source: §5 · p. 24

    Predictive benefits were not consistent across pollutants and sites, and statistically significant performance degradations occurred in some cases.

    Source: §3.4 · p. 22

Abstract

from arXiv · show

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 examines whether traffic, temporal, and meteorological data can estimate O3, NO2, PM10, and PM2.5 locally, and whether traffic information adds value across sites and alongside neighbouring-station measurements.

1 Introduction

  • Traffic contributes to CO, PM10, and PM2.5 emissions and emits VOCs and NO2 involved in O3 formation; studies commonly pair traffic predictors with meteorological variables.Traffic measures include vehicle counts, speed, flow, volume, density, and congestion; meteorological predictors include temperature, wind speed, precipitation, humidity, and pressure.Source: §1 · p. 1
  • Traffic’s effect on air quality varies by location and pollutant: lockdown-related NO2 decreases were reported alongside variable O3 and particulate responses, motivating study of when traffic data improve predictions.The cited studies also reported significant decreases in NO, NO2, and NOx but no significant association for PM10 during COVID-19 traffic reductions.Source: §1 · pp. 1–2
  • Prior studies found traffic-data benefits but left local variation and traffic’s independent contribution with multiple neighbouring stations insufficiently characterised.One study reported aggregate improvements across UK sites; another found benefits at only one of two locations for NO2, while emphasizing neighbouring-station measurements.Source: §1 · p. 2
  • The study compares pollutant estimates from traffic, temporal, and meteorological predictors with scenarios that add neighbouring-station observations, with and without traffic variables, across locations.The scenarios assess whether traffic data can complement monitoring infrastructure or support air-quality assessment where monitoring stations are unavailable.Source: §1 · p. 2
  • Traffic stations are the target sites, while background stations provide complementary pollutant measurements; traffic-site estimates represent conditions along a street segment of at least 100 m.Traffic stations are selected because nearby road traffic is expected to make traffic information more relevant; background stations sample a mix of area sources.Source: §1 · p. 2
  • The study uses RF, Extra Trees, LightGBM, and XGBoost, assesses traffic-variable contributions through ablation and feature importance, and compares performance with Ridge, Kriging, and IDW baselines.The spatial baselines use nearby-station measurements, while Ridge provides a linear-regression baseline.Source: §1 · p. 2

2 Materials and Methods

The methods combine traffic, meteorological, temporal, and neighboring-station data to estimate four pollutants at three London monitoring sites. Six predictor scenarios, chronological model validation, spatial interpolation baselines, and multiple performance metrics structure the comparisons.

2.1 Data Collection and Pre-Processing

  • Hourly NO2, PM2.5, PM10, and O3 at three London sites were paired with traffic, weather, temporal, and neighboring-station data for prediction.Traffic predictors used current travel time and speed-derived traffic level, not direct vehicle counts; chronological splitting prevented future-to-past leakage. Preprocessing left 3,232–3,912 hourly observations per station–pollutant pair, and OpenWeather temperature agreed strongly with station observations, while wind-speed estimates were more uncertain.Source: §2.1 · pp. 3–6

2.2 Feature Scenarios

  • Six predictor scenarios crossed traffic inclusion with three network settings: no neighboring data, one nearest background station, or two background plus two traffic stations.All scenarios included target-location temporal and meteorological variables; the neighboring-station settings added pollutant measurements to those predictors.Source: §2.2 · p. 6

2.3 Machine Learning Workflow for Pollutant Estimation

  • Four tree ensembles and Ridge regression were trained using nested chronological cross-validation, with algorithm selection by mean RMSE and final testing on an untouched fifth split.SHAP characterized feature contributions in the best-performing models, but was not interpreted as evidence of causal effects.Source: §2.3 · pp. 6–7

2.4 Spatial Interpolation Baselines

  • Kriging and inverse distance weighting served as spatial baselines using the same four neighboring stations as the densest machine-learning scenario.Kriging selected among spherical, exponential, and Gaussian variograms by chronological-fold RMSE; IDW used distance-squared weights (p = 2).Source: §2.4 · p. 7

2.5 Performance Metrics

  • Test performance was assessed with RMSE and MAE in µg/m3, supplemented by R2; residuals and percentage RMSE changes supported model and scenario comparisons.Positive percentage improvement denotes lower RMSE when the alternative scenario adds predictors; RMSE penalizes larger errors more than MAE does.Source: §2.5 · pp. 7–8

3 Results

The results assess how traffic information and neighbouring-station measurements relate to RMSE across pollutant, site, and scenario combinations, and examine variation in cross-site transferability.

3 Results

  • NO2 and O3 predictions correlated with observations at about 0.9, versus 0.62–0.83 for PM10 and PM2.5; concentration extremes and pollutant peaks were less accurately reproduced.NO2 predictions generally followed observed temporal patterns, while O3, PM10, and PM2.5 had larger magnitude discrepancies, particularly at peaks.Source: §3 · p. 8, Figure 2 · p. 9
  • Adding traffic generally reduced RMSE—consistently for NO2—while adding nearest- or four-station pollutant data improved every combination versus Scen 1; traffic effects varied by site and pollutant.NO2 residuals were more negatively skewed in Scen 1 than in Scen 2 and Scen 3. Gains from moving from Scen 2 to Scen 3 were sometimes small or negative.Source: Table 3 · p. 10, §3 · pp. 10–11

3.1 Analysis of Predictor Contributions

  • SHAP ranked traffic among the six leading features and, for NO2 in Scen 3, traffic often outranked non-NO2 pollutant predictors; PM traffic contributions were generally small.Wind direction and temporal features, especially month in Scen 1 and Scen 2, also contributed strongly; for O3, traffic improved RMSE only in Scen 3, where traffic ranked above several neighbouring pollutant measurements.Source: §3.1 · pp. 11–12

3.2 Comparison with Spatial Interpolation and Linear Regression Baselines

  • Tree-based models generally outperformed Ridge regression across pollutants, stations, and scenarios, while IDW and Ordinary Kriging underperformed even the traffic-only Scen 1 models for NO2.Ridge sometimes had negative R2, and interpolation particularly failed to capture observed variability at higher NO2 concentrations.Source: §3.2 · p. 12, Table 4 · p. 16, Table 5 · p. 16

3.3 Model Validation for Different Target Points with a Cross-Site Model

  • Cross-site transfer depended on location: at Euston, the Marylebone-trained model beat local models in five of six scenarios, whereas at Oxford Street its RMSE was 36.78%–61.65% higher than local models.At Euston, Scen 3 RMSE fell from 15.28 for the local model to 10.60 for the cross-site model; across scenarios, cross-site average RMSE was 12.12 µg/m3 at Euston and 14.22 µg/m3 at Oxford Street.Source: Table 8 · p. 20, §3.3 · p. 20
  • Cross-site transferability findings are limited by the small number of target sites, and geographic proximity alone may not determine transferability.Road characteristics, fleet composition, traffic patterns, meteorology, nearby pollution sources, and local urban characteristics may contribute to differences between sites.Source: §3.3 · p. 20

3.4 Statistical Analysis

  • In Scen 1, significant RMSE reductions of approximately 5.95% to 10.19% occurred for NO2 at Oxford Street and PM10 and PM2.5 at Camden Kerbside.Source: Table 9 · p. 21, §3.4 · p. 21
  • Scen 2 produced significant NO2 RMSE reductions of approximately 4.53% to 9.28% at Marylebone Road, Putney High Street, and Oxford Street, with no significant gains for other pollutants.The other pollutants tested—O3, PM10, and PM2.5—showed no statistically significant improvements.Source: Table 10 · p. 21, §3.4 · p. 21
  • Scen 3 yielded significant RMSE reductions of 4.63% to 11.42% for NO2 at Camden Kerbside, Putney High Street, and Oxford Street, and O3 at Marylebone Road.Traffic variables significantly worsened PM2.5 at Camden Kerbside, while NO2 at Euston Road also had a substantial RMSE increase.Source: Table 11 · p. 21, §3.4 · p. 21
  • Significant traffic-related RMSE reductions were heterogeneous, concentrated mainly in NO2 under Scen 2 and Scen 3, and ranged approximately 4.5–11.4%.PM10 showed only isolated gains, PM2.5 had mixed results including a significant deterioration, and O3 improved significantly only at Marylebone Road under Scen 3, the only traffic-type station with O3 measurements.Source: §3.4 · p. 22

4 Discussion

Discussion finds that traffic’s contribution varies by pollutant and site, and that adding neighboring-station data does not guarantee better predictions. It also highlights data-acquisition costs and positions the framework as a complement to fixed-site monitoring.

4.1 Understanding the Role of Traffic and Spatial Context

  • Traffic’s contribution was pollutant- and site-dependent; NO2 showed the greatest average RMSE improvement across scenarios, while PM10, PM2.5, and O3 responses varied.The study reports that traffic variables improved some estimates but worsened others for these pollutants.Source: §4.1 · p. 22
  • Traffic variables benefited O3 prediction under Scen 3, alongside pollutants involved in O3 chemistry, but O3 was assessed at only one station.The authors suggest traffic may be more informative when combined with predictors related to photochemical formation and titration.Source: §4.1 · p. 22
  • For NO2 and O3, SHAP analysis found traffic variables could outweigh auxiliary-station pollutant measurements, whereas for PM2.5 they did not improve predictions and sometimes worsened them.The authors suggest testing whether some auxiliary-station measurements can be omitted for NO2 or O3, but note traffic variables may add noise for PM2.5.Source: §4.1 · p. 22

4.2 Data Requirement–Performance Trade-Offs in Monitoring Design

  • Scen 3 outperformed Scen 1 across all analysed locations and pollutants, while its comparison with Scen 2 showed deterioration in some cases and only small gains in others.The authors argue that these performance differences should be weighed against the additional data required from multiple monitoring stations.Source: §4.2 · pp. 22–23
  • For PM10, adding stations improved RMSE by below 5.37% across three sites, including one deterioration; for PM2.5, both cases had higher RMSE with four neighbouring stations.The results show that more monitoring-station information does not necessarily improve predictive performance.Source: §4.2 · p. 23
  • Cross-site models may extend estimates to some unmonitored locations with reasonable accuracy, but the limited validation cases prevent identifying which site characteristics enable transfer.The approach is intended to complement, not replace, fixed-site monitoring, which the authors describe as the most reliable source of accurate pollutant measurements.Source: §4.2 · p. 23

4.3 Scalability and Spatial Deployment Considerations

  • Although the framework uses existing infrastructure, commercial traffic and meteorological data can incur query limits or substantial subscription costs, motivating spatial strategies to reduce acquisition expense.Suggested future approaches include Voronoi partitioning, minimum-distance criteria, and spatial coverage optimisation informed by road networks and urban morphology.Source: §4.3 · p. 23

5 Conclusions

Traffic and meteorological data added predictive value for several pollutants, while traffic’s contribution and cross-site transferability depended on context. The study also provides a scenario-based evaluation framework and identifies practical uses, data constraints, and directions for future validation.

5 Conclusions

  • Traffic and meteorological data added predictive value for NO2, O3, and PM10; traffic variables sometimes matched or exceeded neighbouring-station pollutant measurements in feature importance.This was particularly evident for NO2 at Marylebone Road and Putney High Street under Scen 2; under Scen 3, traffic variables ranked above auxiliary pollutants other than NO2 across three main-analysis stations.Source: §5 · p. 23
  • Cross-site models may help at locations without historical pollution data, but performance varied: degradation versus locally trained models ranged from 5.76% to 61.65%, while some scenarios improved by 3.36% to 30.63%.Transferability was assessed at only two additional stations and only for NO2; results cannot be generalised, and locally trained models remain preferable when sufficient historical observations are available.Source: §5 · pp. 23–24
  • The scenario-based evaluation combines four tree-based ensembles with interpolation and regression baselines to quantify how traffic’s predictive contribution shifts as available information changes.Traffic provided complementary information, particularly in traffic-dominated environments, but its contribution was context-dependent; traffic- and background-station pollutant measurements and meteorology generally had greater predictive value and feature importance.Source: §5 · p. 24
  • The framework uses publicly accessible or commercial traffic, weather, and air-quality data, potentially enabling reproducible virtual monitoring without additional sensing infrastructure.Its potential application in low- and middle-income urban contexts depends on equivalent data services being available.Source: §5 · p. 24
  • The London study’s ULEZ sites and June-to-December data period limit generalisability across cities and seasons, particularly for O3 predictions.Future work should test other cities and full-year performance, expand cross-site validation to more sites and pollutants, and evaluate dynamic spatio-temporal models and emission-inventory features at multi-source sites.Source: §5 · p. 24
  • The conclusions support incorporating information linked to pollution sources to approximate their influence without additional sensors or extensive monitoring networks.Potential extensions include other source-related data, satellite imagery, traffic-flow video, and descriptors of green areas, industrial facilities, and other urban characteristics.Source: §5 · p. 24

5.1 Data Availability

  • Traffic and meteorological data cannot be redistributed at large scale; users can query provider APIs under their own credentials and licensing conditions, while historical air-pollution data are publicly available.Code is available on Zenodo.Source: §5.1 · pp. 24–25

5.3 Author contributions

The authors report their individual contributions, funding sources, and no competing interests.

5.3 Author contributions

  • M.B., V.L.-S., and J.R. led conceptualisation, analysis, investigation, and methodology, while V.L.-S. handled data curation, software, validation, visualisation, and original drafting.M.B. managed the project; A.A., M.B., and M.S.-P. contributed resources and supervision; all authors participated in editing and approved the manuscript.Source: §5.3 · p. 25

5.4 Funding

  • The authors report funding from SIP-IPN, STFC, SECIHTI, Fundación Politécnico, QMUL, and COFAA-IPN, among other listed sources and grants.Support included a QMUL Principal’s PhD Studentship for J.R. and a dual doctoral programme between IPN and QMUL for V.L.-S.Source: §5.4 · p. 25

5.5 Competing interests

A Supplementary Material

The supplementary material provides information for all study monitoring stations, including separately analysed validation stations, and accompanying figures and tables documenting the data, models, and analyses.

A Supplementary Material

  • Supplementary material covers all monitoring stations, including validation stations analysed separately from the main experimental dataset.Its figures are provided as ancillary files.Source: A Supplementary Material · p. 29
  • Figure S1 compares OpenWeather meteorological data with air-quality-station observations at Camden Kerbside and Marylebone Road.Source: A Supplementary Material · p. 29
  • Correlation matrices relate temporal, remotely acquired traffic, and meteorological variables to criteria pollutants at five London monitoring stations.The stations are Marylebone Road, Wandsworth – Putney High Street, Camden Kerbside, Westminster – Oxford Street, and Camden – Euston Road.Source: A Supplementary Material · p. 29
  • Figures S7–S10 show observed and predicted PM10, NO2, O3, and PM2.5 concentrations on the final chronological test set, comparing the best-performing tree-based model with Ridge regression.Source: A Supplementary Material · p. 29
  • Supplementary tables document source observation counts, preprocessing and interpolation missingness, model optimisation search spaces, kriging parameters, model performance, and feature-group contributions.Feature-group contributions are grouped by pollutant, station, and scenario.Source: Table S1 · p. 30, Table S2 · p. 30, Table S3 · p. 31

Related papers

Loading 2608.23219v1…