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

cs.LGAug 21, 2026

Leveraging Remote Traffic Data for Local Air Pollutant Estimation: A Scenario-Based Machine Learning Study Across London Monitoring Sites

Authors: Valeria Legaria-Santiago, Amadeo Arguelles, Magdalena Saldana-Perez, Jocelyn Richardson, Marcella Bona

Organizations: Department of Physics and Astronomy, Queen Mary University of London, United Kingdom · Centro de Investigacion en Computacion, Instituto Politecnico Nacional, Ciudad de Mexico, Mexico

Abstract

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.

Figures & tables

Figure 1: Locations of the pollutant prediction sites and the validation sites (all corresponding to traffic-type monitoring stations), together with the auxiliary stations (including both background and traffic-type stations).

Distance Between Stations	Camden Kerbside	London Marylebone Road	Wandsworth–Putney High Street	Camden–Eusto Road	Westminster–Oxford Street
1st nearest background station	London N. Kensington (3.69 km)	London Bloomsbury (1.99 km)	Wandsworth–Putney (0.18 km)	London Bloomsbury (0.66 km)	London Bloomsbury (2.08 km)
2nd nearest background station	London Bloomsbury (4.20 km)	London Westminster (3.47 km)	London N. Kensington (6.41 km)	London Westminster (3.71 km)	London Westminster (2.58 km)
1st nearest traffic station	Camden High Street (2.23 km)	Camden High Street (2.10 km)	Richmond Upon Thames–Castelnau (2.38 km)	Camden High Street (1.70 km)	Camden High Street (2.98 km)
2nd nearest traffic station	Brent–ARK Franklin Primary Academy (3.22 km)	RBKC Knightsbridge (2.69 km)	Wandsworth–Lavender Hill (3.41 km)	RBKC Knightsbridge (4.04 km)	RBKC Knightsbridge (1.83 km)

Table 1: Distances between the target locations used in the main and validation analyses and their corresponding neighbouring background and traffic-type stations.

Station	Pollutants Measured
Wandsworth–WA9	PM 10 , NO 2
London Bloomsbury	O 3 , Nitric oxide (NO), NO 2 , Nitrogen oxides as NO 2 , SO 2 , CO, PM 10 , PM 2.5
London N. Kensington	NO, NO 2 , Nitrogen oxides as NO 2 , SO 2 , CO, PM 10 , PM 2.5
London Westminster	O 3 , NO, NO 2 , Nitrogen oxides as NO 2 , PM 2.5
Wandsworth–Lavender Hill	NO 2 , PM 10
Richmond Upon Thames–Castelnau	NO 2 , PM 10

Table 2: Pollutants Measured in the neighbouring stations considered in Scen 2 and Scen 3.

Figure 2: Observed vs predicted concentrations for the best-performing model (scenario and model) per station–pollutant pair. RMSE and the Pearson correlation coefficient ( $r$ ) are shown in the upper-left corner of each panel.

Figure 3: Distribution of residuals for each scenario (with and without traffic) of the best-performing model. Distributions of Scen 1 with and without traffic are highlighted in blue and light green, respectively, for improved visual distinction (creating dark green where the two overlap).

Case	Pollutant	Station	S1-noT	S1	S2-noT	S2	S3-noT	S3	$\Delta$ S1	$\Delta$ S2	$\Delta$ S3	$\Delta$ S2/S1	$\Delta$ S3/S1	$\Delta$ S3/S2
1	NO 2	Camden Kerbside	11.43	11.27	8.94	8.86	7.50	7.16	11.46	0.91	4.63	21.34	36.48	19.25
2	NO 2	Marylebone Road	9.73	8.72	8.77	7.96	6.80	6.43	10.38	9.28	5.44	8.77	26.24	19.15
3	NO 2	Putney High Street	11.66	11.52	9.81	9.37	9.45	8.80	11.21	4.53	6.86	18.68	23.61	6.06
4	O 3	Marylebone Road	10.58	11.20	8.04	8.21	7.22	6.91	$-$ 5.79	$-$ 2.12	4.21	26.69	38.25	15.77
5	PM 10	Camden Kerbside	11.34	10.53	9.72	9.70	9.31	9.23	7.14	0.14	0.88	7.88	12.38	4.88
6	PM 10	Marylebone Road	8.36	8.43	5.85	5.84	5.57	5.52	0.90	0.26	0.91	30.82	34.53	5.37

Table 3: RMSE values per scenario and percentage improvement across scenarios (

\Delta

RMSE(%) defined in Equation ( 4 )).

Figure 4: Percentage improvement in RMSE (calculated with Equation ( 4 )) obtained by incorporating additional spatial pollutant information compared with using the simplest predictor set that includes traffic-related variables (Scen 1). The x-axis shows the percentage improvement obtained by adding observations from the nearest auxiliary monitoring station (Scen 2 vs. Scen 1), while the y-axis shows the improvement achieved by incorporating multi-station information (Scen 3 vs. Scen 1). Each point corresponds to a station–pollutant combination.

Figure 5: SHAP beeswarm plots of the best-performing models for NO 2 predictions across stations and scenarios. Each point represents an observation, and its colour indicates the feature value, ranging from low (blue) to high (red). Features are ranked according to their global SHAP importance, from most important (top) to least important (bottom). The horizontal spread indicates the magnitude of each feature’s impact on the predicted pollutant concentration.

Figure 6: SHAP beeswarm plots of the best-performing models for O 3 predictions across stations and scenarios.

Figure 7: SHAP beeswarm plots of the best-performing models for PM 2.5 predictions across stations and scenarios.

Figure 8: SHAP beeswarm plots of the best-performing models for $PM_{10}$ predictions across stations and scenarios.

Station	Scenario	Best ML	Best ML			Ridge
		Model	RMSE	MAE	$R^{2}$	RMSE	MAE	$R^{2}$
Marylebone Road	Scen1-noT	LGBM	9.73	7.36	0.53	11.43	9.15	0.35
	Scen1	XGB	8.72	6.54	0.62	11.74	9.41	0.32
	Scen2-noT	ETR	8.77	6.82	0.62	8.41	6.81	0.65
	Scen2	XGB	7.96	6.17	0.69	9.14	7.32	0.59
	Scen3-noT	LGBM	6.80	5.21	0.77	7.16	5.77	0.75

Table 4: NO 2 model performance: RMSE, MAE, and

R^{2}

for each station and scenario for the best-performing tree-based model and the Ridge regression model.

Station	Scenario	Best ML	Best ML			Ridge
		Model	RMSE	MAE	$R^{2}$	RMSE	MAE	$R^{2}$
Marylebone Road	Scen1-noT	LGBM	10.58	8.47	0.38	15.49	12.82	$-$ 0.34
	Scen1	LGBM	11.20	9.05	0.30	15.93	13.31	$-$ 0.42
	Scen2-noT	LGBM	8.04	6.32	0.64	8.55	6.82	0.59
	Scen2	LGBM	8.21	6.41	0.62	8.89	7.01	0.56
	Scen3-noT	LGBM	7.22	5.66	0.71	7.20	5.94	0.71

Table 5: O 3 model performance: RMSE, MAE, and

R^{2}

for each station and scenario for the best-performing tree-based model and the Ridge regression model.

Station	Scenario	Best ML	Best ML			Ridge
		Model	RMSE	MAE	$R^{2}$	RMSE	MAE	$R^{2}$
Marylebone Road	Scen1-noT	RF	8.36	6.11	0.22	9.20	6.85	0.05
	Scen1	RF	8.43	6.27	0.20	9.20	6.84	0.05
	Scen2-noT	LGBM	5.85	4.01	0.62	6.52	4.39	0.52
	Scen2	LGBM	5.84	4.02	0.62	6.52	4.39	0.52
	Scen3-noT	LGBM	5.57	3.89	0.65	5.83	4.15	0.62

Table 6: PM 10 model performance: RMSE, MAE, and

R^{2}

for each station and scenario for the best-performing tree-based model and the Ridge regression model.

Station	Scenario	Best ML	Best ML			Ridge
		Model	RMSE	MAE	$R^{2}$	RMSE	MAE	$R^{2}$
Marylebone Road	Scen1-noT	LGBM	4.38	3.30	0.18	4.55	3.24	0.12
	Scen1	XGB	4.54	3.44	0.12	4.61	3.28	0.09
	Scen2-noT	ETR	3.37	2.44	0.52	3.39	2.55	0.51
	Scen2	RF	3.32	2.44	0.53	3.49	2.65	0.48
	Scen3-noT	LGBM	3.46	2.52	0.49	3.36	2.58	0.52

Table 7: PM 2.5 model performance: RMSE, MAE, and

R^{2}

for each station and scenario for the best-performing tree-based model and the Ridge regression model.

Figure 9: Observed versus predicted NO 2 concentrations for each station, comparing machine learning models (Scen 1 and Scen 3) with spatial interpolation approaches (Ordinary Kriging and IDW). Interpolation was performed using measurements from four nearby stations of the same pollutant, providing a spatial baseline for comparison.

Figure 10: Distribution of pollutant concentrations at the validation and training stations.

Figure 11: Monthly distribution of NO 2 concentrations across monitoring stations. Boxplots represent the concentration variability for each month, while points and connecting lines indicate the monthly mean concentrations at each station.

Figure 12: Mean hourly NO 2 concentrations throughout the study period at Marylebone Road and the validation stations.

Station	Scenario	RMSE Cs	RMSE L	RMSE R	RMSE IDW	RMSE OK	$D_{g}$ Cs vs L (%)	$D_{g}$ Cs vs R (%)	$D_{g}$ Cs vs IDW (%)	$D_{g}$ Cs vs OK (%)
Euston	S1	12.59	14.33	14.39	18.67	16.09	$-$ 12.14	$-$ 12.51	$-$ 32.57	$-$ 21.74
Euston	S1-noT	13.58	12.84	13.09	18.67	16.09	5.76	3.74	$-$ 27.26	$-$ 15.58
Euston	S2	12.24	13.20	13.85	18.67	16.09	$-$ 7.27	$-$ 11.62	$-$ 34.44	$-$ 23.91
Euston	S2-noT	12.67	13.11	12.10	18.67	16.09	$-$ 3.36	4.71	$-$ 32.14	$-$ 21.24
Euston	S3	10.60	15.28	12.62	18.67	16.09	$-$ 30.63	$-$ 16.01	$-$ 43.22	$-$ 34.11
Euston	S3-noT	11.03	12.98	10.03	18.67	16.09	$-$ 15.02	9.97	$-$ 40.92	$-$ 31.43

Table 8: Performance comparison of the cross-site model with the local model (L), Ridge regression (R), inverse distance weighting (IDW), and Ordinary Kriging (OK) for NO 2 estimation.

Pollutant	Station	Best Model	$\Delta\text{RMSE}$ (%)	95% CI	Holm-Adjusted $p$ -Value	Significant Improvement
NO 2	Marylebone Road	XGB	10.38	[0.56, 16.70]	0.058	No
NO 2	Camden Kerbside	XGB	1.46	[ $-$ 0.37, 3.08]	0.214	No
NO 2	Putney High Street	XGB	1.21	[ $-$ 2.07, 4.57]	0.471	No
NO 2	Oxford Street	XGB	5.95	[0.89, 10.18]	0.030	Yes
NO 2	Euston Road	ETR	$-$ 11.61	[ $-$ 21.48, $-$ 1.96]	0.072	No
O 3	Marylebone Road	LGBM	$-$ 5.79	[ $-$ 10.71, $-$ 0.85]	0.045	No

Table 9: Paired 24-h block-bootstrap comparison between Scen 1 and Scen 1-noT.

Pollutant	Station	Best Model	$\Delta\text{RMSE}$ (%)	95% CI	Holm-Adjusted $p$ -Value	Significant Improvement
NO 2	Marylebone Road	XGB	9.28	[2.62, 14.21]	0.020	Yes
NO 2	Camden Kerbside	XGB	0.91	[ $-$ 2.14, 3.91]	0.555	No
NO 2	Putney High Street	LGBM	4.53	[1.71, 7.78]	0.007	Yes
NO 2	Oxford Street	LGBM	8.78	[3.93, 12.98]	0.002	Yes
NO 2	Euston Road	XGB	$-$ 0.71	[ $-$ 9.22, 7.36]	0.873	No
O 3	Marylebone Road	LGBM	$-$ 2.12	[ $-$ 5.48, 1.62]	0.233	No

Table 10: Paired 24-h block-bootstrap comparison between Scen 2 and Scen 2-noT.

Pollutant	Station	Best Model	$\Delta\text{RMSE}$ (%)	95% CI	Holm-Adjusted $p$ -Value	Significant Improvement
NO 2	Marylebone Road	LGBM	5.44	[ $-$ 1.50, 10.59]	0.093	No
NO 2	Camden Kerbside	LGBM	4.63	[1.09, 8.11]	0.043	Yes
NO 2	Putney High Street	XGB	6.86	[4.08, 9.20]	¡0.001	Yes
NO 2	Oxford Street	XGB	11.42	[5.30, 17.34]	0.002	Yes
NO 2	Euston Road	LGBM	$-$ 17.67	[ $-$ 31.94, $-$ 2.92]	0.072	No
O 3	Marylebone Road	XGB	4.21	[1.11, 6.64]	0.027	Yes

Table 11: Paired 24-h block-bootstrap comparison between Scen 3 and Scen 3-noT.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Station	Pollutant	Pollution	Traffic	Pollution + traffic	Weather	Bkg0	Bkg1	Trf0	Trf1	Final merge
Euston Road	NO2	4152	4510	3747	4151	4151	4151	4151	4151	3747
Camden Kerbside	NO2	4152	4706	3943	4152	4151	4151	4151	4151	3943
Camden Kerbside	PM10	4152	4706	3943	4152	4151	4151	4151	4151	3943
Camden Kerbside	PM25	4152	4706	3943	4152	4151	4151	4151	4151	3943
Marylebone Road	NO2	4152	4707	3944	4152	4151	4151	4151	4151	3944
Marylebone Road	O3	4152	4707	3944	4152	4151	4151	4151	4151	3944

Appendix

Table S1: Number of observations available from each data source before and after merging all data sources.

Station	Pollutant	After cleaning	Missing target	After target removal	Missing predictors	Rows with missing	Interpolated values	Rows with missing (%)	Missing values interpolated (%)
Euston Road	NO2	3747	515	3232	424	264	236	8.17	55.66
Camden Kerbside	NO2	3943	31	3912	866	517	443	13.22	51.15
Camden Kerbside	PM10	3943	103	3840	871	503	449	13.10	51.55
Camden Kerbside	PM25	3943	74	3869	880	507	464	13.10	52.73
Marylebone Road	NO2	3944	53	3891	612	336	304	8.64	49.67
Marylebone Road	O3	3944	334	3610	560	316	293	8.75	52.32

Appendix

Table S2: Number of observations and missing values during data preprocessing and temporal interpolation.

Model	Hyperparameter	Search space	Scale / condition
RF	n_estimators	[200, 800], step 100	integer
RF	max_depth	[4, 30]	integer
RF	min_samples_split	[2, 30]	integer
RF	min_samples_leaf	[1, 15]	integer
RF	max_features	{sqrt, log2, 0.3, 0.5, 0.7, 0.9}	categorical
RF	bootstrap	{True, False}	categorical

Appendix

Table S3: Hyperparameter search space used for model optimisation. RF: Random Forest; ETR: Extra Trees Regressor; LGBM: LightGBM; XGB: XGBoost.

Station	Variogram	Partial sill	Sill	Range (km)	Nugget
Marylebone Road	Spherical	31.3479	34.4045	1.8608	3.0566
Camden Kerbside	Spherical	10.2275	13.8756	2.2328	3.6481
Putney High Street	Exponential	0.0011	42.7250	0.0373	42.7239
Oxford Street	Spherical	22.0926	31.0504	5.0849	8.9578
Euston Road	Exponential	62.5974	66.6085	0.0355	4.0111

Appendix

Table S4: Selected variogram models and fitted parameters for Ordinary Kriging of

NO_{2}

in the Kriging interpolation baseline for each station.

Pollutant	Station	Scenario	Best model	RMSE	MAE	$R^{2}$
NO 2	Marylebone Road	S1	XGB	10.22 $\pm$ 1.71	7.73 $\pm$ 1.23	0.33 $\pm$ 0.17
NO 2	Marylebone Road	S1-noT	LGBM	10.20 $\pm$ 1.93	7.73 $\pm$ 1.26	0.34 $\pm$ 0.14
NO 2	Marylebone Road	S2	XGB	9.17 $\pm$ 1.30	6.85 $\pm$ 0.86	0.47 $\pm$ 0.07
NO 2	Marylebone Road	S2-noT	ETR	9.30 $\pm$ 1.57	7.02 $\pm$ 0.95	0.46 $\pm$ 0.09
NO 2	Marylebone Road	S3	LGBM	8.01 $\pm$ 1.43	5.96 $\pm$ 1.06	0.59 $\pm$ 0.09
NO 2	Marylebone Road	S3-noT	LGBM	8.16 $\pm$ 1.78	5.92 $\pm$ 1.02	0.58 $\pm$ 0.10

Appendix

Table S5: Best-performing model and performance metrics by pollutant, station, and scenario.

Pollutant	Station	Scenario	Model	Temporal	Meteorological	Traffic	N-S pollutants
NO 2	Euston Road	Scen 1	ETR	36.58	37.14	26.28	–
NO 2	Euston Road	Scen 2	XGB	18.62	44.95	25.57	10.85
NO 2	Euston Road	Scen 3	LGBM	9.08	22.54	13.97	54.42
NO 2	Camden Kerbside	Scen 1	XGB	19.64	55.78	24.58	–
NO 2	Camden Kerbside	Scen 2	XGB	18.70	32.64	19.11	29.55
NO 2	Camden Kerbside	Scen 3	LGBM	11.02	17.16	13.05	58.77

Appendix

Table S6: Percentage contribution (%) of feature groups to the model predictions, grouped by pollutant, station, and scenario. N-S pollutants refer to the pollutants measured at the neighbouring stations from the corresponding scenario and the site to be estimated.

Air Quality Prediction

Air quality prediction estimates future pollutant levels and maps local exposure to support health warnings, planning, and pollution control.

See all 42 papers

From Deterministic to Generative Deep Learning for Urban Air Quality Reconstruction from Sparse Observations
Jul 28, 2026Abhishek A. Sabnis, Mihai Mitrea, Lya Lugon +5Conditional Generative Modeling Inverse Problems
CEREA, ENPC, Institut Polytechnique de Paris, ˆIle-de-France, France. · Department of Computing, Imperial College London, London, UK. · Dynamic Systems Lab, University College London, London, UK. +1
AI for Social Good: An Investigation of the Causal Relationship Between Environmental Regulations and Their Effects on Air Pollution in London, UK
Jun 13, 2026Yang Han, Jacqueline CK Lam, Victor OK Li +1Causal Effect Estimation Policy Evaluation
Department of Electrical and Electronic Engineering, The University of Hong Kong, Hong Kong
Interpretable Machine Learning for Air Pollution and Respiratory Health Prediction: A Socioeconomic Subgroup Analysis
Jul 19, 2026Maede Azani Hassan Abadi, Shouyi WangInterpretable ML Air Quality Prediction
Department of Industrial, Manufacturing, & Systems Engineering, The University of Texas at Arlington

Jul 28, 2026J/K move · Enter open · S save

From Deterministic to Generative Deep Learning for Urban Air Quality Reconstruction from Sparse Observations

Abhishek A. Sabnis, Mihai Mitrea, Lya Lugon, Karine Sartelet+4

CEREA, ENPC, Institut Polytechnique de Paris, ˆIle-de-France, France. · Department of Computing, Imperial College London, London, UK. · Dynamic Systems Lab, University College London, London, UK.+1

Conditional Generative Modeling Inverse Problems Diffusion Models

Open paper AskPDF arXiv

Full-field reconstruction of air pollution is essential for evaluating pollution exposure and supporting public health decision-making. However, the complex interactions among pollutants, hard-to-predict weather patterns, and limited monitoring station coverage make this a complex task. We apply deep learning techniques to provide fast and accurate reconstructions from sparse observations of four key pollutants: NO2, O3, PM2.5 and PM10. Models are trained on full-field simulation data and evaluated on real-world observations collected from 9 to 28 monitoring stations in the city of Paris. We introduce a diffusion-based generative framework for multi-pollutant reconstruction and benchmark its performance against deterministic deep learning models. Despite noisy observations and strong spatial variability, the models achieve high structural similarity on simulated validation data and produce realistic spatial patterns on real-world observations, as indicated by power-spectrum analysis. We introduce data augmentation methods that enable transfer to real-world observations without retraining, allowing the models to generalise beyond the training period. These findings highlight the potential of ML models for reliable real-world deployment in air pollution reconstruction tasks.

Abstract

Air Quality Prediction

Explore similar work

From Deterministic to Generative Deep Learning for Urban Air Quality Reconstruction from Sparse Observations

AI for Social Good: An Investigation of the Causal Relationship Between Environmental Regulations and Their Effects on Air Pollution in London, UK

Interpretable Machine Learning for Air Pollution and Respiratory Health Prediction: A Socioeconomic Subgroup Analysis