Nowcasting street-level NO₂ concentrations using Gaussian processes

Area of Science:

• Environmental Science
• Data Science
• Urban Planning

Background:

• Street-level air quality monitoring is crucial for public health.
• Dense sensor networks provide valuable data for understanding pollution.
• Predictive models are needed to map air quality across urban environments.

Purpose of the Study:

• To develop a nowcasting model for street-level nitrogen dioxide (NO2) concentration.
• To predict NO2 distribution across an entire street network using Gaussian Processes (GPs).
• To identify key features influencing NO2 levels in urban areas.

Main Methods:

• Utilized data from 225 low-cost air quality sensors in Camden, UK.
• Employed Gaussian Processes (GPs) for predictive modeling.
• Incorporated street geometry, road network metrics, and coordinates as features.
• Conducted a global sensitivity analysis to determine feature importance.

Main Results:

• Street width, betweenness centrality, and road length were the most significant predictors, explaining 81% of variance.
• The GP model achieved a median absolute error of 7.31 μg/m³ for NO2 nowcasting.
• The model provides uncertainty estimates (standard deviation) for each prediction.
• 70.3% of streets had predictions with uncertainty below the mean.

Conclusions:

• Gaussian Processes offer a robust and scalable framework for NO2 prediction in complex urban networks.
• The model accurately forecasts real-time NO2 concentrations with quantifiable uncertainty.
• Feature importance analysis highlights critical urban characteristics affecting air quality.