Abstract
This study introduces a flexible Bayesian regression model, specifically a lognormal-normal convolution (logNNC) within the bamlss package in R, for the statistical analysis of mobile measurement (MM) data. The model demonstrates significant potential for predicting pollutant concentrations in urban areas. Our findings highlight that for pedestrian exposure, monitoring and regulating Black Carbon (BC) mass concentration is crucial in areas dominated by traffic-related air pollution.
Introduction: The Challenge of Street-Level Monitoring
Most urban air quality data is derived from fixed monitoring stations. While these provide reliable long-term trends, their sparse distribution often fails to capture the "hot spots" of pollution that pedestrians encounter. Mobile measurements (MM)—collected via sensors in backpacks or on bicycles—provide the high-resolution data needed to fill this gap. However, MM data introduces significant statistical challenges, particularly high-frequency measurement noise.
Black Carbon (BC) sensors are especially prone to noise, which can result in negative concentration readings at low levels. Traditional statistical models, such as lognormal regression, cannot handle negative values, often forcing researchers to discard data or apply biasing corrections. This paper presents a novel approach to extracting meaningful spatial trends from noisy mobile data without compromising data integrity.
Methods: The logNNC Bayesian Framework
The authors propose a distributional regression model within a Bayesian framework. Unlike standard regression, which only predicts mean pollution levels, this approach models the entire probability distribution—including mean and variance—as a function of environmental factors.
The core innovation is the lognormal-normal convolution (logNNC):
- Mechanism: It assumes the "true" pollutant concentration follows a lognormal distribution, while the "observed" value is contaminated by Gaussian (normal) noise from the instrument.
- Advantage: This allows the inclusion of negative readings and low-concentration data, providing a superior fit for the "tails" of the distribution where exposure risks are most acute.
- Spatiotemporal Modeling: Gaussian processes are utilized to account for the inherent correlation in measurements taken close together in time or space.
Datasets and Spatial Variability
The study utilized two extensive mobile measurement datasets:
- Leipzig, Germany: A 5.5 km route sampled over a full year (2016), yielding ~40,000 data points across various microenvironments.
- Rome, Italy: A 9 km route sampled in February 2017, yielding ~55,000 data points focused on touristic and low-emission zones.
Analysis of these datasets revealed a critical insight: BC is highly localized, while PM$_{2.5}$ is regional. BC concentrations are extremely sensitive to immediate surroundings like street type and building height. In contrast, PM$_{2.5}$ levels are relatively homogeneous across the city, suggesting they are driven by regional background levels rather than local traffic.
Key Insights: The "Street Canyon" Effect
The physical configuration of urban environments significantly impacts exposure. Moving from an open street to a "traffic street canyon"—a street flanked by tall buildings—increased BC concentrations by 20% to 55%. These structures trap vehicle emissions and hinder dispersion. In both cities, transitioning from a park or pedestrian zone to a primary street more than doubled the mean BC concentration (+114% in Rome and +159% in Leipzig).
Model Performance and Validation
By explicitly modeling instrument noise rather than filtering it, the logNNC model provided a significantly better fit than standard models. It demonstrates that "messy" mobile data is highly valuable when the physics of measurement error is properly accounted for.
The model's quantile residuals remained much closer to the theoretical ideal compared to standard lognormal or linear models, which failed to capture the high-skewness and extreme values characteristic of urban pollution data.
Conclusion and Future Directions
The logNNC model proves to be a robust tool for analyzing mobile measurement data. While the study highlights the importance of localized BC monitoring, it also identifies areas for improvement:
- Traffic Data: Future models would benefit from real-time traffic counts rather than annual averages.
- Wind Complexity: Incorporating localized anemometers could better capture the complex turbulent flows within street canyons.
- Seasonality: Expanding mobile campaigns to cover more diverse meteorological conditions across different urban morphologies will enhance generalizability.
Reference: Alas, H.D., Stöcker, A., Umlauf, N. et al. Pedestrian exposure to black carbon and PM$_{2.5}$ emissions in urban hot spots: new findings using mobile measurement techniques and flexible Bayesian regression models. J Expo Sci Environ Epidemiol 32, 604–614 (2022).
Summary updated