Examining Land Use Regression modeled NO2 as a proxy for air pollutants measured by CRUISER mobile laboratory: A study in Montreal, Quebec

ILAN LEVY (1), Jeff R Brook (1), Julie Narayan (1), Cris Mihele (1), Gang Lu (1), Dan L Crouse (2), Mark S Goldberg (2), Nancy A Ross (2)

(1) Environment Canada (2) McGill University

     Abstract Number: 330
     Last modified: November 9, 2009

Abstract
Over the last decade, Land Use Regression (LUR) models have provided valuable information on intra-urban variability of air pollutants. LUR models are based on the physical characteristics of the city (e.g., road network, population, land use) that are statistically correlated to pollution levels measured often over a period of several weeks using a dense network of passive samplers deployed simultaneously.

LUR models are often used in health studies to estimate the long term exposure to traffic-related pollution at the home addresses of the study population. The question arises as to how well LUR models estimate NO$_2 at independent locations and to what extent is the LUR a surrogate for other (traffic-related) pollutants (e.g., PM$_(2.5), ultrafine particles, Black Carbon (BC))? Validation of LUR models is typically carried out by comparing model predictions to a “hold-back” set of model development measurements or other independent point measurements such as ambient monitoring sites. However, this method of validation uses a limited number of monitors with a given spatial coverage and therefore reduces statistical power and cannot account for the full range of pollution levels.

We compare predictions of levels of NO$_2 from a LUR model [Crouse et al., 2009] developed for Montreal to measurements taken by Environment Canada’s CRUISER (Canadian Regional and Urban Investigation System for Environmental Research) mobile laboratory.

The LUR model was developed separately for three seasons (winter, spring and summer) and juxtaposed to produce an estimate of the annual mean. Measurements from CRUISER were obtained in Montreal in the summer and winter. The sampling strategy of CRUISER was to travel along pre-defined routes, passing through highways, main roads and small streets, as well as residential, commercial and industrial areas. This allowed for multiple samplings of the same road on different days.

As expected, the range of CRUISER’s NO$_2 measurements was higher than that derived from the LUR model (0-470ppb vs. 2.6-31.5ppb, respectively). We spatially aggregated CRUISER’s data by averaging along road segments, first averaging by day and then averaging all days, so as to give an equal weight for every day and avoid bias towards days with multiple measurements. This was done for NO$_2 as well as other measured pollutants.

Preliminary results indicate that the correlation (R$^2) between the independently measured and the modeled NO$_2 are on the order of 0.45. Correlation for other pollutants ranged between 0.55 (BC) and 0.0 (SO$_2). However, these results are very sensitive to the number of measurements used for each road segment and the minimal number of days allowed in the average. Further analysis will include sensitivity analysis to these two parameters, as well as examining the correlation of the annual means.


Reference:
Crouse, D. L., M. S. Goldberg, and N. A. Ross (2009), A prediction-based approach to modelling temporal and spatial variability of traffic-related air pollution in Montreal, Canada, Atmospheric Environment, 43(32), 5075-5084, doi:10.1016/j.atmosenv.2009.06.040.