Assessing the transferability of landuse regression models for ultrafine particles across two Canadian cities

Jad Zalzal; Ibrahim Alameddine; Celine El Khoury; Laura Minet; Maryam Shekarrizfard; Scott Weichenthal; Marianne Hatzopoulou

doi:10.1016/j.scitotenv.2019.01.123

Assessing the transferability of landuse regression models for ultrafine particles across two Canadian cities

Sci Total Environ. 2019 Apr 20:662:722-734. doi: 10.1016/j.scitotenv.2019.01.123. Epub 2019 Jan 14.

Authors

Jad Zalzal¹, Ibrahim Alameddine², Celine El Khoury¹, Laura Minet³, Maryam Shekarrizfard³, Scott Weichenthal⁴, Marianne Hatzopoulou³

Affiliations

¹ Department of Civil and Environmental Engineering, American University of Beirut, Beirut, Lebanon.
² Department of Civil and Environmental Engineering, American University of Beirut, Beirut, Lebanon. Electronic address: ia04@aub.edu.lb.
³ Department of Civil & Mineral Engineering, University of Toronto, Toronto, Ontario, Canada.
⁴ Department of Epidemiology, Biostatistics & Occupational Health, McGill University, Montreal, QC, Canada.

PMID: 30703730
DOI: 10.1016/j.scitotenv.2019.01.123

Abstract

Land use regression (LUR) models have been increasingly used to predict intra-city variations in the concentrations of different air pollutants. However, limited research assessing the transferability of these models between cities has been published to date. In this study, LUR models were generated for Ultra-Fine Particles (UFP) (<0.1 um) using data collected from mobile monitoring campaigns in two Canadian cities, Montreal and Toronto. City-specific models were first generated for each city before the models were transferred to the second city with and without recalibration. The calibrated transferred models showed only a slight decrease in performance, with the coefficient of determination (R²), dropping from 0.49 to 0.36 for Toronto and from 0.41 to 0.38 for Montreal. Transferring models between cities with no calibration resulted in low R²; 0.11 in Toronto and 0.18 in Montreal. Moreover, two additional models were generated by combining data from the two cities. The first combined model (CM1) assumed a spatially invariant effect of the predictors, while the second (CM2) relaxed the assumption of spatial invariance for some of the model coefficients. The performance of both combined models (R² ranged between 0.41 for CM1 and 0.43 for CM2; root mean squared error (RMSE) ranged between 0.34 for CM1 and 0.33 for CM2) was found to be on par with the Toronto city-specific model and outperformed the Montreal model. The results of this study highlight that the UFP LUR models appear to support transferability of model structures between cities with similar geographical characteristics, with a minor drop in model fit and predictive skill.

Keywords: Air pollution; Land use regression; Transferability; Ultrafine particles.