Robustness of Land-Use Regression Models Developed from Mobile Air Pollutant Measurements

Abstract

Land-use regression (LUR) models are useful for resolving fine scale spatial variations in average air pollutant concentrations across urban areas. With the rise of mobile air pollution campaigns, characterized by short-term monitoring and large spatial extents, it is important to investigate the effects of sampling protocols on the resulting LUR. In this study a mobile lab was used to repeatedly visit a large number of locations (∼1800), defined by road segments, to derive average concentrations across the city of Montreal, Canada. We hypothesize that the robustness of the LUR from these data depends upon how many independent, random times each location is visited (N_vis) and the number of locations (N_loc) used in model development and that these parameters can be optimized. By performing multiple LURs on random sets of locations, we assessed the robustness of the LUR through consistency in adjusted R² (i.e., coefficient of variation, CV) and in regression coefficients among different models. As N_loc increased, R²_adj became less variable; for N_loc = 100 vs N_loc = 300 the CV in R²_adj for ultrafine particles decreased from 0.088 to 0.029 and from 0.115 to 0.076 for NO_2. The CV in the R²_adj also decreased as N_vis increased from 6 to 16; from 0.090 to 0.014 for UFP. As N_loc and N_vis increase, the variability in the coefficient sizes across the different model realizations were also seen to decrease.