ANNOUNCEMENTS
Ordinary Least Squares (OLS) regression was employed to develop traditional Land Use Regression (LUR) models using 236 predictor variables spanning across different categories: land use, traffic, geographic, meteorological, vegetation and demographic, to predict annual average concentrations of particulate matter (size ≤ 10 um; PM10), carbon monoxide (CO) and nitrogen oxides (NOx) in one of the most polluted cities in the world, Delhi, India. The statistically significant (α = 5%) predictors, following bivariate linear regression, were subjected to exploratory correlation analysis and backward elimination to produce parsimonious annual LUR models with the most significant predictors. The final models revealed contributions of traffic variables across all the three pollutants. Commercial sources were also significantly correlated with NOx and CO concentrations in the final models. Meteorological variables appeared only in the PM10 (wind speed) and CO (air temperature) models. Industrial variables showed no significant correlation with any of the pollutants. The adjusted R2 value and the Index of Agreement (IOA), considered as indices to measure the model performance, revealed contrasting results. PM10 and NOx models had low adjusted R2 value of 0.315 and 0.318, respectively, while the IOA revealed high agreement with a value of 0.7 for both. The CO model performed much better than the other two counterparts with an adjusted R2 value of 0.497 and IOA of 0.839. 5-fold cross validation revealed good validation IOA values for all the three pollutants. Besides, residual analysis revealed no violation of OLS assumptions. Finally, spatial interpolation of model predicted pollutant concentrations were done to obtain LUR maps for each of the pollutants.
KEYWORDS: air pollution, land use regression, air quality modelling, Geographic Information System, Delhi.
