Linear and Nonlinear Land Use Regression Approach for Modelling PM2.5 Concentration in Ulaanbaatar, Mongolia during Peak Hours

Abstract

In recent decades, air pollution in Ulaanbaatar has become a challenge regarding the health of the citizens of Ulaanbaatar, due to coal combustion in the ger area. Households burn fuel for cooking and to warm their houses in the morning and evening. This creates a difference between daytime and nighttime air pollution levels. The accurate mapping of air pollution and assessment of exposure to air pollution have thus become important study objects for researchers. The city center is where most air quality monitoring stations are located, but they are unable to monitor every residential region, particularly the ger area, which is where most particulate matter pollution originates. Due to this circumstance, it is difficult to construct an LUR model for the entire capital city’s residential region. This study aims to map peak PM2.5 dispersion during the day using the Linear and Nonlinear Land Use Regression (LUR) model (Multi-Linear Regression Model (MLRM) and Generalized Additive Model (GAM)) for Ulaanbaatar, with monitoring station measurements and mobile device (DUST TRUK II) measurements. LUR models are frequently used to map small-scale spatial variations in element levels for various types of air pollution, based on measurements and geographical predictors. PM2.5 measurement data were collected and analyzed in the R statistical software and ArcGIS. The results showed the dispersion map MLRM R2 = 0.84, adjusted R2 = 0.83, RMSE = 53.25 µg/m3 and GAM R2 = 0.89, and adjusted R2 = 0.87, RMSE = 44 µg/m3. In order to validate the models, the LOOCV technique was run on both the MLRM and GAM. Their performance was also high, with LOOCV R2 = 0.83, RMSE = 55.6 µg/m3, MAE = 38.7 µg/m3, and GAM LOOCV R2 = 0.77, RMSE = 65.5 µg/m3, MAE = 47.7 µg/m3. From these results, the LUR model’s performance is high, especially the GAM model, which works better than MRLM.