Insights Gained from Characterizing Hyper-Local Air Pollution Trends Using a Multi-Model Approach

Mongolia’s capital city, Ulaanbaatar (elevation: 1,350 meters, area: 4,700 km²), experiences significant air pollution challenges, yet extensive periods of missing data across Ulaanbaatar’s air quality monitoring network limit research aiming to study trends between urban air quality and health. We sought to impute missing PM_2.5 (particulate matter with aerodynamic diameters of 2.5 microns or less) across Ulaanbaatar’s air quality monitoring network during 2013-2022. This allowed for PM_2.5 predictions at hyperlocal scales as opposed to generalized city-wide predictions, especially given that a majority of Ulaanbaatar’s monitoring stations and population are located within a much smaller region within the city’s center (area: ~400 km²). We achieved our objective by using data fusion techniques to combine datasets containing environmental, industrial, and societal indicators. Ultimately, over 200 predictors were included in our modelling dataset and these predictors varied greatly in temporal and spatial scales. Next, we developed a multi-model prediction approach, consisting of station-unique, hybrid machine learning models, to impute missing PM_2.5 data for each station. Our multi-model prediction approach yielded results comparable to other generalized, city-wide approaches developed to predict PM_2.5 in Ulaanbaatar (Average R² = 0.73, Average RMSE = 44.25 µg/m³). This shows that a multi-model approach can be effective for modelling air pollution in heterogenous urban settings. Over 7,000 daily PM_2.5 means were imputed using are models. The increase in data ranged from 24% to 420% and was on average 145% per monitoring station, underscoring the ability of our method to add significantly to limited air pollution exposure datasets. This substantial increase in data availability will allow for more comprehensive global health research investigating the relationship between PM_2.5 exposure and human health. Our multi-model approach also identified the most useful variables for predicting PM_2.5 exposure in Ulaanbaatar: industrial production and meteorological data. Assimilated remote sensing data and air quality legislation indicators were also found to enhance predictions. These results further highlighted the need for specialized modeling of PM_2.5 across Ulaanbaatar, given the variation of predictors most relevant in each station’s optimal model. Overall, our methodology and findings may be useful for researchers aiming to develop air pollution exposure prediction models for urban settings of varying scales around the globe.