Air pollution is a major global risk causing a large number of illnesses and deaths every year. Many literatures have shown the robust causal relations between health and outdoor exposure to various air pollutants. The complex urban environment causes the uneven distribution of air pollution concentrations, which can change sharply within a short distance. Therefore, understanding the within city air pollution gradients is crucial for various studies including exposure assessment, urban planning, air pollution monitoring, and environmental equity.

Mobile-based air pollution monitoring has been proposed to tackle this challenge since it can typically achieve higher spatial resolution measurement of air pollution concentrations than other methods. However, the inherently different nature of measurement from mobile monitors makes it difficult to apply methods designed for stationary sensors. This dissertation focuses on developing methods that are suitable for mobile sensor data to take advantage of the high spatial resolution nature for exposure assessment, air pollution monitoring, and socioeconomic impact studies.

For the exposure assessment study, we calculate exposure concentrations of traffic-related air pollutants with three different travel modes in the complex urban environment. We simulate bicycle, transit, and vehicle trips within Oakland CA. based on the local road network. With highly resolved mobile sensor data, we calculate the average concentration and the cumulative exposure of nitric oxide (NO), nitrogen dioxide (NO2), and black carbon (BC) for each bicycle, transit, and vehicle trip that we simulate. The results show that cumulative exposure may be a better metric than the more typical average ambient concentration when evaluating the air pollution exposure with different travel modes. The average concentrations of each trip are not significantly different among bicycle, transit, and vehicle. However, the cumulative exposure varies dramatically because it takes trip duration, route variations for different travel modes, and inhalation rates into consideration. Vehicle passengers tend to experience the lowest cumulative exposure, as well as have the lowest average per meter and per minute exposure. Because of the increased inhalation rates for bicyclists and longer trip duration for public transit users, they tend to experience higher cumulative exposure. Our study also compares the importance of trip duration and trip distance influencing exposure, which turns out that total trip duration is more influential than the total trip distance in terms of cumulative exposure. Our work finds better metrics to assess travel air pollution exposure by using big data and modern simulation techniques.

In another study, we combine the land use model with different regression methods to estimate black carbon (BC) concentrations in Oakland, CA. The regression methods used in this study include linear regression, Random Forest (RF), Support Vector Regression (SVR), and Neural Network (NN). The least absolute shrinkage and selection operator (LASSO), principle component analysis (PCA), conditional independence feature ordering (FOCI), and genetic algorithm (GA) are used for feature selection and dimension reduction of the SVR method to reduce overfitting and improve prediction accuracy. The tuning of RF and SVR are automatically conducted with the Bayesian Optimization method, while we manually tune the NN method. Among all these regression methods, RF performs the best with the highest prediction accuracy and robustness. Even though SVR shows much better prediction accuracy than linear and NN methods, the complex feature selection and dimension reduction processes make it less efficient than RF. NN has the highest prediction accuracy on the train set, but the lowest accuracy on the independent validation set, which suggests an overfitting issue. With the one-factor-at-a-time (OAT) sensitivity analysis and localized hotspots identifications, our study shows that the LURs with a common approach are not efficient at identifying localized hotspots. However, LUR coupling RF can achieve higher air pollution prediction accuracy and robustness using mobile sensor measurements. This approach can be used in air pollution exposure assessment to more accurately identify vulnerable population groups or communities and better highlight environmental justice issues.

For the socioeconomic impacts of air pollution, we study the effects of air pollution on housing price in Oakland CA. We evaluate the ambient air quality on a parcel by parcel basis with the high-resolution mobile-based air pollution measurements of NO, NO2 and BC. In this study, a hedonic price model is constructed with a spatial lag model and instrumental variable method to cover both spatial autocorrelation and endogeneity effects between air pollution concentrations and housing price. The results indicate the air pollution influences housing price positively, which is surprising. The results could be explained in two ways: people are not sensitive to air pollution when the overall ambient air quality is good; the low variability of air pollution concentrations leads to false positive results. The explanations could be verified with the high-resolution mobile-based air pollution measurements covering more diversified regions.