Modifiable Factors Driving Air Pollution: Low-Cost Monitoring Networks and Cookstove Emissions

Exposure to air pollution causes a wide variety of adverse health effects. In order to protect public health, the modifiable factors driving air pollution must be identified, characterized, and ultimately controlled. In this dissertation, we investigate some of the modifiable factors influencing air pollution exposure. The instruments used by researchers and regulatory agencies to monitor urban air quality frequently cost tens or even hundreds of thousands of dollars, limiting how many can be deployed. These sparse monitoring networks are not sufficient to capture the intra-urban pollutant variability. It is hypothesized that the increased spatial density enabled by networks of time-resolved lower-cost sensors (LCS) can be used to better capture the modifiable factors contributing to urban air pollution. However, the measurement uncertainty of LCSs is larger than regulatory-grade instruments.

In this dissertation, we characterize the measurement uncertainty for the O3, PM2.5, CO, and NO2 sensors in the Real-time Affordable Multi-Pollutant (RAMP) monitor developed at Carnegie Mellon University. The overall uncertainties ranged from 62%-133%; and the slopes, indicating the existence of concentration-dependent biases, ranged from 0.34 to 0.78. We show that careful calibration, temporal averaging, and reference site corrections can reduce the PM2.5 sensor uncertainty to 1 μg/m3, ~10% of typical long-term average PM2.5 concentrations in Pittsburgh. The usefulness of LCS depends on the balance of sampling and instrument errors. We performed Monte-Carlo simulations to evaluate situations in which LCS networks can be deployed effectively. When instrument error is unbiased, the random noise can be averaged out over long averaging times, and the performance of LCSs can be similar to regulatory monitors.Our analysis highlights that using LCS networks to estimate long-term averages mainly provides advantages for pollutants with large spatial variability, such as NO2. However, a large sensor network is unnecessary for a relatively spatially invariant pollutant like ozone. Concentration-dependent biases can cause substantial challenges when comparing two concentrations. For long-term average differences, the smallest enhancement that could be distinguished were ~7%, and ~9%, and ~15% for LCS with slopes of 0.75 and 0.5, and 0.25, respectively. However, at larger   enhancements, LCS with concentration-dependent biases had similar performance to unbiased monitors. For short-term enhancements, LCS with random noise levels of 100% and greater were not able to successfully distinguishing short-term concentration enhancements at a single location. When the LCS had concentration-dependent biases, more samples were needed at any given enhancement, but they were able to successfully differentiate statistically significant enhancements. LCS with prices in the range required to be deployed in a cost effective manor are already on the market. If they can be calibrated to minimize instrument error, there are promising applications for their future deployment. We use the high spatiotemporal resolution of a network of 64 RAMP PM2.5 LCS deployed across Pittsburgh, PA, to demonstrate some of these applications. We quantify the contribution of different modifiable and non-modifiable factors to urban PM2.5 concentrations.

Episodic and long-term enhancements to urban PM2.5 due to a nearby large metallurgical coke manufacturing facility were 1.6 ± 0.36 μg/m3 and 0.3 ± 0.2 μg/m3, respectively. Daytime landuse regression models identified restaurants as an important local contributor to urban PM2.5. We show that with proper management, a large network of lower-cost sensors can identify statistically significant trends and modifiable factors in urban exposure.