Low-cost Techniques to Measure and Predict Air Pollution Exposure

 Exposure to air pollution is the fifth leading cause of death worldwide. This risk is dominated by exposure to fine particulate matter (PM_2.5) and is deemed even more critical in low- and middle-income countries due to a lack of data available to build effective policies. The scarcity of actionable air quality data in these regions is due to thehigh cost of research-grade air pollutant monitors. The core objectives of my thesis are to (1) develop a low-cost method to measure atmospheric black carbon, a major contributor to PM_2.5 in developing nations, at existing PM_2.5 monitoring sites, (2) use the method to measure the BC component of PM2.5 to identify leading combustion sources contributing to pollutant concentrations in African cities, and (3) develop a framework for developing a robust deep learning-based generalized model to predict PM_2.5 pollutant concentrations that is transferable to other locations. 

US Department of State collects air pollutant data at selected US embassies and consulates around the world to inform embassy personnel and citizens about the local air quality. These measurements are made by Beta Attenuation Monitors (BAMs) to measure hourly ambient PM_2.5 concentrations. These BAMs at the embassies collect PM_2.5 onto glass-fiber filter tapes at a circular spot with a fixed flow rate of 1m³/h^-1 for approximately an hour and pass beta ray through the particle deposit. Hourly PM_2.5 concentration is calculated with Beer-Lambert law by utilizing the attenuation of beta rays through the spot. The BAM moves the spot to sample on a clean area of the filter tape and continues this process until the tape is exhausted. A new tape is installed and the used tapes are usually discarded after a brief period of storage. Chapter 3 of this thesis investigates a cost-effective method to leverage used tapes from these existing BAMs to measure ambient black carbon concentration at an hourly resolution. We developed an image reflectance-based method to determine hourly black carbon (BC) concentrations from red light reflectance using cell phone cameras. The method relies on the working principle that the BC loading on a filter sample is correlated to intensity of reflected red light from the sample. The hourly effective detection limit of the method is estimated to be around 0.15 μg/m^-3 of BC, which makes it suitable to use in most micro-environments, especially the ones with high combustion emissions. This method only requires a reference card sheet and a cellphone camera. Both of these items are easily accessible making this method practically a "zero" cost measurement technique. 

We used this method to measure hourly BC concentrations at multiple locations in Africa and compared it to our measurements in Pittsburgh. Unregulated emission sources are one of the major factors for the high pollution levels in developing nations. For example, primary sources of air pollution in Africa are unregulated transport emissions, burning of solid fuels for cooking and heating in winter seasons, open burning of crop residues and burning fossil fuel for electricity production. Air quality measurements can provide evidence on these high emitting sources for effective policymaking, further alleviating the air pollution scenario in these regions. Chapter 4 presents measurements on BC component of hourly PM_2.5 data with our image reflectance-based method applied on used BAM tapes from multiple cities in Africa, which will be used in investigating and identifying primary sources of pollutants at these locations. The cities in this study include Abidjan (Côte d’Ivoire), Accra (Ghana) and Addis Ababa (Ethiopia). The measurements show a significantly high BC levels across the cities and seasons in Africa. Availability of hourly BC information, in addition to PM_2.5 measurements, reveal high contribution of combustion emissions in the local air pollution. In Addis Ababa, BC composed 20% of PM_2.5 both in summer and winter seasons at two locations within the city. BC measurements in Abidjan showed a higher BC level during wet season due to increased burning of solid fuels to heat the houses as this period experiences lower temperatures. Hourly data also allowed us to derive diurnal patterns of BC for all cities. Accra showed a unique peak at 2 am in the night. This peak was later identified to be from diesel generators used by US embassy at Accra to meet their power demands during power outage incidents and/or illicit waste burning at night. Thus, our measurements reveal tons of information on pollutant levels as well as emission sources. 

A step further in informed policymaking for reducing exposure during hazardous levels of air pollution is to develop a forecast model that can predict air pollutant levels with the use of temporally dynamic variables. A major benefit of these forecast models is their ability to predict high pollution episodes so that both precautionary and preventive measure can be taken in advance to alleviate the exposure levels. Precautionary measures include wearing masks or staying home in case of severe pollution episodes, whereas preventive measures encompass coordinating with local industries and commercial plants to dampen their activities to reduce their emissions. 

Studies have developed variants of land-use regression-based models to forecast air pollution, but these models show poor prediction capabilities and their performance saturate despite inclusion of myriad features for training models. Most of the high pollution episodes overlap with period of temperature inversions, which is period of a very low planetary boundary layer height with minimal atmospheric circulation trapping air pollutants close to the ground. Therefore, we built models that use boundary layer height and related meteorological parameters among training features. Chapter 5 investigates deep learning models to build a novel neural network-based forecast model that utilizes training covariates, including air pollutant and weather-based parameters forecast from GEOS-CF model to generate a more robust and generalized empirical prediction model to forecast PM_2.5 a day in advance. We use 5 years of low-cost sensor-based PM_2.5 data as the ground truth in training the models to predict PM_2.5 for locations in Monongahela Valley, Pennsylvania.