Abstract

Epidemiological studies on the detrimental health impacts of exposure to fine particulate matter (PM2.5) from different sources of emission can inform regulatory policy and identify vulnerable communities. Though PM2.5 has decreased in the U.S. in the two past decades, the increasing frequency and severity of wildfires contribute to episodically impair air quality in wildfire-prone regions and beyond. Monitoring air quality extensively is challenging. Since government-operated monitors are sparsely located across California and the U.S., several regions and populations remain unmonitored. Current approaches to estimate PM2.5 concentrations in unmonitored areas often rely on gathering large amounts of data, such as satellite-derived aerosol properties and meteorological variables. and direct use of low-cost air sensor measurements that may be associated with substantial uncertainty Furthermore, modelling wildfire-specific PM2.5 is often based on chemical transport model predictions, which results in highly computationally intensive efforts. Our study used an ensemble model that integrated multiple machine learning algorithms and a large set of predictor variables to estimate daily PM2.5 at the ZIP code level, a relevant spatio-temporal resolution for epidemiological and public health studies. Our models achieved comparable results to previous machine learning studies for PM2.5 prediction, but avoided processing larger, computationally intensive datasets. In addition, we use machine learning to estimate the wildfire-specific PM2.5 concentrations through a novel multiple imputation approach.