A novel ensemble-based statistical approach to estimate daily wildfire-specific PM2.5 in California (2006-2020)

Abstract

Though fine particulate matter (PM_2.5) has decreased in the United States (U.S.) in the past two decades, the increasing frequency, duration, and severity of wildfires significantly (though episodically) impairs air quality in wildfire-prone regions and beyond. Increasing PM_2.5 concentrations derived from wildfire smoke and associated impacts on public health require dedicated epidemiological studies. Main sources of PM_2.5 data are provided by government-operated monitors sparsely located across U.S., leaving several regions and potentially vulnerable populations unmonitored. Current approaches to estimate PM_2.5 concentrations in unmonitored areas often rely on big data, such as satellite-derived aerosol properties and meteorological variables, apply computationally-intensive deterministic modeling, and do not distinguish wildfire-specific PM_2.5 from other sources of emissions such as traffic and industrial sources. Furthermore, modelling wildfire-specific PM_2.5 presents a challenge since measurements of the smoke contribution to PM_2.5 pollution are not available. Here, we aim to use statistical methods to isolate wildfire-specific PM_2.5 from other sources of emissions. Our study presents an ensemble model that optimally combines multiple machine learning algorithms (including gradient boosting machine, random forest and deep learning), and a large set of explanatory variables to, first, estimate daily PM_2.5 concentrations at the ZIP code level, a relevant spatiotemporal resolution for epidemiological studies. Subsequently, we propose a novel implementation of an imputation approach to estimate the wildfire-specific PM_2.5 concentrations that could be applied geographical regions in the US or worldwide. Our ensemble model achieved comparable results to previous machine learning studies for PM_2.5 prediction while avoiding processing larger, computationally intensive datasets. Our study is the first to apply a suite of statistical models using readily available datasets to provide daily wildfire-specific PM_2.5 at a fine spatial scale for a 15-year period, thus providing a relevant spatiotemporal resolution and timely contribution for epidemiological studies.

Keywords: Air pollution; Human health; Machine learning; PM(2.5); Wildfire.