Estimating daily PM2.5 and PM10 concentrations in Alaska using a random forest model

Wildfire is a dominant source of particulate matter (PM) pollution during the Alaska (AK) fire season. Although high concentrations of PM have been linked to many negative health outcomes, the geographic coverage of PM monitoring across AK has been very limited. There are well-known challenges to using ground-based regulatory PM measurements to monitor statewide exposure in AK. Regulatory stations are limited in number, located only in more densely populated areas and are often missing data. For example, for the 2019 fire season across the over 900,000 km² state, Environmental Protection Agency (EPA) Air Quality System data included 15 sites (24 monitors), two-thirds of sites were in 3 highly populated cities and nearly 48 percent of observations were missing. To address the low spatiotemporal resolution of ground PM estimates in AK, citizen scientists, communities, and researchers are increasingly using low-cost air quality monitors (LCAQMs) as they provide near-real-time, high resolution measurements. PurpleAir (PA) sensors are some of the most popular LCAQM sensors. In AK, the (PA) network supplements the EPA monitoring data with nearly 100 sensors, with data primarily available since 2018. This study seeks to derive daily estimates of fire generated PM_2.5 and PM₁₀ concentrations over AK for 2019 at a 1 km resolution using a random forest algorithm. The concentration estimates were generated using ground measurements from EPA and PA stations, meteorological fields, land cover/use variables and the satellite-derived Moderate Resolution Imaging Spectroradiometer (MODIS) Multi-Angle Implementation of Atmospheric Correction Collection 6 aerosol optical depth (AOD) product. A major new contribution to AOD/PM modeling effort presented here includes the development of a conceptual framework for proper fire activity characterization in random forest models. We created a single composite variable using a satellite-derived fire activity product (MCD14ML) and wind data that accounted for the distance, position and intensity of the fire. A 10-fold cross-validation approach was used to build the model and assess its predictive capability on new data. The model fit was assessed by the out-of-bag error, the percent of variance explained by the model, the root mean square prediction error and median absolute error.