Continuous estimations of daily PM2.5 chemical components from temporally sparse monitoring data using a machine learning approach

Abstract

Listen

Translate article

Accurate estimates of the concentrations of fine particulate matter (PM2.5) components at high spatial and temporal resolution is essential for assessing their impact on human health. In this study, we developed prediction models of daily concentrations of major PM2.5 components (sulfate, nitrate, ammonium, elemental carbon, and organic carbon) in the Kansai region, Japan, from 2010 to 2017 using the random forest algorithm. The objective is to establish a modeling approach for obtaining accurate daily estimates of PM2.5 component concentrations using temporally sparse monitoring data covering only 15% of the study period. We used gaseous and particulate pollutant concentrations, chemical transport model outputs, simulated meteorological parameters, and conventional land use and traffic-related variables to produce daily estimations at 1 km × 1 km resolution. We evaluated our models via spatial and temporal cross-validation and obtained R2 values of 0.59–0.86 for individual components. The model reproduced the day-to-day variations well, with Pearson’s correlation coefficients of 0.75–0.88 between estimates and independent data collected at continuous monitors. We estimated the daily concentrations of PM2.5 components from 2010 to 2017 at a 1 km × 1 km resolution. The annual trends of the components were obtained by temporally and spatially aggregating the daily estimations. Our modeling approach enabled accurate estimates of the daily PM2.5 component levels using temporally sparse monitoring data. The estimated concentrations will be further utilized in a birth cohort study to assess the potential health impacts of PM2.5 components.