Estimating daily concentrations of PM2.5 in Western Australia from satellite-derived aerosol optical depth, fire radiative power, and meteorological variables using a random forest model

Madeleine R Behn

Fine particulate matter (PM2.5) is an air pollutant that has gained increasing attention globally as industrialisation and population growth have led to higher concentrations within the atmosphere. In Western Australia (WA), due to the limited number of ground-based monitoring systems, measurement of PM2.5 is spatially restricted, which limits understanding of the human health related epidemiological impacts associated with PM2.5 in areas that are not monitored. Considering the advancements in remote sensing techniques, the utilization of satellite products in top-down approaches has emerged as a valuable means to estimate emissions of PM2.5. However, the varying spatial and temporal resolution of satellite-retrieved Aerosol Optical Depth (AOD) introduces significant barriers when using AOD to predict PM2.5. This study aims to develop a prediction model for PM2.5 concentrations in WA by utilizing satellite-retrieved AOD from five different satellite products, Fire Radiative Power (FRP) from MODIS, geolocation and season, as well as meteorological variables from best available reanalysis. The study employs a random forest algorithm, which is a machine learning technique capable of handling complex datasets and capturing nonlinear relationships, to construct the prediction model. This method has been shown to be useful elsewhere, however this is the first study which incorporates random forest with the range of variables listed for prediction of PM2.5 over WA. Four models were tested, two of which made use of single AOD products, a third which used a combined AOD data-set from all five AOD products, and finally a model which did not include AOD. The results indicate that the integration of various satellite-retrieved AOD products did not significantly improve the model’s ability to predict PM2.5. The four models tested yielded similar results when compared to measured PM2.5 (R2 ~ 0.23), including the model which did not include AOD. Meteorological variables including Boundary Layer Height (BLH), Wind Speed (WS) and Surface net Solar Radiation (SSR) were the most important predictors of PM2.5 for majority of models and subsequent analysis over WA should consider these variables. The integration of FRP also contributed to the model performance and is an important factor considering the emission profile of WA. Despite the low overall performance, this study provides a valuable contribution towards improving the understanding of prediction models for PM2.5 in WA using remotely sensed AOD.

Estimating daily concentrations of PM2.5 in Western Australia from satellite-derived aerosol optical depth, fire radiative power, and meteorological variables using a random forest model

Files and links (1)

Abstract

Details

Metrics