Spatiotemporal prediction of PM2.5 concentrations from satellite data across Metro Manila using eXtreme Gradient Boosting
Abstract
Exposure to fine particulate matter (PM2.5) has been recognized as a global public health issue. Air quality monitoring across the Philippines is mainly based on data gathered by on-ground stations, which are too sparse to accurately assess the exposure effects of air pollution for the entire archipelago. Thus, the demand for exposure assessment models that estimate physical parameters of ambient air at extensive spatiotemporal resolutions has rapidly grown. We investigate the potential of satellite-derived products to improve PM2.5 estimates across Metro Manila. An eXtreme Gradient Boosting (XGBoost) model was developed incorporating ground monitoring data with Moderate Resolution Imaging Spectroradiometer (MODIS) aerosol optical depth (AOD) data, meteorological parameters, and auxiliary predictors from 2015 to 2018. Variable importance measures suggest day of year, MERRA-2 PM2.5, and 10-meter wind velocities are among the most important features for predicting PM2.5 concentrations. Conventional five-fold cross-validation (CV) results for the model achieved a mean absolute error of 7.61 μg/m3, RMSE of 11.67 μg/m3, sMAPE of 30.45%, and R2 of 0.89, indicating similar or better prediction performance versus previous studies done in other countries. The annual mean predicted PM2.5 concentration from 2015 to 2018 in the study domain was 56.61 μg/m3, and the cities of Makati, Mandaluyong, and Malabon were the most polluted areas.
