Improving satellite aerosol optical Depth-PM2.5 correlations using land use regression with microscale geographic predictors in a high-density urban context

Estimating the spatiotemporal variability of ground-level PM2.5 is essential to urban air quality management and human exposure assessments. However, it is difficult in a high-density and highly heterogeneous urban context as ground-level monitoring stations are most likely sparsely distributed. Satellite-derived Aerosol Optical Depth (AOD) observation has made it possible to overcome such difficulty due to its advantage of spatial coverage. In this study, we improve the AOD-PM2.5 correlations by combining land use regression (LUR) modelling and incorporating microscale geographic predictors and atmospheric sounding indices in Hong Kong. The spatiotemporal variations of ground-level PM2.5 over Hong Kong were estimated using MODerate resolution Imaging Spectroradiometer (MODIS) AOD remote sensing images for the period of 2003-2015. An extensive LUR variable database containing 294 variables was adopted to develop AOD-LUR models by seasons. Compared to the baseline models (fixed effect models include only basic weather parameters), the prediction performance of all annual and seasonal AOD-LUR fixed effect models were significantly enhanced with approximately 20-30% increases in the model adjusted R-2. On top of that, a mixed effect model covers time-dependent random effects and a group of geographically and temporally weighted regression (GTWR) models were also developed to further improve the model performance. As the results, compared to the uncalibrated AOD-PM2.5 spatiotemporal correlation (adjusted R-2 = 0.07, annual fixed effect AOD-only model), the calibrated AOD-PM2.5 correlation (the GTWR piecewise model) has a significantly improved model fitting adjusted R-2 of 0.72 (LOOCV adjusted R-2 of 0.65) and thus becomes a ready reference for spatiotemporal PM2.5 estimation.