A review of statistical methods used for developing large-scale and long-term PM2.5 models from satellite data

Research of PM2.5 chronic health effects requires knowledge of large-scale and long-term exposure that is not supported by newly established monitoring networks due to their sparse spatial coverage and lack of historical measurements. Estimating PM2.5 using satellite-derived aerosol optical depth (AOD) can be used to fill the data gap left by the ground monitors and extend the PM2.5 data coverage to suburban and rural areas over long time periods. Two approaches have been applied in large-scale and long-term satellite remote sensing of PM2.5, i.e., the scaling and statistical approaches. Compared to the scaling method, the statistical approach has greater prediction accuracy and has been widely used. There is a gap in the current literature and review papers on how the statistical methods work and specific considerations to best utilize them, especially for large-scale and longterm estimates. In this critical review, we summarize the evolution of large-scale and long-term PM2.5 statistical models reported in the literature. We describe the framework and guidance of large-scale and long-term satellite based PM2.5 modeling in data preparation, model development, validation, and predictions. Sample computer codes are provided to expedite new model-building efforts. We also include useful considerations and recommendations in covariates selection, addressing the spatiotemporal heterogeneities of PM2.5-AOD relationships, and the usage of cross validation, to aid the determination of the final model.

Automated summary

Research on the chronic health effects of PM2.5 requires understanding of large-scale and long-term exposure, which is not supported by new monitoring networks. The use of satellite-derived AOD can fill the data gap left by ground monitors and extend PM2.5 data coverage. Statistical approaches have greater prediction accuracy compared to scaling methods and have been widely used. However, there is a gap in the current literature on how these statistical methods work and how to best utilize them for large-scale and long-term estimates.