Assessment and statistical modeling of the relationship between remotely sensed aerosol optical depth and PM2.5 in the eastern United States
Christopher J Paciorek,Yang Liu,HEI Health Review Committee
Abstract:Research in scientific, public health, and policy disciplines relating to the environment increasingly makes use of high-dimensional remote sensing and the output of numerical models in conjunction with traditional observations. Given the public health and resultant public policy implications of the potential health effects of particulate matter (PM*) air pollution, specifically fine PM with an aerodynamic diameter < or = 2.5 pm (PM2.5), there has been substantial recent interest in the use of remote-sensing information, in particular aerosol optical depth (AOD) retrieved from satellites, to help characterize variability in ground-level PM2.5 concentrations in space and time. While the United States and some other developed countries have extensive PM monitoring networks, gaps in data across space and time necessarily occur; the hope is that remote sensing can help fill these gaps. In this report, we are particularly interested in using remote-sensing data to inform estimates of spatial patterns in ambient PM2.5 concentrations at monthly and longer time scales for use in epidemiologic analyses. However, we also analyzed daily data to better disentangle spatial and temporal relationships. For AOD to be helpful, it needs to add information beyond that available from the monitoring network. For analyses of chronic health effects, it needs to add information about the concentrations of long-term average PM2.5; therefore, filling the spatial gaps is key. Much recent evidence has shown that AOD is correlated with PM2.5 in the eastern United States, but the use of AOD in exposure analysis for epidemiologic work has been rare, in part because discrepancies necessarily exist between satellite-retrieved estimates of AOD, which is an atmospheric-column average, and ground-level PM2.5. In this report, we summarize the results of a number of empirical analyses and of the development of statistical models for the use of proxy information, in particular satellite AOD, in predicting PM2.5 concentrations in the eastern United States. We analyzed the spatiotemporal structure of the relationship between PM2.5 and AOD, first using simple correlations both before and after calibration based on meteorology, as well as large-scale spatial and temporal calibration to account for discrepancies between AOD and PM2.5. We then used both raw and calibrated AOD retrievals in statistical models to predict PM2.5 concentrations, accounting for AOD in two ways: primarily as a separate data source contributing a second likelihood to a Bayesian statistical model, as well as a data source on which we could directly regress. Previous consideration of satellite AOD has largely focused on the National Aeronautics and Space Administration (NASA) moderate resolution imaging spectroradiometer (MODIS) and multiangle imaging spectroradiometer (MISR) instruments. One contribution of our work is more extensive consideration of AOD derived from the Geostationary Operational Environmental Satellite East Aerosol/Smoke Product (GOES GASP) AOD and its relationship with PM2.5. In addition to empirically assessing the spatiotemporal relationship between GASP AOD and PM2.5, we considered new statistical techniques to screen anomalous GOES reflectance measurements and account for background surface reflectance. In our statistical work, we developed a new model structure that allowed for more flexible modeling of the proxy discrepancy than previous statistical efforts have had, with a computationally efficient implementation. We also suggested a diagnostic for assessing the scales of the spatial relationship between the proxy and the spatial process of interest (e.g., PM2.5). In brief, we had little success in improving predictions in our eastern-United States domain for use in epidemiologic applications. We found positive correlations of AOD with PM2.5 over time, but less correlation for long-term averages over space, unless we used calibration that adjusted for large-scale discrepancy between AOD and PM2.5 (see sections 3, 4, and 5). Statistical models that combined AOD, PM2.5 observations, and land-use and meteorologic variables were highly predictive of PM2.5 observations held out of the modeling, but AOD added little information beyond that provided by the other sources (see sections 5 and 6). When we used PM2.5 data estimates from the Community Multiscale Air Quality model (CMAQ) as the proxy instead of using AOD, we similarly found little improvement in predicting held-out observations of PM2.5, but when we regressed on CMAQ PM2.5 estimates, the predictions improved moderately in some cases. These results appeared to be caused in part by the fact that large-scale spatial patterns in PM2.5 could be predicted well by smoothing the monitor values, while small-scale spatial patterns in AOD appeared to weakly reflect the variation in PM2.5 inferred from the observations. Using a statistical model that allowed for potential proxy discrepancy at both large and small spatial scales was an important component of our modeling. In particular, when our models did not include a component to account for small-scale discrepancy, predictive performance decreased substantially. Even long-term averages of MISR AOD, considered the best, albeit most sparse, of the AOD products, were only weakly correlated with measured PM2.5 (see section 4). This might have been partly related to the fact that our analysis did not account for spatial variation in the vertical profile of the aerosol. Furthermore, we found evidence that some of the correlation between raw AOD and PM2.5 might have been a function of surface brightness related to land use, rather than having been driven by the detection of aerosol in the AOD retrieval algorithms (see sections 4 and 7). Difficulties in estimating the background surface reflectance in the retrieval algorithms likely explain this finding. With regard to GOES, we found moderate correlations of GASP AOD and PM2.5. The higher correlations of monthly and yearly averages after calibration reflected primarily the improved large-scale correlation, a necessary result of the calibration procedure (see section 3). While the results of this study's GOES reflectance screening and surface reflection correction appeared sensible, correlations of our proposed reflectance-based proxy with PM2.5 were no better than GASP AOD correlations with PM2.5 (see section 7). We had difficulty improving spatial prediction of monthly and yearly average PM2.5 using AOD in the eastern United States, which we attribute to the spatial discrepancy between AOD and measured PM2.5, particularly at smaller scales. This points to the importance of paying attention to the discrepancy structure of proxy information, both from remote-sensing and deterministic models. In particular, important statistical challenges arise in accounting for the discrepancy, given the difficulty in the face of sparse observations of distinguishing the discrepancy from the component of the proxy that is informative about the process of interest. Associations between adverse health outcomes and large-scale variation in PM2.5 (e.g., across regions) may be confounded by unmeasured spatial variation in factors such as diet. Therefore, one important goal was to use AOD to improve predictions of PM2.5 for use in epidemiologic analyses at small-to-moderate spatial scales (within urban areas and within regions). In addition, large-scale PM2.5 variation is well estimated from the monitoring data, at least in the United States. We found little evidence that current AOD products are helpful for improving prediction at small-to-moderate scales in the eastern United States and believe more evidence for the reliability of AOD as a proxy at such scales is needed before making use of AOD for PM2.5 prediction in epidemiologic contexts. While our results relied in part on relatively complicated statistical models, which may be sensitive to modeling assumptions, our exploratory correlation analyses (see sections 3 and 5) and relatively simple regression-style modeling of MISR AOD (see section 4) were consistent with the more complicated modeling results. When assessing the usefulness of AOD in the context of studying chronic health effects, we believe efforts need to focus on disentangling the temporal from the spatial correlations of AOD and PM2.5 and on understanding the spatial scale of correlation and of the discrepancy structure. While our results are discouraging, it is important to note that we attempted to make use of smaller-scale spatial variation in AOD to distinguish spatial variations of relatively small magnitude in long-term concentrations of ambient PM2.5. Our efforts pushed the limits of current technology in a spatial domain with relatively low PM2.5 levels and limited spatial variability. AOD may hold more promise in areas with higher aerosol levels, as the AOD signal would be stronger there relative to the background surface reflectance. Furthermore, for developing countries with high aerosol levels, it is difficult to build statistical models based on PM2.5 measurements and land-use covariates, so AOD may add more incremental information in those contexts. More generally, researchers in remote sensing are involved in ongoing efforts to improve AOD products and develop new approaches to using AOD, such as calibration with model-estimated vertical profiles and the use of speciation information in MISR AOD; these efforts warrant continued investigation of the usefulness of remotely sensed AOD for public health research.