|Home | About | Journals | Submit | Contact Us | Français|
Satellite-based remote sensing provides a unique opportunity to monitor air quality from space at global, continental, national and regional scales. Most current research focused on developing empirical models using ground measurements of the ambient particulate. However, the application of satellite-based exposure assessment in environmental health is still limited, especially for acute effects, because the development of satellite PM2.5 model depends on the availability of ground measurements. We tested the hypothesis that MODIS AOD (aerosol optical depth) exposure estimates, obtained from NASA satellites, are directly associated with daily health outcomes. Three independent healthcare databases were used: unscheduled outpatient visits, hospital admissions, and mortality collected in Beijing metropolitan area, China during 2006. We use generalized linear models to compare the short-term effects of air pollution assessed by ground monitoring (PM10) with adjustment of absolute humidity (AH) and AH-calibrated AOD. Across all databases we found that both AH-calibrated AOD and PM10 (adjusted by AH) were consistently associated with elevated daily events on the current day and/or lag days for cardiovascular diseases, ischemic heart diseases, and COPD. The relative risks estimated by AH-calibrated AOD and PM10 (adjusted by AH) were similar. Additionally, compared to ground PM10, we found that AH-calibrated AOD had narrower confidence intervals for all models and was more robust in estimating the current day and lag day effects. Our preliminary findings suggested that, with proper adjustment of meteorological factors, satellite AOD can be used directly to estimate the acute health impacts of ambient particles without prior calibrating to the sparse ground monitoring networks.
Environmental epidemiology studies have established a robust association between acute and chronic exposure to airborne fine particulate matter with diameter <2.5 µm (PM2.5) and adverse health effects such as increased overall mortality, as well as cause-specific mortality, cardiovascular and pulmonary diseases, asthma, and lung cancer (Dockery et al., 1993; Laden et al., 2006; Pope and Dockery, 2006). In contrast to chronic health effect studies of PM which primarily rely on spatial heterogeneity in mean PM concentrations to estimate the effects (Yanosky et al., 2008), day-to-day variations of PM levels are much more important in short-term health effect studies (Dominici et al., 2006; Samet et al., 2000). Many previous studies relied on central monitors to assign uniform exposure to population living within a certain distance to the monitor. Besides the exposure misclassification related to this approach, its application is limited to the spatial and temporal availability of ground measurements from a monitoring network. For example, most U.S. PM2.5 monitors are operated on an every-3-day or every-6-day sampling schedule. In addition, most monitors are located in urban area with sparse or no coverage in suburban and rural areas even in the U.S. Lack of routine ground monitoring is a major factor limiting both chronic and acute PM2.5 health effects research in developing countries.
Various modeling approaches have been explored to improve the spatial and temporal coverage of PM2.5 concentrations. For example, land use regression models have been developed to provide spatially resolved PM2.5 levels to support chronic health effect studies (Jerrett et al., 2005). Model simulated PM2.5 levels have been evaluated as exposure estimates (Bravo et al., 2012). Given its broad spatial coverage, satellite-based monitoring data can greatly supplement and expand ground monitoring networks to study the spatial and temporal variations of PM, particularly in suburban and rural areas far from ground monitoring sites. Satellite-derived aerosol optical depth (AOD), retrieved at visible wavelengths such as the green bands (550 nm), is more sensitive to PM2.5 and can be used as a quantitative measure of PM2.5 abundance in the atmospheric column (Gupta et al., 2006; Koelemeijer et al., 2006; Liu et al., 2005; Liu et al., 2007a; Paciorek et al., 2008). Although satellite-derived AOD has been successfully used to document pollution episodes (Al-Saadi et al., 2005; Wang and Christopher, 2003), the application of satellite-based exposure assessment in environmental health is in its infant stage. Most research focuses on a pre-calibration approach of developing simple empirical models of AOD based on ground PM2.5 measurements, and then evaluates the health effects of built models (Kloog et al., 2012). However, the success of model building is limited by temporal mismatch between 24-h average PM2.5 and daytime (often single snapshot) AOD and various factors impacting on the measurement accuracy from ground or space. To date, there are only a few studies that have examined the relationships between pre-calibrated AOD and long-term health effect (Hu, 2009; Hu and Rao, 2009).
Considering the limited success in model building using pre-calibration approach, we tested the hypothesis that AOD is directly associated with acute and/or chronic health effects without pre-calibration, as observed using ground data, under the condition that AOD is an indicator of ground-level PM concentrations. In this exploratory study, we tested this hypothesis using a healthcare database of hospital admissions collected from the entire geographic region of Beijing in 2006, and further evaluated using two independent healthcare databases, including unscheduled outpatients visits and mortality, from sub-geographic regions of Beijing. We focused on daily health outcomes extracted from these databases to evaluate the acute effects associated with PM exposure. Our main objective was to compare air pollution-associated health effects monitored from space (by AOD) and ground (by PM), thus, ground measurement of PM10 particles was used as reference to evaluate the AOD application in environmental health research even though there was no PM2.5 data available in Beijing. To our knowledge, this study was the first quantitative application of satellite aerosol remote sensing data to estimate air pollution-health effects.
Three databases covering different geographic regions of Beijing (Fig. 1) were used to compare the effects of air pollution assessed using ground PM10 and AOD, including unscheduled outpatient visits, hospital admissions, and mortality. The daily events were extracted according to the disease categories in Table 1. Strongly influenced by weekday/weekend schedule, major national holidays, and administrative interruptions, the daily events of unscheduled outpatient visits and hospital admissions had regular patterns of weekly fluctuations (Fig. 2). The days at the weekends, holidays, and administrative interruptions had the lowest daily events, and the days immediately followed the weekends, holidays, and administrative interruptions had the highest daily events. Using a 7-day moving average, we were able to completely remove weekly fluctuations resulting in a few major gaps corresponding to holidays and administrative interruptions longer than five weekdays. In these two databases, we investigated lag effects of air pollution within one-week period by using 7-day moving average of daily events starting from the current day or the next day.
Air Quality Index (AQI) is a color-coded reporting system commonly used by government agencies to characterize the air quality for a number of pollutants. It is a piecewise linear function to convert air pollutant concentration into AQI, which is divided into ranges with a descriptor and a color code assigned to each range. China implemented a modified AQI system, according to the guideline issued by the US Environmental Protection Agency (http://www.epa.gov/ttn/oarpg/t1/memoranda/rg701.pdf).
Chinese AQI level is based on the levels of 5 atmospheric pollutants, including sulfur dioxide (SO2), nitrogen dioxide (NO2), particulates (PM10), carbon monoxide (CO), and ozone (O3). A daily AQI score is assigned to the level of each pollutant and the final daily AQI is the highest of those 5 scores. The type of pollutant is only specified for the day with AQI> 50.
Daily AQI covering Beijing area was obtained from the Beijing Environmental Protection Bureau. In 2006, PM10 was considered as the major air pollutant because PM10 had the highest concentrations among all monitored pollutants for all of the days with AQI above 50, which accounted for over 92% of the days in 2006. Daily AQI values at all 35 ground monitors in Beijing were converted to PM10 concentrations, and geometric means of PM10 were used in this study. For those days without pollutant specified (AQI<50), we treated PM10 as the major air pollutant. To convert from concentration to AQI the equation:
was used, where: I=AQI value, C=the PM10 concentration, Clow =the lower limit of PM10 concentration of corresponding range of C, Chigh = the higher limit of PM10 concentration of corresponding range of C, Ilow=the lower limit of AQI corresponding to Clow, Ihigh=the lower limit of AQI corresponding to Chigh.
Local meteorological parameters collected at one station located at Beijing International Airport (32 km northeast of city center) were extracted from Worldwide Surface Observations Database (DS3505, integrated surface hourly) at National Climatic Data Center. Daily averages of meteorological parameters were calculated from hourly values, including temperature, relative humidity (RH), precipitation, air pressure, wind, and visibility. Absolute humidity (AH) was calculated from RH and temperature data as described previously (Shaman and Kohn, 2009).
The Moderate Resolution Imaging Spectroradiometer (MODIS) instruments, aboard on the EOS Terra (since February 2000) and Aqua (since June 2002) satellites, cross the equator at approximately 10:30 a.m. and 1:30 p.m. local time, respectively (Remer et al., 2005). Both MODIS instruments provide AOD retrievals at a spatial resolution of 10 km every 1–2 days at mid-latitudes. AOD values of Collection 5 MODIS data in 2006 covering northern China including Beijing were downloaded from the Goddard Space Flight Center MODIS Level 1 and Atmosphere Archive and Distribution System (http://ladsweb.nascom.nasa.gov). Daily AODs were retrieved as described previously (Liu et al., 2009). Briefly, we first selected MODIS pixels whose centroids fall within Beijing city boundaries, then calculated city-average Terra MODIS AOD and Aqua MODIS AOD respectively, and finally took the average of two satellite AOD values as daily mean AOD. Given our single-city study domain and the relatively coarse resolution of MODIS AOD, satellite data in our analysis is used to improve the temporal coverage of fine particle exposure estimates only.
To relate short-term effects of air pollution to the daily health outcome, we used the time series procedures. Because daily counts of health events typically follow a Poisson distribution with large dispersion, we implemented generalized linear models (GLM) with negative binomial distribution as described previously (Dominici et al., 2000; Katsouyanni et al., 1996; Schwartz et al., 1996). The effects on the same day or lagged days of PM exposure (daily PM10 or AOD) were investigated for each health outcome. Since AOD is dimensionless, we used log2-transformed values in modeling to obtain uniformed comparison of the PM-related health effects assessed by ground and space monitoring. To accommodate weekly fluctuations in hospital admission and unscheduled outpatient visit, we created a categorical variable to represent weekdays or weekends/holidays, and incorporate this variable in analysis. This variable was not included in mortality analyses. Since the data structure is relatively simple, which were collected with-in one calendar year and one geographic region, we did not use smoothing function for seasonality and weather control. Instead, a parametric season variable and meteorological variables were included in the models. We also conducted several sensitivity analyses to assess the impacts of different meteorological variables. Model selection, as well as evaluation between health effect models of AOD and PM exposure, was based on likelihood test, Akaike Information Criterion, and residual predictions. In order to reduce weekly fluctuation we also used the 7-day moving average of daily health events in the analysis of hospital admission and unscheduled outpatient visit, without categorical variable representing weekdays or weekends/holidays. All statistical analyses were performed using the SAS package (version 9.1, SAS Inc., Cary, NC). Relative risks were calculated, and all statistical tests were two-sided and values of P<0.05 were considered significant.
In 2006, PM10 daily average was 161 µg/m3 with the 25th percentile at 94 µg/m3 and the 75th percentile at 196 µg/m3. There were 265 days (73%) with valid AOD measurements, and the missing days were not significantly clustered in any season (χ2 test, p=0.082). The correlation between ground PM10 and AOD was low (Pearson correlation coefficient, 0.22; p=0.0003). Compared with PM10, the weather condition had more effects on the AOD as there were stronger correlations between AOD and meteorological factors (Supplemental material, Table 1). In contrast to higher levels of air pollution observed normally in winter and spring, AOD tended to be higher in summer (data not shown).
AOD values can be inversely affected by atmospheric water content. Relative humidity (RH) is the most frequently used meteorological factor for measuring the atmospheric water content, contrary to that the absolute humidity (AH, absolute measure of water vapor in the air) was seldom used in previous studies. Although there was a strong correlation between the two measurements (Supplemental material, Table 1), their daily average showed significantly different patterns (Fig. 3A and B). RH tended to have large day-to-day variations and was more variable in winter, spring, and fall, because high values could be found in all seasons but low RH was seldom observed in summer. AH had much smaller day-to-day variation (mean daily changes, 1.4±1.3 g/m3), with an annual daily average of 8.2 g/m3 (median: 5.9 g/m3; range, 0.4–25.9 g/m3). Annual distribution of daily AH had a distinct bell-shaped pattern and fitted well with the seasonality (Fig. 3A). Driven by low ambient temperature, winter in Beijing had very low AH. AH started to increase gradually in the spring, peaked in the summer, decreased gradually in the fall, and returned to low levels in the next winter. Compared with AH, daily RH seemed to overestimate the atmospheric water content (Fig. 3B) and less coped with seasonal variations (Fig. 3A). Moreover, AH had a strong but non-linear correlation with daily dew point (Fig. 3C), another meteorological measure of atmospheric moisture. Furthermore, the impacts of AH on AOD was supported by a moderate positive correlation (Pearson correlation coefficient, 0.454; p<0.0001), contrary to a small negative correlation between PM10 and AH (Pearson correlation coefficient, −0.225; p<0.0001). Therefore, we calibrated daily AOD values by simply dividing them by corresponding daily average AH, named. The AH-calibrated AOD had an annual trend which agreed well with that of PM10 (Fig. 3D). The correlation coefficient between the AH-calibrated AOD and the PM10 was 0.323 (p<0.0001), considerably higher to that estimated for the uncalibrated AOD.
Certain disease categories, especially cardiovascular and respiratory diseases, demonstrated a clear seasonal trend, with summer having fewer daily events but winter having more events (Fig. 3E and Supplemental material, Fig. 1). Annual patterns of PM10 and AOD were similar (Fig. 3D), and aligned well with health outcomes (Fig. 3F). Thus, seasonality had clear impacts on both air pollution and health outcomes in Beijing. Major meteorological factors, such as temperature, atmospheric pressure, dew point, and humidity, were highly correlated (Supplemental material, Table 1). Except RH, all meteorological factors had seasonal variations with either bell-shaped (AH, temperature, and dew point) or inverted bell-shaped (atmospheric pressure) patterns (data not shown). Since the atmospheric water content as measured by AH is determined by both ambient temperature and pressure, and had direct effects on the AOD, we selected AH as the representative variable of seasonality in this study.
Since AH was used to calibrated AOD, we did not adjusted AH or other meteorological factor as an independent covariate of seasonality in GLM analysis as we did for PM10. In the hospital admission database, both AH-calibrated AOD and PM10 were associated with significantly increased daily admissions of cardiovascular diseases including ischemic heart diseases and heart failure, as well as respiratory tract infections, but not associated with heart rhythm disturbances, strokes, neoplasm, and gastrointestinal disorders (Fig. 4A). AH-calibrated AOD was also significantly associated with higher daily admissions related to peripheral vascular disease, COPD, diabetes, and hypertension. Furthermore, AH-calibrated AOD was associated with decreased daily admissions caused by injury, whereas, the PM10 did not have significant associations with them. In GLM models, all relative risks predicted by PM10 had wider confidence intervals. When we examined the lag effects using 7-day moving average, AH-calibrated AOD had robust predictions on all disease categories except heart rhythm disturbances (Supplemental material, Fig. 2A). Instead of having no immediate effects, AH-calibrated AOD seemed to have significant lag effects on heart rhythm disturbances. In contrast, PM10 predictions on the current day effects and the lag effects were not consistent (Supplemental material, Fig. 2B).
We did not observe any immediate effects by the current day exposure in the database of daily unscheduled outpatient visits. Instead, we found significant lag effects during a 7-day period on cardiovascular diseases, ischemic heart diseases, COPD, and respiratory tract infections by both AH-calibrated AOD and PM10 (Fig. 5A). In contrast to the PM10, AH-calibrated AOD was not significantly associated with strokes.
Because daily mortalities did not have weekly fluctuations, we could investigate lag effects on single day instead of in the 7-day period. As showed in Fig. 6, both AH-calibrated AOD and PM10 performed well in GLM models. The most robust associations were observed with one lag day for daily mortalities. AH-calibrated AOD appeared more sensitive in detecting lag effects.
Across all databases, we found that both AH-calibrated AOD and PM10 were consistently associated with the elevated daily events for cardiovascular diseases, ischemic heart diseases, and COPD. The relative risks estimated by AH-calibrated AOD and PM10 were similar, with the largest effects associated with COPD related outcomes. Overall, AH-calibrated AOD had narrower confidence intervals and was more sensitive in detecting adverse outcomes due to short-term exposure.
We also tested AH as a stand-alone risk factor for daily health outcomes (Figs. 4A, ,5A,5A, and and6).6). In all databases, AH demonstrated protective effects at compatible levels for all categories of cardiovascular and respiratory diseases. Additionally, it had protective effects on diabetes and hypertension, but had increased risk for injury, in the databases of hospital admissions and unscheduled outpatient visits. AH had no effects on daily events related to neoplasm. To address the question whether AH-calibrated AOD associated effects were solely driven by AH component, we further conducted the analyses using uncalibrated AOD and AH in the same GLM models, as in the analyses of PM10. We observed some discrepancies between AH-calibrated AOD and uncalibrated AOD mostly among the disease categories with small number of daily events (Figs. 4B, ,5B,5B, and and6).6). However, we still observed that uncalibrated AOD had significant increased risks associated with cardiovascular diseases, ischemic heart diseases, and COPD. Relative risks estimated for these diseases were compatible to those from AH-calibrated AOD, as well as from PM10. In an additional set of analyses, we calculated residual AOD after linear regression by AH, and applied residual AOD and AH in the same models. Since AH only accounted for a small portion of AOD variations (adjusted R2=0.203, p<0.0001), the results of residual AOD were very similar to the results of uncalibrated AOD (data not shown).
Because of high degree of correlations, other meteorological factors were also associated with daily health outcomes in all databases (Supplemental material, Fig. 3). Further, using other meteorological factors to replace AH, we obtained similar results for PM10 (data not shown). Among all meteorological factors, AH always gave the best models. However, we could not substitute AH with other meteorological factors in uncalibrated AOD models.
In this study, we observed that AH-calibrated AOD were consistently associated with elevated daily events on the current day and/or lag days for cardiovascular diseases, ischemic heart diseases, and COPD. The results were robust across three independent databases. The associated disease patterns and overall effects estimated by relative risks were similar to ground monitoring by PM10. Additionally, compared to ground PM10, we found that AH-calibrated AOD had narrower confidence intervals for all models and was more robust in estimating the current day and lag day effects. The short-term effects associated with AOD were also consistent with previous studies using ground PM2.5 on the acute health effects of unscheduled outpatient visit, hospital admission and mortalities (Chang et al., 2005; Dominici et al., 2006; Fung et al., 2005; Ostro et al., 2006), including studies in Beijing (Xu et al., 1994, 1995b).
Water absorption can increase the size of hydrophilic particles containing sulfate, nitrate, ammonium, and certain species of organic carbon (Tang and Munkelwitz, 1994). The atmospheric water content has direct but inverse impacts on the optical measurement of PM from space, as AOD is a measure of particle light extinction (Malm and Day, 2001). Compare to RH, AH is a specific and direct measure of water vapor density, and is more relevant to the physical characteristics of the optical measurement. Thus, AH-calibrated AOD could provide a more accurate assessment of fine PM-related effects. Our findings suggest that adjusting AOD with other meteorological measures of water vapor content, such as RH and dew point, did not result in robust associations were consistent this (data not shown). We had also tried various ways to adjust AOD using AH, including square or square root of AH, cube or cube root of AH, or log transformed AH, and found that the simple dividing AOD by AH gave the best and most robust models (data not shown).
Since AOD dimensionless, we used log2-transformed values to obtain uniformed comparison of PM-related health effects assessed by ground and space monitoring. Our results can be interpreted as the levels of increased health risks when the ambient PM level is doubled. Beijing is a highly polluted megacity with PM to be the most severe air pollution issue (Okuda et al., 2004; Song et al., 2006). Besides high daily levels, Beijing also had large day-to-day PM variations, as we observed 40 occasions (~every 10 days) that daily PM10 increases were above 2 folds throughout 2006. In addition, the significant linear relationships between log2-transformed PM values and the adverse health effects suggested a nonlinear exposure-response with larger effects at low PM exposure, consistent with previous studies in Beijing (Xu et al., 1995a) and other regions (Pope et al., 2009).
Daily PM10 can be substantially reduced by occasional precipitation and strong wind, but it was less influenced by temperature and humidity. Without adjusting meteorological factor, we found PM10 still had significant associations with cardiovascular and respiratory diseases with similar estimated risks (data not shown). Daily averages of meteorological factors were highly correlated, and were independently associated with daily health outcomes. Similar observations were previously reported in Beijing (Xu et al., 1994), as well as in other countries (Braga et al., 2001; Morabito et al., 2005; Sharovsky et al., 2004). Among meteorological factors, AH seemed to have a robust performance in estimating weather-related health effects, attributing to its inseparable physical characters with atmospheric temperature and pressure. For most of disease categories, AH caused less than 2% changes on daily health events with one unit change (g/m3). Mechanisms leading to the possible influence of weather are most likely multifactorial, with a complex relationship between seasons and pathophysiological exogenous and endogenous factors (Abrignani et al., 2009). Shaman et al. recently reported that AH had stronger effects than RH on influenza virus survival within aerosolized droplets (Shaman and Kohn, 2009), and was a major determinant for seasonal variations of virus transmission and associated with influenza-related mortality in the United States (Shaman et al., 2010). Furthermore, a previous study reported that influenza infection was a major cause of winter increase of all-cause and cardiovascular diseases mortality in the United States over a 40-year period (Reichert et al., 2004). Therefore, the weather-related health effects might be indirectly caused by the increased burden of infections in cold seasons.
Several factors might account for a poor correlation between PM10 and AOD. Firstly, they are measuring different categories of particles between PM10 and AOD (Engel-Cox et al., 2004; Liu et al., 2007b). One limitation of this study was the lack of ground PM2.5 data. Secondly, the sampling periods are different with PM10 representing 24-h average, whereas, AOD corresponding to a snap-shot of daytime PM level. Thirdly, ground monitoring stations have insufficient coverage of PM and meteorology over the entire metropolitan area of Beijing. Finally, various factors have different impacts on the measurement accuracy from ground or space. Being aware of these potential impacts, all of previous studies focused on developing empirical models of AOD by calibrating with ground measurement and local meteorological information (Koelemeijer et al., 2006; Liu et al., 2005). This approach, depending on the available ground data, worked well for long-term average of PM over a large geographic region (Al-Saadi et al., 2005; van Donkelaar et al., 2010; Wang and Christopher, 2003). Currently, there are limited studies that adopt empirical AOD models in environmental health research. By merging AOD with ground measurements using geographically weighted regression, the standardized county-level biennial mortality rates (2003–2004) of chronic heart diseases in the United States were associated with two-year average re-calculated PM2.5 levels (Hu, 2009), or with two-year average satellite-derived AOD raster data (Hu and Rao, 2009). It is worth to note that, in these studies, data for cold seasons (October to March) were not used in the two-year average calculation and model analysis.
The intrinsic differences of healthcare data and associated quality attributed to the variations of PM and/or weather related effects. Since the admission database covered the entire metropolitan area with large daily events and more accurate diagnosis, models for both PM10 and AOD performed best in this database. The mortality data was limited to eight districts of downtown Beijing with small daily events and a lower quality on diagnosis. In contrast, cohort outpatient data was collected from a small, older (mean age±SD: 60±15), and female dominant (67%) population. Although it had large daily events, the quality was affected by large amounts of regular visits for medicine refill and less accuracy on diagnosis.
To our knowledge, this is the first study of directly assessing PM-induced acute adverse health effects by satellite remote sensing, without a prior calibration using ground monitoring data. AH-calibrated AOD is an integrated measurement of satellite-based AOD and local weather conditions. By bypassing a pre-calibration step of empirical model building, it avoids the problems caused by the lack of ground monitoring networks. In contrast to ground monitoring data that often lack spatial and temporal coverage and suffer unbalanced spatial distribution, the repetitive and broad-area coverage of satellites allows atmospheric remote sensing to offer a unique opportunity to monitor air quality at global, continental, national and regional scales. Satellite-based remote sensing could help fill pervasive data gaps that impede efforts to study air pollution and protect public health. In addition, the evidence of satellite-based environmental health research supports targeting of policy interventions on high-risk regions to reduce pollution levels. Addition studies are warranted to further explore the potential health benefits of satellite remote sensing.
This work was supported by grant ES000002 from National Institute of Environmental Health Sciences, by Research Project Awards from Harvard University Center for the Environment, and by pilot project from Harvard-NIEHS Center for Environmental Health.
We thank Dr. Xihong Li at Biostatistics Department, Harvard School of Public Health for organizing Environmental Statistics Retreat in 2006, where the idea of this study was initiated. We thank Dr. Jeffrey Shaman at Oregon State University for his insightful seminar at Harvard School of Public Health, which enlightened us to apply absolute humidity in AOD calibration. We also thank our colleague, Dr. Douglas Dockery, for discussion and encouragement.
Appendix A. Supplementary data
Supplementary data to this article can be found online at http://dx.doi.org/10.1016/j.envint.2012.10.011.