Supporting Information

Transcription

1 1 Supporting Information Daily estimation of ground-level PM 2.5 concentrations over Beijing using 3 km resolution MODIS AOD Yuanyu Xie 1, Yuxuan Wang* 1,2,3, Kai Zhang 4, Wenhao Dong 1, Baolei Lv 1, Yuqi Bai 1 1 Ministry of Education Key Laboratory for Earth System Modeling, Center for Earth System Science, Tsinghua University, Beijing, China. 2 Department of Marine Science, Texas A&M University at Galveston, Galveston, TX, 77553, USA 3 Department of Atmospheric Science, Texas A&M University, College Station, TX, 77853, USA 4 Department of Epidemiology, Human Genetics and Environmental Sciences, University of Texas School of Public Health, Houston, Texas, 77030, USA; * Corresponding Author 15 Yuxuan Wang: Phone: ; yxw@tsinghua.edu.cn Contents 15 pages 4 Figures 5 Tables 21 S1

2 Text S1. MODIS Aqua 3 km AOD calibration We calibrated the MODIS Aqua 3 km AOD against AERONET AOD at three surface sites in Beijing: Beijing-CAMS, Beijing, and Beijing_RADI (Figure S1). The total number of AOD pairs for the three sites are 35, 61 and 49 respectively with the corresponding Pearson correlation coefficient r of 0.93, 0.93 and 0.94 at the three sites. The correlation slopes at the three sites are 1.32, 1.33 and 1.23, respectively, indicating a persistent high bias of the MODIS AOD compared to AERONET AOD at all ranges of AOD values. The mean AOD differences (MODIS minus AERONET) at the three sites are 0.33, 0.30 and 0.25 respectively Figure S1. Validation of MODIS Aqua 3 km AOD with three surface AERONET sites (Beijing-CAMS [39.933N, E], Beijing [39.977N, E], Beijing_RADI [40.005N, E]). 36 S2

3 Text S2. Site-specific statistics and seasonal comparisons Descriptive statistics. Descriptive statistics for PM 2.5 concentrations and site-collocated MODIS AOD at each of the 35 site are shown in Table S1. The total number of valid observations during the study period is for PM 2.5 at the ground-based monitors and 2818 for the satellite AOD. The number at individual site locations ranges from 286 to 338 for PM 2.5 and 32 to 148 for the satellite AOD. The mean concentration of PM 2.5 over the study period has a maximum of μg/m 3 at site #1 in Daxing district and a minimum of μg/m 3 at site #35 in Miyun district. The standard deviation at individual sites ranges from μg/m 3 to μg/m 3, indicating large PM 2.5 variability in Beijing. The maximum AOD of 0.89 is found at site #6 in Daxing district and minimum AOD of 0.33 is found at site #32 in Yanqing district. The standard deviation of AOD is also found to be large (0.34 to 0.71). Average PM 2.5 from the ground-based monitors is μg/m 3 (SD = μg/m 3 ) in the warm season and μg/m 3 (SD =80.39 μg/m 3 ) in the cold season. Average site-collocated AOD is 0.75 (SD=0.51) for the warm season and 0.34 (SD = 0.26) for the cold season. Data availability for ground-based PM 2.5 monitors does not differ much between the warm and the cold season, but the number of available satellite AOD data in the cold season is only one third of that in the warm season. However, ground-based AOD at the three AERONET sites has a higher percent of available data in the cold season (62.1%) than in the warm season (41.3%). This indicates that the limitation of satellite AOD retrieval over snow is a more important reason than cloud cover in causing the lower percent of available AOD retrievals from MODIS in the cold season. The seasonal difference of mean AOD of 54.7% from MODIS is much higher than 18.7% from the AERONET. The seasonal difference of AERONET AOD could be attributed to lower planetary boundary layer (PBL) depths and lower relative (RH) humidity in the cold season, while the seasonal differences of MODIS AOD is caused, additionally, by larger percent of missing satellite data in the cold season. The median PM 2.5 to AOD ratios derived during the periods when both data are available are higher in the cold season than the warm season for both MODIS and AERONET S3

4 AOD. This further verifies the effect of lower PBL and lower RH in the cold season in causing lower AOD and higher concentrations of PM 2.5 at the surface. Model performance statistics. Model performance statistics for the linear regression and mixed effects models at each of the 34 site (#12 and #14 were averaged) are shown in Table S2. The total available pairs of data (N) at each site range from 19 to 97 for the linear regression and 16 to 67 for the mixed effects models during the whole study period. The slightly smaller N of the mixed effects models can be mainly attributed to the data selection process which requires at least two pairs of AOD-PM data for a single day. Average R 2 of the 34 sites is 0.51 (SD=0.15) with the maximum value of 0.81(site #35) and minimum of 0.23 (site #8) for linear regression model. Site-specific MPE is between μg/m 3 and μg/m 3 (SD=4.17), and RMSE is between μg/m 3 to μg/m 3 (SD=7.01). The average R 2 for the mixed effects models with and without site effect is 0.83 at the 34 sites. MPE and RMSE range from 4.6 μg/m 3 to μg/m 3 for the mixed effects models. The site-specific statistics indicate large improvements in model performance of the mixed effects models compared with the linear regression model. Compared with the standard models (i.e., the linear and mixed effects models fitted with data from the whole study period), the warm-season-only models show no improvement in model performances, yet the cold-season-only models appear to perform better with higher R 2 and 5% ~ 12% reduction in MPE and RMSE. However, the standard models also show better performance statistics when applied to the cold season data only, with their R 2, MPE, and RMSE all similar to those from the cold-season-only models. Based on this observation, it is unlikely that the model itself can be improved through separate fitting by season. 90 S4

5 91 Table S1. Descriptive statistics for PM 2.5 concentrations (μg/m 3 ) at 35 surface sites 92 and site-collocated satellite AOD from March 26 th 2013 to Apr 23 th PM 2.5 Site-collocated AOD Site N a Mean SD b Min Max Median N Mean SD Min Max Median a:n denotes total valid observations; b:sd represents the standard deviation of the data S5

6 95 Table S2. Model performance statistics of R 2, MPE (μg/m 3 ) and RMSE (μg/m 3 ) at 96 each site for the linear regression and mixed effects models. Linear regression Mixed effects Mixed effects (without site effect) (with site effect) Site # N a R 2 MPE b RMSE c N R 2 MPE RMSE N R 2 MPE RMSE * *Site # 12 and # 14 are averaged for model input and results shown are for the two sites averaged value. a:n denotes total available pairs of data; b:mpe is estimated as the absolute differences between predicted and measured PM 2.5 concentrations. S6

7 c:rmse is estimated as the root mean squared differences between predicted and measured PM 2.5 concentrations. Text S3. Sensitivity test for linear regression and mixed effects models. We tested the model sensitivity to site location by excluding the sites that are not represented by the MODIS 3 km AOD product. In one sensitivity test we excluded the two southern sites (#1, 2) and in the other test we excluded the two traffic sites (#8, 12). Model performance statistics for these sensitivity tests as well as the original model (i.e., model fitted with all the sites) are compared in Table S3. The R 2 with site #1 and 2 excluded is 0.47 for the linear regression and 0.80 for the mixed effects models, similar to the original models. The MPE and RMSE are slightly smaller compared to the original models of the linear regression and mixed effects model without site effect. For the mixed effects model with site effect, the MPE and RMSE become slightly higher after exclusion of the two southern sites, indicating that these two sites, when given a location specific value, would be better represented by the mixed effects model. When site #8 and 12 are excluded, the model performance statistics are similar to the original models for the linear regression and mixed effects model without site effect, while MPE and RMSE for the mixed effects model with site effect is higher than those of the original model. This also shows that these two traffic sites could be better represented provided a consideration of spatial location. 123 S7

8 Table S3. Model performances of R 2, MPE (μg/m 3 ) and RMSE (μg/m 3 ) at each site for the linear regression and mixed effects models. Model type N a Slope b Intercept c R 2 MPE d (μg/m 3 ) RMSE e (μg/m 3 ) MODIS 3 km AOD Linear regression Mixed effect Mixed effect (site effect) MODIS 3 km AOD (sites # 1,2 excluded) Linear regression Mixed effect Mixed effect (site effect) MODIS 3 km AOD (sites # 8,12 excluded) Linear regression Mixed effect Mixed effect (site effect) a:n denotes total available pairs of data; b:fixed regression slope derived from the models c:fixed regression intercept derived from the models d:mpe is estimated as the absolute differences between predicted and measured PM 2.5 concentrations. e:rmse is estimated as the root mean squared differences between predicted and measured PM 2.5 concentrations. S8

9 Text S4. Comparison between satellite-derived and monitor-derived district-mean PM 2.5 The district-mean PM 2.5 derived from surface monitors and satellite predictions are compared in Figure S2. The relative differences between satellite-derived (green) and monitor-derived (orange) district-mean PM 2.5, percentages calculated as the satellite-mean minus monitor-mean and divided by the monitor-mean, span from -52% (Mentougou district) to +20% (Xicheng district). The monitor-derived district-mean PM 2.5 is higher than the satellite-derived value for all the suburban districts and three urban districts. The much higher value derived from surface monitors are mainly due to uneven site distributions, with fewer sites located in remote and mountainous regions. For the densely populated urban districts such as Xicheng, Chaoyang and Fengtai, the satellite-derived mean PM 2.5 is ~20%, 12% and 4% higher than that derived from the ground-based monitors, respectively. This suggests the problem of under-sampling by surface monitors over those urban regions Figure S2. District-mean PM 2.5 derived from surface monitors (orange) and the satellite (green) as well as total population (gray) for the 16 districts in Beijing. The red line indicates China s NAAQS annual mean PM 2.5 standard. 154 S9

10 Text S5. Comparison between model performance using MODIS Aqua 3 km AOD and coarser resolution products. We performed a set of comparisons between the MODIS Aqua 3 km AOD product, regridded 6 km and 9 km AOD products averaged from the 3 km product, and the standard Level 2 (L2) 10 km AOD product. Comparison between the regridded 6 km and 9 km AOD products and the original 3 km AOD product. The 6 km and 9 km AOD products were derived from averaging the original MODIS Aqua 3 km AOD product. The criteria set for regridding is that at least half of the 3 km AOD is available for a coarser grid (i.e. 2 out of 4 for 6 km grid, or 4 out of 9 for 9 km grid). The total available pairs of AOD and PM 2.5 data are 1332 for the 6 km and 1568 for the 9 km products. Compared to the original 3 km product, the data availability of those regridded coarser-resolution products do not show significant changes (within 8%). We developed both linear regression and mixed effects model using the regridded 6 km and 9 km AOD products. The model results are shown in Table S4. The R 2 ranges from 0.44 to 0.46 for the linear regression models and 0.79 to 0.83 for the mixed effects models that are developed from the regridded 6 km and 9 km products. MPE and RMSE are higher than 22.6 μg/m 3 and 34.8 μg/m 3 for the linear regression model, and those for the mixed effects models are higher than 11.3 μg/m 3 and 18.5 μg/m 3 respectively. The mixed effects model developed from the 3 km AOD product has better performances than that developed from the 6 km and 9 km products, with R 2 higher by 0.001~0.02 and MPE and RMSE lower by ~10% (Table S4). Comparison between the L2 10 km AOD products and the 3 km AOD product. The standard MODIS Aqua L2 10 km AOD product was downloaded from the NASA website ( The L2 10 km AOD was validated with AERONET and the Pearson correlations coefficient r are 0.91, 0.92 and 0.91 respectively at three sites in Beijing (Figure S3). By comparison, the 3 km product is more consistent with AERONET with higher R 2 (Figure S1). Basic statistics for the S10

11 L2 10 km and 3 km AOD products are compared in Table S5. The total available data (N) decreases by 32.9% from 2818 to 1890 when changing from the 3 km to 10 km AOD product. The daily mean time series of the L2 10 km AOD at collocated PM 2.5 sites and spatial distribution pattern for the study period mean are compared with those of the 3 km AOD in Figure S4. The 10 km product has more scattered points in the time series, which reflects the large decrease of data availability. Spatially, the 10 km products share similar pattern with the 3 km products, but provides much less details in terms of spatial variability (Figure S4). The performance statistics of the linear regression and mixed effects models using the L2 10 km AOD products are shown in Table S4 for comparison. The model R 2 from the 10 km product ranges from 0.43 (linear regression model) to 0.81 (mixed effects model). MPE and RMSE of the linear regression model are higher than 23.9 μg/m 3 and 36.6 μg/m 3 respectively; those from the mixed effects models are higher than 12.1 μg/m 3 and 17.9 μg/m 3 respectively. The model R 2 from the 10 km AOD product is ~0.02 lower compared to that from the 3 km AOD product, with mean MPE and RMSE higher by ~13% and ~17% respectively. The regression performance for the L2 10 km product is also found to be worse than that of the regridded 6 km and 9 km products. The better performance of the 3 km AOD product demonstrates the advantage of its finer resolution, which can better represent the variability of site-observed PM 2.5 concentrations. This also indicates large PM 2.5 spatial variability within small scales (several kilometers) over urban areas. The comparison based on MODIS AOD products at different resolutions leads to a conclusion of better performance statistics when using higher-resolution AOD products. 206 S11

12 Table S4. Model performances of R 2, MPE (μg/m 3 ) and RMSE (μg/m 3 ) at each site for the linear regression and mixed effects models. Model type N a Slope b Intercept c R 2 MPE d (μg/m 3 ) RMSE e (μg/m 3 ) MODIS 3 km AOD Linear regression Mixed effect Mixed effect (site effect) regridded 6 km AOD Linear regression Mixed effect Mixed effect (site effect) regridded 9 km AOD Linear regression Mixed effect Mixed effect (site effect) MODIS L2 10 km AOD Linear regression Mixed effect Mixed effect (site effect) a:n denotes total available pairs of data; b:fixed regression slope derived from the models c:fixed regression intercept derived from the models d:mpe is estimated as the absolute differences between predicted and measured PM 2.5 concentrations. e:rmse is estimated as the root mean squared differences between predicted and measured PM 2.5 concentrations. S12

13 Table S5. Descriptive statistics for site-collocated MODIS Aqua 3 km and 10 km AOD during the study period from March 26 th 2013 to Apr 23 th Averaging Period MODIS L2 10 km AOD b MODIS 3 km AOD N a Mean SD Min Max Median N Mean SD Min Max Median All Warm season (Apr 15 th - Oct 14 th ) Cold season (Oct 15 th to Apr 14 th ) a. N denotes total available data; b. All data are unitless Figure S3. Validation of MODIS Aqua L2 10 km AOD with three surface AERONET sites (Beijing-CAMS [39.933N, E], Beijing [39.977N, E], Beijing_RADI [40.005N, E]). 228 S13

14 Figure S4. Daily time series of AOD from the MODIS Aqua L2 10 km and 3 km product averaged at collocated PM 2.5 sites (a); spatial distribution of the L2 10 km (b) and 3 km (c) AOD products averaged for the entire study period over Beijing. 233 S14

15 References: (1) Remer, L. A.; Kaufman, Y. J.; Tanre, D.; Mattoo, S.; Chu, D. A.; Martins, J. V.; Li, R. R.; Ichoku, C.; Levy, R. C.; Kleidman, R. G.; et al. The MODIS aerosol algorithm, products, and validation. J. Atmos. Sci. 2005, 62 (4), (2) Levy, R. C.; Mattoo, S.; Munchak, L. A.; Remer, L. A.; Sayer, A. M.; Patadia, F.; Hsu, N. C., The Collection 6 MODIS aerosol products over land and ocean. Atmos. Meas. Tech. 2013, 6 (11), (3) Remer, L. A.; Mattoo, S.; Levy, R. C.; Munchak, L.A. MODIS 3 km aerosol product: algorithm and global perspective. Atmos. Meas. Tech. 2013, 6 (7), (4) Lee, H. J.; Liu, Y.; Coull, B. A.; Schwartz, J.; Koutrakis, P. A novel calibration approach of MODIS AOD data to predict PM 2.5 concentrations. Atmos. Chem. Phys. 2011, 11 (15), (5) Fu, J. Y.; Jiang, D.; Huang, Y.H. 1 KM Grid Population Dataset of China ( PopulationGrid_China ). Global Change Research Data Publishing & Repository. 2014, DOI: /geodb V1, S15