AOD products from satellites have been used to infer surface PM2.5 concentrations at scales ranging from urban to regional and to global (Liu et al., 2007; H. Zhang et al., 2009; Lee et al., 2011; Hu et al., 2014; Boys et al., 2014; van Donkelaar et al. 2014; Xie et al., 2015). To circumvent data scarcity of longer-term in situ surface measurement over China, here we used AOD retrieved from the Moderate Resolution Imaging Spectroradiometer (MODIS) sensor aboard both NASA EOS-Terra and Aqua satellite as the proxy data to represent the distribution and variability of PM2.5 air quality. Terra and Aqua are both polar-orbiting satellites launched in December 1999 and May 2002, respectively. They have been providing data every 1 to 2 days since February 2000 (Terra) and July 2002 (Aqua). MODIS retrieves aerosol properties in seven wavelengths from 0.47 to 2.13 µm and separate algorithms are applied over land and ocean (Tanré et al., 1997; Remer et al., 2005; Levy et al., 2007). To improve the retrieval over bright-reflecting source regions, the Deep Blue AOD algorithm was developed using multiple narrow-band channels at near-UV wavelengths (Hsu et al., 2004). Although the AOD uncertainty over land (±0.05 ± 0.2 × AOD) is higher than that over ocean (±0.03 ± 0.05 × AOD) (Remer et al., 2005; Chu et al., 2002), previous comparisons of MODIS AOD and ground-based AOD measurements from AErosol RObotic NETwork (AERONET) sites over land have shown tight correlations between the two, indicating that the MODIS AOD product is capable of providing quantitative information on the spatial and temporal variations of AOD over land (Levy et al., 2010; Prados et al., 2007).

Previous studies have indicated good correlations between the MODIS AOD and surface PM2.5 concentrations over selected sites in China (Wang et al., 2003; Xie et al., 2015). Here we used the MODIS level-3 monthly gridded AOD (550 nm) data (Version 5.1) from December 2000 to February 2013 with a 1∘ × 1∘ resolution. The AOD values over bright surfaces were replaced by the Deep Blue aerosol retrieval (550 nm) at the same grid.

To verify the robustness of our analysis using MODIS AOD, we also analyzed level-3 monthly gridded AOD from Multi-angle Imaging SpectroRadiometer (MISR) aboard of Terra. The MISR standard AOD products have a 0.5∘ × 0.5∘ resolution at 558 nm for 2001–2013. MODIS has a large number of spectral bands, while MISR has the multi-view-angle capabilities (Lyapustin et al., 2007).

The meteorological variables used to explore the mechanism behind the variations of SH and AOD are obtained from National Centers for Environmental Prediction (NCEP) reanalysis (Kalnay et al., 1996), including sea level pressure (SLP), relative humidity (RH), geopotential heights, and winds. The NCEP/NCAR reanalysis data provide a historical record of more than 50 years (Kistler et al., 2001) and are available on the 2.5∘ × 2.5∘ grid globally.

To verify the robustness of NCEP reanalysis in characterizing large-scale circulation patterns, we also analyzed the reanalysis data from European Centre for Medium-Range Weather Forecasts (ECMWF) Re-analysis Interim (ERA-Interim), the latest global atmospheric reanalysis produced by ECMWF (Simons et al., 2007). NCEP and ERA-Interim are the two widely used reanalysis products with relatively long periods.

Figure 3 depicts the time series of winter AOD averaged over NC, showing a significant increase in AOD from about 0.5 in 2001 to about 0.8 in 2013. A linear regression of the time series gives a trend of 1.5 % year-1 (r= 0.65, p < 0.05). Since the meteorological variables and atmospheric circulation patterns are not expected to drive such a large linear trend during this period, this AOD trend is mostly likely caused by increasing anthropogenic emissions over this region (Lu et al., 2010, 2011; Zhang et al., 2012; Streets et al., 2009). The departure of each winter's AOD from that depicted by the linear trend is assumed to represent the influence of meteorology. The years in which winter AOD lies above 30 % of the residual confidence interval of the linear trend line are referred to as the high-AOD winters (including 2001, 2003, 2007, 2008, 2013) and those below 30 % of the residual confidence interval as the low-AOD winters (including 2002, 2004, 2006, 2009, 2010, 2012). Since the high- or low-AOD is defined relative to the trend line, the corresponding high- or low-AOD winters are expected to be driven by the interannual variability of meteorology.

Mean meteorological conditions between the high- and low-AOD winters were compiled and compared to identify any significant differences in large-scale circulation patterns between them. The differences in winter-mean SLP and 850 hPa wind fields are shown in Fig. 4 (high-AOD winters minus low-AOD winters). Surprisingly, Fig. 4 does not reveal any significant decrease of SLP from low-AOD to high-AOD winters over Mongolia where the climatological center of the Siberian High is located (cf. Fig. 1a). Instead, significant changes of SLP are located over west of Mongolia (negative differences) and over Japan (positive differences). The high-AOD winters also have a stronger component of southeasterly winds at 850 hPa over North China. This change of wind directions not only suppresses the northwesterly flow that brings cleaner continental background air, but also reduces the transport of pollution from NC to SC, both of which lead to higher pollution levels over NC.

The index widely used in the literature to describe the SH variability is the Siberian High intensity (SHI), defined as the mean SLP over northern Mongolia between 80–120∘ E and 40–65∘ N (black rectangle in Fig. 1a and 4) (Jeong et al., 2011; Hasanean et al., 2013). However, as shown by Fig. 4, there is no significant difference in SLP over northern Mongolia between the high- and low-AOD winters, suggesting that this conventional index of SH may not be able to explain the interannual variability of PM over North China. As an example, Fig. 5 compares winter SLP and 850 hPa wind fields between 2003 (a high-AOD winter) and 2004 (a low-AOD winter). While winter-mean AOD over NC was significantly higher in 2003 (0.68) than that in 2004 (0.45), the SHI was almost the same between the two winters. The noticeable difference, however, is that the high-pressure isobars in the 2003 winter extended further east over the continent than those in the 2004 winter. Through linear regression, we found a poor correlation between SHI and detrended winter-mean AOD over NC (Fig. 6a), with SHI explaining only 4 % of the AOD variance. There is no significant (p < 0.05) trend in SHI during 2001–2013.

Figure 4 manifests the displacement of the high SLP center during the high-AOD winters from northern Mongolia where the conventional SHI is defined. Figure 5 further illustrates that the main difference in SH between the two specific winters of largely varying AODs lies in its spatial extension. Given this feature, we further hypothesized that the position of the Siberian High is a more important factor than its intensity in terms of affecting PM concentrations over NC. We thus proposed a Siberian High position index (SHPI) as the weighted mean of the longitudes of all the grids within the 1023 hPa isobar over the broad region of 60–145∘ E and 30–65∘ N (black rectangle in Fig. 5). The SHPI is defined by Eq. (1): SHPI=∑(Pi×Li)∑Pi where Li is the longitude of any eligible grid i within the 1023 hPa isobar and the definition domain, and Pi is the SLP of the corresponding grid i. The unit of SHPI is degree in longitude. Our definition of SHPI is similar to the longitude index of SH defined by Hou et al. (2008), but differs with regards to the region over which SHPI is calculated. They defined the index as the weighted mean longitudes of all the grids within the 1023 hPa isobar which may extend westward to Europe and northward to the Arctic. Our definition of SHPI limits the spatial domain over which the 1023 hPa isobar is considered in the SHPI calculation because of our focus on East Asia and particularly China (Fig. 5). The 2001–2013 time series of winter SHPI is displayed in Fig. 6b (black line) and the wintertime mean SHPI during this period is 98.9∘ E. A larger SHPI indicates that the center of the Siberian High is located further east of its normal position. Referring back to Fig. 5, the 2003 winter has a significantly higher value of SHPI (102.3∘ E) than that of 2004 (SHPI = 96.3∘ E); so does the AOD over NC but not SHI (cf. Fig. 6a).

Figure 6b shows the time series of winter-mean SHPI and NC AOD from 2001 to 2013. They exhibit a positive correlation of 0.39, which is not significant due to the confounding effect of the increasing trend in AOD. Since the focus here is on variability, the AOD time series were detrended by removing any significant linear trend (detrended AOD) and the SHPI time series were normalized by their climatological mean and standard deviation. As shown in Fig. 6c, the detrended NC AOD and normalized SHPI display a strong correlation of 0.76 (p < 0.01), which means that the position-based SHPI captures 58 % of the interannual variance in winter AOD over NC. This indicates that on the interannual scale, winter AOD over NC can be better explained by SHPI, an index of the SH position, than the conventional SHI, an index of the SH intensity. According to Hou et al. (2008), the longitude index and intensity index of the SH may not be significantly correlated. In support of this point, we found the SHI and SHPI have a weak correlation of only -0.32 during the study period (Fig. S1 in the Supplement).

Figure 6d displays the time series of normalized SHPI and detrended NC AOD on the monthly scale. The corresponding raw data prior to the detrending and normalization are provided in Fig. S2. Here the normalization of SHPI is conducted separately for November, December, and January to retain its intraseasonal variability. At the monthly scale, the correlation between normalized SHPI and detrended NC AOD is also significant at 0.45 (p < 0.01). Some extremely high values of monthly AOD over NC show clear associations with higher values of SHPI. Taking January 2013 as an example, which has the highest AOD over NC among all the 39 winter months studied here, the SHPI of that month is also the highest (106.5∘ E), lying 2.6 times the standard deviation away from the 2001-2013 January mean (99.8∘ E). This association indicates that the anomalous feature of the Siberian High in January 2013 was not only the weakening of its strength (cf. Fig. 1b) but also its more eastward extension, the latter being the primary factor contributing to high PM levels over NC. Another example is February 2011. Both AOD and SHPI of that month are among the highest values of the study period (Figs. 6d and S2). We thus conclude that the SHPI indicator of the SH variability is able to explain extremely high PM pollution over NC on the monthly scale.

To understand the mechanistic connection between SHPI and winter AOD over NC, we examine in this section how the SHPI variability is linked with the change of large-scale circulation patterns using the NCEP reanalysis data which span 30 years (1982–2011). The years with extremely high SHPI (beyond one standard deviation of the mean) in winter are defined to be high-SHPI years and those below one standard deviation of the mean as low-SHPI years. Figure 7a displays the climatological distribution of 850 hPa wind fields during 1982–2011. The northwesterly winds larger than 5 m s-1 over North China and Japan indicate the strong influence of the Siberian High and East Asian winter monsoon. The area covered by the prevailing northwesterly winds and the mean speed of those winds exhibit interannual variability that correlates with SHPI to some extent. For example, the winter of 1990 has the highest SHPI (105.9∘ E) during the 30-year study period and that of 2004 has the lowest SHPI (96.3∘ E). As shown in Fig. S3, the area covered by northwesterly winds larger than 5 m s-1 is smaller in 1990 than that in 2004, and the average wind speed over that area is also smaller in 1990. On average, 850 hPa wind speeds over NC are about 0.5 to 1 m s-1 lower during the high-SHPI winters than during the low-SHPI years (Fig. 7b). Table 1 summarizes wintertime-mean zonal and meridional wind speeds over NC at different vertical levels for the 30-year average, high-SHPI average, and low-SHPI average. In the high-SHPI winters, both zonal and meridional wind speeds are lower not only at 850 hPa but also at the upper levels. Lower wind speeds are conducive for pollution accumulation over the source region, which partly contributes to higher AOD in the high-SHPI winters. To further illustrate the connections between SHPI and wind changes, Figure 7c depicts the spatial distribution of correlation coefficients between SHPI and surface RH from 1982 to 2011. SHPI shows a significant positive correlation with RH over NC, indicating enhanced water vapor convergence over NC in the high-SHPI winters. This positive correlation arises because weaker northerly winds lead to reduced transport of dry air masses from the cold Siberian landmass, compensated by enhanced transport of moist air masses through the anomalous southerly winds. Higher RH during the high-SHPI winters leads to higher mass concentrations and extinction of aerosols as a result of hygroscopic growth of aerosol species (Mu et al., 2014; Tai et al., 2010). Although higher SHPI is always associated with lower northwesterly wind and higher RH over NC, local wind speed or RH itself is not as good an indicator as SHPI in explaining the interannual variation of NC AOD. One explanation is that SHPI represents the combined effects of large-scale circulation change on local meteorological conditions. In addition, systematic errors have been found for lower-level wind fields from NCEP reanalysis (Shi et al., 2006).

To verify the above analysis of the mechanism, we tested the utility of SLP over Japan (SLPJ, defined over 130–145∘ E and 40–50∘ N) as an alternative indicator of the large-scale circulation in explaining the interannual variations of AOD over NC. The reason why the SLPJ is used for comparison is because the high-AOD winters also feature significant positive changes of SLP over Japan (cf. Fig. 4). The time series of SLPJ is shown in Fig. S4. SLPJ shows a positive correlation with NC AOD and explains 38 % of the variance in detrended NC AOD (Fig. S4a). By comparison, SHPI explains 58 % of the variance of detrended NC AOD. SLPJ also correlates well with SHPI (Fig. S4b), which indicates that in the high-SHPI years the eastward extension of the SH leads to an increase of SLP over Japan and as a result SLPJ is not independent from SHPI. The anomalously high SLP over Japan influences the PM level over NC by reducing the prevailing northwesterly winds and increasing RH over NC, which is consistent with the mechanism provided above.

To summarize, the SHPI indicator developed here is able to capture the interannual variations of winter-mean and monthly-mean NC AOD to a large extent. Comparing to the climatology, 850 hPa wind speeds over NC during the high-SHPI years are suppressed by 13 % and the surface relative humidity is enhanced by 12 % as a result of the eastward extension of the SH. Since the suppressed wind speed is unfavorable for the dispersion of air pollution and higher surface relative humidity enhances secondary aerosol formation and hygroscopic growth, both factors lead to higher PM levels over NC in the high-SHPI years.

Our above analysis suggests that the suppression of prevailing northwesterly winds and the enhancement of surface RH are the key meteorological features during the high-SHPI winters. The implication of such conditions for wintertime PM over SC, the domestic receptor region of wintertime NC outflow, is not straightforward. On one hand, suppressed northwesterly winds are unfavorable meteorological conditions for the export of pollution from NC, which may lead to reduced PM levels over SC. On the other hand, the Siberian High variability is expected to have an influence on local meteorological conditions over SC. In this section, we examine the extent to which the SHPI indicator developed in the previous section can explain the interannual variability of AOD over SC.

Figure 8 displays the time series of winter mean AOD over SC from MODIS. The multi-year mean AOD over SC is about 0.4, with a positive but not significant trend of increase of 0.13 % year-1. The two highest AOD winters for SC are 2004 (0.46) and 2008 (0.48), both corresponding to the lowest SHPI. The overall correlation between SC AOD and normalized SHPI is -0.82, suggesting that SHPI explains 67 % of the variance in SC AOD. In the high-SHPI winters, the meridional wind speed over NC is reduced by 17, 16 and 19 % at 850, 700, and 500 hPa, respectively, compared to the low-SHPI winters (Table 1). The suppressed northerly winds over NC lead to the direct effect of reduced southward transport of pollution from NC to SC, resulting in lower AOD over SC during the high-SHPI winters. Meanwhile, the 850 hPa wind speeds over SC do not show a significant difference between the high-SHPI and low-SHPI winters (Fig. 7b). Although there is a 7.5 % enhancement of surface relative humidity over SC during the high-SHPI years (Fig. 7c), the overall significantly negative correlation between SC AOD and SHPI indicate that the suppressed pollution transport from NC to SC is the dominant factor to explain the influence of SHPI on AOD over SC.