Large‐scale meteorological control on the spatial pattern of wintertime PM2.5 pollution over China

The frequent episodes of severe air pollution over China during recent years have posed serious health threats to densely populated eastern China. Although several studies investigated the linkage between enhanced severity and frequency of air pollution and the large‐scale weather patterns over China, the day‐to‐day covariability between them, as well as its local and remote mechanisms, has not been systematically documented. The wintertime synoptic covariability between PM2.5 and large‐scale meteorological fields is studied using surface observations of PM2.5 in 2013/2014–2016/2017 and ERA‐Interim meteorological fields through maximum covariance analysis (MCA). The first MCA mode (MCA1) suggests a consistent accumulation of ambient PM2.5 as a result of weakened winds that block the pollutant removal passage in heavily polluted areas of eastern China, as well as moist air from southeast coast favoring haze formation. A northeast–southwest belt that extends into northeastern China and central China on each end is more sensitive to MCA1. The second MCA mode (MCA2) shows a north–south dipole in PM2.5 linked to the contrast of boundary layer height and surface wind speed between northern and southern regions of China. Spatial patterns of both modes are supported by the GEOS‐Chem chemistry transport model with realistic emission inventory. The spatial patterns of the two modes are robust on the interannual time scales. Based on that, we investigate the variability of the first two modes of the identified modes on the multidecadal scale by projecting GPM_500 pattern to 1981–2010. Correlation analysis of the projected time series and climate indices over 30 years indicates the possible linkage of Arctic oscillation, ENSO indices, Pacific decadal oscillation and east Atlantic/western Russia to regional air pollution patterns over China.


| INTRODUCTION
Severe air pollution episodes struck China quite frequently during recent years, posing serious threats to public health (Cohen et al., 2005;Silva et al., 2013;Geng et al., 2015;Wang et al., 2017). The key air pollutant of concern during the haze period is particulate matter (PM). Recent work showed that stagnant weather conditions and its persistency, besides other factors such as high emissions, secondary formation of PM and so forth, contributed substantially to the accumulation of fine PM (e.g., PM 2.5 with a diameter of 2.5 μm or less), deteriorating the visibility and thus affecting transportation and economic activity (Tao et al., 2014;Sun et al., 2017). These stagnant weather conditions tend to become more persistent under climate change, supported by multimodel ensembles (Horton et al., 2014).
With respect to the relationship between haze episodes and atmospheric circulation, previous studies focused on sensitivity of air pollution to regional scale meteorology (He et al., 2017;Megaritis et al., 2014). Zhao et al. (2013) also examined the vertical profile of particles, temperature and winds during a haze episode over the North China Plain on a local scale, demonstrating that heavy pollution might be caused by temperature inversion and weakened transport (Li et al., 2015). Most recently, Wang et al. (2018) suggested that 10-m wind speed as an index of horizontal dispersion and boundary layer height as an indicator of the strength of vertical mixing should be considered as the threshold of air stagnation conditions, indicating that these two variables should be examined when defining stagnant weather conditions.
From the perspective of large-scale meteorology, Zhang et al. (2016) showed that increased frequency of air pollution over Beijing and its neighboring region is associated with the shallowed east Asia trough and weak Siberian high. Leung et al. (2018) showed dominant synoptic weather parameters differ from one subregion to another, based on day-to-day variability. Some recent studies show that Arctic sea ice loss in the preceding autumn, which favors a reversed northeastsouthwest pressure gradient that weakens the east Asian winter monsoon (EAWM) (Wang and Chen, 2016;Zou et al., 2017), might also contribute to intensified pollution episodes, although a large fraction of the observed trend can be attributed to atmospheric internal variability (Xue et al. 2017).
Although several studies have shed light on the linkage between enhanced severity and frequency of air pollution and the large-scale weather patterns (Chen et al., 2008;Niu et al., 2010), the day-to-day covariability between air pollution and large-scale meteorology fields over the entire China, as well as its local and remote mechanisms, have not been systematically documented. In this study, we apply maximum covariance analysis (MCA) to the covariance matrix of PM 2.5 and meteorological factors, and obtain leading modes of variability for the wintertime PM 2.5 and investigate the long-term large-scale meteorological controls over entire China. The observed patterns are further supported by the simulations from the GEOS-Chem chemistry transport model (CTM) with realistic emission inventory. We extend the time series of the MCA modes to the period of 1981-2010, and thus get the projected variability on the decadal time scale to further explore whether past climate variability plays a role in the stagnant weather conditions and hence the air pollution over China.

| DATA AND METHODOLOGY
We use the MCA to identify the leading modes of covariability between high concentration of PM 2.5 in China and stagnant meteorological conditions. We focus on the winter months when pollutants are most severe, and the analysis is conducted for DJF (December-February) from 2013/2014 to 2016/2017. Because the MCA method only requires a common temporal dimension of two fields, the meteorological fields we analyzed cover a larger domain, including entire east Asia, India and also southern part of , in order to capture the large-scale teleconnection pattern influencing air pollution in China. Details of MCA could be found in Appendix S1.
The PM 2.5 observational data are obtained from China National Environmental Monitoring Center (http://www. cnemc.cn/). We applied quality control and interpolation and get a subset of 364 stations with continuous record during winters from December 2013 to February 2017. Details could be found in Appendix S2.
Meteorological data are obtained from ERA-Interim (ERA-I) reanalysis data, available from https://www.ecmwf. int/. It has a spatial resolution of 0.75 × 0.75 , with a temporal resolution of 6 hr. The set of data includes sea level pressure (SLP), geopotential, zonal and meridional winds at 10 m, 850 hPa and 500 hPa. Twelve-hourly boundary layer height data is obtained from ERA-I. PM 2.5 observation at Beijing time has been converted to UTC to match that of meteorological data. Both datasets have been averaged to daily mean before applying MCA, and 21-day running mean is removed. Daily mean records are used here because it is the time scale that captures synoptic weather events interaction with air pollution.

| RESULTS AND DISCUSSION
3.1 | Spatial patterns of PM 2.5 in China and large-scale meteorological control In this section, we show MCA results using observational data (station record for PM 2.5 , ERA-Interim reanalysis for meteorology fields) over winters from 2013/2014 to 2016/2017. With a focus on large-scale synoptic meteorology, we select SLP and geopotential at 850 and 500 hPa (GPM_850 and GPM_500). We perform MCA for PM 2.5 with SLP, GPM_850 and GPM_500, separately. The PM 2.5 singular vectors of the three MCA analyses are nearly identical (i.e., correlation coefficient greater than 0.98) as well as their PM 2.5 time series, so we regressed MCA patterns of GPM_850 and GPM_500 based on that of PM 2.5 and SLP. Figure 1 shows the mean state and the leading spatial covariance patterns of PM 2.5 and large-scale meteorology. As shown in the first column of Figure 1, PM 2.5 mean concentration over four winters are relatively higher over northern part of China (to the north of 34 N). Relatively high SLP occurs over eastern China, Mongolia and southern Siberia. East Asian trough can be easily identified from GPM_500. Winter monsoon over east Asia is characterized by the geostrophic wind flowing through the trough, which is believed to be vital in transporting the pollutants toward tropics (Liu et al., 2003). The first two modes combined explain 40-50% of the total squared covariance, and is assured that uncertainty does not compromise the particularity of each mode ( Figure S1). Monte Carlo method is applied to test the significance of spatial regression patterns, please see Appendix S4 for details. In MCA1 (second column of Figure 1), the leading meteorological pattern resembles that reported by Cai et al. (2017) based on severe haze episodes in Beijing. Through the covariance of PM 2.5 and meteorological fields, it is found that the leading PM 2.5 patterns reflect a consistent accumulation of pollutant over the entire domain of China.
The anomalous values reflect the sensitivity of ambient PM 2.5 concentration to meteorological parameters. More specifically, an area of higher concentration over Northern China Plain leads to a northeast-southwest belt that extends into northeastern China and central China on each end. At sea level, we see negative SLP anomaly over the highpressure center, which may have led to weaker horizontal transport and comparatively weaker northerlies at the surface, resulting in the accumulation of pollutants over China. If we consider all three altitudes from surface to 500 hPa, it is obvious that there is a baroclinic structure of circulation anomaly tilting toward west, which is typical for synoptic weather systems. Such a structure contributed to the accumulation of air pollution by inducing a geostrophic southerly wind at 850 hPa near east coast and easterly wind with a southerly component near Yangtze River Delta at 500 hPa, by weakening northerlies and bringing warm and moist air to northern China (NC). In conclusion, at all altitudes, the winds induced by large-scale meteorological patterns weaken the climatological trough as shown in the mean state in Figure 1, which, in turn, blocks the transport of pollution through the passage. Furthermore, the southerly wind from coast brings humid air to polluted areas in NC, providing moisture for the formation of fog-haze events .
In MCA2 (third column of Figure 1), a north-south dipole pattern exists for spatial distribution of PM 2.5 , with positive values over the northern region (34 N-55 N, 70 E-140 E) of China and negative values over the southern region (15 N-34 N, 70 E-140 E). Corresponding meteorological fields (SLP and geopotential) all present a pattern with positive anomaly to the north of Japan and Siberia and a negative anomaly over central and western China. The contrast of PM 2.5 over northern and southern region of China can be explained by the anomalous northerly winds around 45 N at 850 hPa, indicating stronger intrusion of cold air from Siberia is able to reach NC. In particular, the region of cold air advection during the development of a typical mid-latitude cyclone is associated with anomalous sinking motion, thus creating stable conditions at boundary layer in northern region. Similar pattern has been proposed by Cheng et al. (2016), finding "north-higher south-lower" dipole pattern through composite analysis of stronger versus weaker winter monsoon years, which can be attributed to the difference of dry deposition.
We regressed surface wind speed and planetary boundary layer height back to the PM 2.5 time series, to see how horizontal and vertical dispersions contributed to the spatial distribution of PM 2.5 . The regression patterns outside China are masked out to highlight the dispersion conditions over China. As shown in Figure S2, positive PM 2.5 values in northern region of China as shown in MCA1 are associated with anomalously low wind speed and boundary height, indicating stagnant weather conditions conducive to air pollution. Liu et al. (2013) also pointed out that weakened high-pressure center, which is accompanied by large-scale atmospheric movement, leads to subsiding airflow on local scale and suppressed PBL development. For MCA2, the southern region is dominated by relatively stronger surface wind and higher boundary layer height compared to northern region, partly explaining the relatively low values of PM 2.5 .
In order to examine whether the patterns we found represent the actual pollution concentrations, domain averaged time series for northern and southern regions are calculated for comparison (black curves in Figure 2a,b). Gray stripes highlight the days when PM 2.5 concentration exceeds 150 μg/m 3 , which is the standard of moderate pollution according to the regulation of air quality issued by Chinese government. The PM 2.5 time series of MCA1 (red curve in Figure 2a) captures the variation of both regions of China well, with correlation coefficients of 0.82 and 0.46 for northern and southern regions, respectively. For MCA2, the PM 2.5 time series is strongly correlated with PM 2.5 averaged over the northern region with a coefficient of 0.65, and negatively correlated with PM 2.5 over the southern region by −0.22, corresponding to the north-south contrast of the PM 2.5 singular vector pattern. This indicates the two MCA modes can work constructively to severe pollution episodes in NC and simultaneously destructively to pollution in southern China (SC).
For both modes, the time series of the three meteorological fields are well correlated (Figure 2c,d). This corroborates that the spatial structures of atmospheric circulation identified in Figure 1 are vertically coherent. Similar spatial structures can be recovered if the MCA is applied to PM 2.5 and a combined matrix of three meteorological fields.

| Model simulation
To confirm that the covariance patterns identified from observations are robust, we conducted model simulations using GEOS-Chem CTM v9-02, a regional CTM simulation using comprehensive chemistry processes and realistic emission inventory. Details of model setup could be found in Appendix S3.
In the analysis of the modeling results, we masked out simulated PM 2.5 concentrations for the grids outside mainland China, and used SLP, meridional and zonal winds at 10 m from the GEOS-5 reanalysis data that drive the model for the meteorological analysis. For grids above 90% confidence level, PM 2.5 grids are marked with "+", and only wind vectors over those grids are shown.
The mean state and MCA patterns of model simulation results match well with those obtained from observations ( Figure 3). The PM 2.5 singular vector pattern of MCA1 retains positive and is associated with anomalous southerly wind, leading to accumulation of pollutants over the entire domain, especially over Northern China Plain. The PM 2.5 spatial pattern of MCA2 presents a dipole over the northern and southern regions, although the positive PM 2.5 values are confined in a narrower belt compared to observational results. Besides the similarity in spatial distributions, the amplitudes of both PM 2.5 and meteorological singular vector patterns are comparable to observational results shown in Figure 1.
It is noteworthy that the MCA patterns are similar even though only two winters of model simulations are analyzed. We have also divided observational dataset into first and latter 2-year periods, and performed the MCA for each group separately. We find that their spatial patterns ( Figure S3) are relatively independent of the period examined. The weak dependence on the year of interest is consistent with the notion that the MCA modes represent the daily interactions between synoptic weather (e.g., east Asian trough) and pollution episodes. This gives us a sound basis to investigate the variability of patterns identified over a longer time scale, and to see if the stagnant patterns are potentially related to climate variabilities. southern China (SC). In the top subplots are domain average time series of NC and SC in the first and third row, respectively, with the second row the normalized PM 2.5 time series for the covariability of PM 2.5 and SLP. Numbers in parentheses are correlation coefficients, with the former/latter one for NC/SC averaged time series. Gray stripes denote the days of which PM 2.5 concentration are over 150 μg/m 3 , of which the darkest ones highlights the days for both NC and SC, the medium dark ones for NC only, and the lightest ones for SC only. In the bottom subplots are MCA time series of SLP, GPM_850 and GPM_500, respectively, with numbers in parentheses for correlation coefficients with SLP time series

| Possible linkage between climate variability and stagnant weather conditions
We have also selected some climate indices to investigate whether the observed large-scale circulation patterns show significant correlation with any climate index influencing east Asia. Here we assume the MCA singular vectors of meteorology will hold beyond the 4 years we have examined due to its relatively weak dependence on the year examined, and thus we can extend the GPM_500 time series of each MCA mode by projecting the geopotential height over the period of 1980/1981-2009/2010 based on singular vectors. We choose Arctic oscillation (AO), ENSO indices (Nino1+2 and Nino3.4), Pacific decadal oscillation (PDO) and east Atlantic/western Russia (EA/WR) for our research, which are climate indices for mid-to high-latitudes of Northern Hemisphere (Table S1). Table 1 shows the correlation of the extended GPM_500 time series with different climate indices. EA/WR is strongly correlated with MCA1 (with a correlation coefficient of 0.53), which makes sense as it almost synchronizes with EAWM and Siberian high (Lim and Kim, 2016). AO has a statistically significant positive correlation of 0.33 with MCA2. When AO is at a positive phase, the westerly component of jet stream around the Arctic Circle grows stronger, which in turn obstructs the cold polar air from striking toward mid-latitudes, weakening the east Asian trough and helping build up the patterns of MCA2 (Thompson and Wallace, 1998). This is supported by the recent work of Martineau et al. (2017), pointing out that the positive phase of AO is associated with reduced frequency in persistent cyclonic wave events over east Asia, significantly lessening the chance of reaching northern parts. For ENSO and PDO, we found a statistically significant negative correlation of −0.24 of Nino1+2 and −0.32 of PDO with MCA2. ENSO and PDO interact together and modify EAWM patterns by an obvious zonal shift of wind path, and turns out to be closely linked to temperature dipole patterns over eastern China, resulting in the contrast of stability of boundary layer and thus pollution dissipation (Zhu and Yang, 2003;Chen et al., 2013a;2013b).

| Historical records of projection from 1981 to 2010
Based on previous results, we want to know whether intensity of preceding MCA patterns has substantially changed over several decades. We projected the MCA1 and MCA2 F I G U R E 3 As in Figure 1, but for the model simulation results using GEOS-Chem with realistic emission inventory. PM 2.5 pollutants outside mainland China is masked out for the MCA modes, considering the meteorological control over mainland China merely. PM 2.5 grids above 90% confidence interval are marked with "+", and only wind vectors over grids that pass 90% significance test are shown. Surface winds (at 10 m level) are plotted over shadings as vectors. Note the model domain is slightly smaller than the reanalysis domain used in Figure 1 T A B L E 1  (3) (projection, see Appendix S1), which reflects the intensity of such geopotential patterns. Linear regression for our pattern intensities are shown (Figure 4), and a modest increasing trend in projection is identified. The 30-year period is divided into the first period (1981)(1982)(1983)(1984)(1985)(1986)(1987)(1988)(1989)(1990)(1991)(1992)(1993)(1994)(1995) and second period (1996)(1997)(1998)(1999)(2000)(2001)(2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010), to see if the modes behave differently over the two periods. All of the linear regressions passed two-tailed 10% Student's t test. We find a stronger increase in MCA1 during the first period, at 0.09 per year. Jiang and Tian (2013) suggested that EAWM tend to be strengthened compared with 1980-1999 period for areas to the north of 25 N, consistent with the upward trend of MCA1 during the first period. This contrast in dominant increasing time period between MCA1 and MCA2 suggests the trough-shallowing southerlies' contribution to the stagnant weather condition over eastern China was strongly enhanced during the first period, and a contrast of relatively higher concentration in the northern region enhanced slowly during the second period.

| CONCLUSIONS
In this study, we examined the spatial patterns of wintertime PM 2.5 over China and determined what patterns of largescale meteorology field contributed to the accumulation (or removing) of PM 2.5 based on both observational data and model simulation results. Using station observational PM 2.5 data and reanalysis meteorological fields from winters of 2013/2014 to 2016/2017, we have identified two important modes using MCA method. We successfully reproduced large-scale meteorology patterns proposed by Cai et al. (2017) in MCA1, in which the pattern-induced winds shallows the trough, weakening the northwesterly wind and building up a stagnant weather condition over eastern China. Beyond that, we showed that the spatial pattern of PM 2.5 is coherent over much of China due to the large-scale structure of synoptic weather systems. In MCA2, PM 2.5 presents a north-south dipole pattern with enhanced/reduced PM 2.5 concentrations in NC/SC. The contrast is caused by relatively stronger surface wind and higher boundary layer height in the southern region. Model simulations using GEOS-Chem with realistic emission inventory verified the robustness of these patterns, reproducing both pattern and amplitude identified in previous section.
We projected MCA patterns to extend DJF GPM_500 time series with different climate indices in 1980/1981-2009/2010, and investigate how the patterns correlate with climate indices. The historical record of geopotential time series shows high positive correlation with EA/WR for MCA1 only, highlighting the synchronization with Siberian high and strength of EAWM, while correlation with AO at MCA2 denotes cold air intrusion to NC. A statistically significant negative correlation with Nino1+2 and PDO also exists for MCA2, linked to temperature dipole F I G U R E 4 Monthly historical projections of MCA1 and MCA2 and their trend during 1981-1995and 1996-2010. Solid black lines are temporal series of GPM_500 historical monthly record of MCA1 and MCA2 divided by standard deviation showing in panel (a) and (b), respectively. Dashed lines are correspondent linear regression results of 1981-1995 and 1996-2010, expressed in numerical form at the topright corner. The slope calculated from monthly projection is multiplied by 3 months and translated to increase (or decrease) per year over eastern China and thus boundary layer stability. The differences in correlation imply different regional fingerprints of these climate indices, and the mechanisms for these relationships had been documented and supported by various papers.
The MCA modes identified in this study may be important for future research and air quality prediction. They connect the PM 2.5 pollution pattern with corresponding large-scale meteorological fields, manifesting that anomalies in high pressure at surface level over East China Sea and those in low pressure at 500 hPa over west Siberia play a role in controlling the pollution concentration over eastern China. It is worth investigating how future climate change would affect the intensity of the patterns, and how the patterns would respond to different representative concentration pathways (RCPs).