We focus on pollution over Eastern China (25–41∘ N, 110–123∘ E). Guided by an EOF analysis, we contrast pollution over the southern (SEC, south of 35∘ N) and northern (NEC) parts to address the regional differences in day-to-day pollution variability. Such latitudinal separation coincides with the Huai River climate transitional zone (Ye and Li, 2017). The orange lines in Fig. 1 separate the two regions.

Our study period is from 25 October to 25 December 2013, with a total of 1488 h in 62 days. Most air pollution data are missing in January and February 2014 because of instrumental failure or data retrieval failure, and data before 25 October are not available.

We retrieve hourly measurements of NO2 and PM2.5 from 193 air quality monitoring stations of the MEP. Most stations are located in urban areas, and only six stations are suburban. As almost every station has missing values in more than one day, we exclude stations that have missing values at ≥30 % of the 1488 h or during a consecutive 72 h period. We thus select 163 stations for NO2 and 159 stations for PM2.5 in the same 42 cities. The dots in Fig. 1a and b depict the stations and cities, respectively. The blue dots show stations and cities with both valid NO2 and PM2.5, the green dots with NO2 only, and the purple dots with PM2.5 only. The slight difference between NO2 and PM2.5 stations does not affect our analysis of the regional pattern of pollutants.

We use 3-hourly measurements of 2 m air temperature, 2 m relative humidity, and 10 m wind speed from meteorological stations recorded at the National Oceanic and Atmospheric Administration National Centers for Environment Information (NOAA NCEI). We do not use surface pressure additionally because it is highly correlated with air temperature and relative humidity on the day-to-day scale. The locations of these stations do not always coincide with air pollution stations. Thus, we select 36 meteorological stations within 10 km of air pollution stations (red hollow dots in Fig. 1). Despite the difference (in number and location) between pollution and meteorological stations, an analysis of the regional temporal patterns of pollutants and meteorology is still informative (see Sect. 3.2).

To fill in missing values, we apply an interpolation process that accounts for diurnal variability using information for an adjacent day. For example, if the temperature on 26 October at 12:00 is missing, we calculate the temperature difference between 09:00 and 12:00 on the 25th as well as the difference between 15:00 and 12:00 on the 25th. We then use these differences to adjust the temperatures at 09:00 and 15:00 on the 26th, and finally use the mean of the two adjusted temperatures as the temperature on the 26th at 12:00.

For consistency with the hourly pollution data, we linearly interpolate the 3-hourly meteorological measurements to each hour. This interpolation does not distort the EOF–EEMD analysis, as confirmed by comparing the statistical analysis on 1-hourly GEOS-FP meteorological parameters versus an analysis on 3-hourly GEOS-FP data. Note that the GEOS-FP meteorology is used to drive GEOS-Chem.

We use the nested GEOS-Chem CTM version 9-02 (L. Zhang et al., 2016) to simulate NO2, PM2.5, and other pollutants over China in October–December 2013. The model resolution is a 0.3125∘ long. × 0.25∘ lat. grid with 47 vertical layers, and the lowest 10 layers are of ∼130 m thickness each. The model is driven by the GEOS-FP assimilated meteorology from the National Aeronautics and Space Administration (NASA) Global Modeling and Assimilation Office, with the full Ox–NOx–VOC–CO–HOx gaseous chemistry (Mao et al., 2013) and online aerosol calculations. Vertical mixing in the PBL adopts a nonlocal scheme (Holtslag and Boville, 1993; Lin et al., 2010). Model convection is simulated with the relaxed Arakawa–Schubert scheme (Rienecker et al., 2008).

Chinese anthropogenic emissions of NOx and other pollutants adopt the monthly MEIC inventory with a base year of 2010 (http://www.meicmodel.org, last access: 1 December 2015) (Geng et al., 2017). We further use the monthly DOMINO v2 NO2 data to scale monthly anthropogenic NOx emissions from 2010 to the simulation year (Lin et al., 2015). The emission scaling improves the simulation of NO2 (Cui et al., 2016). Other model setups are referred to Lin et al. (2015) and Yan et al. (2016).

GEOS-Chem modeled PM2.5 includes secondary inorganic aerosols (sulfate, nitrate, and ammonium), black carbon, primary organic carbon, natural dust, and sea salt. Secondary organic aerosols are not included in this study, considering the severe underestimate in China due to missing precursor emissions and formation pathways (Fu et al., 2012; L. Zhang et al., 2016). Anthropogenic dust is also not included.

The nested model simulation is from 15 October to 25 December in 2013, allowing for a 10-day spin-up period. Its lateral boundary conditions of chemicals are updated every 3 h by results from a corresponding global simulation on a 2.5∘ long. × 2∘ lat. grid. Modeled NO2 and PM2.5 in the first layer are sampled at city centers and times with valid observations, unless stated otherwise.

As shown in Fig. 3, our EOF–EEMD analysis visualization package consists, in order, of an EOF analysis (Lorenz, 1956), an EEMD analysis (Wu et al., 2009), a Hilbert transform (HT) with marginal spectrum analysis (MSA), and a visualization step to quantitatively depict the spatial–temporal scales of measurement or model data.

The basic purpose of our package is to quickly and simultaneously identify and visualize various spatial and temporal scales of interest in the observation or model datasets. As shown by Feng et al. (2014) and Wu et al. (2016), combining EOF with EEMD to decompose the datasets leads to a faster calculation than MEEMD by 1 or 2 orders of magnitude because here the EEMD is applied to the temporal components (i.e., PCs) out of an EOF analysis rather than to all dimensions. Also, our EOF–EEMD package conducts additional HT–MSA and provides visualization of all spatial and temporal scales of interest.

EOF analysis to decompose a two-dimensional dataset (time series at multiple locations) into spatial and temporal components.

Suppose there are n locations, each having a time series of length p. The associated dataset Z is an n×p matrix. An EOF analysis of Z givesZ=U∑WT.Here Σ is a diagonal q×q matrix containing the first q singular values of Z, and it represents the contribution of each pattern to the total variance of Z. The diagonal values of Σ are in a descending order, and thus the first several modes are the dominant ones. U is an n×q matrix representing the spatial component, and each column of U represents a spatial mode. W is a p×q matrix representing the temporal component, and each column of W represents a principal component (PC) for temporal variation associated with the corresponding spatial mode.

The colored dots in Fig. 5a and b show the observed spatial distributions of city mean NO2 and PM2.5 averaged over the time period. Both NO2 and PM2.5 are largest over Beijing–Tianjin–Hebei (BTH) in the north and the Yangtze River Delta (YRD) in the east. NO2 concentrations exceed 60 µg m-3 at many sites. The range of PM2.5 is larger, from below 10 µg m-3 in some northern and coastal cities to about 200 µg m-3 in several cities of BTH.

Figure 6a and b show the diurnal variations of NO2 and PM2.5 over Eastern China, NEC, and SEC averaged over all days. Similarly, over the three regions, NO2 peaks around 19:00 due to evening rush hour emissions, reduced PBL mixing, and a lengthened lifetime. NO2 reaches a minimum at 14:00 because of the shortest lifetime and strongest PBL mixing. The diurnal range (maximum minus minimum) is about 30 µg m-3. The PM2.5 level also reaches a minimum in the early afternoon. It has a much smaller diurnal range at 10 µg m-3. The vertical error bars in Fig. 6a and b depict the standard deviation for the day-to-day variation of NO2 and PM2.5 at any given hour. At a given hour, the PM2.5 level is much more variable across the days than NO2. In particular, the day-to-day standard deviation for PM2.5 at a given hour is as large as the diurnal range of PM2.5.

Figure 6c and d further show the time series of daily mean NO2 and PM2.5. All data are de-trended (trends are at 0.01 µg m-3 h-1 for NO2 and 0.05 µg m-3 h-1 for PM2.5). Although local maxima and minima (peaks and troughs of the time series) occur every several days, there is no single period or amplitude for the variation of each species. For NO2 over Eastern China (black line in Fig. 6c), the local maxima vary from 60 to 100 µg m-3, and the local minima vary from 20 to 40 µg m-3. For PM2.5 over Eastern China (black line in Fig. 6d), the local maxima vary from 100 to 300 µg m-3, and the local minima vary from 20 to 120 µg m-3. Furthermore, comparing the green and blue lines reveals that pollutants over NEC and SEC synchronize on some days but are out of phase on others; this feature is quantitatively analyzed in Sect. 3.2. These day-to-day variation patterns are associated with meteorological conditions and pollutant lifetimes.

Figure 7 shows day-to-day anomalies of observed pollutant concentrations and meteorological parameters over NEC and SEC. All data are de-trended. Over NEC, wind speed is clearly anticorrelated with pollutant levels. The correlation coefficient reaches -0.73 between NO2 and wind speed and -0.60 between PM2.5 and wind speed. Over this region, stronger winds are often associated with lower RH and lower temperature, characteristic of a cold air passage that brings cleaner, colder, and drier air from the north to NEC and transports the NEC pollution out of the region. Correspondingly, RH is strongly positively correlated with NO2 (R=0.62) and PM2.5 (R=0.69). The meteorology-associated day-to-day variability is more apparent after mid-November, when the variations of the two pollutants are more synchronous.

Over SEC (Fig. 7), the relationship between pollutant levels and meteorological parameters is more complex. The correlation between daily mean PM2.5 and wind speed is relatively weak (R=-0.44 compared to -0.60 over NEC), and its correlation with RH is even weaker (R=0.29). This indicates that the northerly air does not reduce PM2.5 levels over SEC as effectively as over NEC, as PM2.5 from NEC may be transported to SEC. By comparison, NO2 is still highly anticorrelated with wind speed (R=-0.77) over SEC, likely a result of the short lifetime of NO2. Compared to PM2.5 whose lifetime is sufficiently long (several days) for transport from NEC to SEC (Hu et al., 2014), NO2 has a much shorter lifetime (below 1 day; Lin et al., 2012) and cannot undergo effective long-distance transport. However, almost all pollution measurement sites are urban, and weaker (stronger) winds allow for rapid accumulation (removal) of urban NO2 pollution.

Although informative, the time series analyses of regional mean pollution in Sect. 3.1 do not provide adequate quantitative information on the spatiotemporal variability and embedded scales. In fact, the separate discussion on NEC and SEC in Sect. 3.1 is largely inspired by the following EOF–EEMD analysis that suggests distinctive features between these two subregions. In this section, we use the EOF–EEMD package to distinguish and visualize the quantitative contributions of individual spatial and temporal modes to variations in the pollutant and meteorological data.

The columns in Fig. 8 show the EOF–EEMD results for the observed temperature, RH, wind speed, NO2, and PM2.5. The first two rows show the first two spatial patterns (EOF1 and EOF2) from the EOF analysis. The third row visualizes the EEMD–HT–MSA results for PC1 and PC2, the temporal counterparts of EOF1 and EOF2. For all variables, the first two PCs contribute more than 50 % of the total variance of the original data. The following PCs (PC3, PC4…) contain small variances and are not discussed here.

The color contours in Fig. 5a–d show the horizontal distributions of NO2 and PM2.5 simulated by GEOS-Chem and CMAQ. The model results here are averaged from all days over the time period rather than sampled from days with valid observations. Both models capture the general spatial patterns of observed NO2 and PM2.5, with the heaviest pollution over the north and east.

Figure 9 evaluates the regional mean diurnal and day-to-day variations of modeled pollutant levels over NEC and SEC. Here model data are sampled from days and locations with valid observations. All trends are negligible and have been removed, consistent with the observational analysis. GEOS-Chem underestimates the observations by about 17 µg m-3 (21 µg m-3 over NEC and 13 µg m-3over SEC) for NO2 and by 35 µg m-3 over Eastern China (31 µg m-3 over NEC and 41 µg m-3 over SEC) for PM2.5 averaged over the whole period. The model bias is relatively consistent across individual hours. GEOS-Chem captures the observed diurnal variability for both pollutants as well as the day-to-day variability of PM2.5, although it greatly underestimates the day-to-day variability of NO2. More model evaluation statistics are shown in Table 3.

Figure 9 also shows that WRF/CMAQ overestimates the nighttime observations by about 30 µg m-3 for NO2 and 60 µg m-3 for PM2.5 averaged over Eastern China, although it reproduces the daytime pollutant levels. This means an overestimate of the diurnal range, as is also revealed by the EOF–EEMD analysis in Sect. 4.2. CMAQ captures the day-to-day variability of daily mean NO2 and PM2.5 much better than GEOS-Chem (R=0.63–0.84 versus 0.25–0.37 over NEC and SEC for NO2; and 0.87–0.88 versus 0.55–0.75 for PM2.5). Note that the correlations shown here mainly reflect the model capabilities to capture Eastern China synchronous day-to-day variation, and they do not imply the model performance in simulating the NEC–SEC contrast, which is revealed in Sect. 4.2. More model evaluation statistics are shown in Table 3.

Figures 10 and 11 evaluate the EOF–EEMD results for modeled NO2 and PM2.5, respectively. Prior to the EOF–EEMD analysis, modeled NO2 and PM2.5 were sampled at times and locations with valid observations and then underwent the same interpolation procedure to fill in the missing values. In these figures, the last three rows visualize the EOF–EEMD–HT–MSA results in different ways (manifested in different lengths of the horizontal bar for each IMF). In the third row, the variance of each IMF is normalized to the total variance of the original data (NO2 or PM2.5). In the fourth row, the variance of each IMF is normalized to the variance of its respective PC in order to better visualize the signals from PC2 (which has a much smaller variance than PC1); as such, only the IMFs from the same PC are intercomparable. The fifth row visualizes the absolute variance of each IMF without any normalization.

The first two rows in Fig. 10 show EOF1 and EOF2 of NO2. Both GEOS-Chem and CMAQ exhibit a synchronous pattern (EOF1) and an NEC–SEC contrasting pattern (EOF2), consistent with the observation. However, the CMAQ-simulated NEC–SEC contrast in EOF2 is much weaker than the observed. Table 1 shows that for modeled NO2, PC1 is highly correlated with Eastern China mean NO2 for diurnal (R=1.0 for GEOS-Chem and CMAQ) and day-to-day (R=0.56–0.97) variability and that PC2 is correlated with NEC NO2 (R=0.74–0.81) and anticorrelated with SEC NO2 (R=-0.47 to -0.32) in terms of day-to-day variability, in line with the observational analysis.

The last three rows in Fig. 10 show that both models underestimate the contribution of day-to-day variability to the total variance of NO2 (with a shorter length of horizontal bar). For PC1, CMAQ captures the RPR (position of “error bar”) and variance (length of horizontal bar) of the observed IMFs fairly well. By comparison, GEOS-Chem underestimates the day-to-day variance (too-small horizontal length) and does not capture its RPR. These results are consistent with the analysis in Sect. 4.1 (Fig. 9) showing that CMAQ is correlated with the observed Eastern China synchronous NO2 time series much better than GEOS-Chem. For PC2, which reflects the NEC–SEC contrasting pattern, GEOS-Chem outperforms CMAQ in capturing the RPR and variance of the observed day-to-day IMFs (red colored in fourth row). This model characteristic is not seen from the time series discussion in Sect. 4.1.

Figure 11 shows that both GEOS-Chem and CMAQ capture the synchronous pattern (EOF1) and the NEC–SEC contrasting pattern (EOF2) of PM2.5. For PC1, GEOS-Chem captures the variance of each IMF but not its RPR (especially for the day-to-day IMFs). CMAQ simulates too-strong diurnal IMFs, consistent with its overestimated diurnal cycle discussed in Sect. 4.1. CMAQ outperforms GEOS-Chem in capturing the RPR of day-to-day IMFs of PC1, in line with its better correlation with the observations (Fig. 9). For PC2, GEOS-Chem captures the variance and RPR of the observed day-to-day IMFs better than CMAQ.

WRF/CMAQ overestimates the diurnal variation of NO2 and PM2.5. The causes are multifaceted. The ACM2 PBL mixing scheme in WRF v3.5.1 and CMAQ v5.0.1 (used here) assumes the same value for the eddy diffusivity of momentum (Km) and heat (Kh), which implies a Prandtl number (Pr=Km/Kh) of unity and too-weak mixing under stable atmospheric conditions (i.e., at night). This deficiency has been alleviated in WRF v3.7 and CMAQ v5.1. Also, there is inconsistency between CMAQ and WRF in the Monin–Obukhov length in the surface layer module. This error has been corrected in CMAQ v5.1. For more model update details, please refer to the online document (https://www.airqualitymodeling.org/index.php/CMAQ_version_5.1_(November_2015_release)_Technical_Documentation, last access: 3 September 2018).

GEOS-Chem (the first model layer) underestimates surface NO2 by about 17 µg m-3 and PM2.5 by 35 µg m-3 averaged over Eastern China. The underestimate of PM2.5 is in part because this simulation of GEOS-Chem does not include secondary organic aerosols, which likely contribute as much as 21 % of PM2.5 over Eastern China (Fu et al., 2012). Also, the model does not include anthropogenic dust. Furthermore, although the observation stations are close to the ground, the first layer of GEOS-Chem is too thick (130 m) to fully capture the vertical gradient of pollution concentrations. Figure 12 shows Eastern China mean vertical profiles of NO2 in the two models. The center of the first layer of CMAQ (40 m) is closer to the ground, and the center of its second layer is located at a height similar to the center of the first layer of GEOS-Chem. CMAQ shows a strong vertical gradient of NO2 from its first to second layer. Had we used the CMAQ-simulated ratio of the first over the second layer to extrapolate GEOS-Chem first-layer NO2 to 40 m, this would significantly increase the model's “ground-level” NO2 (by 24 % over NEC and 17 % over SEC) and PM2.5 (by 45 % and 17 %). However, the extrapolation does not improve the day-to-day correlation to the observations, indicating the important roles played by other factors. See Table 3 for more evaluation statistics.

GEOS-Chem (the first model layer) also underestimates the Eastern China synchronous day-to-day variation of NO2. When averaged over the 10 lowest layers (below 850 hPa), GEOS-Chem NO2 captures the day-to-day variability of observed surface NO2. This suggests that the model deficiency in day-to-day variability may be specific to the first layer. Moreover, the first layer of GEOS-Chem captures the day-to-day variation of observed NO2 in the afternoon (12:00–15:00 LT, R=0.9 over NEC and 0.8 over SEC), but the model performance is rather poor in the evening (20:00–23:00 LT, R=0.1 over NEC and SEC), suggesting nighttime-specific model inadequacies. A further analysis of nighttime ozone and the NO:NO2 ratio suggests that GEOS-Chem greatly underestimates the observed nighttime ozone by 49.2 % on average over NEC and 54.6 % over SEC, particularly on days when its NO:NO2 ratio is much greater than the CMAQ-modeled ratio. The mean NO:NO2 ratio in GEOS-Chem is 1.8 over NEC and 1.4 over SEC, greater than the ratio in CMAQ (1.0 over NEC and 0.4 over SEC) by a factor of 2–3. Overall, it appears that the nighttime chemistry is poorly represented in the first layer of GEOS-Chem, the causes of which warrant further investigations.

The magnitude of emission differences between the two models plays an insignificant role in the differences between their simulated NO2 or PM2.5 concentrations. The Chinese anthropogenic emissions in 2010 used in GEOS-Chem (except for NOx) are close to the emissions in 2013 used in CMAQ (within 10 % for both gases and primary aerosols, mostly within 5 %; see Zheng et al., 2018). NOx emissions in GEOS-Chem are scaled to 2013 using satellite NO2 data, which further eliminates the differences from those used in CMAQ. The difference in the spatial distribution of emissions is also small (Geng et al., 2017; Zheng et al., 2018).

We further use CMAQ simulations to investigate whether the inclusion of SOA affects our analysis of the spatiotemporal patterns of PM2.5. Supplement Fig. S1 compares the time series of CMAQ-simulated PM2.5 with versus without including SOA. Although SOA contributes about 8–9 µg m-3 of PM2.5 averaged over the days, inclusion of SOA does not affect the temporal variability. The EOF–EEMD results in Supplement Fig. S2 further confirm that the spatiotemporal scales are very consistent whether or not SOA is included.