OMI is onboard the EOS-Aura satellite. The satellite measurements have a pixel size of 13 × 24 km2 at nadir with a local overpass time around 13:40. VCDs of tropospheric NO2 are derived in three major steps, including derivation of slant column densities (SCDs), separation of stratospheric and tropospheric SCDs, and calculation of tropospheric air mass factors (AMFs) for deduction of the tropospheric VCDs. On a regional and monthly mean basis, the error of retrieved VCDs is about 30 % (a relative error) plus 0.7×1015 molecules cm-2 (an absolute error) (Boersma et al., 2011; Lin and McElroy, 2011). More detailed algorithms and error descriptions involved in retrieving tropospheric NO2 VCDs can be found in Boersma et al. (2007, 2011).

We mapped the level-2 DOMINO v2 product (http://www.temis.nl/airpollution/no2.html, Boersma et al., 2015) to a 0.25 ∘ × 0.25∘ grid, and then averaged daily data to produce monthly mean VCD values. We used data from October 2004 to May 2014 for the present analysis. For data quality control, we excluded pixels with a cloud radiance fraction > 50 % or affected by row anomaly (Boersma et al., 2011). We filled the missing monthly mean values in some grid cells using values in the adjacent years; the impact on the trend analysis is found to be small by sensitivity analyses on the respective GEOS-Chem simulation results (see Sect. 4.2.2).

We used the nested GEOS-Chem chemical transport model version 9-02, on a 0.667∘ long.× 0.5∘ lat. grid with 47 vertical layers, to simulate the tropospheric NO2 and other pollutants over Asia (Chen et al., 2009). The model is run with the full Ox-NOx-VOC-CO-HOx gaseous chemistry and online aerosol calculations, and it is driven by the GEOS-5 assimilated meteorology from the NASA Global Modeling and Assimilation Office. Vertical mixing in the planetary boundary layer follows the non-local parameterization scheme implemented by Lin et al. (2010b). Convection is simulated with a modified Relaxed Arakawa-Schubert scheme (Rienecker et al., 2008). Lateral boundary conditions of the nested model are updated every 3 h by results from corresponding global modeling on a 5∘ long. × 4∘ lat. grid.

Chinese anthropogenic emissions of NOx and other species adopt the monthly MEIC inventory with a base year of 2008 (www.meicmodel.org). The spatial resolution of MEIC used in the simulation is 0.667∘ long. × 0.5∘ lat., according to the model grid. We further scaled monthly anthropogenic NOx emissions to other years, by applying the ratios of monthly DOMINO v2 NO2 VCDs in those years over the VCDs in the respective months of 2008. The scaling with OMI data was done at the model resolution, after regridding the satellite data from the 0.25∘ long. × 0.25∘ lat. resolution. Emissions for other Asian regions follow the INTEX-B inventory (Zhang et al., 2009a). Other model setups are described in Lin et al. (2015).

Due to limited meteorological inputs, model simulations were conducted from 2004 to April 2013. The first simulation year was used for model spin-up, and results from 2005 onward were analyzed in the present analysis. In the following order, modeled vertical profiles of NO2 were averaged over 13:00–15:00 local time, regridded to a 0.25∘ × 0.25∘ grid, sampled in locations and days with valid OMI data, applied with the DOMINO averaging kernel (AK), and then averaged to derive monthly mean VCD values. The use of AK was to eliminate the effect of differences in NO2 vertical profiles between GEOS-Chem and TM4 (that provides the a priori profiles for the DOMINO retrieval). Following our previous work (Lin et al., 2010a; Lin, 2012), we regridded the pixel-specific AK to the 0.25∘ × 0.25∘ grid. Modeled VCDs data without applying the AK are also analyzed in Sect. 4.2.2 to test the effects of data sampling and temporal interpolation.

We took Chinese official provincial-level NOx emission inventories for 2007 and 2010–2013 to compare with trends in OMI NO2. Chinese central government commenced its official estimate of anthropogenic NOx emissions following the first nationwide pollution census in 2007. (The first nationwide pollution census committee, 2011). NOx emissions in 2010–2013 were also based on the estimating system of the first pollution census, allowing for a consistent comparison throughout time. We also included the official emission targets aimed for 2015 from the 12th Five-Year Plan (2011–2015), a well-known socioeconomic planning step of China. We obtained all socioeconomic data from the China Statistical Yearbooks Database (http://tongji.cnki.net/overseas/engnavi/navidefault.aspx).

Figure 1 highlights the study area in China. We extracted provincial and regional NO2 data according to their administrative divisions. We separated Western China into two sub-regions, including Northwestern China (Gansu, Inner Mongolia, Ningxia, Qinghai, Shaanxi and Xinjiang) and Southwestern China (Chongqing, Guangxi, Guizhou, Sichuan and Yunnan). Tibet is excluded from the present analysis due to lack of socioeconomic data. We also selected three key regions from Eastern China for comparisons with Western China: the Beijing–Tianjin–Hebei region (BTH, including Beijing, Tianjin and Hebei Province), the Yangtze River Delta (YRD, including Shanghai, Jiangsu Province and Zhejiang Province) and the Pearl River Delta (PRD, part of Guangdong Province).

This study is focused on areas that have been subjected to significant changes in anthropogenic NOx emissions. Since NOx are emitted from both anthropogenic and natural sources (Lin, 2012), we exploited their distinctive seasonal patterns to determine areas dominated by anthropogenic sources.

Over China, anthropogenic emissions tend to maximize in winter, although the seasonal variation is often within 20 % (Zhang et al., 2009a). Soil and lightning emissions exhibit summer maxima with very low values in winter. Biomass burning emissions of NOx are negligible over China (Lin, 2012). In addition, the lifetime of NOx in winter is several times longer than in summer. Therefore the NO2 VCDs are the lowest in summer and the highest in winter over the areas dominated by anthropogenic sources, while the opposite seasonality occurs over the regions dominated by natural emissions (Lin, 2012). Furthermore, lightning and soil emissions are mostly independent of direct anthropogenic influences for 2005–2013, albeit with certain effects from changes in climate and/or land use. There is no evidence that these natural emissions underwent significant trends from 2005 to 2013. By comparison, anthropogenic emissions have exhibited dramatic changes along with the rapid socioeconomic development, and these changes have affected the seasonality of NO2.

Figure 2a shows the seasonal variation in OMI NO2 VCDs, averaged over 2005–2013, for each 0.25∘ × 0.25∘ grid cell in Western China; all grid cells are sorted according to their 9-year mean NO2 values. Once a grid cell is ordered, its monthly NO2 values are averaged over 2005–2013 to obtain a 9-year mean monthly climatological data set with 12 values. Finally, the monthly climatological values are converted to their reverse ranks (from 1 to 12), for improved illustration across all grid cells. Figure 2a shows that for grid cells with 9-year mean NO2 VCDs below 1.0 × 1015 molecules cm-2, NO2 generally experiences summer maxima and winter minima, reflecting the dominance of natural sources. In contrast, grid cells with 9-year mean NO2 VCDs above 1.0 × 1015 molecules cm-2 exhibit winter maxima, due to the dominance of anthropogenic emissions as well as a longer lifetime. Van der A et al. (2006) also found that over 1996–2005, polluted Eastern China experienced NO2 maxima in winter due to large anthropogenic emissions while much cleaner western China experienced summer maxima due to natural sources. Similar results were shown by Lin (2012) who compared polluted and cleaner regions in Eastern China in 2006.

Figure 2b further shows the standard deviation (SD) of monthly OMI NO2 VCDs year by year for each 0.25∘ × 0.25∘ grid cell in Western China; again, all grid cells are sorted according to their 9-year mean NO2 values. Once a grid cell is ordered, the SD is calculated for each year to obtain a data set with nine values over 2005–2013. Finally, the SD values are converted to their reverse ranks (from 1 to 9), for better illustration across all grid cells. Figure 2b shows that grid cells with 9-year mean NO2 VCDs above 1.0 × 1015 molecules cm-2 exhibit a large growth in SD especially since 2009, as a result of large growth in anthropogenic emissions that amplified the seasonality. By comparison, grid cells with 9-year mean NO2 VCDs below 1.0 × 1015 molecules cm-2 did not experience such significant changes in SD between 2005 and 2013.

Based on the above seasonality analysis, we determined the regions dominated by anthropogenic emissions as those with 2005–2013 mean NO2 VCDs exceeding 1.0 × 1015 molecules cm-2.

To obtain the sole anthropogenic NO2, we further subtracted all NO2 VCDs by certain “background” values representing the natural influences. Removing the “background” influences is meaningful for Western China where the NO2 VCDs are currently not at an extremely high level (see Sect. 4.2.1).

We identified six “background” areas that are away from cities and are supposed to be dominated by natural emissions (see the hatched areas in Fig. 1), and we assumed NO2 VCDs there are all natural. Russell et al. (2012) used the same method to identify “background” NO2 over the United States. When calculating the trends of NO2 in the grid cells of the chosen human-dominant areas, we subtracted NO2 VCDs at these grid cells by the NO2 value averaged over the nearest “background” region. For all grid cells in a given province, the corresponding background region is the same and is indicated in Table 1. The background subtraction was done on a monthly basis to account for natural variability. We processed the model NO2 data with the same method.

Figure 1 shows that the “background” regions in Western China are normally the uninhabited areas. Over there, the NO2 VCDs are only about 0.4–0.5 × 1015 molecules cm-2 in 2005 (with little interannual variability), lower than NO2 in the polluted areas by a factor of 2–5 (Table 1). For Eastern China (Table 1), the “background” values are higher (0.7–1.2 × 1015 molecules cm-2 in 2005); whereas these values are 6–13 times lower than the NO2 VCDs over the three polluted eastern regions (BTH, YRD and PRD).

Note that the chosen “background” values may not fully represent the actual natural contributions to the targeted human-dominant areas. For example, soil emissions may vary in space due to differences in temperature, radiation, land cover and land use type, and other climatic factors. Lightning emissions of NOx may have spatial dependence as well. The “background” regions may not be totally free from anthropogenic influences, as a certain amount of NOx in the polluted areas may be oxidized to produce peroxyacyl nitrates (PANs), which can be transported to “background” areas and converted back to NOx. For these reasons, our choice of background values is relatively rough. Nevertheless, unless the actual natural contributions differ substantially from the chosen values, which we do not expect to occur on a provincial average, the resulting effect on our trend calculations should be small, because the chosen background values are smaller than NO2 over their corresponding polluted areas by a factor of 2–13 (Table 1). Future work is needed to fully separate the anthropogenic from natural NO2 for individual locations.

Due in part to the short lifetime of NOx, the tropospheric NO2 VCDs respond quickly to emission changes at various temporal scales, from a general growth along with socioeconomic development to short-term perturbations such as the Chinese New Year holidays and the economic recession (Lin and McElroy, 2011; Lin et al., 2013). Also, uncertainties and sampling biases in the satellite data may introduce additional noises in the NO2 monthly time series. If not separated, these short-term variability and noises may affect linear trend calculations.

Here we conducted discrete wavelet transform (DWT) (Daubechies, 1992; Partal and Küçük, 2006) to distinguish temporal variability of NO2 at multiple scales. The wavelet transform is a useful tool for diagnosing the multi-scale and non-stationary processes over finite space and time periods, with the advantage of localization in the time and frequency domain (Echer, 2004; Percival and Walden, 2006), suitable for our analysis of NO2 trends and variability. Different from the approaches adopted by previous NO2 studies (e.g., van der A et al., 2006), our wavelet analysis does not require prior assumptions about seasonality and other temporal scales. As shown in Sect. 3.1, the magnitude of NO2 seasonality is correlated to the amount of annual mean NO2 and anthropogenic sources, and this information is captured by the wavelet analysis here.

The multi-scale analysis in DWT is able to decompose a time series f(t) into n scale components (n is the decomposition level): f(t)=∑i=1nDi+An, where Di is a detail signal (high frequency) at level i and An is the approximation signal (low frequency) at the set of maximum level n. The detail and approximation signals were generated based on the convolution of time series of wavelet functions and scaling functions. We chose Meyer orthogonal discrete wavelets as the wavelet functions which have been used to study ozone column, NDVI and land-cover changes (Abry, 1997; Echer, 2004; Freitas and Shimabukuro, 2008; Martínez and Gilabert, 2009). Specifically, the approximation and detail signals were derived through an iterative multi-layer decomposition process. In the first layer of decomposition, f(t)=A1+D1. Then, A1=A2+D2 and f(t)=A2+D2+D1; and so on. The iteration stops at level n=5 for all provinces. At level 5, the period of the A5 time series is longer than the length of the data set (116 months). This criterion is typically used in investigating the long-term trend of a time series (Echer, 2004; Chen et al., 2014). As an example, Fig. 3 presents the wavelet transform result for one grid cell (34.5∘ N, 108.9∘ E) in Xi'an City, Shaanxi Province.

As a result, the approximation signal A5 represents the long-term trend of the original NO2 time series (with a period longer than the length of the data set). Although similar to the 12-month moving average time series, the A5 time series is much smoother with no short-term variability (see the example in Fig. 3). The detail components D1–D5 indicate higher-frequency variations, which are not analyzed in this study.

Figure 4 shows the spatial distributions of annual average OMI NO2 VCDs over China in 2005, 2012 and 2013. Here the “background” values have not been subtracted. The NO2 VCDs exceed a high value of 6 × 1015 molecules cm-2 in many areas of Central-East China and parts of Western China. Chengdu-Chongqing, Urumqi and Shaanxi-Guanzhong city clusters are well-known pollution “hot spots” of Western China (see Fig. 1 for region definitions). These “hot spots” have intensified since 2005, as well as other polluted western areas including Gansu–Ningxia and Inner Mongolia industrial city clusters. The annual and regional average NO2 VCDs over Western China has increased by 51 % between 2005 and 2013, higher than the increase at 41 % in Central-East China. The large growth of NO2 over Western China highlights the necessity of understanding potential human influences in these regions.

Figure 4 also compares the OMI derived and GEOS-Chem modeled annual average NO2 VCDs in 2005 and 2012. OMI and model NO2 share similar spatial and temporal patterns. Linear regression for model NO2 as a function of OMI NO2 reveals that for any given year, model NO2 are highly correlated with OMI values in space. Table 2 shows that for all of China in 2008, the magnitudes of model NO2 are close to OMI NO2 (slope = 1.09, R2=0.88). For other years, the slopes are larger (1.11–1.26), indicating positive model biases, while the R2 ranges from 0.86 to 0.90. The scatterplot in Fig. 4 further confirms the model-OMI consistency in 2012. Similar results are found for Western China, although the R2 is smaller, at 0.68–0.76 over 2005–2012. The model biases in years other than 2008 reflect the nonlinear relation between changes in NOx emissions and changes in NO2 VCDs (Martin et al., 2003; Valin et al., 2011; Lin, 2012) that we did not account for when linearly scaling model emissions from MEIC 2008 to other years based on the interannual variation in OMI NO2.

Figure 7 shows Chinese official bottom-up provincial anthropogenic emission inventory for 2007 and 2010–2013, together with the provincial emission targets for 2015 (as a goal of the 12th Five-Year Plan) (The State Council of the People's Republic of China, 2011a). Provincial mean OMI NO2 VCDs are also shown for comparison. Both emission and VCD data sets were normalized to their 2007 mean values to remove the effect of regional dependence in the relation between NOx emissions and NO2 VCDs. Ningxia, Xinjiang and Inner Mongolia had the largest increases in NOx emissions from 2007 to 2010, consistent with their growth of NO2 VCDs. NOx emissions in most provinces grew significantly from 2007 to 2010 and peaked in 2011–2012, also in general consistency with the trends in OMI NO2. On the other hand, the emission inventory suggests a reduction since 2011 for Xinjiang and Yunnan, inconsistent with the notable growth in NO2 VCDs. This likely suggests an underestimate in the official emission inventory.

As described in Sect. 4, the tropospheric NO2 VCDs over Western China have grown notably since 2005. The growth occurred not only over cities but also over many suburban and rural regions, indicating an expansion of human influences from urban to remote areas. This scale of pollution growth was associated with the rapid urbanization and industrialization over Western China following the “Go West” movement. Table 4 shows that the urban population (i.e., the percentage of total population living in urban areas) increased by 10 % or more from 2005 to 2013 in all provinces of Western China except Xinjiang. Over the same period, Western China experienced steep economic growth with industrial GDP growth rates of 12.4–20.3 % yr-1 across the provinces.

On the other hand, the NO2 VCDs declined or stabilized since 2011 in many provinces (see Fig. 6), partly reflecting some improvements in environmental strategies. China's air pollution control strategy has been transformed from a traditional end-of-pipe control strategy (i.e., only using low NOx combustion technologies in some power plants) into a combined energy saving and emission reduction strategy after 2006 (Gu et al., 2013; Zhao et al., 2013). In particular, total NOx emissions have become a major target of national pollution control in the 12th Five-Year Plan (2011–2015), with a legally binding goal to reduce the national emissions by nearly 10 % in 2015 compared to 2010 (The State Council of the People's Republic of China, 2011b). Furthermore, the Chinese central government has also decided to consider the effectiveness of this reduction in evaluating local governments' performance (The State Council of the People's Republic of China, 2012). Regarding energy saving measures, great efforts have also been made to improve energy efficiency, to slow down growth of energy demand, and to adjust structure in various sectors (power plants, transportation, industries, and residential use) over the past few years (Wang and Hao, 2012; Zhao et al., 2013).

Northwestern China (Inner Mongolia, Xinjiang, Qinghai, Gansu and Shaanxi) has an average NO2 growth rate at 11.3 ± 1.0 % yr-1 from 2005 to 2013, about twice the average growth rate (5.9 ± 0.6 % yr-1) in Southwestern China (Sichuan, Chongqing, Guizhou, Guangxi and Yunnan). The contrast in NO2 growth rate between Northwest and Southwest reflects their distinctive states of socioeconomic development. According to the nationwide pollution census, Northwestern China generates much more NOx emissions per unit of GDP (11.94 t/billion RMB in 2007) than the Southwest (6.98 t/billion RMB) (The first nationwide pollution census committee, 2011). The difference in pollution intensities also reflects their dissimilar economic structures. In particular, Northwestern China has recently become an important energy producer (due to the “West to East Power Transmission” project) and a heavy industry base (in terms of mining, fossil fuels and raw materials)(Chen et al., 2010; Deng and Bai, 2014), and these industries are often associated with significant NOx emissions. The electricity consumption of heavy industries in Northwestern China grew by 152.5 % from 2005 to 2011, greater than the growth at 99.6 % in Southwestern China.

About 70 % of China's industrial and residential energy consumption is supplied by coal burning in 2009 (Li and Leung, 2012), and the value did not changed drastically in later years. Figure 8 shows that Northwestern China has consumed more coal than the Southwest since 2005, and by 2012 their difference has increased by a factor of 35 (from merely 9.03 million tonnes in 2005 to as large as 318.3 million tonnes in 2012). For the Northwest, there is an extremely high correlation between NO2 VCDs and coal use across the years (R2=0.95, P value < 0.05), compared to a correlation at 0.84 (P value < 0.05) for the Southwest.

Furthermore, the annual amount of electricity generated by coal-fired power plants in Northwestern China increased by 237 %, from 226.3 billion kWh in 2005 to 763.1 billion kWh in 2013; the annual growth rates are 9.8–22.8 % yr-1 for individual provinces (see Table 4). The growth was much smaller in the Southwest, about 110 % from 165.6 to 347.9 billion kWh, translated to growth rates of 6.0–14.9 % yr-1 for individual provinces. This difference was partly due to the stronger growth in hydropower production in the Southwest (from 147.2 to 471.6 billion kWh over 2005–2013, at the rates of 8.9–21.6 % yr-1 in individual provinces) than the growth in the Northwest (from 40.0 to 107.3 billion kWh, 4.7–18.4 % yr-1).

Transportation plays a more important role in NOx pollution over the Southwest compared to the Northwest. Table 4 shows that transportation contributes to much larger fractions of NOx emissions in the capital cities of Southwestern China than in the Northwestern capital cities except Xi'an, Shaanxi. In addition, the number of vehicles grew faster in the Southwestern capital cities during 2005–2012.

The average NO2 growth rate was 8.6 ± 0.9 % yr-1 for Western China, much larger than the rates in the three key eastern regions BTH (5.3 ± 0.8 % yr-1), YRD (4.0 ± 0.6 % yr-1) and PRD (-3.3 ± 0.3 % yr-1) (see Table 1). This regional contrast reflects both their economic activities and the emission control policies adopted by the Chinese central and local governments. In particular, China's development strategy for its western provinces might have led to unintended westward pollution migration, as many resource- and pollution-intensive industries gradually moved from the East to the West after 2000. Table 4 shows that from 2005 to 2013, the average industrial GDP growth rate in Western China was 17.2 % yr-1 (relative to 2005), higher than the rates in the three key eastern regions (13.2 % yr-1 in BTH, 11.6 % yr-1 in YRD and 12.0 % yr-1 in PRD). The fast economic and pollution growth in Western China in part reflects its growing production to support consumption in other regions (Lin et al., 2014a; Zhao et al., 2015). According to Zhao et al. (2015), NOx emissions over Western China in 2007 were largely attributable to the economic production to supply Eastern China and foreign countries, with 366 Gg related to interprovincial trade and 49.1 Gg related to international trade. Together with atmospheric transport, trade has become a critical mechanism for transboundary pollution transfer at both the global and regional scales (Lin et al., 2014a), with significant consequences on public health (Jiang et al., 2015).

The west–east contrast in NO2 growth also reflected their different pollution control strategies and measures. Although China has a national NOx emission reduction target at 10 % (from 2010 to 2015), the targets are set differently for individual provinces. Table 1 shows that the targets were higher, at 13.9, 17.7 and 16.9 %, for the three key eastern regions (BTH, YRD, and PRD), but they are as low as 5.7 % averaged over Western China (The State Council of the People's Republic of China, 2011a). In particular, an emission increase by 15 % is allowed for Qinghai Province. In addition, although NOx emission reduction measures have been taken in power plants and some other industrial sectors since 2006 (via de-nitrification systems that involve selective catalytic or non-catalytic reduction), by 2010 as much as 57 % of these systems were installed in the three key eastern regions (Zhao et al., 2013). The capacity of small power generators being shut-down in Western China was about 10 808 MW (excluding small diesel generators), only accounting for about 19 % of the capacity of total shut-down small power plants in China (55 630 MW) during the 11th Five-Year Plan period (2006–2010) (NDRC, 2009–2011; Xu et al., 2013).

Furthermore, the vehicle emission control has also been implemented much more stringently in the East than in the West (Li and Leung, 2012). Although large amounts of “Yellow-Label Vehicles” (YLVs, highly-emitting vehicles that fail to meet the National I emission standard) have been banned from entering into big cites in Eastern China, over recent years a considerable number of used YLVs have been brought to the West, where the restrictions on YLVs are much weaker (Qi, 2010). Greater efforts to reduce NOx pollution in Western China, with lessons learnt from the East, will help to achieve its sustainable development.

The DOMINO v2 data are available at the TEMIS web site (http://www.temis.nl/airpollution/no2.html).