SO2 is observed in the UV spectral range of satellite observations of SCIAMACHY (on Envisat), GOME-2 (on METOP-A) and OMI (on EOS-AURA). SO2 retrieval algorithms have been developed earlier for GOME-1 by Eisinger and Burrows (1998), for SCIAMACHY by Lee et al. (2008), and for GOME-2 and OMI by Krotkov et al. (2006). Recently a new retrieval algorithm has been developed (Theys et al., 2015) that improves the precision of the SO2 data for OMI by a factor of 2, allowing us to derive more accurate trends based on OMI. The retrieval method is based on a differential optical absorption spectroscopy (DOAS) scheme to determine the slant columns from measured spectra in the 312–326 nm spectral range, which are then background-corrected and converted to vertical columns using an air mass factor (AMF). The AMF is calculated with the radiative transfer model LIDORT (LInearized Discrete Ordinate Radiative Transfer model). More details about the retrieval procedure are described in Theys et al. (2015). The operational algorithm of NASA for SO2 from OMI has also recently been improved. This algorithm and the algorithm of Theys et al. (2015) have a very comparable performance as shown by Fioletov et al. (2016). For this study, the algorithm of Theys et al. (2015) has been applied to the observations of the OMI instrument (Levelt et al., 2006) for its whole mission from 2004 onwards.

To improve the quality of the OMI SO2 data we exclude observations with a cloud fraction of more than 50 % or with a fitting chi-square higher than 1. The solar zenith angle is limited to 75∘ and the viewing angle to 50∘. Since the OMI instrument has been suffering from the so-called row anomaly since 2007 (KNMI, 2012), we filter the affected rows (24–49, 54–55) in the same way for all years in the time series.

As we focus on anthropogenic SO2, the SO2 data for 15 June–9 July 2011 have been removed because of contamination with volcanic SO2 from the eruption of the Nabro volcano in Africa and the transport of its plume to China (Brenot et al., 2014).

As a first step in our study we have made monthly means for the whole data set by averaging and gridding the data to a resolution of 1/8∘ by 1/8∘. The gridding algorithm takes into account the area of each satellite footprint overlapping the grid cell. The resulting data set is a time series of monthly means for the time period October 2004 to December 2015.

For comparison we also use the official ESA SCIAMACHY/Envisat SO2 version product SGP 5.02; and the standard data from the GOME-2/Metop-A version GDP 4.7, as developed within the EUMETSAT Satellite Application Facility for Atmospheric Composition and UV radiation (O3MSAF) project and distributed by http://atmos.caf.dlr.de/gome2/. The data of these instruments are noisier than the OMI data sets because of the lower spatial coverage, different fit window, and the lower signal-to-noise ratio of the SCIAMACHY and GOME-2 instruments. Therefore, their quality-controlled monthly mean SO2 data have been recalculated by spatially averaging, for each grid cell, the data from the eight surrounding neighbouring cells, hence creating a smoothed SO2 field. For details on the methodology and findings, refer to Koukouli et al. (2016).

For NOx emission data we use the results of an update (version 4) of the DECSO (Daily Emission estimates Constrained by Satellite Observation) algorithm developed by Mijling and van der A (2012). DECSO calculates emissions by applying a Kalman filter for the inversion of satellite data and a regional chemical transport model (CTM) for the forward model calculation. It takes transport from the source into account with a semi-Lagrangian approach. The CTM we use is CHIMERE v2013 (Menut et al., 2013) with meteorological information from the European Centre for Medium-Range Weather Forecasts (ECMWF) with a horizontal resolution of approximately 25 km× 25 km. The DECSO algorithm is applied to OMI NO2 observations derived by the DOMINO v.2 algorithm (Boersma et al., 2011). The latest improvements of the DECSO algorithm resulting in version 4 are described by Ding et al. (2015, 2016a). The monthly average emission data over China we use are available at 0.25∘ resolution for the period 2007–2015 on the web portal www.globemission.eu.

The multi-annual mean of SO2 for 2005–2015 is shown in Fig. 2. As the lifetime of SO2 is relatively short (typically 4–48 h) (Lee et al., 2011; Fioletov et al., 2015); the observed SO2 concentrations are a good proxy for the location of SO2 emissions. Regions with large SO2 concentrations are south Hebei, the province of Shandong (around the city of Zibo) and the region around Chongqing. South Hebei is a region with many power plants just east of the mountainous coal-mining area in Shanxi. The hotspot in the Shandong province is related to a strongly industrialized area with substantial coal-using industry. Both coal mines and heavy industry are located in the Chongqing region.

Rather than located at hotspots, high NO2 concentrations are more distributed over the east of China, mainly because traffic is an important source of NOx emissions (see Fig. 3a). The underlying NOx emissions are shown in Fig. 3b. Like the SO2 concentrations, NOx emission spots can be found at the location of large power plants. Also clearly visible are the megacities of China, ship tracks along the coast, and sources along the large rivers.

To construct time series of SO2 we have averaged the data to annual means of the vertical columns derived from OMI. From these annual mean SO2 data we constructed time series for each province (see Table A1). Figure 4 shows the mean normalized time series for the 10 provinces with the highest total SO2 column densities (i.e. Tianjin, Shandong, Hebei, Shanxi, Henan, Beijing, Jiangsu, Shanghai, Anhui, and Liaoning), together responsible for 60 % of all ambient SO2 in China. The individual time series are drawn as thin grey lines. The minimum and maximum of these time series for each year are shown in the grey shaded area to indicate the variability. The time series of Shanghai is the lowest grey line of the 10 series; thus the reductions have been strongest in this province since 2005. Apart from Ningxia province, all provincial time series show very similar patterns. In general, the SO2 concentrations were at a maximum in the year 2007, when the start of a decreasing trend is visible in China. Despite some fluctuations the SO2 concentrations remain relatively constant from 2010 until 2013, whereafter they are decrease again.

A different trend is observed for Ningxia, a province in the mid-northern region of the country with a relative low population density and large coal resources. Here an increasing trend emerges for the years starting from 2010, when several new coal power plants were put into operation. A list of the largest power plants (with a capacity of more than 600 MW) and the start year of their operation is shown in Table 3. From 2012 onward, the more stringent SO2 emission regulations also started to have an effect in Ningxia.

National NOx emission trends show a different pattern than those of SO2. We observe an increasing trend until about 2012, with the exception of the year 2009, which is related to regulations started at the Olympic Games in 2008 (Mijling et al., 2009) and the global economic crisis, which briefly slowed down Chinese economic growth. Total NOx emissions in East China reached their peak levels in 2012, and have stopped increasing since this year. While the economy kept growing after 2012, the emission of NOx slowly decreased again as a result of the air quality regulations described in Sect. 1. According to the DECSO emission inversion, in 2015 the NOx emissions were 4.9 Tg N yr-1, which is 22.8 % lower than in the peak year, 2012. However, the 2015 emissions were still 14.1 % higher than in the reference year, 2007. The trends per province (see Table A2) show very similar patterns, with only the starting year (the year with maximum NOx emissions) of the decrease in emissions varying over the provinces. Events like the Olympic Games in Beijing in 2008 and the World Expo in Shanghai in 2010, when temporary strict air quality regulations were enforced, can be recognized in this table as years with significantly lower emissions for these provinces. In Fig. 5a, the normalized (to the year 2007) time series of annual NOx emissions for East China (102–132∘ E, 18–50∘ N) is shown in similar way to SO2 in Fig. 4. The mean, minimum and maximum of the 10 provinces with the highest NOx emission are shown (Shandong, Hebei, Henan, Jiangsu, Guangdong, Shanxi, Zhejiang, Anhui, Sichuan, and Hubei), together responsible for 65 % of all Chinese NOx emissions. The thin grey lines show the times series for the individual 10 provinces, where the lower line represents Guangdong. Figure 5b shows the peak year for each province. Provinces where air pollution regulations, e.g. for traffic, received a lot of attention at an early stage, like Beijing and Shanghai, reached their maximum before 2011. Most industrialized regions show their peak in the years 2011–2013. Some of the less developed and populated provinces show a maximum in 2014, which means that their decrease in NOx emissions is very recent. Regional variations are mainly due to the fact that regulations for the NOx emission reductions, for instance in traffic or power plants, are determined and implemented on a provincial level (Liu et al., 2016b). For the province of Ningxia we see a pattern occuring that is very similar to that of SO2, which shows for this low-population-density province that traffic plays a small role and the trend is determined by the operation of newly built power plants.

To relate the observed SO2 and NOx reduction to environmental regulations we have to take into account the coal and oil consumption in the same time period. The total coal consumption in standard coal equivalent (SCE) units per year for China and the total oil consumption (also in SCE units) are shown in Fig. 6, based on data of NBSC (2015). According to Guan et al. (2012) and Hong et al. (2016) the sum of coal consumption of all provinces is more accurate than the number provided for the whole of China; thus we use the provincial totals for coal consumption. For NOx emissions the transport sector plays an important role; ships especially are one of the largest NOx emitters per fuel unit in the transport sector. The total freight transport almost doubles every 6 years in China.

Since the burning of coal and oil constitutes the dominant sources of SO2 and NOx emissions, we can consider the total emissions of these air pollutants as the product of the national use of coal and oil (activity) and the average emission factor of one unit of coal/oil. The effectiveness of environmental regulation will be reflected in a decrease in this emission factor. Therefore, we divide the annual SO2 column measured from satellites and the annual NOx emissions by the annual coal and oil consumption in China. In this way we get a measure of the emitted SO2 or NOx per unit (SCE) of fossil fuel consumption reflecting the Chinese environmental policy. The results are shown in Fig. 7. One might argue that SO2 is more related to coal than oil, but division by only coal yields the same results. In our analysis we omit gas consumption since this is very limited in China and hence does not affect the results significantly.

We focus here mainly on the results for OMI, because of the instrument's high spatial resolution and lack of instrumental degradation. However, SO2 data of the SCIAMACHY and GOME-2 instrument are also added in Fig. 7 to be able to further look into the past (starting in 2003) and to verify the results of OMI. The SO2 data of SCIAMACHY and GOME-2 are averaged over the summer months (April–September). The remaining monthly means are excluded from the analysis due to a lower accuracy at higher latitudes, and a large part of the higher latitudes is missing due to snow cover. For OMI each data point is averaged over 12 months and the total area of China, which reduces the root-mean-square error to a negligible level. Biases among all instruments are removed by normalizing the values to those in reference year 2007. Up to 2009, the results agree fairly well. After 2009, we see the results of GOME-2 and OMI for SO2 slowly diverge in time, which might be a result of the instrument degradation of the UV spectra of GOME-2 after 2009 (Munro et al., 2016).

Changing weather conditions from year to year can affect the results for SO2 concentrations, and when the weather conditions are different during the overpass of SCIAMACHY and GOME-2 (around 09:30 LT, local time) than those at the overpass of OMI (around 13:30 LT), this can lead to differences between the instruments. The global coverage of SCIAMACHY is once every 6 days and for GOME-2 and OMI almost daily. The limited number of samples for SCIAMACHY makes these data more sensitive to weather conditions. Note that, due to the nature of the inversion algorithm, the NOx emission data are in general not sensitive to meteorological variability.

For SO2 we see a large decrease in the years 2008 and 2009, while the desulfurization programme of the 11th five-year plan had already started in 2005/2006, when the authorities began to reduce SO2 emissions by installing desulfurization devices in many power plants (Lu et al., 2010). In 2006 SO2 monitoring devices were also installed in the chimneys of the power plants. This resulted in a decrease in SO2 emissions from 2006, while the much larger decrease in SO2 in 2008–2009 reflects the stronger government control at that time on the actual use of the equipment (Xu et al., 2011). After 2009, the SO2 content per consumed coal unit only slowly decreases until 2011. From 2012 onwards we see a stronger annual decrease in SO2. This coincides with the 12th five-year programme; when new measures were taken to upgrade the coal quality, to modernize the industry, and to put more effort on law enforcement. The enforcement of laws in the last few years concerning the prohibition of flue gas bypass and the use of desulfurization devices in the steel industry played an especially important role.

For the NOx emissions the total annual emissions are used and divided in the same way as for SO2 by the total coal and oil consumption. Here, however, we should keep in mind that the transport sector (especially by shipping) emits much more NOx per fuel unit than the power and industrial sectors (see, e.g., Zhao et al., 2013). Thus, the percentage of the total fuel used by transport is relevant for the graph of NOx. In the early years we see, in general, a small increase in NOx emissions per fuel unit due to the increasing fraction of the transport sector in fuel usage. Exceptions are the year 2009 and the recent year 2015. The year 2009 coincides with the global economic crisis (Lin and McElroy, 2011), when there was less export of goods from China. This especially affected the transport sector, mostly transport over water, as shown in De Ruyter de Wildt et al. (2012). Faber et al. (2012) and Boersma et al. (2015) showed that the economic crisis also resulted in a significant reduction in the average vessel speed to save fuel used by ship transport. This not only caused a shift in source sectors but also in general led to lower NOx per fuel values. This explains the dip in pollution per fuel unit in 2009. After 2009 the NOx per fossil fuel is slowly increasing because the transport sector is growing faster than the energy sector and has a higher emission factor. Statistics of the NBSC (2015) show that transport is growing by a factor of 2 every 5–6 years (Wu et al., 2017). After 2012 the gradual increase in NOx per fuel slowly stops, and the year 2015 shows a sharp decline in NOx per fossil fuels unit. This can be directly related to the rapidly growing installation of SCR equipment at power plants since 2012 and to a lesser extent to the introduction of new emission standards for cars, as shown by Liu et al. (2016). This strong reduction in NOx for 2015 and the equally strong reduction for SO2 in 2014 and 2015 are a result of very effective recent environmental regulations in the last years in China. By comparing the efficiency level in 2015 with earlier levels, we can conclude from Fig. 7 that, without these air quality regulations, SO2 concentrations would today be about 2.5 times higher. For NOx per fossil fuel we were expecting a gradual growth after 2012 because of the continuing relative growth of the transport sector. Keeping this in mind we compare the years 2015 with 2012 and conclude that without air quality regulations the NO2 concentrations would be at least 25 % higher in China today.

On a provincial scale we can, in principle, do similar analyses, but unfortunately the provincial energy consumption related to coal and oil has a very high uncertainty due to inconsistencies in interprovincial imports and exports (Hong et al., 2016). We see this reflected in a high variability in the annual provincial data and sometimes missing data. The data have high uncertainties especially for oil consumption (Guan et al., 2012; Hong et al., 2016). Therefore, we have only analysed the five provinces with dominating coal consumption as shown in Fig. 8. In this graph we excluded Guizhou province because of its difficult-to-interpret coal consumption in 2011 as a result of large power shortages (NBSC, 2015; Sun and Zhou, 2011). For SO2 per fossil fuel unit we see that all provinces follow the national trend. For NOx per fossil fuel we see more variation per province, depending on the role of transport. Most of these coal-consumption-dominated provinces start their decreasing trend from 2011, reflecting the national programme on SCR installations starting that same year. It is interesting to see that the commissioning of new power plants in 2011 causes a strong increase in both SO2 and NOx in Ningxia province (see Figs. 4 and 5a). However, when compensating for fossil fuel usage, one can see that the same national air quality regulations are applied here, as the trend in Fig. 8 shows the same pattern as for other provinces. This shows the strength of the presented method to assess the efficiency of air pollution regulations.