Observations and modeling of air quality trends over 1990 – 2010 across the Northern Hemisphere : China , the United States and Europe

Trends in air quality across the Northern Hemisphere over a 21-year period (1990–2010) were simulated using the Community Multiscale Air Quality (CMAQ) multiscale chemical transport model driven by meteorology from Weather Research and Forecasting (WRF) simulations and internally consistent historical emission inventories obtained from EDGAR. Thorough comparison with several ground observation networks mostly over Europe and North America was conducted to evaluate the model performance as well as the ability of CMAQ to reproduce the observed trends in air quality over the past 2 decades in three regions: eastern China, the continental United States and Europe. The model successfully reproduced the observed decreasing trends in SO2, NO2, 8 h O3 maxima, SO 2− 4 and elemental carbon (EC) in the US and Europe. However, the model fails to reproduce the decreasing trends in NO−3 in the US, potentially pointing to uncertainties of NH3 emissions. The model failed to capture the 6-year trends of SO2 and NO2 in CN-API (China – Air Pollution Index) from 2005 to 2010, but reproduced the observed pattern of O3 trends shown in three World Data Centre for Greenhouse Gases (WDCGG) sites over eastern Asia. Due to the coarse spatial resolution employed in these calculations, predicted SO2 and NO2 concentrations are underestimated relative to all urban networks, i.e., US-AQS (US – Air Quality System; normalized mean bias (NMB) =−38 % and −48 %), EU-AIRBASE (European Air quality data Base; NMB =−18 and −54 %) and CN-API (NMB =−36 and −68 %). Conversely, at the rural network EU-EMEP (European Monitoring and Evaluation Programme), SO2 is overestimated (NMB from 4 to 150 %) while NO2 is simulated well (NMB within±15 %) in all seasons. Correlations between simulated and observed O3 wintertime daily 8 h maxima (DM8) are poor compared to other seasons for all networks. Better correlation between simulated and observed SO 4 was found compared to that for SO2. Underestimation of summer SO 2− 4 in the US may be associated with the uncertainty in precipitation and associated wet scavenging representation in the model. The model exhibits worse performance for NO−3 predictions, particularly in summer, due to high uncertainties in the gas/particle partitioning of NO−3 as well as seasonal variations of NH3 emissions. There are high correlations (R > 0.5) between observed and simulated EC, although the model underestimates the EC concentration by 65 % due to the coarse grid resolution as well as uncertainties in the PM speciation profile associated with EC emissions. The almost linear response seen in the trajectory of modeled O3 changes in eastern China over the past 2 decades suggests that control strategies that focus on combined control of NOx and volatile organic compound (VOC) emissions with a ratio of 0.46 may provide the most effective means for O3 reductions for the region devoid of nonlinear response potentially associated with NOx or VOC limitation resulting from alternate strategies. The response of O3 is more sensitive to changes in NOx emissions in the eastern US because the relative abundance of biogenic VOC emissions tends to reduce the effectiveness of VOC controls. Increasing NH3 levels offset the relative effectiveness of NOx controls in reducing the relative fraction of aerosol NO−3 formed from declining NOx emissions in the eastern US, while the control effectiveness Published by Copernicus Publications on behalf of the European Geosciences Union. 2724 J. Xing et al.: Observations and modeling of air quality trends over 1990–2010 was assured by the simultaneous control of NH3 emission in Europe.


Introduction
The last 2 decades have witnessed significant changes in air pollutant emissions across the globe.Developed countries in North America and Europe have implemented emission reduction measures which have led to a continuous improvement in air quality.Conversely, in developing regions of the world, in Asia in particular, though control actions have been taken, their effectiveness has been outweighed by the sharp increase in emissions resulting from increased energy demand associated with rapidly growing economies and populations.The striking contrast in the trends in air quality between developed and developing countries has been well discussed in recent years (e.g., Richter et al., 2005).It is also believed that the observed "dimming" and "brightening" trends over the past 2 decades is primarily related to the changes of emission patterns over the Northern Hemisphere (e.g., Wild, 2009;Gan et al., 2014).Therefore, an accurate description of the decadal variations in emissions and associated aerosol burden in the atmosphere is the basis of any attempt to explain the causes of decadal changes in surface solar radiations and short-term climate forcing issues arising from human activities.
Improving air quality and protecting the health and welfare of the population is an important goal for any country.Studies on historical trends in air quality can provide an indication of progress in the direction as well as an assessment of future steps towards the goal.On the basis of long-term records, the effectiveness of past or current control policy can be evaluated and suitable control strategies can be designed for the future.In Europe and North America, several monitoring networks have been in operation for decades and observational records available at some networks are long enough to be used in trends analysis studies (e.g., Sickles and Shadwick, 2007).Such records are vital not only because they reflect the changes in air quality over time, but also because they can be used to evaluate long-term trends in air quality arising from estimated changes in historical emissions, simulated by air quality models.Colette et al. (2011) analyzed the air quality trends during 1998-2007 over Europe by using observations of European Monitoring and Evaluation Programme (EU-EMEP, http://www.emep.int)and the European Air quality data Base (EU-AIRBASE, http://acm.eionet.europa.eu/databases/airbase/)records as well as model simulations.Hogrefe et al. (2009) adjusted 6-year model simulations (2000)(2001)(2002)(2003)(2004)(2005) by using the observed PM 2.5 species concentrations from the observations of Interagency Monitoring of Protected Visual Environments (US-IMPROVE, http://vista.cira.colostate.edu/improve/) and Chemical Speciation Network (CSN) sites in the northeastern US.Trends in O 3 concentration and SO 2− 4 , NO − 3 depositions from 1988 to 2005 simulated by the same model were also compared with long-term observations (Civerolo et al., 2010;Hogrefe et al., 2011).However, due to the large computational cost, very few studies have examined the decadal trend in air pollution over large regions such as northern hemisphere.Koumoutsaris and Bey (2012) evaluated the global model performance of O 3 trends simulation (1991)(1992)(1993)(1994)(1995)(1996)(1997)(1998)(1999)(2000)(2001)(2002)(2003)(2004)(2005) through comparison with long-term observed records from EMEP, the World Data Centre for Greenhouse Gases (WD-CGG, http://ds.data.jma.go.jp/gmd/wdcgg/) and the Clean Air Status and Trends Network (US-CASTNET, http://epa.gov/castnet/).Long-term records of lower troposphere O 3 concentrations from selected sites which are believed to represent baseline conditions in Europe (Logan et al., 2012) and the US (Parrish et al., 2009(Parrish et al., , 2012) ) were used to make quantitative comparisons of simulation results from three chemistry-climate models (NCAR CAM-chem, GFDL-CM3 and GISS-E2-R) (Parrish et al., 2014).To date, however, limited attempts have been made to systematically assess longterm trends in multiple linked atmospheric pollutants (oxidants, particles and acidifying substances) across regional to hemispheric scales.As a regional chemistry transport model (CTM), the Community Multiscale Air Quality (CMAQ) modeling (version 5.0) system (Binkowski and Roselle, 2003;Byun and Schere, 2006;Foley et al., 2010) has previously been successfully applied for several quality studies over North America (Eder and Yu, 2006;Appel et al., 2007Appel et al., , 2008;;Mathur et al., 2008), Europe (Matthias et al., 2012;Kukkonen et al., 2012) and eastern Asia (Yamaji et al., 2006;Wang et al., 2011a;Xing et al., 2011a).However, the need for time-varying lateral boundary conditions (LBCs) which are usually derived from global CTM simulations limits its applications in trend analysis over decades.Recently, the applicability of CMAQ model has been successfully extended to hemispheric scales (Mathur et al., 2012(Mathur et al., , 2014)), so that the application of hemispheric CMAQ provides a consistent approach to generate LBCs for nested regional domains employing finer resolution.
Changing emission patterns across the globe over the past 2 decades have influenced background air pollution levels for different regions across the Northern Hemisphere.To examine air quality trends in different regions over the Northern Hemisphere, we used a multiscale chemical transport model (i.e., CMAQ) driven by historical emission inventories and a meteorological data set to simulate air quality from 1990 to 2010.The ability of the multiscale model to reproduce observed trends over the Northern Hemisphere, including North America, Europe and eastern Asia, was assessed.A brief description of the model configuration, emission processing and observations is given in Sect. 2. The evaluation of model performance through comparison with long-term observation records is presented in Sect.3.1.The trends in both observed and simulated air quality are provided in Sect.3.2 and further discussed in Sect. 4.

Model configuration
Unlike the traditional regional studies with CMAQ, this study used a simulation domain extended to cover the entire Northern Hemisphere with a grid of 108 km × 108 km resolution and 44 vertical layers of variable thickness between the surface and 50 mb (Mathur et al., 2012(Mathur et al., , 2014)).We selected three sub-regions, i.e., eastern China (20-40 • N, 100-125 • E), eastern US (28-50 • N, 100-70 • W) and Europe (35-65 • N, 10 • W-30 • E), for further analysis and comparison with measurements.These three sub-regions are parts of the original Northern Hemispheric domain and no nested simulations were conducted.
The meteorological inputs for 21-year Weather Research and Forecasting (WRF) simulations were derived from the NCEP/NCAR Reanalysis data which has 2.5 • spatial and 6 h temporal resolution.NCEP ADP Operational Global Surface Observations were used for surface reanalysis which is used for indirect soil moisture and temperature nudging (Pleim and Xiu, 2003;Pleim and Gilliam, 2009) in the Pleim-Xiu Land Surface Model (PX LSM) (Pleim and Xiu, 1995;Xiu and Pleim, 2001).The WRF configurations also used MODIS land-use types with 20 categories, RRTMg shortwave and longwave radiation scheme (Iacono et al., 2008), and the ACM2 planetary boundary layer (PBL) model (Pleim, 2007a, b).WRF performance for the simulation of hourly surface temperature (T ), relative humidity and wind speed and direction was evaluated through comparison with observations from NOAA's National Climatic Data Center (NCDC) Integrated Surface Data (ISD, with lite format) which provides hourly (or with 3 h interval) meteorological observations over a long historical period across the globe.The mean bias of T , wind speed and direction over the simulation domain is −0.4 K, 0.4 m s −1 and −3 • , respectively, over the 21-year period.The ranges of biases meet the model performance criteria recommended by Emery et al. (2001) for retrospective regional-scale model applications which are ≤ ±0.5 K, ≤ ±0.5 m s −1 and ≤ ±10 • , respectively, suggesting that meteorology simulations in this study are acceptable.The evaluation of WRF performances ensures that there is no significant bias in the meteorological fields used in the coupled model.

Emission inventories from 1990 to 2010
Figure 1 presents a flow chart of the approach to emission processing employed in creating model inputs spanning the 21-year period.EDGAR (Emission Database for Global Atmospheric Research, version 4.2) (European Commission, 2011) provides a consistent global emission inventory for 1970-2008 for 17 anthropogenic sectors at a 0.1 • × 0.1 • resolution.In this study, we used year-specific EDGAR emissions for the period 1990-2008.Estimates for 2009 and 2010 were derived from projections based on the three most recent references for the United States (Xing et al., 2013), Europe (EEA, 2012) and China (He, 2012).In Europe and North America, pollutant emissions, SO 2 and NO x in particular, have seen continuous reductions during 1990-2010 (refer to Fig. 2).In contrast, NO x and volatile organic compound (VOC) emissions in China have continuously increased, while SO 2 increased during 1990-2006 then decreased from 2007 to 2010 due to more recent strict controls (Zhao et al., 2013;Wang et al., 2014).Emissions in other areas during 2009-2010 were kept the same as the 2008 values.Additionally, since EDGARv4.2provides only PM 10 emissions, PM 2.5 emissions were estimated by deriving the ratio of PM 2.5 to PM 10 from the 2000-2005 EDGAR HTAP (Hemispheric Transport of Air Pollution, version 1) inventory (Janssens-Maenhout et al., 2012), which provides both PM 10 and PM 2.5 emissions, and then applying this ratio to split EDGARv4.2PM 10 emissions into PM 2.5 and PM 2.5−10 .Biogenic VOC and lightning NO x emissions were obtained from GEIA (Global Emission Inventory Activity) (Guenther et al., 1995;Price et al., 1997) and were kept the same for all years during 1990-2010.The 0.1 • resolution gridded data were spatially allocated to the CMAQ grid ensuring conservation of mass.Vertical profiles for anthropogenic sectors and lightning were based on Simpson et al. (2003) and Ott et al. (2010), respectively.The annual mean emissions in each sector were distributed into each hour for each simulated day using the EDGAR default temporal profiles which are primarily based on western European data (http://themasites.pbl.nl/tridion/en/themasites/edgar/documentation/content/Temporal-variation.html).Emissions of PM 2.5 and non-methane volatile organic compounds (NMVOCs) were further speciated into AERO6 and CB05 species based on default profiles in the Sparse Matrix Operator Kernel Emissions modeling system (SMOKE, http://cmascenter.org/smoke/), which is primarily based on data for the United States.Uncertainties are expected when region-specific temporal and speciation profiles are applied to all other counties; however, this approach is reasonable given the lack of any additional information.Further improvement and data are needed to develop more representative profiles for other countries.EU-EMEP are specifically designed for trend assessments since most of their sites are located in rural background areas in order to represent regional atmospheric pollution.Sites in US-AQS and EU-AIRBASE are typically closer to urban areas and may be impacted by local pollution and features sub-grid to the model resolution, thus representative of much smaller regions.To obtain a more valid analysis, the US-AQS and EU-AIRBASE data were averaged over the 108 km grid cells before comparison with the model.CN-API is the average of observed air pollutant concentrations from urban monitoring sites in each city and represents records in seven Chinese cities (i.e., Beijing, Shanghai, Guangzhou, Xi'an, Wuhan, Guiyang and Guilin, which are located in the North China Plain, Yangtze River delta, Pearl River delta, northwest China, central China, southwest China and southern China, respectively) where long-term observations are available starting from 2005.(Jiang et al., 2004;Wang et al., 2011a).In addition, three selected WD-CGG sites were used for O 3 trends analysis in eastern Asia.Only data at sites that covered the 75 % of entire 21-year period (i.e., at least 18 available years with > 75 % coverage for each year) are considered except in the case of CN-API which was only recently setup in the early 2000s and in the case of US-CASTNET (for O 3 only) because most sites have no O 3 records in winter (criteria set as at least 15 available years with > 75 % coverage from March to November for each year).Details about the time-period covered, the number of sites selected for analysis and the record frequency for each network can be found in Table 1.Model results at each monitor location were matched in time to the available record; thus, model data were not considered during periods of missing observations, in either the statistical evaluation or in the trend analysis.

Observed long-term trends
To evaluate the model's performance, model-observed comparisons were conducted by network and pollutant.Five statistical measures, correlation coefficient (R), mean bias (MB), normalized mean bias (NMB), root mean squared error (RMSE) and normalized mean error (NME), are employed for evaluation.In consideration of the limited length of record, this study only focuses on linear trends (Colette et al., 2011).The linear least square fit method was employed, and significance of trends was examined with a Student t test at the 95 % confidence level (p = 0.05).

Model performance
Table 2 summaries the statistics of model performance for gaseous species (Table 2a) and fine particles (Table 2b).

SO 2 and NO 2 concentration
Model performance characteristics for SO 2 , primarily emitted from point sources, can largely be attributed to artificial dilution effects over the large grid volumes employed here.As expected, a hemispherical simulation with relatively coarse spatial resolution is unable to accurately capture the peak values.As seen in Table 2a, SO 2 is underestimated for all urban networks characterized by higher concentrations than rural networks, i.e., US-AQS underestimated by 38 %, EU-AIRBASE by 17 % and CN-API by 36 %.For the rural network EU-EMEP, SO 2 is overestimated in all seasons (4-150 %).A small bias is evident for US-CASTNET annual concentrations since the overestimation in fall is compensated by the underestimation in spring and winter.
Similar performance is noted for simulated NO 2 .The model significantly underestimates NO 2 at urban networks: US-AQS by 48 %, EU-AIRBASE by 54 % and CN-API by 68 %.However, much better performance is noted at sites in the rural network EU-EMEP with bias within ±15 % in all seasons.Though the model-observation correlation coefficients (R) are low for EU-AIRBASE (0.4) and CN-API (0.08) on annual basis, the MB in EU-AIRBASE (−13.9 µg m −3 ) is comparable with previous modeling as reported by Colette et al. (2011) (−6.5 to −18.1 µg m −3 ) and the magnitude of NMB in CN-API (67.5 %) is comparable with Wang et al. (2009) (−61.2 to −81.3 %) but in the opposite direction.It is expected that the performance should be better when simulations are conducted with finer horizontal resolution and with more accurate spatially resolved emissions.

O 3 concentration
Model performance for O 3 is examined through comparisons of seasonal or annual maxima of the daily maxima 8 h (DM8) average or 1 h values since those are the metrics most relevant to air quality standards and health assessments.Correlation coefficients in EU-AIRBASE (0.4) are lower than Colette et al. ( 2011) (0.6-0.8) because the frequency of the observed record used in this study is annual, and therefore, the correlation coefficients calculated here do not benefit from the fact that the model simulations generally capture the observed seasonal cycle.However, the MB (14.4 µg m −3 ) is comparable with that reported in Colette et al. ( 2011) (−4.3 to 18.5 µg m −3 ).Simulations in winter (R = 0.3-0.5)have the worst correlation with observations for all networks compared to those in other seasons (R = 0.6-0.8).On the other hand, both NMB (−13.6 to 16.9 %) and NME (< 25.9 %) are fairly small in all seasons and comparable with that reported by Zhang et al. (2009) (NMB: −10.6 to 15.9 %; NME: < 25.4 %) and Wang et al. (2009) (|NMB| < 37.9 %).

SO 2−
4 , NO − 3 and NH + 4 concentration SO 2− 4 which is formed from the oxidation of SO 2 , is the predominant inorganic aerosol component.In general, SO 2− 4 concentrations show a strong positive response to the changes in SO 2 emissions (Butler and Lakens, 1991), though the SO 2 effective cloud oxidation rate can be affected by NH 3 (Pandis and Seinfeld, 1989;Tsimpidi et al., 2007).As a secondary species, SO 2− 4 is widely spread over the region, unlike SO 2 which is usually more localized to source areas.As seen in Table 2b, correlation coefficients for SO 2− 4 simulation (0.5-0.9) are higher than those for SO 2 (0.4-0.8).The NMBs for US-CASTNET (−8 to −45 %) and US-IMPROVE (−29 to 22 %) are comparable with the results reported by Zhang et al. (2009), which are −23 to 22 % and −8 to 16 %, Eder and Yu ( 2006), which are −10 and −5 % on annual level, and Wang et al. (2009) (|NMB| < 55 %).Significant SO 2− 4 underestimation is noted during summer at both US-CASTNET (by 45.2 %) and US-IMPROVE (by 28.9 %).Some studies also found similar underprediction in their simulations and they attributed such low biases to the uncertainty in precipitation and overestimation of wet scavenging (Luo et al., 2011;Zhang et al., 2014).However, precipitation simulated in this study is underestimated domain-wide by 4 % (in summer) and 65 % (in winter).Wang et al. (2009) found similar underestimation of precipitation from −31 to −41 %, but SO 2− 4 was overpredicted because higher SO 2 emissions were used.Future investigation of the low bias in predicted SO 2− 4 is still necessary.Better performance is shown in EU-EMEP, with NMB within ±30 %.The difference in sulfate biases between the US networks and the European networks might be associated with the different SO 2 biases, i.e., a moderate bias (NMB = −9.4%) in US-CASTNET but a relatively larger bias (NMB = +67 %) in EU-EMEP.The transition rate from SO 2 to SO 2− 4 is likely underestimated in both regions, lead-ing to the underestimation of SO 2− 4 in the US and the better estimates of SO 2− 4 in Europe.Worse performance for NO − 3 prediction is expected because of higher uncertainties in representing the gas/particle partitioning of airborne nitrate (Mathur and Dennis, 2003;Eder and Yu, 2006), especially in summer when SO 2− 4 concentrations are higher and available NH 3 preferentially react to form ammonium sulfate, leading to low ambient NO − 3 level.Simulated and observed NO − 3 have the lowest correlations for both US-CASTNET and US-IMPROVE sites (R = 0.31 and 0.10 respectively) during summer compared those in other seasons (R=0.7).Similar magnitudes of NMB (−56 to 59 %) and NME (89 to 197 %) at US-IMPROVE sites were reported by Wang et al. (2009) and Zhang et al. (2009).The underestimation in summer and overestimation in spring/winter are found relative to both CASTNET (NMB: −48 and 93/75 %) and IMPROVE (NMB: −41 and 107/95 %) and comparable to previous CMAQ analysis of Eder and Yu (2006) (|NMB| > 40 %).Uncertainties in NH 3 emission particularly in the seasonal temporal profile may also contribute to such bias characteristics.Slightly better performance is noted for NO − 3 at EU-EMEP sites, with a higher R (> 0.6) and smaller bias (NMB: −67 to 23 %) for all seasons.

Elemental carbon (EC) concentration
As elemental carbon (EC) is a primary pollutant, its spatial distributions exhibit a strong correlation to its emissions.The correlation between the observed and simulated EC concentrations is high, with R > 0.5, though the model significantly underestimates the concentrations.NMB is up to −74 %, which is worse than previous modeling studies utilizing relatively higher spatial resolution (Zhang et al., 2009 2009) (NMB = 101.7 %), which also utilized coarse spatial resolution.Some previous CMAQ modeling studies (Tesche et al., 2006;Appel et al., 2008) with higher spatial resolution also found a similar underestimation of EC, indicating other factors besides model resolution, such as uncertainties of PM speciation profiles used to estimate the EC emissions, might also contribute to such low biases.

Trend analysis
Simulated trends in SO 2 , NO 2 , O 3 , SO 2− 4 , NO − 3 , NH + 4 and EC concentrations in three regions (eastern China, eastern US and Europe) are given in Table 3.To help understand the changes, trends in input emissions used in this study are also provided in Table 3 as well as depicted in Fig. 2. Capability of the CMAQ model to capture the observed trends was examined through comparisons with network measurements, and both simulated and observed trends are quantified in Table 4 and Figs.3-9.

SO 2 and NO 2 trends
Simulated trends in both SO 2 and NO 2 concentrations over the Northern Hemisphere reflect trends in SO 2 and NO x emissions, respectively (see Figs. 2a-b, 3a and 4a), with a pronounced increasing trend in Asia and decreasing trend in Europe and North America.In particular, in China, annual change rates of SO 2 and NO 2 concentration are about 2.7 and 4.1 %, which are comparable to their corresponding emission rates (SO 2 and NO x ) of 3.2 and 4.3 %, respectively.Annual change rates of SO 2 and NO 2 concentrations in the US (−5.7 % and −1.4 %) and Europe (−5.1 % and −1.2 %) are also close to the rates of emission changes in both regions, at −5.4 % and −1.8 %, and −5.4 % and −1.5 %, respectively.
Such decreasing trends in the US and Europe are comparable with those inferred from observations at the different networks.The annual change rates of SO 2 observed from US-CASTNET and US-AQS are −5.0 and −5.3 %, close to that simulated by the model as −6.6 and −6.5 %.Most of the reductions are located in the eastern US as seen in Fig. 3e-f.The model was unable to capture the increasing trend at two of the eastern AQS sites and also the large decreasing trend at a few sites in the Midwest.It should be noted that the AQS SO 2 measurements predominantly represent urban conditions, and the ability of a coarse resolution model to capture SO 2 levels and trends is influenced both by its inability to accurately represent sub-grid variability as well as changes in local emissions.For instance, the monitor in Kansas City, MO, shows a sharp increase in SO 2 levels starting 2003; in contrast the grid-averaged SO 2 emissions in the corresponding model cell show systematic decreasing trends over the 21-year period resulting in the simulated decreasing SO 2 trend at this location.Also, as seen in the scatter plots in these panels, the pathway of such reductions from 1990 to 2010 is in good agreement between observation and simulation.Stronger trends are noted in winter when SO 2 concentrations are higher compared to other seasons in both observed (−0.368 µg m −3 yr −1 ) and simulated trends (−0.366 µg m −3 yr −1 ) at US-CASTNET (see Table 4).Annual change rates of SO 2 observed from EU-AIRBASE and EU-EMEP are −8.9 and −7.3 %, which are close to that simulated by the model at −5.9 and −6.1 %, with higher rates in winter when SO 2 concentration are at their highest level.Significant reductions are found at locations in the southern UK, the Benelux countries, Germany, Italy, Czech Republic, Poland, Hungary and Romania.
The overall reductions in NO 2 from 1990 to 2010 are also in good agreement between the observations and model simulations.Observed decreasing trends of NO 2 concentrations (and annual change rate) are shown in urban networks, i.e., US-AQS and EU-AIRBASE are −0.63 µg m −3 yr −1 (−2.3 %) and −0.64 µg m −3 yr −1 (−1.9 %), respectively.Model-simulated trends (and annual change rate) at these two urban network, −0.32 µg m −3 yr −1 (−2.2 %) and −0.14 µg m −3 yr −1 (−0.9 %), respectively, are however underestimated.The reason might be associated with the underestimation of NO 2 concentrations.The model slightly overestimated the trends (annual change rates as well) at the rural EU-EMEP network (−0.16 µg m −3 yr −1 (−2.0 %) from the model, compared to the observed trends of −0.13 µg m −3 yr −1 (−1.7 %)).Such decreasing trends are more pronounced over the eastern US, California, the southern UK, northern France, the Benelux countries and Germany.
Large increases in the remotely sensed NO 2 vertical column density (VCD) over eastern China over the past decade have been noted in many studies (Richter et al., 2005;Irie et al., 2005;Akimoto et al., 2006;Zhang et al., 2007) but very limited in situ data are available.Trends in SO 2 and NO 2 inferred from available CN-API data (for 6 years) were not significant (Table 4 and Fig. 3-4b); the model was unable to capture these trends, yielding trends more similar to those of the emissions.These discrepancies could likely arise from uncertainties in local emissions as well as the coarse spatial resolution which limits the model's ability to represent pollution distribution at a finer scale which is likely captured at these monitors.Some industries were moved out of city centers to rural areas nearby so that the improvement of local air quality observed in city centers cannot be captured by large-scale simulations.However, the model results agree with the findings from studies analyzing satellite information over Asia.For example, Zhang et al. (2012) analyzed SCIAMACHY-SO 2 VCD during 2004-2009, suggesting a continuous increase in tropospheric SO 2 loading in western China, but transition from increase to decrease in 2007 in eastern China resulting from stricter controls.

O 3 trends
Ozone concentrations are sensitive to the control of NO x and VOC emissions and studies have indicated that the control in NO x emissions without a simultaneous significant reduction of VOC might lead to an increase of daily O 3 due to the switch from a VOC-limited to NO x -limited regime (e.g., Chameides et al., 1992;Sillman, 1999).However, O 3 chemistry is likely to occur at NO x -limited regimes during periods of heavy photochemical pollution (Trainer et al., 1993;Xing et al., 2011b), suggesting that NO x controls are more effective in reducing annual maxima (rather than averages) of DM8 O 3 .Therefore, trends in NO x emission are more likely to have positive correlations with trends in annual maxima (rather than averages) of DM8 O 3 .As expected, the simulated trends of annual maxima of DM8 O 3 concentration (see Fig. 5a) look quite similar to the NO x and VOC emission trends (Fig. 2b-c).The simulated increasing rate of annual maxima of DM8 O 3 in eastern China is 1.49 %, which is associated with the increase in NO x and VOC emissions (by 4.3 and 2.3 % per year).In contrast, due to reductions of emissions, substantial decreasing trends in annual maxima of DM8 O 3 are apparent in both the eastern US and Europe, with magnitudes of −0.66 and −0.54 % per year, respectively (see Table 3).Significant increases of O 3 are also shown in northern India, western Asia and sub-Saharan Africa where both NO x and VOC emissions have increased during this period (see Fig. 2b-c).Observed decreasing trends in annual maximum of DM8 O 3 concentrations (and annual change rate) in EU-EMEP, EU-AIRBASE and US-CASTNET are −1.07 µg m −3 yr −1 (−0.7 %), −1.35 µg m −3 yr −1 (−0.8 %) and −1.86 µg m −3 yr −1 (−1.1 %), respectively.Similar trends are estimated by the model simulation for both networks, i.e., −1.31 µg m −3 yr −1 (−0.9 %), −2.13 µg m −3 yr −1 (−1.1 %) and −0.95 µg m −3 yr −1 (−0.6 %) (see Table 4).The failure to capture the slightly increasing trends in observations in the urban network (i.e., EU-AIRBASE) might be associated with the limitation by coarse spatial resolution that causes the model to fail to represent the VOC-limited regime at these urban locations and a likely switch of O 3 chemistry from a VOC-to NO x -limited regime which usually goes along with the transition from urban to rural area (e.g., Xing et al., 2011b).Such decreasing trends are noted in all seasons except during winter, when O 3 is at the lowest level.In contrast, the most significant reduction occurred in summer when O 3 concentrations are at the highest.The spatial pattern of O 3 trends is quite similar to that of NO 2 , with more pronounced decreases in regions downwind of urban areas across the eastern US and California as well as southern UK, northern France, the Benelux countries and Germany.The reason for increasing trends shown in both observations and the model in the Midwest US might be explained by the changes in local emissions (less or no controls in Midwest) as well as increasing long-range transport of pollutants across the Pacific (Mathur et al., 2014).Analysis of long-term observations at remote sites along the western US (e.g., Jaffe and Ray, 2007;Parrish et al., 2009) also show increasing trends in O 3 within the boundary layer attributable to inflow from the Pacific.
Though long-term observation records of O 3 are not available in China, recent studies have suggested increasing trends similar to those found here.For instance, Xu et al. (2011) suggested significant increasing trends in tropospheric ozone  ing 1990-2010 is estimated to be 2 % per year, which is comparable to that inferred from observations in these two recent studies.
Observation records at three sites in WDCGG network were used to investigate trends in O 3 distribution in eastern Asia.One of these sites, Minamitorishima (noted as S1, 24.28 • N, 153.98 • E), is located far from land and can be considered to be a representative of clean conditions, while two sites located on Honshu, i.e., Tsukuba (noted as S2: 36.05 • N, 140.13 • E) which is to the northwest of Tokyo and closest to urban regions, and Ryori (noted as S3: 39.03 • N, 141.82 • E) which is in the north and representative of rural conditions.The model generally captured the observed pattern of O 3 trends at each site.For the clean site (S1), no significant trends are inferred either in the observed or the simulated maximum of DM8 O 3 .However, for the urban site (S2), a significant reduction, particularly during summer, is noted in the observed values and is reflective of emission reductions in Japan during the past 2 decades (e.g., Wakamatsu et al., 2013).In contrast, increasing trends are inferred at the rural site (S3) in all seasons expect fall, presumably representing transport from upwind locations in eastern Asia.The model produces similar magnitudes (though of smaller significance) of the decreasing/increasing trends at S2/S3.The contrasting trends at sites S2 and S3 likely result from different controls in local emissions as well as transboundary transport.

SO 2− 4 , NO − 3 and NH + 4 trends
Simulated SO 2− 4 shows a pronounced increasing trend in eastern China (2.8 % per year) and decrease in the US (−3.2 % per year) and EUROPE (−3.7 % per year) which is consistent with, though slightly smaller in magnitude, trends in SO 2 emissions in these regions (see Table 3 and Fig. 6).
Simulated SO 2− 4 trends are in a good agreement with observed trends inferred from all three networks.Simulated trends in SO 2− 4 concentrations (and annual change rate) at US-CASTNET, US-IMPROVE and EU-EMEP are −0.09µg m −3 yr −1 (−3.5 %), −0.03 µg m −3 yr −1 (−2.1 %) and −0.09 µg m −3 yr −1 (−3.6 %), which are comparable with the observed trends of −0.10 µg m −3 yr −1 (−2.9 %), −0.03 µg m −3 yr −1 (−2.4 %) and −0.10 µg m −3 yr −1 (−4.1 %), respectively.More significant trends are noted in summer compared to other seasons because of relatively higher summertime SO 2− 4 concentrations.Average trends at US-CASTNET are more significant than those at IMPROVE because the majority of CASTNET sites are located in the eastern US which witnessed stronger reductions in SO 2 emissions.In Europe, most SO 2− 4 reductions are found in central to eastern Europe, i.e., Germany, Czech, Poland, Hungary, the Benelux countries, Italy, and Romania.NH 3 emissions play an important role in NO − 3 formation (Mathur and Dennis, 2003;Wang et al., 2011b).Growth in NH 3 emissions or reduction in SO 2 emissions (consequently more free NH 3 due to less association with SO 2− 4 ) without simultaneous reduction in NO x emissions can enhance NO − NO − 3 is at the highest level.A similar observed increasing trend is noted during winter at the EU-EMEP monitors which is not captured by the model.The decreasing trend at the EU-EMEP locations during other seasons is, however, captured by the model.Successful reproduction of NO − 3 trends depends on an accurate baseline emission as well as an accurate representation of changes in historical NH 3 emission.Unfortunately, both current NH 3 emissions and their historical trends over the globe still suffer from large uncertainties (e.g., Heald et al., 2012) and likely contribute to the significant bias in the simulated NO − 3 trend.NH + 4 is simulated based on the thermodynamic equilibrium between the NO x , SO x and NH x species.It shows a similar increasing trend in China (3.4 %) and a decreasing trend in the US (−0.7 %) and Europe (−2.9 %), as illustrated in Fig. 8. NH + 4 simulation suffers the same uncertainties as NO − 3 , which leads to difficulties in reproducing the trend in observations (see Table 4).

Elemental carbon (EC) trends
Growth of human activities such as biomass burning and open fires results in the simulated increasing trends in EC levels in China (1.0 %; see Table 3), India and sub-Saharan Africa (see Fig. 9).In contrast, continuous increasing controls have led to a decreasing trend in EC concentrations in the US (−3.4 %) and Europe (−2.5 %).The observed trend in EC at US-IMPROVE, i.e., −0.006 µg m −3 yr −1 (−2.6 %) is well reproduced by the model, i.e., −0.003 µg m −3 yr −1 (−3.3 %).Both observations and the model suggest higher magnitudes of trends during fall and winter, and are likely associated with higher ambient levels during these seasons.
A decreasing trend of EC in Europe has also been observed in other studies (Järvi et al., 2008).The model estimates a consistent decreasing EC trend in the Canadian Arctic (see Fig. 9) which is mainly impacted by emissions from Europe and Russia during winter and spring as demonstrated by Sharma et al. (2004) who analyzed in situ ground-level observations of aerosol black carbon between 1989 and 2002.The increasing trend of EC in southern Asia is corroborated by the evidence found from the Nam Co Lake (located in the central Tibetan Plateau) sediments indicating a recent rise in BC deposition flux (Cong et al., 2013).

O 3 chemistry
As discussed in Sect.3.2.2, the response of O 3 concentration depends on changes in NO x and VOC emissions, and the nonlinear chemistry associated with the subsequent VOCor NO x -limited environment.The response of O 3 to changing levels of NO x and VOC have previously been examined through a variety of methods ranging from isopleths created from chemistry box-model calculations to detailed spatially varying response surfaces developed from output of hundreds of simulations with detailed air pollution modeling systems (e.g., Xing et al., 2011b).Exploration of the changes in O 3 levels in response to historical (and geographically varying) changes in NO x and VOC emissions, as captured by the multi-decadal simulations presented here, provide a unique opportunity to develop insights into factors controlling changes in O 3 production and distributions.
Figure 10 attempts to summarize the changes in NO x and VOC emissions as well as the surface O 3 response during the 1990-2010 period for the three regions; the figures in the left panel illustrate the changes in emissions relative to the 1990 values and the figures in the right panel show the corresponding percentage change in both the maximum and the average of the DM8 O 3 for each year.As can be noted, the relative changes in NO x and VOC emissions vary significantly over different time-period for different regions.Based on the emission estimates, simultaneous growth of VOC and NO x emissions is noted in China with a ratio of 0.46 (i.e., x% NO x growth along with 0.46 x% VOC growth on the basis of 1990 emission level).The modeled increases in both maximum and average of DM8 O 3 values in China during this period are significant.The almost linear response seen in the trajectory of modeled O 3 changes in the region over the past 2 decades suggests that control strategies that focus on combined control of NO x and VOC emissions with a ratio of 0.46 may provide the most effective means for O 3 reductions for the region devoid of nonlinear response potentially associated with NO x or VOC limitation resulting from alternate strategies.The ratio suggested is less than 1, indicating greater sensitivity of ozone to NO x emissions than VOC emissions.It is also obvious to see that the rate of O 3 increase was much smaller during 1995-2002 which was the period when VOC emission growth was much greater than that of NO x emissions in China.
In contrast, trends in emissions over the eastern US indicate significant reductions in VOC emissions compared to NO x prior to 2000.NO x emissions increased slightly during 1996-2000, and then decreased significantly resulting from regional control measures.The change of O 3 during the first decade (1990)(1991)(1992)(1993)(1994)(1995)(1996)(1997)(1998)(1999)(2000) when VOC controls were dominant (reduction ratio of VOC and NO x of −42 and −4 %, respectively) is smaller (−2 %) than that in the subsequent decade (2000)(2001)(2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010) when NO x controls were dominant (reduction ratio of VOC and NO x is −13 and −33 %, respectively), leading to an estimated reduction of −11 % in ambient O 3 .Additionally, model simulations also show an increase in O 3 during 1997-1999, when NO x emissions were estimated to increase.Thus, the response of O 3 is more sensitive to changes in NO x emissions in the eastern US.The relative abundance of biogenic VOC emissions that tend to reduce the effectiveness of VOC controls contributes to this differing response.

PM chemistry
The nonlinear response of NO − 3 concentration to SO 2 , NO x and NH 3 emissions is well documented (e.g., Mathur and Dennis, 2003;Tsimpidi et al., 2007;Makar et al., 2009).Figure 11 attempts to summarize the changes in emissions and factors driving the NO x -SO x -NH x system and its influence on changing inorganic particulate matter composition for the three regions.Contrasting trends in emissions over the past 2 decades in the three regions are apparent: while China and many growing regions of Asia have witnessed significant increases in emissions of NO x , SO 2 , and NH 3 , significant reductions in emissions of all these species have occurred in Europe.In contrast, in the eastern US, while combustion-related emissions of NO x and SO 2 have declined, growth in agricultural animal husbandry have resulted in significant increases in NH 3 emissions.To examine the impact of the varying emissions patterns on inorganic particulate matter formation and composition in these regions, we examined trends in two metrics relative to their 1990 values: (i) the degree of sulfate neutralization, an estimate of the neutralization of sulfate by ammonium (Pinder et al., 2008b ), and (ii) a new metric, the "nitration ratio (NR)" (i.e., NO − 3 concentration divided by NO x emission) to represent the relative amount of oxidized-N emissions that is eventually transformed to aerosol NO − 3 ; changes in the ratio could thus be viewed as an indicator of the relative effectiveness of NO x controls for given conditions.Figure 11 presents the response of PM chemistry to the changes in emissions as indicated by the trends in these metrics during the period 1990-2010.
In eastern China, simultaneous growth of NH 3 emission with SO 2 and NO x plays a very important role in the increases of SO 2− 4 and NO − 3 concentrations (Wang et al., 2011b).During the period 1993-2002, the rate of increase in NH 3 emissions is greater than that of NO x + 2 × SO 2 emissions (representing the amount of NH 3 needed for complete neutralization), with a ratio of 1.1 (i.e., x% (NO x + 2SO 2 ) growth along with 1.1 x% NH 3 growth on the basis of 1990 emission level).In these NH 3 -rich conditions, both DSN and NR consequently exhibit an increasing trend, suggesting that sufficient NH 3 was available to neutralize the available and increasing aerosol SO 2− 4 and also enable formation of particulate NO − 3 .The increasing trend in NR for this region also indicate that the simultaneous growth in emissions of both reduced and oxidized nitrogen results in a greater fraction of NO x being eventually transformed to particulate NO − 3 .After 2002, both DSN and NR decline when the growth of NO x + 2 × SO 2 emissions is faster than that of NH 3 (ratio of 0.9), resulting in the decline of the DSN and NR, eventually back to the 1990-levels.
In contrast, in the eastern US, both DSN and NR exhibit a steady increase during the entire 21-year period, suggesting progressively richer NH 3 conditions stemming from both increased NH 3 emissions as well as more free NH 3 being available due to reduced SO 2− 4 levels associated with declining SO 2 emissions.Steadily increasing trends in NR values also suggest that increasing NH 3 levels offset the relative effectiveness of NO x controls in reducing the relative fraction of aerosol NO − 3 formed from declining NO x emissions.
Interestingly, in Europe simultaneous control of NH 3 along with NO x and SO 2 emissions yields an emission change ratio of 0.6 (i.e., x% (NO x + 2SO 2 ) reduction along with 0.6 x% reduction of NH 3 on the basis of 1990 emission level).Though a slight increase of DSN is simulated during 1992-2003 resulting from faster growth of NO x and SO 2 compared to NH 3 , there is no discernable trend in the estimated NR suggesting comparatively greater control effectiveness in this region compared to the other two, due to the simultaneous control of NH 3 with combustion-related emissions of NO x and SO 2 .

Conclusions
Trends in air quality across the Northern Hemisphere from 1990 to 2010 have been simulated by the WRF-CMAQ model driven with a representation of historical emission inventories derived from the EDGAR.Thorough comparison with several surface observation networks mostly in Europe and North America has been conducted.Significant contrasting changes in emissions have occurred across the Northern Hemisphere over the past 2 decades with reductions in North America and western Europe resulting from control measures on combustion-related sources and increases across large parts of Asia associated with economic and population growth.Model calculations show associated contrasting trends in air pollution across the Northern Hemisphere emphasizing the changing tropospheric composition of trace pollutants as well as the potentially changing background pollution levels in different regions resulting from changes in the amounts of long-range transported pollution.is generally able to capture the observed trends in air pollution, and performance statistics are comparable with results from other studies in regions across the Northern Hemisphere.However, the model estimates still suffer from uncertainties in emissions (in regards to temporal variation and speciation), coarse spatial resolution and subsequent impacts on representation of nonlinear atmospheric chemistry.The lightening NO x emissions used in this study (Price et al., 1997) are likely overestimated compared to a more recent study (Schumann and Huntrieser et al., 2007) and may contribute, to some extent, to the overestimation of NO x , O 3 and nitrate concentrations.The trend of biogenic emissions, which has not been considered in this study, might also impact the analysis.The lack of long-term observations in Asia, particularly over China and India, limits a robust model performance evaluation and O 3 and PM chemistry assessment in these polluted areas.To explore the limitation of coarse spatial resolution in the future, we are currently conducting a study with a finer-scale simulation over the continental United States domain for the same simulated period as from 1990 to 2010.A detailed description and comparison will be provided in a separate paper (Gan et al., 2015).Model-simulated air quality trends over the past 2 decades largely agree with those derived from observations.Significant reductions in ambient levels of most pollutants have been seen in the US and Europe resulting from emission controls implemented during 1990-2010, while levels of all pollutants in China show pronounced increasing trends during the same period.Examining the simulated and observed historical trends in atmospheric chemistry can help guide development of future air pollution abatement strategies.Model calculations over the 1990-2010 period suggest that in the relative amounts of VOC and NO x emission controls in different regions across the Northern Hemisphere (eastern US, Europe and China), have led to significantly different trends in tropospheric O 3 in these regions.In particular, steady increases in NO x and VOC emissions (with a ratio of 0. increase in surface O 3 concentrations in the region, suggesting that possible control strategies that maintain this relative ratio could potentially be most effective in avoiding a nonlinear response resulting from VOC-limitation of alternate approaches.Differences in the historical changes in the relative amounts of NH 3 , NO x and SO 2 emissions in these regions also impact the trends in inorganic particulate matter amounts and composition in these regions.In particular, the amount of particulate nitrate formed per unit of NO x emissions is influenced by changing NH 3 emissions and could be important in assessing the relative effectiveness of different control strategies.Simultaneous growth of NH 3 emission along with those of NO x and SO 2 in China over the past 2 decades has resulted in the increasing particulate nitrate formation trends in the region.In contrast, in the eastern US, the relative fraction of NO x converted to particulate nitrate exhibits a steady increase over the past 2 decades, suggesting an offset in the relative effectiveness of control measures on particulate nitrate levels in the region.Simultaneous reduc-tions in NH 3 emissions along with those of NO x , and SO 2 in western Europe over the past 2 decades resulted in no significant trend in nitration ratio, suggesting the effectiveness of the overall measures in terms of particulate nitrate levels in the region.

Figure 2 .
Figure 2. EDGAR emission trend over 1990 to 2010 for SO 2 , NO x , NMVOC and NH 3 (kg km −2 yr −1 , computed on the basis of annual means over the 1990-2010 period with a linear least square fit method).

Figure 3 .
Figure 3. (a) simulated SO 2 trend from WRF-CMAQ (µg m −3 yr −1 ); (b) map: simulated SO 2 trend in eastern China overlaid with observed SO 2 trends from China-API, dots represent each observation site, computed on the basis of annual means over the 2005-2010 period with a linear least square fit method, dot size is determined by the significance of the trend, i.e., larger symbols denote more significant trends at the p = 0.05 level (µg m −3 yr −1 ); scatterplot: observed and simulated SO 2 concentration, network mean for each year's corresponding grid cells from the model simulation are selected for comparison (µg m −3 ); (c) same as (b) for EU-AIRBASE; (d) same as (b) for EU-EMEP; (e) same as (b) for the US-AQS; (f) same as (b) for the US-CASTNET.

Figure 10 .
Figure 10.Changes in O 3 chemistry from modeling results.

Figure 11 .
Figure 11.Changes in PM chemistry from modeling results (calculation based on molecular units; grid-averaged for three regions; (NO x + 2*SO 2 ) represents the amount of NH 3 needed for complete neutralization; DSN -degree of sulfate neutralization; nitration ratio = NO − 3 concentration / NO x emission).

Table 1 .
Summary of long-term observations used for trends analysis in this study.
* There are few O 3 records from CASTNET in winter; thus, criteria is set as at least 15 available years with > 75 % coverage from March to November for each year.

Table 4 .
Comparison of observed and simulated trends (µg m −3 yr −1 , computed on the basis of annual and seasonal means over the 1990-2010 period with a linear least square fit method) and the annual change rate (x%, i.e., concentration in the year Y (C Y ) will be fit as C Y = C 1990 × (1 + x)Y −1990).
Formatted entries are significant at p = 0.05 level: italic = significant decrease; bold = significant increase.* Trend in O 3 is computed on the basis of annual or seasonal maximum of DM8 (daily 8 h maxima) values, except that for AIRBASE which is computed on the basis of annual maximum of DM1 (daily 1 h maxima).