HTML
--> --> -->Aerosol optical depth (AOD) can be used to test calibration of satellite retrieval data, and AOD is the key factor to determine the aerosol climate effect (Kaufman et al., 2002). Currently, AOD can be obtained from ground-based equipment [e.g., the Aerosol Robotic Network (AERONET) (Holben et al., 1998), airborne AOD observations (Shinozuka et al., 2011), and satellite sensors (Kaufman et al., 2002)]. However, there is little data from ground-based observations of aerosols in East Asia (Che et al., 2014), so it is important to improve the accuracy of atmospheric chemical model predictions by combining satellite observations.
Data assimilation (DA) can integrate observational information and background fields, and it can give the optimal estimation of the atmospheric state (Guan et al., 2011). In the 1970s, DA was introduced into the field of air quality prediction and became a new direction for atmospheric environmental science research (Bai et al., 2007). However, the assimilation of aerosol data was conducted in the 1990s (Jacobson, 2001; Otte et al., 2005). With the enrichment of observational data and model improvement, efforts have also focused on satellite AOD DA. Atmospheric chemical models contain both aerosol particulate matter and gaseous chemical quantities, and current assimilation studies have focused on gaseous chemical quantities, such as ozone, methane, carbon dioxide, and sulfur dioxide (Zhang et al., 2008; Kukkonen et al., 2011). However, there are some limitations in the assimilation of aerosol particulate matter data. Firstly, the observational data are limited, and data reliability is still improving. Secondly, there is uncertainty in pollutant emission estimations, and the related chemical reaction processes in the model are imperfect. Compared with weather forecast models, atmospheric chemical prediction models have many chemical variables, and the aerosol particles are divided into different particle size segments, which increases the computational cost.
At present, there are several kinds of AOD DA technology, including optimal interpolation (OI), two-dimensional variational DA, three-dimensional variational (3D-Var) DA, the Ensemble Kalman filter, and four-dimensional variational DA. (Niu et al., 2008) assimilated Fengyun-2C satellite-derived dust and sand concentration data into the Global/ Regional Assimilation and Prediction System-Chinese Unified Atmospheric Chemistry Environment for Dust by using the 3D-Var method. They found that the 3D-Var method could effectively improve short-term sand and dust weather forecasting. (Adhikary et al., 2008) used the OI technique with a regional chemical transport model, while (Collins et al., 2001) used Model of Atmospheric Transport and Chemistry to simulate aerosol distributions; (Schroedter-Homscheidt et al., 2010) and (Nieradzik and Elbern, 2006) used variational approaches with European Acid Deposition models. (Liu et al., 2011) used the Gridpoint Statistical Interpolation (GSI) system to assimilate Moderate Resolution Imaging Spectroradiometer (MODIS) AOD data and compared with AOD observations from the Aerosol Robotic Network (AERONET) and Cloud-Aerosol Lidar with Orthogonal Polarization (CALIOP) instrument. More recently, (Yumimoto et al., 2016) showed the first application of aerosol optical properties derived from Himawari-8 data for aerosol DA. Their results showed that AOD data can improve air quality prediction.
The Fengyun-3A (FY-3A) meteorological satellite, a Chinese second-generation polar-orbiting meteorological satellite, has been widely used in meteorology, environmental protection, and other national sectors. It has ultraviolet, visible, infrared, and microwave multispectral instruments, and it has accumulated aerosol optical thickness inversion data since 2008 (Yang et al., 2009). In this study, we performed a DA experiment with AOD data derived from the Chinese FY-3A Medium-resolution Spectral Imager (MERSI) targeting transboundary pollutants and dust outflows over East Asia in April 2011.
2.1. AOD data
The FY-3A satellite, launched on 27 May 2008, is the first satellite of the second-generation polar orbital series in China, and it passes over the equator between 1000 and 1020 LST (Yang et al., 2009). In this study, the assimilated AOD data were obtained from FY-3/MERSI (Li et al., 2008). The FY-3A aerosol product includes AODs at three wavebands: 470 nm, 550 nm, and 650 nm. We used AOD data at the 550 nm wavelength, which has a spatial resolution of 0.01°. The assimilation time window was set to 3 h, to obtain maximum coverage.AOD data with a spatial resolution of 10 km from MODIS sensors onboard the Terra and Aqua satellites were used in our study. We used AOD retrievals derived from the Dark Target and Deep Blue products according to the inversion algorithm (Hsu et al., 2004; Hsu et al., 2006; Remer et al., 2005). In this study, we used collection 051 of level 2 total AOD retrievals from both Terra and Aqua. MODIS aerosol inversion products provide seven wavelengths of AOD data: 470 nm, 550 nm, 660 nm, 870 nm, 1240 nm, 1630 nm, and 2130 nm. In this study, we used AOD at 550 nm, to compare with the assimilation results of the FY-3 aerosol product.
An observation error specification for MODIS AOD data was suggested by (Remer et al., 2005), and we used the MODIS AOD retrieval uncertainty of 5% for AOD over oceans and 15% for AOD over land. We estimated FY-3A observation errors to be the retrieval uncertainty attached to the FY-3 AOD data plus a standard deviation calculated as the representative error in the regridding (Zhang et al., 2008). The FY-3 retrieval uncertainty ranged from 0.0001 to 0.93, with an average of 0.05. We only assimilated the highest quality AOD retrievals and thinned to the same resolution as the model grid. To reduce cloud contamination and noise in the data, pixels adjacent to missing values were discarded and only AOD values below 2.5 were used (Saide et al., 2014). In addition, ground-based AOD data acquired by AERONET (Holben et al., 1998) and the China Aerosol Remote Sensing Network (CARSNET) (Che et al., 2015) were used to evaluate the assimilation results. AERONET uses the CE318-type solar photometer for aerosol ground-based observations, and it has more than 500 sites worldwide. It provides the AOD ?ngstr\"om index, inversion parameter products, and precipitation data for different aerosol types in the world (Holben et al., 1998). Due to its high precision (error from 0.01-0.02), the AERONET data play an important role in the validation of aerosol satellite remote sensing and model products. In this study, we used the level-2 AOD data acquired by AERONET in Beijing, Lahore, Jaipur, and Taihu (Fig. 1) from 28 April to 3 May 2011. CARSNET is a ground-based network for monitoring aerosol optical properties, and it uses
Figure1. The experimental domain. Blue labeling indicates the ground-based AOD data acquired by the AERONET sites, and red the locations of the CARSNET sites, used in this study.the same types of instruments as AERONET (Che et al., 2009). It has 60 stations that are now operated by the China Meteorological Administration and local meteorological administrations, institutes, and universities throughout China (Che et al., 2015). It has become a national resource for studying aerosol optical properties for different regions in China, and it is used for validating satellite retrievals and aerosol numerical models (Xie et al., 2011; Zhao et al., 2013; Che et al., 2014; Lin et al., 2014). To evaluate the dust storm in mainland China, we used the ground-based AOD data acquired by CARSNET in Dunhuang, Datong, Xi'an, and Lanzhou from 28 April to 3 May 2011.
2
2.2. Aerosol DA system
In this study, we used a GSI 3D-Var meteorological DA system to assimilate FY-3 AOD data. The work in this study is based on Liu's (2011) work, and we added new interface to the FY-3A/MERSI AOD data. The 3D-Var assimilation method (Lorenc, 1986) is the process of minimizing the objective function, which can be expressed as: \begin{equation} \label{eq1} J({\textbf{x}})=\frac{1}{2}\{(x-x_{\rm o})^{\rm T}B^{-1}(x-x_{\rm b})+[H({\textbf{x}})-y_{\rm o}]^{\rm T}O^{-1}[H(x)-y_{\rm o}]\}, \ \ (1)\end{equation} where x o is the state vector composed of the model variables to be analyzed at every grid point of the 3-D model computational grid; x b is the background state vector; y o is usually provided by a previous forecast; y b is the vector of observations; J(x) is the cost function; H is an observation operator that establishes the relationship between model variables and observations with specific maps from the space of the model state vector to the space of the observation vector; and B and O are the background and the observation error covariance, respectively. By adjusting the weight between the background and observational data, the analysis fields achieved the best fit; the analysis fields were produced when the objective function reached the minimum value.2
2.3. Experimental design
A dust storm that started in Gansu blasted Beijing on 29 April 2011 and covered large areas of China in the following days. Our experiments were performed for the period of 28 April 2011 to 3 May 2011, during which a sand-dust storm affected China and surrounding areas. This storm influenced visibility, air quality, and human health in the eastern Tibetan Plateau, the southern Xinjiang Basin, the eastern part of Northwest China, mid-western Inner Mongolia, and North and Northeastern China. On 28 April, poor visibility of less than 100 m appeared in Jiuquan, while the Air Pollution Index measurements reached 500 in Lanzhou and Jinchang. The dust was lofted by strong winds accompanying a cold front that crossed China on 30 April. The winds passed over the regions of Mongolia, Xining, and Yan'an. From 2 May, the dust storm arrived in the Yangtze River Delta region, and then severe air pollution occurred in Shanghai, Nantong, Ningbo, Suzhou, Nanjing, and other cities.We performed four experiments to evaluate the impacts of FY-3 AOD DA on aerosol analysis and forecasting over China. The first experiment was a control and it had no DA (CNT). The other experiments had DA: the second and third experiments employed the FY-3 AOD DA (FY3 DA) and the MODIS AOD DA (MODIS DA), respectively. As we knew the FY-3 overpass time was similar to the Terra satellite, we hypothesized that the combined MODIS and FY-3 data would have a better fit to the observational data than the MODIS AOD data. Therefore, in the last experiment, both FY-3 AOD DA and MODIS AOD DA were assimilated (FY3 + MODIS DA).
Each experiment initialized the Weather Research and Forecasting with Chemistry (WRF/Chem) forecasting every 24 h, from 0600 UTC 27 April to 0600 UTC 3 May, using NCEP FNL data as the initial condition and then making 24-h forecasts. This group of experiments cycled for six days from 0600 UTC 28 April to 0600 UTC 3 May 2011. The assimilation frequency was 24 h. So, six assimilation cases were tested. In addition, the initial aerosol fields were produced by a three-day forecast. In other words, the spin-up period was three days, which was needed to overcome the unrealistic initial fields of the WRF/Chem forecasting.
The WRF/Chem model grid settings were set at 200 grids in the longitudinal direction and 150 grids in the latitudinal direction, with a grid interval of 40 km, and 40 layers in the vertical direction with a 50 hPa top level. The initial condition and lateral meteorological boundary conditions were obtained from the NCEP's global 11 reanalysis data, and the initial lateral chemical boundary conditions were obtained from 6-h simulation data from the NCAR's MOZART-4 model (Pang, 2012). The Goddard Chemistry Aerosol Radiation and Transport (GOCART) aerosol scheme (Chin et al., 2000) and the Regional Atmospheric Chemistry Mechanism-Kinetic Preprocessor gas chemistry scheme (Stockwell et al., 1997) were used as the aerosol and chemical process parameterization schemes for the model. In the WRF/Chem model, the following physical parameterizations (Stockwell et al., 1997, and references therein) were used: the Madronich photochemical process option, the Goddard shortwave radiation scheme, the RRTM longwave radiation scheme, the Lin microphysical scheme, the Noah land surface parameterization scheme, the Kain-Fritsch convective parameterization scheme, the Yonsei University planetary boundary layer scheme, and the Monin-Obukhov scheme for meteorological processes in the near-surface layer.
2
2.4. Background error covariance
In our AOD DA system, the background field was obtained from 14 aerosol species: hydrophobic black carbon (BC1) and hydrophilic black carbon (BC2) (particle size: 0.036 μm); hydrophobic organic carbon (OC1) and hydrophilic organic carbon (OC2) (particle size: 0.087 μm); four sizes of sea-salt particles (sea-salt 1-4; particle sizes: 0.3 μm, 1.0 μm, 3.25 μm and 7.5 μm, respectively); five sizes of dust particles (dust 1-5; particle sizes: 0.5 μm, 1.4 μm, 2.4 μm, 4.5 μm and 8.0 μm, respectively); and sulfate particles (particle size: 0.242 μm). The background error covariance was obtained by using the National Meteorology Center (NMC) method (Parrish and Derber, 1992). The NMC method uses the forecasts from two different time periods at a common time to estimate the background error covariance. We used differences of 24-h and 12-h WRF/Chem forecasts of the aerosol species valid at the same time for 62 valid pairs at either 0000 or 1200 UTC from 14 April 2011 to 28 April 2011 to compute the aerosol background error covariance.In this study, we used WRF data assimilation system's GEN_BE package (Barker et al., 2012) to calculate the NMC statistics. Only standard deviation and horizontal and vertical length scales of the background error for GSI analysis variables were needed to apply recursive filters both horizontally and vertically (Liu et al., 2011). Therefore, Figs. 2 and 3 show the regional average vertical profile with standard deviation and the regional average horizontal correlation length scale of background error covariances for 14 aerosol species, respectively. The standard deviations were closely related to the aerosol particle species at different levels, which reflected the uncertainty of the model predictions. The shape of the regional average vertical profile with standard deviation was consistent with (Benedetti and Fisher, 2007) and (Liu et al., 2011). In addition, the regional average horizontal correlation length scale, which represented the range of influence between the deviations of the observations and the background, differed among the 14 aerosol species. The values of the horizontal correlation length scale were 1-2.5 times the grid interval, which also matched the conclusions of (Kahnert, 2008).
Figure2. Regional average vertical profiles with standard deviations (units: μg kg-1) of the background error covariances for BC1, BC2, OC1, OC2, sea-salt 1-4, dust 1-5, and sulfate with different effective diameters.
Figure3. Regional average horizontal correlation length-scale (units: km) of the background error covariance for BC1, BC2, OC1, OC2, sea-salt 1-4, dust 1-5, and sulfate with different effective diameters.3.1 Comparison to FY-3 AOD and MODIS AOD
To examine the impacts of AOD DA on the aerosol analysis, we plotted the bias (Figs. 4a-c) and root-mean-square error (RMSE) (Figs. 4d-f) of the simulated AOD when using the FY-3 and MODIS observation data as reference data. The DA experiments were performed at 0600 UTC, when both the FY-3 and the MODISAOD data were present, from 28 April to 3 May 2011. As shown in Fig. 4 and Table 1, it is clear that the model underpredicted AOD (biases of -0.3 to -0.7). After DA, the model low bias was substantially reduced, with bias (RMSE) values consistently near -0.1 (0.2) in FY3 DA, while bias (RMSE) values were consistently near -0.2 (0.2) in MODIS DA and 0.1 (0.1) in FY3+MODIS DA. Overall, the bias and RMSE were reduced by an average of 30% compared to the CNT experiment during the dust storm period, which verified the positive effects of AOD DA systems. In addition, assimilation using different satellite data showed different effects: the bias and RMSE values calculated based on the FY3 DA experiment were reduced more than the values from the MODIS DA experiment. Furthermore, in the FY3+MODIS DA experiment, the bias and RMSE values were reduced more than the two individual data systems.We plotted the distributions of AOD observations and model simulations from the FY3 DA and MODIS DA experiments in Fig. 5. AOD represents contributions from all aerosol types, so we also plotted the distributions of the dust field (Fig. 6), which gives a direct indication of the dust storm event. Both the AOD and dust field give the same indication regarding the performance of the experiments. It is clear that the CNT experiment only simulated a large AOD distribution in central and southeastern Asia, an area with a tropical monsoon climate where fires on farmland and forest land occur frequently during the spring, leading to air pollution. However, the CNT experiment contained very little information on the dust weather in northern China. The FY-3 AOD data (Fig. 5a) contained more information than the MODIS AOD data (Fig. 5b) over East Asia. In particular, more information was captured about the dust outflows over most parts of China, while the MODIS AOD data (Fig. 5b) possessed information on the eastern regions. After assimilating satellite AOD data, the main region of the dust storm was added to the analysis fields (Figs. 5c and d). The FY-3 analysis field (Fig. 5c) showed that AOD data were distributed in northwestern Xinjiang, northeastern and southern Mongolia, southern Gansu, Qinghai, and most parts of Tibet. The MODIS analysis field (Fig. 5d) presented the distribution of AOD data in northeastern India, northern Burma, southeastern China, and Bohai Bay. After adjustment of the DA, the observational information in the analysis fields increased and the aerosol distribution of the analysis fields were closer to the satellite observation values. The results showed that there were good adjustment effects to the background field in our DA systems, and the FY-3 analysis field captured more information from the dust storm.
Figure4. Time series of the (a-c) bias and (d-f) RMSE of the simulated AOD. The DA systems involved (a, d) FY-3 AOD data, (b, e) MODIS AOD data, and (c, f) both FY-3 AOD and MODIS AOD data.
Figure5. Observations of 
Figure6. (a) Observations of 
Figure7. Observations at 550 nm from (a) FY-3 and (b) MODIS. True-color imagery, which is formed via a weighted combination of red, green, and blue (RGB) spectral information from (c) FY3 and (d) MODIS. The area in the rectangle represents the upstream area of the dust storm.Table 2 summarizes the number of satellite observations in the MODIS DA and FY3 DA experiments statistics over all six times. As shown in Table 2, the number of satellite observations in MODIS DA was more than that in the FY3 DA experiment. In addition, as shown in Fig. 6 and Fig. 7, we plotted true color images from both MERSI and MODIS on 1 May 2011 (0600 UTC). Both the true color images from MERSI and MODIS captured the dust storm over the northwestern areas as framed in Fig. 7, while FY-3A had more retrievals than MODIS. (Deng et al., 2016) and (Zhang et al., 2016) studied the quality of FY-3 AOD products in Guangdong and Shenyang, respectively. Overall, most of the areas covered by FY3 were dust-source areas, and also located upstream of the environmental field; although the MODIS satellite covers more in the east, it is neither the source of dust nor the upstream area of the environment. These two factors worked together to yield the benefits in the FY3 DA experiment.
2
3.2. Comparison with AERONET AOD
As shown in Fig. 8, under normal conditions without pollution, the impacts of DA were not significant. For example, when the AOD calculated in the CNT experiment was approximately 0.3, the AOD calculated in the assimilation experiment ranged from 0.4-0.5 (Figs. 8b and c). However, during the dust storm, the DA significantly improved the AOD simulation, resulting in much closer agreement between the analysis and ground-based AERONET AOD data at all stations. After assimilation, the AOD values in the experiment assimilating both FY-3/MERSI AOD data and MODIS AOD data were closer to the ground-based observations than the individual assimilation systems. However, all experiments did not adequately capture the observed AOD peaks. The emissions used for this study were suggested by (Zhang et al., 2009), who used the 2006 Asia emissions inventory. GOCART prognoses a global distribution of sulfate and its precursors, organic carbon, black carbon, mineral dusts and sea salt. AOD is determined from the dry mass concentrations and the mass extinction coefficients, which are functions of the size distributions, refractive indices, and RH-dependent hygroscopic growth of individual aerosol types (Chin et al., 2000). The aerosol species experience the processes of emission, advection, convection, diffusion, dry deposition and wet deposition. The dust was modeled in the WRF/Chem-GOCART by running the code of module_gocart_dust.F. Clearly, the larger AOD values from the DA experiment agreed more closely with the AERONET observations.2
3.3. Comparison with CARSNET AOD
As shown in Fig. 9, after assimilation of the AOD data, the trend of the analysis fields was generally more consistent with ground-based measurements than the control experiment. The control experiment underestimated AOD values, which is consistent with the results of (Zhang and Reid, 2006) and (Liu et al., 2011).China is the most populated and largest developing country in the world, and it has become one of the largest global sources for aerosol particles (Huebert et al., 2003; Seinfeld et al., 2004). The AOD values in the MODIS DA analysis also underestimated the dust storm, likely due to less coverage of the MODIS AOD data (e.g., Fig. 3b). The analysis results in the experiment assimilating both FY-3/MERSI AOD data and MODIS AOD data were more consistent with the ground-based values.
Figure8. Comparisons between AERONET retrievals and modeled results in the four experiments from 28 April to 3 May 2011, at the AERONET sites of (a) Beijing, (b) Lahore, (c) Jaipur, and (d) Taihu.
Figure9. Comparisons between CARSNET retrievals and modeled results in the four experiments from 28 April to 3 May 2011, at the CARSNET sites of (a) Dunhuang, (b) Datong, (c) Xi'an, and (d) Lanzhou.A summary of the statistics for the modeled, AERONET and CARSNET observed AOD comparisons is shown in Table 3. In general, the results showed good assimilation efficiency to improve the capability of the model to simulate the AOD over eastern Asia. The assimilation achieved improvements at all the eight sites as measured by the correlation coefficient and the bias between the model and the observation. Greater improvements were found over the sites that had more available assimilated observations, such as the surrounding area of Lanzhou. Interestingly, the assimilation yielded larger improvements at sites where both FY3 and MODIS had retrievals (Lahore and Jaipur).
We conducted assimilation experiments of FY-3AOD and MODIS AOD data, and the results showed that the assimilation of AOD satellite observation data had a good adjustment effect on the background field. After assimilation, the AOD distribution of the analysis field was closer to the empirical satellite observations. In addition, due to the abundant observational information from mainland China, the FY-3 DA analysis results agreed better with the ground-based values than the MODIS DA analysis results. Furthermore, the analysis results in the experiment assimilating both FY-3/MERSI AOD data and MODIS AOD data agreed better with the ground-based values than the other experiments.
Comparison of the DA of the MODIS and FY-3 aerosol products for the dust storm over East Asia indicated that the Chinese FY series satellite aerosol products are widely applicable in numerical models and dust weather analysis, which may increase the application and popularity of the Chinese FY series satellite aerosol products. Future work may be needed to assimilate and analyze multispectral, multi-sensor aerosol-related data. For example, the Earth Observation System satellite, the Advanced Very High Resolution Radiometer, the Visible Light Infrared Imaging Radiometer Suite, and the Multiangle Imaging Spectroradiometer are all capable of detecting and retrieving AOD. The Chinese FY series satellites FY-3A and FY-3B provide different wavelengths of land aerosols, marine aerosols, and dust detection, with daily and monthly products, which can also be used for further investigation.
