Analysis of Spatial and Temporal Variability of PM 10 Concentrations Using MODIS Aerosol Optical Thickness in the Pearl River Delta Region , China

Characterizing spatial and temporal variations of PM pollution is critical for a thorough understanding of its formation, transport and accumulation in the atmosphere. In this study, Aerosol Optical Thickness (AOT) data retrieved from a Moderate Resolution Imaging Spectroradiometer (MODIS) were used to investigate the spatial and temporal variations of PM10 (particles with aerodynamic diameters of less than 10 μm) pollution in the Pearl River Delta (PRD) region. Seasonal linear regression models between 1-km retrieved MODIS AOT data and ground PM10 measurements were developed for the PRD region with meteorological corrections, and were subjected to a validation against observations from the regional air monitoring network in this region from 2006 to 2008, with an overall error of less than 50%. Consistent with ground observations, the estimated PM10 concentrations from the regression models appeared to be highest in winter, lower in autumn and spring, and lowest in summer. A high PM10 concentration band was detected over the inner part of the PRD region, where heavy industries and dense populations are located. The shape and concentration levels of this band exhibit significant seasonal variations, which shift with synoptic wind direction, indicating different source regions and their contributions to the PM10 pollution in the PRD region. Several discrete “hot spots” were found in the southwest of the PRD region during spring and other seasons, where no ground measurements are available. The reasons for the formation of these hot spots are unclear, and further investigations are needed. Despite the limitations of this work, the results demonstrate the effectiveness of retrieving remote sensing data for characterizing regional aerosol pollution, together with ground measurements. The combination of satellite data and ground monitoring presented in this work can help in better understanding the sources, formation mechanisms and transport process of particulate matters on a regional scale.


INTRODUCTION
Particulate matter (PM) or aerosol is an important air pollutant, which is associated with adverse human health effects (Pope III et al., 2000;Sacks et al., 2011), deterioration in visibility (Cheung et al., 2005) and uncertain impacts on climate change (IPCC, 2007).In the past few decades, substantial efforts have been made to help identify chemical compositions at individual sites and emission characteristics from various sources (Putaud et al., 2004).However, in many regions, the formation and transport of regional aerosol pollution are still under the shadow due to its complexity and limited knowledge on its spatial and temporal variations (Harrison and Yin, 2000;Zhang et al., 2012).
Ground-based monitoring networks can provide important temporal and spatial information of air pollution, however, due to high operational costs, monitoring stations are generally limited in number, which are insufficient to characterize high resolution of spatial and temporal variations of pollutant concentrations, especially on a regional scale (Liu et al., 2010;Li et al., 2011).In recent years, air quality models have been largely developed on different resolution scales, which help to simulate concentrations of air pollutants from local to continental scale (Zhang et al., 2004;Holmes and Morawska, 2006;Che et al., 2011;Rodrigues-Silva et al., 2012).But the highly parameterizations in these models always pose it a challenging work to accurately predict PM concentrations, owing to large uncertainties in emission sources and their complex interplay with chemistry transformation and meteorology transport in the atmosphere (Zhang et al., 2006;Borge et al., 2010).
With the advancement of remote sensing technology and global data assimilation systems, observations made from satellite remote sensors may pose a cost-effective approach for determining ground-level particle concentrations (Liu et al., 2005;Li et al., 2011;Wang et al. 2011).For example, Liu et al. (2005) estimated ground-level PM 2.5 in the eastern United States based upon satellite remote sensing data using an empirical regression model with adjustments by meteorological conditions.Overall, the model can explain 48% of the variability in PM 2.5 concentrations.Li et al. (2011) discussed the potential applications of satellite data in air quality monitoring and forecasting in five Chinese cities, and over 50% of the hindcasts had percentage error ≤ 30%.Van Donkelaar et al. (2010) applied satellite-based aerosol optical depth to estimate global long-term average PM 2.5 concentrations.They found that the World Health Organization Air Quality PM 2.5 Interim Target-1 (35 μg/m 3 annual average) was exceeded over central and eastern Asia for 30% and for 50% of the population.The findings of these studies illustrate the strong potential of satellite remote sensing used in ambient air quality monitoring as a supplemental approach to ground networks and air quality modeling.
Located on the southeast of China, the Pearl River Delta (PRD) region has been regarded as a national cluster of megacities associated with intensive economic development and complicated regional air pollution problems (Chan and Yao, 2008;Zhang et al., 2008;Zheng et al., 2010).Particulate matter has been identified to be an important regional pollutant in the RPD region based upon ground observations and air quality models (Wang et al., 2008;Zheng et al., 2009).A few studies have been conducted previously to help understand the relationship between satellite AOT data and ground-based PM measurements in this region and other regions in China (Li et al., 2004;Kim et al., 2007;Li et al., 2010;Bian et al., 2011;He et al., 2012).For example, Bian et al. (2012) used MODIS-AOD data to trace the dust storm and compare with the numeric model results.Li et al. (2004) identified some meteorological variables to be important for AOT-PM regression model for the PRD region.He et al. (2012) used MODIS-AOD data (2000-2007) to study aerosol spatial and temporal distributions, as well as their variations with local meteorological conditions over East China.However, due to lack of ground-based monitoring data for the study periods, few of them have been validated with surface PM measurements at multiple locations.A regional monitoring network, composed of 16 sites in the PRD region, has been put into operation since 2005 for pollution monitoring (e.g., SO 2 , PM 10 , O 3 , NO 2 , CO), which reports hourly, daily, annual pollutant concentrations.This regional air quality monitoring network has not been fully used for establishing and validating the regression models between the satellite-retrieved AOT data and ground-based PM measurements yet, compared to applications in other regions with denser networks of ground monitoring such as North America (Liu et al., 2005;Li et al., 2011).With the accumulation of substantial information for ground-based PM concentrations, this network poses a good opportunity to characterize spatial and temporal variations of PM 10 concentrations on a regional scale by using satellite remote sensing data.
The objectives of this study are: (1) to develop regression models using 1 km × 1 km AOT data and ground-based PM 10 measurements from the PRD regional air quality network by incorporating meteorological, geographical, and seasonal conditions; (2) to validate the derived regression models by comparing the predicted PM 10 concentrations with ground-based PM 10 measurements; (3) to investigate spatial and temporal variations of PM 10 concentrations in the PRD region using the derived regression models; and (4) to provide suggestions on regional air quality monitoring and management.

km × km AOT Data
Observation data from Moderate resolution Imaging Spectroradiometer (MODIS), which is onboard satellite Terra with the overpass at 10:30 am, were obtained for a three-year period (from January 2006 to December 2008) from NASA's Goddard Earth Sciences Distributed Active Archive Center (http://modis.gsfc.nasa.gov).The observation data obtained from MODIS needed to be retrieved into Aerosol Optical Thickness (AOT) before it can be used in the regression model.Currently there are three operational algorithms to help retrieve aerosol information from MODIS data over oceans (Tanré et al., 1997), dark-target land surface (Kaufman, 1997) and bright land surface (Hsu et al., 2004).The algorithms related to ocean and dark-target land surface were conceived before Terra launch and were developed for deriving aerosols over ocean and land, respectively; while the retrieval algorithm for bright land surface was proposed post-launch focusing on surface with high reflectivity, such as deserts.A comprehensive description of these three algorithms and their global validations to groundbased data can be referred to Remer et al. (2005).Given the surface characteristics of the PRD region (largely covered by vegetation), the dark-target land approach was adopted to retrieve 1-km AOT for this region (Li et al., 2004;Levy et al., 2007).Fig. 1 shows annual averaged spatial distribution of MODIS-AOT over the PRD region from year 2006 to year 2008, and its seasonal variations.Generally, the retrieved AOT concentrated on the central part of this region, with AOT values higher in spring (March, April and May) and summer (June, July and August), lower in fall (September, October and November) and winter (December, January and February).For validation, the new 1-km MODIS AOT was compared with the ground-based AOT from the Aerosol Robotic NETwork (AERONET) within ± 30 min of the overpass time of the satellite.Results showed a good agreement between the 1 km MODIS AOT and AREONET AOT measurements with a correlation coefficient of 0.92 (see Fig. 2).

Ground-Based PM 10 Measurements
The ground-based PM 10 measurements were obtained  areas, and others are located in residential or mixed areas with different functions (see Table 1).Hourly averaged PM 10 concentrations used in this study were collected from 12 sites in the mainland cities of the PRD region from January 2006 to December 2008, while the site in Dongguan cannot be used due to a large amount of missing data arising from station maintenance.As the satellite Terra passes over the South China at 10:30 a.m.every day, hourly PM 10 concentrations between 10:00 a.m. and 12:00 a.m. were extracted every day and were averaged to be groundbased PM 10 concentrations for retrieving and validating the AOT-PM relationship, in order to reduce the effect of diurnal variations on the AOT-PM regression models.The 12 sites were split into two subsets, each of which consisted of 6 sties.One subset (denoted as 'Model development sites') was used for developing the AOT-PM regression models, while the other subset (denoted as 'Model validation sites') was used for validation of the model (see Fig. 3).

Meteorological Data
Meteorological data, including relative humidity, temperature, wind speed, and other parameters from January 1, 2006 to December 31, 2008, were obtained from automatic meteorological monitoring stations at 3-h intervals, and daily averaged meteorological data were extracted from the data at 11:00 a.m. to match the satellite crossing time.
The summary statistics for these meteorological data are given in Table 2.

METHODS
As the satellite-retrieved Aerosol Optical Thickness (AOT) is the integral of extinction coefficient from the surface (ground level) up to top of atmosphere, certain conversion factors need to be determined when retrieving the surface PM level (Liu et al., 2005;Wu et al., 2011).In earlier studies, only simple linear relationships, as shown in Eq. ( 1), Mixed residential/commercial/industrial a DGHG site is not used in this study due to a large lack of data during its maintenance.were established between satellite-retrieved AOT and corresponding ground-based PM concentrations (Chu et al., 2003;Wang and Christopher, 2003).
where (PM 10 ) j is the averaged PM 10 located in the j th grid cell, AOT j is the AOT value retrieved from MODIS observation data at the j th grid cell, c and d are fitted coefficients.
Lately meteorology conditions(such as relative humidity and boundary layer height) were found to be important in the regression models for AOT-PM relationship, as particle extinction properties can change significantly with different vertical mixing and aerosol hygroscopic growth (Kim et al., 2007;Wang et al., 2010;Wu et al., 2012).Gupta et al (2006) assessed particulate matter air quality over different locations across the global urban areas using 1 year of aerosol optical thickness (AOT) retrieved from the MODIS on board NASA's Terra and Aqua satellite along with ground measurements of PM 2.5 mass concentrations spread over 26 locations.They concluded that it was necessary to include meteorological data (e.g., relative humidity (RH), fractional cloud cover and height of the mixing layer) when applying satellite data for air quality research.Wang et al. (2010) retrieved satellite-based estimation of regional particulate matter (PM) in Beijing using vertical-and-RH correcting method, and the results showed a significant improvement on AOT-PM correlations after RH correction with the R 2 increasing from 0.43 to 0.77 for PM 10 , and from 0.35 to 0.66 for PM 2.5 .
In this study, both relative planetary boundary layer (PBL) and humidity (RH) were incorporated into the AOT-PM regression model for vertical correction and humidity correction in order to better characterize the relationships between AOT and PM 10 concentrations.

Vertical Correction
As mentioned before, the AOT is an accumulation of aerosol extinction coefficients along the vertical column in the whole atmospheric layer, while the PM concentration retrieved only represents the near surface property of aerosol, their correlation strongly depends on the vertical distribution of aerosols.Eq. ( 2) illustrates the relationship between the AOT and vertical distribution of aerosols extinction coefficient.
where τ a (λ) is the total Atmospheric Optical Thickness (AOT) at the wavelength of λ; k a (λ, z) stands for the k a at the altitude of z and the wavelength of λ; k a is the aerosol extinction coefficient.Assuming that the vertical distribution of aerosols follows the negative exponent form (Chu et al., 2003): where k a,0 is the aerosol extinction at the surface; H is the scale height of aerosol.Substituting Eq. (3) to Eq. ( 2), we have: Eq. ( 4) shows that the surface aerosol extinction could be calculated from AOT and aerosol scale height (H) approximately (Wang et al., 2010).Space and groundbased measurements have shown that the majority of particle mass loading resides in the lower troposphere, and the particle mass distribution below the planetary boundary layer (PBL) tends to be more homogeneous due to the active mixing.Hence, the scale height can be approximately represented by the PBL height (Liu et al., 2005).In this study, due to lack of daily PBL height measurements, seasonal averaged PBL heights were used.The averaged seasonal PBL heights used in this study were referred to the results from Li et al. (2002) and Xu et al. (2009).

Humidity Correction
Since the hygroscopic growth of particles significantly affects aerosol properties, the humidity correction should be introduced to reduce the impact of variations in humidity on the PM 10 concentrations.Hygroscopic growing factor, f(RH), is usually used to describe variations of the aerosol sizes from the dry circumstance (e.g., at the RH below 40%) to the wet circumstance at the ambient RH (Kotchenruther et al., 1999;Im et al., 2001).f(RH) can be expressed as: where a and b are empirical coefficients, which are determined by the aerosol types.The empirical coefficients a and b are based on previous studies of hygroscopic growing factor in the PRD region (Liu et al., 2008).
The dry aerosol extinction coefficient (κ a, dry ) can be obtained through the RH correction: According to the Mie theory, κ a is a function of the particle size distribution and the extinction characteristics of particles in the ambient air (Liou, 2004).After a series of deduction, the relationship between aerosol extinction coefficient and PM 10 concentration at the ground surface can be described as (Chu et al., 2003;Wang et al., 2010): where Q ext is the size-distribution integrated extinction efficiency, γ eff is the effective radius and ρ is the aerosol mass density.Suppose that the chemical composition and the size distribution of aerosols have little change under certain condition, ρ, Q ext and γ eff of particles are approximately constant, and then κ a will be proportional to the PM concentrations.Because the PM concentrations are measured under a dry condition, it is expected that κ a, dry correlates better with PM 10 concentration than κ a .Combining Eq. ( 6) and Eq. ( 7), the relationship of κ a, dry and PM 10 concentrations can be described as Wang et al. (2010): We applied these formulas to estimate the surface PM 10 concentrations from retrieved MODIS-AOT using vertical and humidity corrections for the PRD region.

AOT-PM 10 Regression Models
Six out of twelve ground monitoring sites in the PRD region (namely CHTH, GZLH, HZJG, SDDX, SXLY, and ZQCZ), were used for the development of AOT-PM 10 regression models, with locations shown in Fig. 3.There was no priority in selecting model development sites in term of data quality, as all of the sites are operated under the same QA/QC system.However, we took into account the locations of the sites to improve the spatial representativeness of the models.The six sites selected for model development are uniformly distributed in the space, so they were considered to be representative datasets to characterize spatial variations of PM concentration in the PRD region.
Three-year daily averaged PM 10 concentrations were extracted simultaneously at these sites with corresponding daily retrieved MODIS-AOT values for regression analysis from January 2006 to December 2008.The correlations between ground-measured PM 10 and satellite-retrieved AOT were investigated at different spatial scales (urban and suburban) and time scales (annual and seasonal).Table 3 summarized the correlation coefficients (R) between groundmeasured PM 10 and satellite-retrieved AOT at the "model development sites" before and after meteorological corrections.There were significant improvements on the correlations between ground PM 10 concentrations and MODIS-AOT after vertical and humidity corrections, with R increasing from 0.163-0.353to 0.234-0.549for all development sites in the PRD region.Fig. 4 showed the regression models for AOT-PM 10 in the PRD region with and without vertical and humidity corrections.The data points were more compact after meteorological correction; with the overall correlation coefficient R increasing from 0.234 (without correction) to 0.458 (with correction).This R value (0.458) is comparable to the correlation coefficients in other studies (with R ranges from 0.09-0.63)conducted previously in China (Guo et al., 2010;Wu et al., 2012).Thus, regression models with meteorological correction were used to investigate the AOT-PM 10 relationship in this study.
As shown in Fig. 5 and Fig. 6, the regression models exhibited different performance characteristics for different locations and seasons.Higher correlation coefficients and slopes were observed at urban sites (R = 0.466; Slope = 68.12)than suburban sites (R = 0.324, Slope = 37.98).Also, higher correlation coefficients were observed in summer (R = 0.453) and fall (R = 0.428) than those in spring (R = 0.392) and winter (R = 0.200).Such variations have also  been reported in other studies (Tian and Chen, 2010;Zha et al., 2010).For example, Tian and Chen (2010) suggested that the correlation between AOT and PM was stronger in the spring and summer while weaker in the fall and winter for the area of southern Ontario.Zha et al. (2010) found a minimum correlation coefficient between AOT and PM for the season of winter (R = 0.47), but a much stronger correlation (R > 0.80) in summer and autumn for the city of Nanjing in China.Despite similar results observed in these studies, different explanations were given for these seasonal variations.In Tian and Chen's study (2010), the wavelength, at which the AOT product was derived from MODIS, was expected to be closely associated with the correlation between AOT and PM; while Zha et al. (2010) reported that high pollution levels may reduce the accuracy of MODIS monitoring data in winter.Previous study also reported that MODIS retrieving algorithms has relatively good capability of retrieving fine mode aerosols (with effective radius between 0.1 µm to 0.25 µm), while poor ability for retrieving coarse mode aerosol (with effective radius between 1.0 µm to 2.5 µm) due to currently lack of MODIS data over high reflectance of natural dust sources (Dubovik et al., 2007;Santese et al., 2007).As the concentration levels of particles with different sizes vary largely over space and time (Fang et al., 2000;Chan and Kwok 2001;Namdeo and Bell, 2005), which may influence the accuracy of retrieving algorithms on PM for different locations and seasons.For example, coarse particles with large size (usually with diameter larger than 2.5 µm) were observed to have higher concentration levels in winter under dry continental winter monsoon and lower in summer under wet oceanic monsoon (Chan and Kwok, 2001), which may account for the relatively low coefficient correlations of AOT and PM in winter for the PRD region.However, this inference needs further investigation in the future with more available PM size measurements in this region.

Validation of AOT-PM 10 Models
The regression models developed in this study were validated against ground measurements at another six sites (denoted as "Model validation sites" in Fig. 3), which were not used for model development.Fig. 7 showed a detailed histogram of MODIS-estimated PM 10 concentrations and their corresponding values measured at these validation sites for each season.Overall, most of predicted PM 10 concentrations in all the seasons were comparable to the observations with the differences in 30 μg/m 3 .The relative errors of MODIS-estimated PM 10 were within ± 50% of the ground measurements for all sites at four seasons, ranging from -37.08% to 42.46%.These results were comparable to those in Li et al. (2011), in which the relative errors were within 100% for cities in China.Similar to other regression models, our model also tended to underestimate PM 10 at higher concentrations (> 100 μg/m 3 ), which might be caused by the poor representativeness of the PBL height in the regression models during the episodes.Most of higher PM 10 concentrations were observed to be strongly associated with unfavorable meteorological conditions, in which pollutants were compacted into a much lower PBL layer in a short time or accumulated due to vertical subsidence and horizontal stagnation (Yang et al., 2012).As we used the seasonal averaged PBL height to predict the ground concentrations of PM 10 , it might not reflect the sudden changes in PBL heights, and thus could not capture these episodes very well.Besides, it was observed that the relative errors of AOT-PM model had seasonal variations among sites in the PRD region.For example, the best prediction of surface PM 10 concentrations at ZHTJ site occurred in spring (relative error = -1.48%),while that for GZWQ happened in fall (relative error = -14.04%).This was probably caused by the different PM compositions, geographic characteristics and meteorological conditions at these sites during different seasons.For example, Wang (2012) reported that equal mass of species in PM (e.g., sulfate, nitrate, organic carbon, and elemental carbon) accounted for different portions to the total light extinction coefficients.In the PRD region, seasonal variations were found for these species at different sites (Yue et al., 2010;Peng et al., 2011;Huang et al., 2012), thus, these variations might lead to different light extinction coefficients of PM at different seasons and locations, which will further influence the image received by MODIS.Given that these validation sites were not included in the development of the regression models, the AOT-PM model established in this study showed an operational capability to predict the groundlevel PM 10 concentrations in terms of the magnitude.

Spatial and Temporal Distributions of PM 10
The validated seasonal linear regression models were then used to characterize the spatial and temporal variations of PM 10 concentration in the PRD region.Fig. 8 showed the MODIS-derived spatial variations of PM 10 concentration in the PRD region, together with ground observations in the form of histograms.The observations from ground monitoring network exhibited the heterogeneity of PM pollution among discrete sites, while the MODIS-retrieved PM 10 provided a more comprehensive map on the spatial distributions of PM pollution over the whole region.Results derived from MODIS showed that the PRD region was shrouded by severe particulate pollutions, with most of areas exceeding new annual Chinese PM 10 National Standard (70 μg/m 3 ) for all seasons expect summer.The PM 10 pollution was heavy at the inner part of the PRD region with concentrations going up to 120 μg/m 3 in winter.This inner part was located with heavy industries and dense population, together contributing to a high production of primary particles (Zheng et al., 2009).Besides, the land-sea breeze circulation over this area also favored the formation and accumulation of the secondary PM pollution (Lo et al., 2006).The shapes and levels of the high concentration band changed with seasons.This band moved northward with relatively low PM 10 concentrations (< 70 μg/m 3 ) in summer, while it tilted to the west with high concentrations in autumn (around 100 μg/m 3 ) and winter (around 120 μg/m 3 ).These seasonal variations can be closely related to the synoptic wind flow in the PRD region.In summer, most of this region is under southwesterly monsoon, which helps to dilute the surface pollution with fresh and clean air from the ocean; while in autumn and winter, this region is subjected to northeasterly monsoon due to continental outflow, which can help transport the pollution from inner continent (Kowk et al., 2010).In spring, a transition period between summer and winter monsoon, the high concentration band was divided into several discrete spots surrounding the PRD estuary (90-110 μg/m 3 ).This phenomenon was argued to be influenced by local circulations such as land-sea breeze, which usually caused an accumulation of the pollution at the coastal areas (Lo et al., 2006).
Other areas of the PRD region exhibited similar seasonal variations as the inner part, with PM levels highest in winter, lower in spring and autumn, and lowest in summer, which were consistent with ground observations and modeling studies in this region as well (Cheung et al., 2005;Louie et al., 2005).In addition, high PM 10 concentrations were also found on the southwest of the PRD region during spring, fall and winter, distributed discretely in the city of Jiangmen.As those areas were not covered by current air quality monitoring network, reasons for the formation of these "high spots" are not clear and need to be further investigated with field measurements.ground observations at these sites from January 2006 to December 2008.The MODIS-derived estimations depicted a clear seasonal trend of PM 10 variations with lowest in summer (June, July, and August) and highest in winter (December, January and February), which were consistent with ground observations.Daily variations of PM 10 concentration were also very significant for the PRD region.The highest daily PM 10 concentration recorded by the ground observations can go up to 250 μg/m 3 while lowered to less than 50 μg/m 3 under some conditions.Our models were capable of reproducing daily variations; however, the scales of the variations were much smaller than those in the observations.The major reason contributing to the small variations is probably that our model overestimated daily PM 10 concentrations under relatively clean conditions (PM 10 < 50 μg/m 3 ) and underestimated for heavy pollution events (PM 10 > 150 μg/m 3 ).The underestimation was more severe in our model prediction, as shown in Table 4, the overall comparison of our model predictions with observations gave negative relative errors at validation sites.Similar situation has been reported in previous studies that MODIS systematically underestimated AOT under heavy pollution conditions partially due to its poor ability to address different aerosol types in the MODIS retrieval algorithms (Li et al., 2011).Presumed aerosol type in the retrieval algorithms may lead to errors in AOT-PM 10 retrievals and thus affect the PM 10 estimation.Incorporation of aerosol composition is necessary in the aerosol retrieval process, to reduce these negative biases under high pollution events.In this study, as mentioned above, lack of daily PBL height data might be another major reason leading to temporal under-estimation or overestimation.

Implications for Air Quality Monitoring and Management
Ground measurements generally conclude the spatial and temporal heterogeneity of PM pollution among discrete sites in the PRD region (Fig. 9).However, the limited number of measurements poses it a challenging work to fully understand the degree of spatial and temporal variability of PM 10 in this region, which is complicated by a number of factors, including local emissions, regional transports, geographic and meteorological conditions.With a wide swath of spatial coverage, MODIS aerosol retrieval has been demonstrated as a useful supplementary to understanding regional and local aerosol pollutions in the PRD region.It can help to depict the spatial distributions and temporal variations of the pollution over the whole region, investigating the influence from emissions, geographic and meteorological conditions in the atmosphere.The retrieval product is especially useful in determining the relative pollution burden in areas without ground measurements, which can be used to support future monitoring construction plan in this region.For example, one potential "hot spot" has been identified in Jiangmen in this study with high PM concentrations  The retrieval product can also play an important role in monitoring the regional pollution transport, which can help identify the contributions from regional and local sources.Due to its large spatial coverage, uninterrupted measurements and long period records, the retrieved MODIS aerosol can provide spatial patterns of surface PM pollutions for both long term and short term, which can help to trace the transport process and identify source regions.Results in our study showed that the high concentration band moved northward under southwesterly monsoon in summer with relatively low PM 10 concentrations (< 70 μg/m 3 ), which indicated that the major contributor to the PM 10 pollution in summer might be local emission sources in the PRD region.While this band tilted to the west with high PM 10 concentrations in autumn (around 100 μg/m 3 ) and winter (around 120 μg/m 3 ) under northeasterly winds due to continental outflow, suggesting the super-regional contributions from outside of PRD region and the importance of cooperation with surrounding regions for the air quality improvement.The retrieval of MODIS aerosol can also provide information for the causation and transport routes for the short term episodes, which needs to be further investigated in the future.

CONCLUSIONS
Characterizing spatial and temporal variations of PM pollution is critical for a thorough understanding of its formation, transport and accumulation in atmosphere, which is important in air quality monitoring and management.Seasonal AOT-PM regression models were established for the PRD region to investigate spatial and temporal variations of PM 10 pollution using three-year observations from satellite and measurements from 6 ground monitoring sites in this region, with corrections by meteorological conditions (relative humidity correction and vertical correction).These models were validated with ground measurements, with overall relative errors less than 50%.
Significant seasonal variations were found for surface PM 10 concentrations in the PRD region, with PM levels highest in winter, lower in spring and autumn, and lowest in summer.A high concentration band was observed over the inner part of this region, where heavy industries and dense population were located.This high concentration band moved northward with relatively low PM 10 concentrations (< 70 μg/m 3 ) under southwesterly clean ocean monsoon in summer; while it tilted to the west with high PM 10 concentrations in autumn (around100 μg/m 3 ) and winter (around 120 μg/m 3 ) under northeasterly winds due to continental air masses.Some "hot spots" were detected on the southwest of the PRD region (inside Jiangmen city) during spring and other seasons, which were not covered by the current monitoring network.Further investigation is needed for a better understanding of their formation via field measurements or modeling work.
Limitations need to be aware of when applying the aerosol retrievals for surface PM monitoring.First of all, no continuous measurements of satellites observations are available.As mentioned before, the on board satellite (Terra or Aqua) only overpass a region at certain time of the day, and thus cannot characterize diurnal variations of PM pollutions.Also, aerosol retrievals can only be made over cloud-free scenes and sometimes bias are introduced when there are undetected clouds (Gupta et al., 2006;Li et al., 2011).Uncertainties also rise in the retrieval process from the assumptions on aerosol properties, surface and meteorological conditions.For example, several meteorological parameters are suggested to be important on aerosol properties, but their impacts on the retrieved surface PM concentrations are location-dependent and not well characterized in the current regression models (Wang et al., 2010).
Despite these limitations, our results demonstrated the capability of using retrieved remote sensing data for monitoring surface PM and their important roles in characterizing spatial and temporal variations of regional aerosol pollution, together with ground measurements.The satellite retrieved aerosol map can capture the synoptic nature of pollution events and help to identify the sources of pollutants on a large scale, which can make up to the limitations of spatial coverage by ground monitoring.Further studies are needed on how to better utilize both remote sensing data and ground monitoring in order to have a better understanding of emission sources, formation mechanisms and transport processes of particulate matters in this and other regions.

Fig. 7 .
Fig. 7. Comparsions of predicted PM 10 concentrations with obversations at the validation sites by seasons.

Fig. 9
Fig. 8. Spatial distributions of predicted PM 10 concentrations and obversations by seasons.

Table 1 .
Location information on monitoring sites in the PRD network.

Table 2 .
Summary statistics of mean AOT, hourly PM 10 concentration, PBL Height, and RH for the "model development sites".Mean 1 km × 1 km MODIS AOT in the PRD region; c Hourly PM 10 mass concentration between 10:00 to 12:00 a.m.local time; d Mean relative humidity at surface between 10:00 to 12:00 a.m.local time; b e Average PBL by seasons.

Table 3 .
Correlation coefficient ranges between observed PM 10 concentrations and AOT.