The spatiotemporal relationship between PM 2.5 and AOD in China: Influencing factors and Implications for satellite PM 2.5 estimations by MAIAC AOD

. Satellite aerosol retrievals have been a popular alternative to monitoring surface-based PM 2.5 concentration due to its extensive spatial and temporal coverage. Satellite-derived PM 2.5 estimation strongly relies on an accurate representation of the relationship between ground-level PM 2.5 and satellite aerosol optical depth (AOD). Due to the limitation of satellite AOD 15 data, most studies examined the relationship at a coarse-resolution (i.e., ≥ 10km) scale; more effort is still needed to better understand the relationship between ‘in-situ’ PM 2.5 and AOD at finer spatial scales. While ground PM 2.5 and AOD could have obvious temporal variations, few studies have examined the diurnal variation in their relationship. Considerable uncertainty therefore still exists in satellite-derived PM 2.5 estimation due to these research gaps. Taking advantage of the newly released fine-spatial-resolution satellite AOD data derived by the Multi-Angle Implementation of Atmospheric 20 Correction (MAIAC) algorithm and real-time ground aerosol and PM 2.5 measurements, this study explicitly explored the relationship between PM 2.5 and AOD and its plausible impact factors including meteorological parameters and topography in mainland China during 2019, at various spatial and temporal scales. Coefficient of variation, Pearson correlation coefficient and slope of linear regression model were used. Spatially, stronger correlations mainly occurred in northern and eastern China and linear slope in northern inland regions was on average larger than those in other areas. Temporally, such as in very high terrain i.e. Tibetan Plateau and north-central China i.e. Qinghai and Gansu. The Ffindings in this study have useful implications for satellite-based PM 2.5 monitoring and will further inform the understanding of the aerosol variation and PM 2.5 pollution status in mainland China. is 0.25%). For individual observational sites, ~96% of monitoring sites had an enhanced PM 2.5 -AOD correlations after both corrections, increasing by 0.01~0.61. The correlation after RH correction slightly increased by an average of ~0.07, and the vertical correction achieved a better improvement on PM 2.5 AOD relationship than the former, which contributed to the average r increase by ~0.18. Specifically, ~22% of monitoring sites, primarily located in Guangdong, Guangxi, Yunnan, and Heilongjiang provinces, had larger r values after RH 280 correction than vertical correction. These results indicate that the spatial difference existing in the influence of RH and vertical correction on the relationship between PM 2.5 and AOD cannot be ignored, due to their significant impact on the relationship in most monitoring sites. It should be noted that there was an obvious negative increment after both corrections, in particular in most monitoring stations in Yunnan province. that AOD a better indicator of fine particle pollution compared with MODIS aerosol product. In this study, surface PM 2.5 measurements in 2019 across China were spatially and temporally matched with MAIAC 1-km AOD, combined with humidity and vertical data and topography to perform explicit correlation analyses, which are the cornerstone for retrieving ground-level PM 2.5 from satellite aerosol loading. We found that, under the MAIAC high-resolution perspective, the relationship between surface PM 2.5 and column AOD was generally in a good agreement with 460 previous results but also shows a little difference in spatiotemporal variability, which should be cautious for applications to both aerosol-related and PM 2.5 -related studies. The findings in this study are helpful to better understand the PM 2.5 pollution status in China and also its relationship with column AOD, facilitating the reconstruction of high-resolution PM 2.5 from satellite aerosol retrievals.


Introduction 35
Aerosols play an important role in regional and global climate change through the direct and indirect effects (King et al., 1992;Li et al., 2007;Liu et al., 2019b;Liu et al., 2020;Liu et al., 2018). Fine particulate matter (PM2.5) has attracted loads of concern from scientists, policymaker and the public due to its negative impact on the environment Yang et al., 2018b;Yim et al., 2019a) and human health (Schwartz, 1996;Pope III et al., 2002;Gu et al., 2018;Gu and Yim, 2016;Gu et al., 2020;Hou et al., 2019;Shi et al., 2020;Yim et al., 2015) . With an extended spatial and temporal 40 coverage, the retrieval of near-surface PM2.5 concentration from satellite aerosol optical depth (AOD) has become a popular approach to bridging the gap left by ground-level monitoring network. The use of satellite AOD retrievals facilitates the detection of the large-scale and long-term aerosol loading and their transboundary transport and assists the determination of the population exposure level for epidemiological and health study Zou et al., 2019). The theoretical fundamental lies on the strong link between satellite AOD retrievals and ground-level PM2.5 concentrations (Wang and 45 Christopher, 2003). Therefore, it is of great importance to examine the PM2.5-AOD relationship, not only for the relationship itself, but also for the underlying physical understanding involved.
With the rapid urbanization and industrialization, China has been suffering from serious air pollution (Zhao et al., 2019;Guo et al., 2011;He et al., 2019). Given the heavy haze pollution in China (Luo et al., 2018;Tong et al., 2018b, a;Yang et al., 50 2019b;Yim, 2020;Yim et al., 2019b), observation studies have demonstrated and investigated the PM2.5-AOD relationship in China (Wang et al., 2014;Guo et al., 2017;Guo et al., 2009;Kong et al., 2016;Shao et al., 2017;Xin et al., 2016;Yang et al., 2019a). For instance, the Moderate Resolution Imaging Spectroradiometer (MODIS) Dark Target (DT) and Deep Blue (DB) 10-km AOD products have been widely used for quantifying the PM2.5-AOD relationship in both regional (Kong et al., 2016;Shao et al., 2017) and national scales (Guo et al., 2017;Yang et al., 2019a). The relationship is found not always 55 concordant and exhibits large spatiotemporal variability. With the development of satellite remote sensing technique, a global daily aerosol product at a high spatial resolution, i.e., the Multi-Angle Implementation of Atmospheric Correction (MAIAC) 1-km AOD product has been released recently, providing local-scale aerosol gradients and details missed by previous coarse-resolution (10 km) products (Lyapustin et al., 2018). Because the ground-level PM2.5 concentration is generally measured at an in-situ point while the satellite AOD is reported at a grid (e.g., with 1 km × 1 km for MAIAC AOD 60 and 10 km × 10 km for DT AOD), the relationship between a "point" PM2.5 and a "gridded" AOD at a coarse resolution might mask their discrepancy or correspondence at a finer resolution. That is, due to the presence of characterizing fine-resolution aerosol heterogeneity as well as the changes in retrieval scheme and accuracy (Guo et al., 2017), the use of the high-resolution AOD data probably introduces different/additional uncertainties into the PM2.5-AOD association, leading to a different spatiotemporal pattern and further affecting the performance of satellite-derived PM2.5 prediction. Thus, further 65 exploring the relationship between near-surface PM2.5 and MAIAC AOD over China seems critically imperative based on much finer AOD products.
The PM2.5-AOD link has been widely recognized to vary in space and time, which is associated with the definition difference between surface PM2.5 measurement and satellite AOD involved in aerosol types, meteorological condition, and topography 70 (Kloog et al., 2014;Lee et al., 2016a;Ma et al., 2014;Wang et al., 2014;Zhang and Li, 2015) . The ground-level air quality monitoring equipment contains a dynamic heating system to evaporate water vapor in particles and thus measures dry mass concentration for PM2.5 concentration, which is inevitably different with satellite AOD that interprets aerosol extinction without considering the humidity condition. The vertical structure of aerosols can induce uncertainty between PM2.5 and AOD because satellite AOD is an integration of aerosol extinction in the entire atmosphere while ambient PM2.5 75 concentration is only measured at the ground surface. Irrespective of different conditions, the meteorological factors may have confounding influence, causing the relationship to a different extent and even leading to a contrary result (i.e., improvement or deterioration). In addition, topography is another major contributor to the uncertainty existed in the PM2.5-AOD relationship but the impact is difficult to be quantified due to its complex structure. Therefore, it deserves further explicit analyses to explore the effect of terrain and meteorological factors, including hygroscopicity and vertical structure of 80 aerosols on PM2.5-AOD relationship, especially for the mainland China with a complex terrain and land cover types.
The diurnal cycle of surface PM2.5 concentration has various applications, e.g., the short-term health impact and aerosolcloud interaction (Arola et al., 2013;Guo et al., 2016;Kim et al., 2010;Lee et al., 2016b;Zheng et al., 2020)). To successfully capture the diurnal variation in the PM2.5-AOD relationship and represent the diurnal cycle of PM2.5, understanding the 85 diurnal variation involved in the data is crucial. However, it is easily ignored in studying the relationship between surface PM2.5 and satellite AOD due to the limitation of hourly satellite aerosol observations and the difficulty in making hourly PM2.5-AOD collocations. To the best of our knowledge, few studies have taken into account the diurnal cycle in the data when investigating the PM2.5-AOD relationship.

90
Therefore, this study aimed at comprehensively investigating the spatiotemporal relationship between ground-level PM2.5 and satellite AOD in the mainland China using the MAIAC 1-km aerosol product with multi-scale analyses in space and time, with an emphasis on the difference in the spatial variation and temporal trend against previous relationships dependent on coarse-resolution data. We carried out correlation analyses measured by Pearson correlation coefficient and slope of linear regression to explore the PM2.5-AOD relationship at 1500 monitoring stations, in more than 360 cities and eleven 95 urban agglomerations, by taking the latent impact of diurnal variation of aerosol particle, relative humidity (RH), planetary boundary layer height (PBLH) and elevation into account. In addition to monthly variation, seasonality and annual pattern, the real-time ground aerosol observations were also employed in this study to uncover the diurnal variability and impact on PM2.5-AOD relationship using hourly comparisons during a day when Terra and Aqua data are available. The remaining of this study is organized as follows: descriptions of the data used, including the ground-level PM2.5 measurements, the MAIAC 100 AOD, RH, PBLH and elevation in China, are given in Section 2. The results regarding correlation and linear slope analysis in PM2.5-AOD association, and its impact factors are presented in Section 3 and discussed in Section 4. The major conclusions are summarized in the final section.

Data collection and preprocessing 115
Hourly ground-level PM2.5 measurements from ~1500 monitoring stations for 2019 were collected from the official website of the Ministry of Ecology and Environment of China (https://www.mee.gov.cn). These observed PM2.5 mass concentrations are measured by the tapered-element oscillating microbalance or beta-attenuation method with quality control (National
The MODIS 1-km AOD data archived in Collection 6 derived by the MAIAC algorithm in 2019 over China were used in 125 this study. This new aerosol dataset not only represents aerosol retrievals at high spatial resolution (Chudnovsky et al., 2013;Lyapustin et al., 2018) but also achieves a comparable accuracy with the previous popular coarse-resolution aerosol product Liu et al., 2019a;Martins et al., 2017), making it possible to detect local aerosol variability. Both Terra (passing over at 10:30 am local time) and Aqua (crossover at 1:30 pm local time) MAIAC AOD data were downloaded from the NASA Level-1 and Atmosphere Archive and Distribution System (LAADS) Distributed Active Archive Center 130 (DAAC). To investigate the aerosol diurnal variation, the ground-level AOD observations from Version 3.0 Aerosol Robotic Network (AERONET) with assured quality (Level 2.0) were obtained.
Meteorological variables including RH and PBLH and terrain were used for impact factor analysis. The RH data were obtained from the China Meteorological Data Service Center (https://data.com.cn) and the PBLH data were downloaded 135 from the reanalysis dataset provided by the European Center for Medium-Range Weather Forecast (ECMWF, https://www.ecmwf.int/). The Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) Global Digital Elevation Model (GDEM) was collected to measure the terrain in terms of elevation.

Data collocation and analytical methods
According to previous studies Liu et al. (2019a), the MAIAC AOD data also confront the missing data problem as its coarse 140 spatial resolution. To tackle this issue and improve the number of daily collocations with ground PM2.5 measurements, an imputation method was conducted to fill the missing AOD values for one satellite when the other was available, and the improved AOD data served as the daily mean AOD for the collocation with daily PM2.5 averages. The imputation is similar to our previous studies (He et al., , 2019, through constructing a relationship between Terra and Aqua retrievals for each observation day and applying the obtained coefficients to predict the missing values. In addition, hourly collocation was 145 also accomplished to support the hourly analyses in this study, which directly collocated the hourly PM2.5 measurements at each observational site with the Aqua/Terra AOD retrievals within the hour of its overpass. All the daily and hourly collocations required the ground-level monitors spatially fell into the 1-km grid cell of AOD. That is, we used the 1-km grid of AOD to match the in-situ PM2.5 measurements and averaged those PM2.5 observations from multiple monitors located within a 1-km grid cell. . 150 The coefficient of variation (CV), defined as the ratio of the standard deviation to the mean to measure the dispersion of data points around the mean, allows us to compare the degree of difference between ground PM2.5 and columnar AOD. The relationship between ground PM2.5 and satellite AOD was quantified by Pearson correlation coefficients (r). The linear regression model was applied using AOD as an independent variable and PM2.5 a dependent variable. The estimated slopes 155 of the linear models were used for the statistical analysis.
To examine the impact of hygroscopicity and vertical structure of aerosols on PM2.5-AOD relationship, the RH correction and vertical correction were applied to obtain RH-corrected PM2.5 and PBLH-corrected AOD for the subsequent correlation analyses. Following previous studies (Kong et al., 2016;Wang et al., 2014), the RH-corrected PM2.5 and PBLH-corrected 160 AOD were calculated as follows, respectively: where 2.5 (μg/m 3 ) represents PM2.5 corrected by relative humidity; PM2.5 is the daily mean PM2.5 concentration measured by the ground-level air quality network; RH stands for relative humidity, ranging from 0 to 100; 165 (m -1 ) represents AOD corrected by PBLH and PBLH stands for planetary boundary layer height with unit of meter. (NCP), in particular in southeastern Hebei, western Shandong and northern Henan provinces with PM2.5 concentration up to 55 μg/m 3 or larger. It is noted that southern and western Xinjiang, the margin of Taklimakan Desert, had very high annual concentration exceeding 75 μg/m 3 , highlighted as one PM2.5 pollution center. In general, the MAIAC AOD shared a similar spatial pattern with PM2.5. That was, areas with high AOD values congruously correspond to those with high PM2.5 175 concentration except for typical mismatch in southwestern China. For example, higher mean AOD with values of 0.5-0.8 is found in Chengdu-Chongqing (CDCQ) agglomeration, which was as high as that in the NCP, contrary to the relatively low PM2.5 concentration with a mean of 38.93 μg/m 3 (mean concentration for NCP was 49.32 μg/m 3 ).

Characteristics of variation in surface PM2.5 and satellite AOD
[ Figure 1 is about here] 180 Figure S3 in SI shows the temporal profile of PM2.5 and AOD in the entire country in 2019. The national PM2.5 average was 38.53 μg/m 3 , and the annual mean AOD was 0.22. The monthly changes of the two indicators show a different temporal pattern. The monthly PM2.5 concentration shows a U-shape trend, which remained low and stable between June and August, and the monthly concentration in summer months was almost a third of those in winter months. Compared with winter, 185 summer is generally associated with lower human air pollutant emissions (e.g., reduced coal and biomass burning for heating during summer) and favorable meteorological conditions for pollution dispersion (e.g., enhanced convection and more precipitation during summer) (Zhang and Cao, 2015), which may account for the summertime minimum. This result is consistent with findings reported in previous researches (Yang et al., 2018a). In contrast, the monthly AOD variation shows a rough inverse U-shape curve with the plateau between March and August. The amplitude of monthly AOD variation was 190 also less than that of PM2.5: the maximum AOD level of 0.28 was occurred in March, which was 0.1 (~35.42%) higher than to the smallest AOD month in November. This further demonstrates the difference between ground-level PM2.5 and satellite AOD observations in Figure 1.

195
The diurnal characteristics of PM2.5 variations in China in the 2019 were investigated. In general, the national maximum PM2.5 diurnal mean concentration occurred in morning and evening, whereas the minimum mean values appeared in afternoon ( Figure S4 in SI), generally consistent with the diurnal pattern revealed in previous studies (Guo et al., 2017). The two peaks appeared in the overall diurnal PM2.5 curves were linked to PM2.5 spatial variation. Figure 2 shows the spatial distribution of the hours at which the maximum and minimum PM2.5 concentration occurred and the corresponding 200 concentration value. Among the 1494 monitoring stations, the largest diurnal PM2.5 concentration appeared in the evening (19:00-23:00) at 638 sites (~42.70%), followed by 567 sites (~37.95%) with the maximum concentration in morning (7:00-11:00). There were only 50 sites (~3.35%) reaching its maximum in afternoon (12:00-18:00), whereas the remaining 239 sites (~16.00%) had their maximum in midnight (00:00-06:00). To be more specific, the significant spatial variability in the highest diurnal PM2.5 concentration can be found. The morning PM2.5 peak generally occurred in the NCP while the evening 205 maximum prevailed in the southern and northeastern China where the afternoon PM2.5 maximum sporadically appeared.
Besides, the highest diurnal mean concentration in morning peak was generally greater than that in the evening, at an average of 46.60 μg/m 3 vs. 42.24 μg/m 3 . Among the 413 sites in NCP, there were more than 94% sites having concentration exceeding 35 μg/m 3 , pinpointing the serious PM2.5-pollution risk in the NCP region. We note that there was almost not any spatial difference in the minimum hourly mean concentration, which uniformly appeared between 14:00 and 17:00. 210 [ Figure 2 is about here] The CV analysis was conducted based on temporally averaged PM2.5 and AOD values to evaluate the spatial variation of aerosol particles over a given region. As shown in the upper panel of Figure 3, the ambient PM2.5 concentration in 2019 over 215 the CP region was the highest (multi-year mean value of 57.69 μg/m 3 ) associated with a moderate spatial variability (CV of 0.79), suggesting that the CP region was the most polluted among the 11 urban agglomerations and the majority of cities within this area were exposed to very dangerous PM2.5 levels. With the lowest PM2.5 concentration (27.43 μg/m 3 ) as well as the smallest CV value (0.58), almost all PRD cities experienced relatively clean ambient air. The PM2.5 concentration in NTS urban agglomeration remained a moderate level (45.96 μg/m 3 ), but the CV value was the largest (CV=1.25), implying a 220 distinct contrast in air quality between cities over the region. Figure 3 also presents a different regional pattern in both mean AOD and CV values.
[ Figure 3 is about here]

Spatiotemporal variability of PM2.5-AOD relationship 225
Figure 4 presents the spatial distribution of correlation coefficients (upper panel) and linear trend (lower panel) between the daily mean PM2.5 concentration and AOD in the entire study period and the four seasons over all monitoring sites. With correlation coefficient (r) values ranging from -0.5 to -0.18, abnormal negative associations between PM2.5 and AOD were observed in the Tibetan Plateau (TP) where the altitudes are almost higher than 4000m. In addition to TP, the correlation coefficient for each monitoring site ranged from 0.15 to 0.84. In general, the stronger correlations (r>0.4) were mostly 230 concentrated in the northern and eastern China, whereas r values for most observational sites dropped below 0.4 in southern areas.
Particularly, Sichuan Basin (SCB) was highlighted in Figure 4 (a) with the strongest correlation (r >0.6), whereas very low correlations (r<0.3) were found in northern Xinjiang and north-central China. Figure 4 (b-e) exhibits a distinct seasonal 235 variability of PM2.5-AOD even though the spatial pattern in each season is generally similar to that shown by all data pairs in Figure 4 (a). The overall correlation coefficient in winter was relatively higher, with overall r of 0.61, which demonstrates that the light extinction in the atmosphere in wintertime may be mainly caused by fine particles near the surface, then followed by spring (0.57) and autumn (0.55). The r value reached the lowest in summer (0.47). The distribution of correlations along with various PM2.5 levels may account for the seasonal difference. As shown in Figure S5 in SI, 240 correlations between PM2.5 and AOD were the strongest in extremely high PM2.5 concentration levels, and coincidentally, higher PM2.5 concentrations typically occurred in winter, thus leading to a better PM2.5-AOD relationship in this season.
Also, the seasonal variation was probably related to lower RH in winter and unstable atmosphere condition in summer.
Figure 4(f-j) shows the spatial distribution of slopes (S) of linear regressions between daily PM2.5 and AOD for each 245 monitoring site. Corresponding to the correlation analysis in Figure 4 (a-e), the slope values in TP were also negative, and for the other area over the mainland China, the slopes of linear regressions ranged from 9.37 to 158.45 over the entire observation period (Figure 4 (f)) with significant spatial and seasonal variabilities (Figure 4 (g-j)).
[ Figure 5 is about here] 250 To better represent the spatiotemporal features, correlation analyses between ground-level PM2.5 and satellite AOD have been conducted over all monitoring stations exhibited in Figure 4, through zooming in to urban-agglomeration scale in Figure 5. In general, the PM2.5-AOD relationships at the regional scale mirrored those at the observational-site scale. In terms of correlation coefficient, it reached higher values in urban agglomerations located in northern and eastern China such as the 255 GZP, BTH and CDCQ agglomerations with r values greater than 0.50; the weakest relationship with r of 0.26 occurred in the NTS agglomeration. As for the slope of the regression function, the maximum slope (81.10) was observed at the GZP region, followed by the CSL (51.48) and BTH (51.37) agglomerations; the lowest value was at the PRD urban agglomeration with a slope of 23.00. 260 Figure S7 in SI demonstrates that the spatial distribution of PM2.5-AOD correlations during the study period using Terra and Aqua AOD matched with PM2.5 measurements at respective overpass hour. It is intriguing that correlations for almost all monitoring sites at 2 pm were greater than those at 11 am, and the overall correlation at 2 pm was 0.54, which was also obviously higher than that of 0.47 at 11 am. To determine whether the difference is caused by aerosol retrieval accuracy from the Terra (i.e. at 11 am) and Aqua (i.e. at 2 pm), we conducted a comparison analysis using hourly ground PM2.5 and 265 the spatiotemporal collocated AERONET AOD observations. Figure S8 in SI indicates that the diurnal correlations between ambient PM2.5 and AERONET AOD were also larger in afternoon.

Influence of meteorology on PM2.5-AOD correlation
To examine the spatiotemporal effect of RH and PBLH on daily PM2.5-AOD relationship, we conducted correlation analyses using RH-corrected PM2.5 data and PBLH-corrected AOD data for monitoring sites ( Figure 6) and for the 11 agglomerations 270 (Table 1). The details of the correction method for PM2.5 and AOD have been provided in Section 2.3. In general, after correction by RH and PBLH (Figure 6), the mean correlation coefficient r reached 0.71, and the r values in ~ 16% sites increased to >0.80, which presents a notable improvement in PM2.5-AOD relationship compared to those based on uncorrected data (mean r=0.44 and the proportion of >0.80 is 0.25%). For individual observational sites, ~96% of monitoring sites had an enhanced PM2.5-AOD correlations after both corrections, increasing by 0.01~0.61. The correlation after RH 275 correction slightly increased by an average of ~0.07, and the vertical correction achieved a better improvement on PM2.5-AOD relationship than the former, which contributed to the average r increase by ~0.18. Specifically, ~22% of monitoring sites, primarily located in Guangdong, Guangxi, Yunnan, and Heilongjiang provinces, had larger r values after RH correction than vertical correction. These results indicate that the spatial difference existing in the influence of RH and vertical correction on the relationship between PM2.5 and AOD cannot be ignored, due to their significant impact on the 280 relationship in most monitoring sites. It should be noted that there was an obvious negative increment after both corrections, in particular in most monitoring stations in Yunnan province.
[ Figure 6 is about here] 285 There was also significant seasonal discrepancy existed in the meteorological effect on improving PM2.5-AOD relationship.
As shown in seasonal patterns of Figure 6, after both corrections, the correlations had the largest improvement in autumn with the mean r value increased by 0.21, followed by spring (0.11) and winter (0.10), and summer had the smallest improvement with an increase by 0.09. Specifically, in spring and autumn, both RH and vertical corrections imposed a positive effect on PM2.5-AOD correlation, but the strength was different. In spring, RH correction had a greater contribution 290 to improving PM2.5-AOD correlation compared with vertical correction, and the r value based on RH-corrected data was 0.02 larger at an average level. In autumn, vertical correction exerted a more important influence than RH correction, with the mean r value of 0.68 vs. 0.65. It should be noted that the correlation revised by PBLH was lower than the original condition in summer with the mean r value reducing by 0.05, especially for monitoring stations in central China such as Jiangxi, Shaanxi, and Henan provinces where significant decreases in r values (changing of <-0.15) occurred. 295 [ Table 1 is about here] Table 1 illustrates the statistics of PM2.5-AOD relationship based on both corrected and uncorrected data over the 11 urban agglomerations. It follows that there are also obvious spatial and seasonal discrepancies in the influence of humidity and 300 vertical distribution on regional PM2.5-AOD association. In summer, the correlation coefficient r of AOD, vertically revised by PBLH, and ground PM2.5 had a significant decrease over the original relationship, except for the NTS and YRD regions.
Despite of a general weak relationship in the NTS region, PBLH-corrected AOD was superiorly related to surface PM2.5 observations than the original correlation, not only in summer (0.29 vs. 0.21), but also in spring (0.42 vs. 0.37) and autumn (0.45 vs. 0.23). 305

Impact of terrain on PM2.5-AOD correlation
The effect of topography on the PM2.5-AOD correlation analyses was investigated by separating the collocated samples into 21 elevation bins with an interval of 200 m. As shown in Figure 7, a distinct trend and pattern can be identified between terrain and PM2.5-AOD relationship. Corresponding to the abnormal negative relationship over Tibetan areas found in Figure   4, the correlation coefficients between PM2.5 and AOD at a higher elevation, namely, greater than ~ 3200 m, were also less 310 than zero. The correlation peaked at 400-600 m with r of 0.60 and p-level<0.05. These samples were then overlaid onto the spatial map of topography ( Figure S9 in SI) and found that those having the maximum r values were primarily located in basin terrain such as Sichuan Basin, Shanxi-Shaanxi Basins and Junggar Basin.
[ Figure 7 is about here] 315

Discussion
Our results show a mismatch between satellite AOD values and ground-level PM2.5 concentrations, especially in southwestern China. The main cause for the PM2.5-AOD mismatch lies at the fundamental discrepancy in physical concept.
AOD is a unitless variable that denotes the total extinction induced by aerosols in the entire atmospheric column, whereas PM2.5 represents the mass concentration of fine aerosol particles measured at the ground level. Another reason was probably 320 attributed to the different available observations between the two metrics of aerosol loading: AOD derived from satellite only offered relatively fewer effective retrievals over this area due to cloud contamination ( Figure S2 in SI) while surface PM2.5 observation has almost no such limitation. Consequently, the limited satellite AOD samples cannot fully represent the actual aerosol loading. If those missing AODs are not filled in, excessive uncertainty will be induced to the subsequent PM2.5-related health research. Thus, a particular caution should be paid to the mismatch between satellite AOD and ground 325 PM2.5 caused by the amount of observations. We found that mean AOD and CV values had a different regional pattern. This result suggests that SDP was the uniformly polluted by aerosol particles throughout the atmosphere because the largest mean AOD value (0.48) was observed in this region with the smallest CV value (0.16). In addition, the regional CV in relation to PM2.5 generally exhibited larger values 330 than AOD, revealing that the extent of spatial non-stationarity shown by PM2.5 was obviously stronger compared to AOD in each individual region. Notably, the national CV of AOD was remarkably larger than the regional ones and contrarily higher than the PM2.5 CV, implying that the overall degree of spatial variability represented by ground-level PM2.5 measurements was weaker than that of satellite AOD at a national scale. It is probably due to the fact that the monitoring stations only sparsely located in urban areas while satellite AOD can almost cover the whole study domain. The PM2.5 measurements 335 represent the pollution level at a particular location while the satellite AOD interprets the average extinction level for 1-km "area". These further corroborate the mismatch between PM2.5 and AOD pattern observed in Figure 1 and Figure S3 in SI.
For the spatiotemporal variability of PM2.5-AOD relationship, Tibetan Plateau shows abnormal negative associations between retrieved PM2.5 and AOD. According to Lyapustin et al. (2018), the MAIAC algorithm did not make reliable aerosol 340 retrievals at very high elevations, and a fixed climatology value should be applied instead, thereby leading to the PM2.5-AOD relationship over TP contrary to other areas. While we found a large spatial discrepancy of correlation coefficient for each monitoring site that was consistent with previous studies using MODIS DT aerosol data (Guo et al., 2017;Yang et al., 2019a), the relative low r values (<0.4) in southern areas were slightly different from previous studies (Guo et al., 2017).
Based on MODIS DT AOD dataset, Guo et al. (Guo et al., 2017) reported that correlation coefficients between PM2.5 and 345 AOD in eastern China were apparently larger than those in the western China because DT algorithm had a larger uncertainty in the western arid and semi-arid areas. Unlike DT algorithm that can only derive reliable AOD values over surface with low albedos, MAIAC could provide aerosol retrievals in both bright and dark targets. As shown in Figure S2 in SI, MAIAC aerosol product can provide more retrievals over the northern arid and semi-arid areas than southern China. Moreover, according to (Liu et al., 2019a), the accuracy of MAIAC AOD in unoccupied land, built-up and mixed regions was evidently 350 better than that by DT. These above-mentioned evidences in turn result in relatively lower biases for the PM2.5-AOD relationship under these conditions and thus partly account for the difference of spatial pattern in PM2.5-AOD association between using MAIAC and DT AOD products. In addition to the accuracy and data availability of aerosol retrieval algorithms, the natural condition including climate and terrain also had obvious influences on the relationship between PM2.5 and AOD. For example, compared with southern China, the RH in the north was generally lower, which contributed to the harmony in the PM2.5-AOD association in terms of humidity condition; in addition, monsoon could cause aerosols unstable in subtropical climate zone, which may weaken correlations in these areas.
Obvious a spatial variation of PM2.5-AOD correlation and slopes of linear regressions between daily PM2.5 and AOD for each monitoring site were observed in our results. The largest correlation in Sichuan Basin was most likely due to its bowl 360 topography. The surrounding terrain is not favorable for pollutant dispersion and thus causes pollutant accumulation in the local environment (Ning et al., 2018;Ning et al., 2019;Ning et al., 2020). This explains a good agreement between groundlevel PM2.5 and total column AOD. For the low correlation in north-central China, e.g. Qinghai and Gansu provinces, MAIAC AOD retrievals were replaced with a mean value over a mesoscale area of 150 km (Lyapustin et al., 2018), which cannot resolve the 1-km aerosol variability as elsewhere and thus bring additional biases into the correlation between PM2.5 365 and AOD, thereby leading to incongruous associations between PM2.5 and AOD over these areas. Regarding the slopes of linear regression between daily PM2.5 and AOD, a larger slope suggests less aerosol extinction. The higher S values in inland areas and during winter indicate that, under these circumstances, AOD was not very sensitive to PM2.5 change. This was probably attributed to the spatial discrepancy and seasonal variation in a humid condition. Due to the increased hygroscopicity, the size of aerosol particles became larger than dry particles (Silva et al., 2015), which allowed more light to 370 be scattered in the context of higher RH, e.g., over the coastal areas and during summer. In addition to humid condition, aerosol type was another reason to account for the spatial variability in slope of PM2.5-AOD relationship. Previous studies have borne out that the regression function between ambient PM2.5 concentration and AOD corresponded with the typical air masses over a specific region (Kong et al., 2016), and aerosol extinction property cannot be as sensitive in areas dominated by coarse-mode aerosols (Kong et al., 2016;Xin et al., 2014). As indicated by the spatial distribution of mean Angström 375 exponent (AE or α) in Figure S6 in SI, over China, aerosol particles with a larger size (smaller AE) such as dust primarily dominated in northern areas while the fine-mode aerosols prevail in the south, thereby leading to a greater slope of the PM2.5-AOD relationship in northern China. It should be noted that the regional pattern of PM2.5-AOD relationship was a little different from the pattern by individual monitors because the regional association was an integration of observational sites within an urban agglomeration other than a simple averaging of all sites within the region. Moreover, we found that most of 380 regional correlations based on MAIAC AOD in this study were closed to (and even slightly higher than) those by MODIS DT AOD in previous studies (Guo et al., 2017;Yang et al., 2019a). For example, the correlation over the BTH, PRD, SDP, and CP region was 0.55, 0.38, 0.42, and 0.39, respectively, while the corresponding correlation in this study was 0.53, 0.30, 0.48, and 0.47, respectively. This was because MAIAC AOD (1 km spatial resolution) had a refined spatial resolution than DT AOD (10 km), thereby reducing the variability in aerosol inside a grid cell collocated with the in-situ measurements. The 385 improved overall retrieval accuracy performed by MAIAC algorithm Liu et al., 2019a) may be another reason, which can reduce the biases for the correlation analysis in PM2.5-AOD relationship. These findings suggest that MAIAC AOD product is a better indicator of fine particle pollution compared with MODIS coarse-resolution aerosol product.
We also found significant effect of meteorology on PM2.5-AOD correlation using Terra and Aqua AOD. Previous study has evaluated the MAIAC retrieval performance of Terra and Aqua AOD datasets against ground-level aerosol observations and found that AODs from the two satellites presented a similar retrieval accuracy . Thus, we deduced that the difference in correlation between the two times (11 am & 2 pm) was most likely due to the diurnal variability of climatology, e.g., PBLH. The diurnal cycle of PBLH was more likely one important factor leading to the noon and afternoon 395 maximum. It is because higher PBLH often observed in the noon and afternoon (Liu and Liang, 2010;Zhang et al., 2018) would definitely allow air pollutants to be mixed vertically such that the PM2.5-AOD relationship should be closer. At nighttime, the relative stable atmosphere may be favorable for the accumulation of aerosol particles near the ground surface, and thus AOD may not be able to reflect the ground-level aerosol concentration as well as they do in daytime. Our results show a significant improvement on PM2.5-AOD correlation after RH and vertical correlation, suggesting the importance of 400 applying the correction for the retrieval work. It should be noted that there was an obvious negative increment after both corrections, in particular in most monitoring stations in Yunnan province. The poorer correlation may be attributed to the vertical correction, because the single correction by humidity can slightly enhance PM2.5-AOD correlation that contributed to the r values rise by an average of 0.01 but the correlations revised by vertical data remarkably reduced by 0.04~0.42 for individual sites. Due to the East Asian and Indian monsoons, extraneous aerosols such as smoke and dust transported from 405 India, Burma and desert regions in northern China can form elevated aerosol layers in the Yunnan-Guizhou Plateau Liu et al., 2016), thus leading to the local AOD vertically corrected by PBLH poorly related to the ground-observed PM2.5 in this region. The effect of meteorology on PM2.5-AOD relationship improvement was also found to vary seasonally.
Nevertheless, our results show vertical correction in summer may not be useful to improve PM2.5-AOD in central China such as Jiangxi, Shaanxi, and Henan provinces. Central China is located in a typical subtropical monsoon climate zone, the 410 summer monsoon causes frequent occurrence of unstable weather and increases the dissipation of local aerosols associated with the scour effect of rainfall and atmospheric convection (Gong et al., 2017;Wang et al., 2011), which may account for the weaker PM2.5-AOD correlation corrected for PBLH. RH correction had a weaker influence in winter during which the averaged coefficient increased less than 0.01, which may be related to the dry climate in winter. These findings signify that seasonal variability in the data plays an important role on improving PM2.5-AOD relationship, and cannot be ignored, despite 415 the existing differences in corrected correlations. On the other hand, PBLH-correction can effectively improve PM2.5-AOD relationship in northwestern China. Contrary to the poor effect of vertical revision in summer and spring, seasonal PM2.5 revised by humidity was more closely linked to satellite AOD than the uncorrected relationship in almost all agglomeration areas, except for summer and autumn in the CDCQ area and winter in the BTH and CP regions. This further corroborates that regional pattern was a little different from the pattern presented by monitoring stations, and the RH and vertical 420 corrections did not improve the correlation between ground measured PM2.5 and satellite AOD with significant scope but should be tailored for specific areas.
Besides meteorology, terrain was also found to have impact on PM2.5-AOD correlation, especially in basin areas. It may be because basin terrain tends to form a local atmospheric environment, aerosol particles can mix well in the entire atmosphere. 425 Moreover, the complex topography protects the basins from transboundary transport of air pollution (Ning et al., 2019;Ning et al., 2020). These two reasons contributed to the consistency between near-surface PM2.5 and column AOD. Figure 7 also shows that the r values tended to become lower as the elevation rises, which indicates a deteriorated PM2.5-AOD correlation with the elevated altitude, especially in the eastern China where the majority of observational sites are in operation. Due to the spatially uneven distribution of air quality monitoring stations, there were only 23.68% pairs of the total matched 430 samples located at elevations higher than 1000 m, leading to Figure 7 and Figure S9 in SI mainly representing the eastern-China effect of terrain on PM2.5-AOD relationship.
At last, it is necessary to discuss the implication for ground PM2.5 derived from satellite AOD. With the release of MODIS MAIAC aerosol product, the high-spatial-resolution AOD dataset has become an emerging application to inferring fine 435 particulate matters in China. This study profiled the relationship between the ground-level PM2.5 and the 1-km satellite AOD in the perspective of spatiotemporal variations and analyzed the underlying factors that contribute to the current pattern.
Generally consistent with previous studies (Yang et al., 2019a;Shao et al., 2017;Wang et al., 2014), the PM2.5-AOD association in this study also presents a significant spatiotemporal variation. The meteorological parameters and topography have a complicated influence on the PM2.5-AOD relationship, posing extreme challenges for ground PM2.5 estimations. For 440 example, PBLH has an adverse impact on the relationship in Yunnan, whereas humidity may not be useful for improving the association in winter. Aerosol type is another contributor accounting for the uncertainties involved. However, as our results show, the pattern based on the 1-km MAIAC AOD is a little different from previous studies using coarse-resolution aerosol products such as the MODIS 10-km DT or DB AOD datasets (Guo et al., 2017;Yang et al., 2019a), which should be noticed for PM2.5 predictions. For example, the aerosol retrieval scheme implemented by the MAIAC algorithm at very high altitudes 445 brings about the unnatural negative PM2.5-AOD correlation over Tibetan region so that it has a direct effect on PM2.5 estimation ability. To clarify this issue, we established two separate multiple linear regression models using the same daily samples: one was for areas with elevation higher than 3200 m and the other was for the remaining area and obtained significant different R 2 of 0.14 vs 0.41 at p-level < 0.01. This result suggests that the PM2.5 prediction model should treat very high-altitude areas with caution; otherwise additional uncertainty within the data may hinder the acquisition of model 450 parameters and further attenuate the prediction accuracy. Therefore, it should be keeping a discreet attitude to apply the newly released aerosol product to PM2.5, especially for large areas with a complicated terrain and land cover. cornerstone for retrieving ground-level PM2.5 from satellite aerosol loading. We found that, under the MAIAC highresolution perspective, the relationship between surface PM2.5 and column AOD was generally in a good agreement with previous results but also shows a little difference in spatiotemporal variability, which should be cautious for applications to both aerosol-related and PM2.5-related studies. The findings in this study are helpful to better understand the PM2.5 pollution 460 status in China and also its relationship with column AOD, facilitating the reconstruction of high-resolution PM2.5 from satellite aerosol retrievals.
The relationship between PM2.5 and AOD varied notably in mainland China, both spatially and temporally, in terms of correlation coefficients and slopes of linear regression. Spatially, stronger correlations mainly occurred in northern and 465 eastern China and linear slope in northern inland regions was on average larger than those in other areas. Temporally, the PM2.5-AOD correlation peaked in the noon and afternoon and varied greatly in seasons, among which the maximum correlation happened in winter (r=0.61). MAIAC 1-km AOD can better represent the ground-level fine particulate matter in most domains with exceptions such as in very high terrain i.e. Tibetan Plateau and north-central China i.e. Qinghai and Gansu, where MAIAC does not interpret the actual 1-km aerosol loading, resulting in unnatural or weaker PM2.5-AOD 470 relationships for these domains. This merits more attention in the application of PM2.5 estimations.
As far as the confounding impact of meteorology on the relationship between PM2.5 and AOD was concerned, we found that simultaneously considering RH and PBLH in the relationship could improve the correlation but the effect of RH and PBLH on the relationship varied spatially and temporally, both in strength and direction. The diurnal cycle of PBLH coincided with 475 that of PM2.5-AOD correlation, and PBLH may impose an adverse impact on the daily relationship in Yunnan province and weaken the seasonal relationship in summer in central China such as Jiangxi, Shaanxi, and Henan provinces. Compared with PBLH, RH exerted a greater contribution to improving PM2.5-AOD correlation in spring and had a weaker influence in winter. In addition, the correlation peaked at 400-600 m primarily in basin terrain such as Sichuan Basin, Shanxi-Shaanxi Basins and Junggar Basin. Nevertheless, despite the publicly available high-resolution AOD data during the satellite era, an 480 accurate reconstruction of high-resolution from satellite remote sensing remains greatly challenging, not only due to the uncertainties existing in the satellite aerosol retrievals but also due to the uncertainties coherent in PM2.5 and AOD which covary with topography and meteorological conditions. Even though the individual effect of humidity, vertical structure of aerosols and terrain has been clarified, the synthetic impact of these factors could also be core to regulating the statistical relationship between near surface PM2.5 and total column AOD, which is necessary to be elucidated in the future. Therefore, 485 more work in this regard is in urgent need. In addition, aerosol properties such as aerosol types are plausible factor that could modulate the PM2.5-AOD relationship, which is worthy of more attention in the future.
Data availability. PM2.5 measurements are from the official website of the Ministry of Ecology and Environment of China