Assessment of MERRA-2 Surface PM2.5 over the Yangtze River Basin: Ground-based Verification, Spatiotemporal Distribution and Meteorological Dependence

A good understanding of how meteorological conditions exacerbate or mitigate air pollution is critical for developing robust emission reduction policies. Thus, based on a multiple linear regression (MLR) model in this study, the quantified impacts of six meteorological variables on PM2.5 (i.e., particle matter with diameter of 2.5 μm or less) and its major components were estimated over the Yangtze River Basin (YRB). The 38-year (1980–2017) daily PM2.5 and meteorological data were derived from the newly-released Modern-Era Retrospective Analysis and Research and Application, version 2 (MERRA-2) products. The MERRA-2 PM2.5 was underestimated compared with ground measurements, partly due to the bias in the MERRA-2 Aerosol Optical Depth (AOD) assimilation. An over-increasing trend in each PM2.5 component occurred for the whole study period; however, this has been curbed since 2007. The MLR model suggested that meteorological variability could explain up to 67% of the PM2.5 changes. PM2.5 was robustly anti-correlated with surface wind speed, precipitation and boundary layer height (BLH), but was positively correlated with temperature throughout the YRB. The relationship of relative humidity (RH) and total cloud cover with PM2.5 showed regional dependencies, with negative correlation in the Yangtze River Delta (YRD) and positive correlation in the other areas. In particular, PM2.5 was most sensitive to surface wind speed, and the sensitivity was approximately −2.42 μg m−3 m−1 s. This study highlighted the impact of meteorological conditions on PM2.5 growth, although it was much smaller than the anthropogenic emissions impact.


Introduction
Since its reform and opening up (1978 onward), with the rapid development of urbanization and industrialization, China has become one of the most polluted areas [1-6]. PM 2.5 (i.e., particle matter with diameter of 2.5 µm or less) is an important metric for air pollution, which has attracted considerable attention. When these particles are breathed in, they can lead to premature deaths [7][8][9][10]. In addition, they can effectively scatter and absorb solar radiation, thereby reducing atmospheric visibility and producing significant climate impacts [11][12][13][14][15][16][17][18][19]. Table 1. Empirical Studies of the impacts of meteorological conditions on PM 2.5 (i.e., particle matter with diameter of 2.5 µm or less) concentrations.
Saudi Arabia (2014-2015) [31] MLR Relative humidity, wind speed and temperature PM 2.5 , PM 10 Both PM 10 and PM 2.5 were anti-correlated with relative humidity, but positively related to wind speed and temperature.
Beijing China (2005-2013) [32] MLR Relative humidity and wind speed Black carbon (BC) High BC concentrations were usually related to poor visibility and low wind speed.
China (2013-2014) [33] Multivariate analysis Relative humidity, wind speed, surface pressure and temperature PM 2.5 PM 2.5 was significantly correlated with meteorological conditions in China. * SO 4 , OC, EC, NOx refer to sulfate, organic carbon, element carbon and nitrogen oxides, respectively.

Data
The MERRA-2 atmospheric reanalysis product was newly released by the NASA Global Modeling and Assimilation Office (GMAO) in 2017. Based on the NASA GMAO Earth system model version 5 (GEOS 5) and the Goddard Chemistry, Aerosol, Radiation and Transport (GOCART) aerosol module, the gridded aerosol data of MERRA-2 were assimilated with satellite and ground observations. More details on MERRA-2 PM 2.5 treatment could be found in [37][38][39][40], and were briefly described here. MERRA-2 PM 2.5 concentrations were calculated by Equation (1): PM 2.5 = 1.375 × SO 4 + 1.6 × OC + BC + Dust 2.5 + SS 2. 5 (1) Here, SO 4 , OC, BC, Dust 2.5 and SS 2.5 represented sulfate, organic carbon, black carbon, dust and sea-salt particulate matter with a diameter of less than 2.5 µm from the GOCART aerosol module, respectively. Notably, the nitrate particulate matter primarily emitted by vehicle exhaust and industrial production was lacking in the MERRA-2 PM 2.5 reanalysis, probably leading to biases when compared with ground measurements. In order to verify the performance of the MERRA-2 PM 2.5 product, we compared MERRA-2 PM 2.5 with two-year (2015-2016) daily mean PM 2.5 ground observations from 476 state-controlled air quality monitoring stations over the YRB (Figure 1). By using the tapered element oscillating microbalance method and/or the beta absorption method, these state-controlled air quality monitoring stations provided hourly and 24-hour average concentrations of six major air pollutants (including PM 2.5 , PM 10 , SO 2 , NO 2 , O 3 and CO) after 2013 [40]. The systematic uncertainty Remote Sens. 2019, 11, 460 5 of 25 of ground PM 2.5 from these air quality monitoring stations was controlled within 15% in accordance with the latest version of China's environmental protection standards [40].
In addition, the spatial and temporal distributions of MERRA-2 PM 2.5 were further compared with another PM 2.5 reanalysis product released by the Atmospheric Composition Analysis Group of NASA. Similarly, based on the GEOS-Chem chemical transport model, the satellite-retrieved PM 2.5 concentrations (0.01 • × 0.01 • ) were estimated by combining column integrated AODs from MODIS (Moderate Resolution Imaging Spectroradiometer), MISR (Multiangle Imaging Spectroradiometer) and SeaWIFS (Sea-Viewing Wide Field of View Sensor) instruments [41,42]. Then, in order to adjust the residual bias, the satellite-derived PM 2.5 concentrations were calibrated by ground-based observations using a geographically weighted regression (GWR). Hereafter, the PM 2.5 reanalysis product was abbreviated as GWR PM 2.5 in this study. As shown in Figure A1, the annual mean GWR PM 2.5 concentrations were compared with ground PM 2.5 measurements from 476 state-controlled air quality monitoring stations over the YRB during 2015-2016. A high correlation coefficient (R = 0.76) demonstrated a good consistency between GWR PM 2.5 and ground PM 2.5 , which provided us enough confidence to use the GWR PM 2.5 product to verify the spatiotemporal distribution of MERRA-2 PM 2.5 .
Since both the non-bias-corrected AODs (MISR and AERONET) and the bias-corrected AODs (MODIS and AVHRR) were assimilated in the MERRA-2 aerosol reanalysis, these above AOD datasets were all not suitable for verifying the MERRA-2 AODs. Therefore, the daily ground AOD observations from five China Aerosol Remote Sensing Network (CARSNET) sites in the YRB during 2013-2016, were selected to be compared with their corresponding MERRA-2 AODs at 550 nm ( Figure 1 and Table 2). The CARSNET was a ground aerosol network with approximately 50 sites across China, established by the China meteorological administrations and university research institutes [9,12]. It used Cimel sun photometers to retrieve the Level-1.0 (unscreened) and Level-1.5 (cloud-screened) AOD data at eight wavelengths (340, 380, 440, 500, 675, 870, 1020 and 1640 nm) every 15 minutes [9]. The total uncertainty in the CARSNET AODs was about 0.01 to 0.02 [12]. The CARSNET AODs used in this study were all cloud-screened data. As shown in Figure 1, from east to west, the five CARSNET sites were Lin'an (rural site, 39 m above sea level), Wuhan (urban site, 15 m above sea level), Changde (rural site, 38 m above sea level), Chengdu (urban site, 485 m above sea level) and Shangri-La (remote site, 3280 m above sea level) [12].
Remote Sens. 2019, 11, x FOR PEER REVIEW 5 of 27 concentrations of six major air pollutants (including PM2.5, PM10, SO2, NO2, O3 and CO) after 2013 [40]. The systematic uncertainty of ground PM2.5 from these air quality monitoring stations was controlled within 15% in accordance with the latest version of China's environmental protection standards [40]. In addition, the spatial and temporal distributions of MERRA-2 PM2.5 were further compared with another PM2.5 reanalysis product released by the Atmospheric Composition Analysis Group of NASA. Similarly, based on the GEOS-Chem chemical transport model, the satellite-retrieved PM2.5 concentrations (0.01° × 0.01°) were estimated by combining column integrated AODs from MODIS (Moderate Resolution Imaging Spectroradiometer), MISR (Multiangle Imaging Spectroradiometer) and SeaWIFS (Sea-Viewing Wide Field of View Sensor) instruments [41,42]. Then, in order to adjust the residual bias, the satellite-derived PM2.5 concentrations were calibrated by ground-based observations using a geographically weighted regression (GWR). Hereafter, the PM2.5 reanalysis product was abbreviated as GWR PM2.5 in this study. As shown in Figure A1, the annual mean GWR PM2.5 concentrations were compared with ground PM2.5 measurements from 476 state-controlled air quality monitoring stations over the YRB during 2015-2016. A high correlation coefficient (R = 0.76) demonstrated a good consistency between GWR PM2.5 and ground PM2.5, which provided us enough confidence to use the GWR PM2.5 product to verify the spatiotemporal distribution of MERRA-2 PM2.5.
Since both the non-bias-corrected AODs (MISR and AERONET) and the bias-corrected AODs (MODIS and AVHRR) were assimilated in the MERRA-2 aerosol reanalysis, these above AOD datasets were all not suitable for verifying the MERRA-2 AODs. Therefore, the daily ground AOD observations from five China Aerosol Remote Sensing Network (CARSNET) sites in the YRB during 2013 -2016, were selected to be compared with their corresponding MERRA-2 AODs at 550 nm ( Figure 1 and Table 2). The CARSNET was a ground aerosol network with approximately 50 sites across China, established by the China meteorological administrations and university research institutes [9,12]. It used Cimel sun photometers to retrieve the Level-1.0 (unscreened) and Level-1.5 (cloud-screened) AOD data at eight wavelengths (340, 380, 440, 500, 675, 870, 1020 and 1640 nm) every 15 minutes [9]. The total uncertainty in the CARSNET AODs was about 0.01 to 0.02 [12]. The CARSNET AODs used in this study were all cloud-screened data. As shown in Figure 1, from east to west, the five CARSNET sites were Lin'an (rural site, 39 m above sea level), Wuhan (urban site, 15 m above sea level), Changde (rural site, 38 m above sea level), Chengdu (urban site, 485 m above sea level) and Shangri-La (remote site, 3280 m above sea level) [12].

Comparison Analysis
According to [40], grid-level daily mean MERRA-2 PM 2.5 concentrations were compared with the corresponding ground measurements during 2015-2016 over the YRB by using a series of linear regressions. The regression parameters include slope, y-intercept, correlation coefficient (R), root mean square error (RMSE, Equation (2)), mean absolute error (MAE, Equation (3)) and the relative mean bias (RMB, Equation (4)): The MERRA-2 PM 2.5 reanalysis was thought to be reasonably good if RMB ≥ 0.5 [27].

Trend Analysis
Multi-year average trends (1980-2017) in PM 2.5 and its major components (including SO 4 , BC, OC, Dust 2.5 and SS 2.5 ) were assessed based on a linear regression. Furthermore, their corresponding significant levels were calculated using the Mann-Kendall (MK) nonparametric method. They were described by Equations (5)-(8): Var(s) = n(n − 1)(2n + 5) − ∑ q p=1 t p t p − 1 2t p + 5 18 Remote Sens. 2019, 11, 460 where S is the statistical value, Var (S) is the variance of the statistical value S and X j and X i represent the annual mean concentrations of the total PM 2.5 or its major components in the jth and ith years, respectively. The trend was significant (i.e., p-value < 0.05) only when the standardized test statistics |Z| > |Z (1 − α/2) |.

Multiple Linear Regression
In order to assess the comprehensive impact of various meteorological factors on PM 2.5 , the following multiple linear regression (MLR) model was used: where the dependent variable y refers to daily mean PM 2.5 concentrations (including PM 2.5 , SO 4 , OC, BC, Dust 2.5 and SS 2.5 ) in each 0.05 • × 0.625 • grid, β 0 is a constant term, x i is the daily mean meteorological factor described in Section 2.1 and ε i is the interaction term. Moreover, the slope coefficient β i can be interpreted as the sensitivity of PM 2.5 to changes in one meteorological factor, assuming that the other meteorological variables remain fixed. This was determined via a least-square adjustment [26,27].

3.1.
Ground-Based Verification of MERRA-2 PM 2.5 3.1.1. Regional and Seasonal Comparisons between MERRA-2 and Ground PM 2.5 Comparisons of daily mean PM 2.5 concentrations derived from MERRA-2 and 476 ground sites over the YRB are presented in Figure 2. The comparison over the YRD (Figure 2b) was superior to other regions of the YRB, with higher R (0.58) and RMB (0.90) values. By contrast, the SB (Figure 2d) produced the worst performance, with the lowest R (0.31) and RMB (0.55) values. The unique basin topography was prone to cloudy weather, which may restrict the assimilation of MERRA-2 PM 2.5 products [37]. Generally, MERRA-2 produced lower daily mean PM 2.5 concentrations (4.77-23.26 µg m −3 ) and the bias was more outstanding for high ground PM 2.5 (>75 µg m −3 ). For instance, MERRA-2 PM 2.5 concentrations were entirely less than ground observations when PM 2.5 > 75 µg m −3 . The lack of nitrate in the MERRA-2 PM 2.5 reanalysis would explain the underestimation [37,40]. In addition, to a certain extent, the bias caused by MERRA-2 AOD assimilation might be another reason for the MERRA-2 PM 2.5 underestimation, which will be verified in detail in the next section. Notably, in Figure 2d, no matter how large the ground PM 2.5 concentrations were, MERRA-2 PM 2.5 remained small. The reason might be that the unique basin terrain was prone to cloudy weather in the SB, and thus probably led to cloud screening bias for MERRA-2 aerosol reanalysis [8,9,40].

Impact of AOD Assimilation
Here, the effect of MERRA-2 AOD assimilation on the PM 2.5 simulation was estimated. Since there were no CARSNET AOD observations in the MERRA-2 aerosol assimilation, we compared daily mean MERRA-2 AOD with those ground measurements from five CARSNET sites of the YRB during 2013-2016 ( Figure 4). As shown in Figure 4a, although MERRA-2 AOD was in good agreement with CARSNET AOD (R = 0.82), approximately 19% of MERRA-2 AOD were still underestimated, especially when CARSNET AOD > 1.0. This was further confirmed by Figure 4b. The high MERRA-2 AOD values in Wuhan and Chengdu (>0.8) were all below their corresponding CARSNET observations. By contrast, Shangri-La with low mean AOD (<0.2) exhibited a good performance. Compared with the remote site of Shangri-La, Wuhan and Chengdu were two urban CARSNET sites with more complex aerosol sources and more diverse underlying surfaces, probably restraining the assimilation of the MERRA-2 AOD reanalysis [37,40]. Overall, the underestimation of MERRA-2 PM 2.5 was partly attributed to the negative bias of MERRA-2 AOD used in the GOCART model.
The high MERRA-2 AOD values in Wuhan and Chengdu (> 0.8) were all below their corresponding CARSNET observations. By contrast, Shangri-La with low mean AOD (< 0.2) exhibited a good performance. Compared with the remote site of Shangri-La, Wuhan and Chengdu were two urban CARSNET sites with more complex aerosol sources and more diverse underlying surfaces, probably restraining the assimilation of the MERRA-2 AOD reanalysis [37,40]. Overall, the underestimation of MERRA-2 PM2.5 was partly attributed to the negative bias of MERRA-2 AOD used in the GOCART model.   Figure 5(a), SO4 dominated the aerosol types of Wuhan for the whole year. All AOD products showed similar daily variations, i.e., high values approximately occurred between the 50th and the 150th days. In addition, both satellite (Aqua and Terra) and ground (CARSNET) AOD observations were higher than MERRA-2 AODs. Additionally, is worth noting that neither the satellites nor sun-photometer provided continuous AOD observations, especially during a heavy pollution episode ( Figure 5(b)). This would make the assimilation of MERRA-2 AOD impossible, implying that the GOCART model for simulating aerosols in MERRA-2 needed further improvement.  All AOD products showed similar daily variations, i.e., high values approximately occurred between the 50th and the 150th days. In addition, both satellite (Aqua and Terra) and ground (CARSNET) AOD observations were higher than MERRA-2 AODs. Additionally, is worth noting that neither the satellites nor sun-photometer provided continuous AOD observations, especially during a heavy pollution episode ( Figure 5b). This would make the assimilation of MERRA-2 AOD impossible, implying that the GOCART model for simulating aerosols in MERRA-2 needed further improvement. PM2.5 components varied by region due to differences in anthropogenic emissions and meteorological conditions. The spatial distributions of annual mean MERRA-2 PM2.5 components were illustrated in Figure 6. In Figure 6(a), high PM2.5 concentrations were observed over the YRD, CC and SB, while low values were concentrated in the SYR. A similar spatial pattern of PM2.5 was also confirmed by the GWR PM2.5 ( Figure A2). Previous studies reported that GWR PM2.5 reanalysis yielded good performance relative to ground observations [41,42]. Therefore, we used it here to verify the spatial distribution of MERRA-2 PM2.5 reanalysis. According to the ambient air quality standard established by the World Health Organization (WHO), the annual mean PM2.5 concentration should not exceed 10 µg/m 3 [23]. Approximately 67.49% and 78.27% of the pixels in the YRB exceeded the threshold for MERRA-2 and GWR, respectively ( Figure A2). This fact suggested that it was necessary to clarify the composition of PM2.5 and its influencing factors over   Figure 6. In Figure 6a, high PM 2.5 concentrations were observed over the YRD, CC and SB, while low values were concentrated in the SYR. A similar spatial pattern of PM 2.5 was also confirmed by the GWR PM 2.5 ( Figure A2). Previous studies reported that GWR PM 2.5 reanalysis yielded good performance relative to ground observations [41,42]. Therefore, we used it here to verify the spatial distribution of MERRA-2 PM 2.5 reanalysis. According to the ambient air quality standard established by the World Health Organization (WHO), the annual mean PM 2.5 concentration should not exceed 10 µg/m 3 [23]. Approximately 67.49% and 78.27% of the pixels in the YRB exceeded the threshold for MERRA-2 and GWR, respectively ( Figure A2). This fact suggested that it was necessary to clarify the composition of PM 2.5 and its influencing factors over the YRB.
In terms of each PM 2.5 component, sulfate PM 2.5 (SO 4 ) occupied an average of 39.87% of the PM 2.5 concentration, and its high values were concentrated in YRD, CC and SB. Similarly, high carbonaceous PM 2.5 concentrations (BC and OC) were also observed in those regions, accounting for 8.02% and 28.32% of the PM 2.5 concentrations, respectively. Emissions from power plants, vehicle exhaust and biomass burning could certainly explain the high PM 2.5 concentrations in those areas. For dust PM 2.5 (Dust 2.5 ), it accounted for 22.75% of the PM 2.5 on average. In particular, high Dust 2.5 concentrations appeared in the northern part of the YRB, which primarily came from local arid sources. Another region with high Dust 2.5 concentrations was the SYR, possibly due to dust transportation from the Taklimakan desert in summer [7,8,43]. MERRA-2 sea salt PM 2.5 (SS 2.5 ) constituted the lowest proportion of the total (approximately 1.02%), and rarely penetrated into land.  (Table 3). This was associated with the prevailing local monsoon climate [31]. In summer, windy and rainy weather was conducive to scavenging PM2.5 particles from the atmosphere, whereas the dry weather affected by the Mongolian anticyclone in winter prevented the diffusion of aerosol particles. The increase in emissions caused by heating in winter could certainly be another reason for the high PM2.5 concentration [31]. For each PM2.5 component, seasonal variations in SO4, BC and OC were consistent with those of PM2.5. Nevertheless, SS2.5 witnessed little seasonal changes over the YRB. Dust2.5 concentration was highest in spring (8.63 µg m −3 ), and most occurred in northern parts of the YRB. This was related to frequent sandstorm events in spring [35,36].   (Table 3). This was associated with the prevailing local monsoon climate [31]. In summer, windy and rainy weather was conducive to scavenging PM 2.5 particles from the atmosphere, whereas the dry weather affected by the Mongolian anticyclone in winter prevented the diffusion of aerosol particles. The increase in emissions caused by heating in winter could certainly be another reason for the high PM 2.5 concentration [31]. For each PM 2.5 component, seasonal variations in SO 4 , BC and OC were consistent with those of PM 2.5 . Nevertheless, SS 2.5 witnessed little seasonal changes over the YRB. Dust 2.5 concentration was highest in spring (8.63 µg m −3 ), and most occurred in northern parts of the YRB. This was related to frequent sandstorm events in spring [35,36].       Figure A3). The results showed that both PM 2.5 products had similar growth trends, and there was a significant turning point from increasing to decreasing around 2007-2009. Similarly, He et al. [10] revealed that an over-increasing trend in MERRA-2 AOD over the YRB was curbed after 2008. The reason was probably due to the effective implementation of energy conservation and emission reduction policies in China [10]. Comparable trends were also found in SO 4 , BC and OC, confirming that reduction in anthropogenic emissions was the dominant factor for the improvement in air quality. In addition, Dust 2.5 and SS 2.5 witnessed slightly increasing trends of 7.42E-5 µg m −3 yr −1 and 1.63E-5 µg m −3 yr −1 after 1980, respectively. Nevetheless, whether meterological factors impacted air quality requires further analysis.  Figure A3). The results showed that both PM2.5 products had similar growth trends, and there was a significant turning point from increasing to decreasing around 2007-2009. Similarly, He et al. [10] revealed that an over-increasing trend in MERRA-2 AOD over the YRB was curbed after 2008. The reason was probably due to the effective implementation of energy conservation and emission reduction policies in China [10]. Comparable trends were also found in SO4, BC and OC, confirming that reduction in anthropogenic emissions was the dominant factor for the improvement in air quality. In addition, Dust2.5 and SS2.5 witnessed slightly increasing trends of 7.42E-5 µg m −3 yr −1 and 1.63E-5 µg m −3 yr −1 after 1980, respectively. Nevetheless, whether meterological factors impacted air quality requires further analysis. The annual mean trends of each MERRA-2 PM2.5 component at grid level were estimated by a linear regression, and their corresponding significant levels were further calculated based on the MK method ( Figure 10). Overall, a significant increasing trend in PM2.5 (0.7µg m −3 yr −1 ) was observed over the YRD, CC and SB. However, GWR exhibited a significant decreasing trend in PM2.5 in the mountains of the northern YRB ( Figure A4), which was not observed in MERRA-2 PM2.5. Recent studies demonstrated that the decreasing PM2.5 trend in those regions might be a response to the implementation of forest ecology projects over the middle and upper reaches of the YRB since the 1990s [44]. In terms of each PM2.5 component, SO4, BC and OC witnessed significant increases, The annual mean trends of each MERRA-2 PM 2.5 component at grid level were estimated by a linear regression, and their corresponding significant levels were further calculated based on the MK method ( Figure 10). Overall, a significant increasing trend in PM 2.5 (0.7 µg m −3 yr −1 ) was observed over the YRD, CC and SB. However, GWR exhibited a significant decreasing trend in PM 2.5 in the mountains of the northern YRB ( Figure A4), which was not observed in MERRA-2 PM 2.5 . Recent studies demonstrated that the decreasing PM 2.5 trend in those regions might be a response to the implementation of forest ecology projects over the middle and upper reaches of the YRB since the 1990s [44]. In terms of each PM 2.5 component, SO 4 , BC and OC witnessed significant increases, especially over the YRD, CC and SB. By contrast, significant growth trends in Dust 2.5 and SS 2.5 occurred over the SYR and the eastern coastal regions, respectively.
Remote Sens. 2019, 11, x FOR PEER REVIEW 15 of 27 especially over the YRD, CC and SB. By contrast, significant growth trends in Dust2.5 and SS2.5 occurred over the SYR and the eastern coastal regions, respectively.

Discussion
To quantify the impact of meteorological variables on PM2.5 changes, the slope coefficients βi between daily mean PM2.5 concentrations and six meteorological factors from 1980 to 2017 were calculated using the multiple linear regression (MLR) model (Equation 9). Assuming that all other meteorological variables remained fixed, these slope coefficients actually represented the sensitivities of PM2.5 to unit changes in each meteorological factor. In the MLR process, there should be no multicollinearity between the independent variables (i.e., meteorological factors), otherwise they would be eliminated [26,27]. In the correlation matrix (Table A1), even though there were

Discussion
To quantify the impact of meteorological variables on PM 2.5 changes, the slope coefficients β i between daily mean PM 2.5 concentrations and six meteorological factors from 1980 to 2017 were calculated using the multiple linear regression (MLR) model (Equation (9)). Assuming that all other meteorological variables remained fixed, these slope coefficients actually represented the sensitivities of PM 2.5 to unit changes in each meteorological factor. In the MLR process, there should be no multicollinearity between the independent variables (i.e., meteorological factors), otherwise they would be eliminated [26,27]. In the correlation matrix (Table A1), even though there were several correlations of more than 0.5, most of them failed to pass the 95% significance test, indicating that each meteorological variable was not significantly related to any other. Therefore, we concluded that collinearity was minimal and the MLR analysis could be performed. Figure 11 showed the adjusted coefficients of determination (adjusted-R 2 ) at grid level. Higher adjusted-R 2 values indicated that the MLR model performed better. The highest adjusted-R 2 values (~0.67) were mostly concentrated over the SYR, indicating that the local meteorological conditions could explain up to 67% of the daily PM 2.5 changes. By comparison, the values were lowest over the YRD and CC, where less than 20% of the daily PM 2.5 changes were due to meteorological variations. Over those regions, anthropogenic emissions from frequent industrialization and urbanization may be the major reason for high PM 2.5 concentrations. In Figure A5, adjusted-R 2 values showed seasonal variations, with high values (adjusted-R 2 > 0.5) in spring and summer over the SYR, suggesting that the local meteorological conditions could account for more than 50% of PM 2.5 changes. This may be closely related to the summer monsoon climate [26,27]. However, the relatively low adjusted-R 2 values (adjusted-R 2 < 0.2) dominated most areas of the YRB in autumn and winter. This may be due to an increase in anthropogenic emissions such as winter heating, which weakened the impact of meteorological conditions on air pollution in winter [26].
Remote Sens. 2019, 11, x FOR PEER REVIEW 16 of 27 several correlations of more than 0.5, most of them failed to pass the 95% significance test, indicating that each meteorological variable was not significantly related to any other. Therefore, we concluded that collinearity was minimal and the MLR analysis could be performed. Figure 11 showed the adjusted coefficients of determination (adjusted-R 2 ) at grid level. Higher adjusted-R 2 values indicated that the MLR model performed better. The highest adjusted-R 2 values (~0.67) were mostly concentrated over the SYR, indicating that the local meteorological conditions could explain up to 67% of the daily PM2.5 changes. By comparison, the values were lowest over the YRD and CC, where less than 20% of the daily PM2.5 changes were due to meteorological variations. Over those regions, anthropogenic emissions from frequent industrialization and urbanization may be the major reason for high PM2.5 concentrations. In Figure A5, adjusted-R 2 values showed seasonal variations, with high values (adjusted-R 2 > 0.5) in spring and summer over the SYR, suggesting that the local meteorological conditions could account for more than 50% of PM2.5 changes. This may be closely related to the summer monsoon climate [26,27]. However, the relatively low adjusted-R 2 values (adjusted-R 2 < 0.2) dominated most areas of the YRB in autumn and winter. This may be due to an increase in anthropogenic emissions such as winter heating, which weakened the impact of meteorological conditions on air pollution in winter [26]. . Red grids represent positive correlations, whereas green grids are negative correlations. In Figure 12(a), the PM2.5 is significantly anti-correlated with surface wind speed throughout the YRB. High wind speeds facilitated circulation and dilution of aerosol particles, leading to a consistent decrease in PM2.5. Negative correlations between PM2.5 concentrations and surface wind speeds were also observed in the United States of America [26], Europe [27][28][29]31] and China [32][33][34]. Yang et al. [33] discovered a negative wind-PM2.5 correlation throughout China except Hainan Island. Compared with the USA and Europe, China had the worst atmospheric dispersion conditions with annual stagnation frequency of 29%, leading to serious air pollutions especially in winter [31]. However, a significant positive correlation occurred in the SYR, where a plenty of dust particles were concentrated (see Section 3.2). It was probably due to the strong wind speed dependence of dust particles, further confirmed by the Dust2.5-wind regression in Figure 12 (e). Similarly, in Saudi Arabia (an arid region), both PM2.5 and PM10 were positively related to wind speed [30]. Additionally, in Figure A6 and Table A2, (Table A2).  Figure 12 depicts correlations of each PM 2.5 component with surface wind speed, as measured by the MLR regression coefficient (β 1 ) in Equation (9). Red grids represent positive correlations, whereas green grids are negative correlations. In Figure 12a, the PM 2.5 is significantly anti-correlated with surface wind speed throughout the YRB. High wind speeds facilitated circulation and dilution of aerosol particles, leading to a consistent decrease in PM 2.5 . Negative correlations between PM 2.5 concentrations and surface wind speeds were also observed in the United States of America [26], Europe [27][28][29]31] and China [32][33][34]. Yang et al. [33] discovered a negative wind-PM 2.5 correlation throughout China except Hainan Island. Compared with the USA and Europe, China had the worst atmospheric dispersion conditions with annual stagnation frequency of 29%, leading to serious air pollutions especially in winter [31]. However, a significant positive correlation occurred in the SYR, where a plenty of dust particles were concentrated (see Section 3.2). It was probably due to the strong wind speed dependence of dust particles, further confirmed by the Dust 2.5 -wind regression in Figure 12e. Similarly, in Saudi Arabia (an arid region), both PM 2.5 and PM 10 were positively related to wind speed [30]. Additionally, in Figure A6 and Table A2, (Table A2).

Sensitivity of PM2.5 Components to Surface Temperature
In Figure 13, surface temperature (T) was significantly positively correlated with each PM2.5 component across the YRB. Similar results were also reported on global [27] and regional [26] scales. Westervelt et al. [27] discovered a robust positive relationship between surface temperature and PM2.5 throughout the continents and oceans except East and South Asia. This positive correlation may be related to temperature-dependent reaction rates and changes in oxidant abundance, that is, the oxidation of aerosol precursors (such as sulfur dioxide) could be enhanced with higher temperature [27]. However, in China, Yang et al. [33] revealed that surface temperature was negatively related to PM2.5 in autumn but positivity correlated with PM2.5 in winter. Dawson et al. [45] found an averaged negative effect of temperature on PM2.5 east of the USA, primarily due to the volatilization of nitrate at high temperature. Since nitrate aerosols were not assimilated in the MERRA-2 PM2.5 reanalysis, there was little negative relationship between PM2.5 and temperature over the YRB. In addition, Figure A6 and Table A2 showed an average increase of 0.03 K per year in temperature from 1980 to 2017. Based on the MLR model between temperature and the PM2.5 (Figure 13(a)), the average sensitivity of 0.11 µg m −3 K −1 could lead to PM2.5 growth of up to 0.003 µg m -3 per year. Overall, the temperature effect on PM2.5 (0.11 µg m -3 K -1 ) was weaker than the influence of wind speed (−2.42 µg m −3 m −1 s) over the YRB (Table A2).

Sensitivity of PM 2.5 Components to Surface Temperature
In Figure 13, surface temperature (T) was significantly positively correlated with each PM 2.5 component across the YRB. Similar results were also reported on global [27] and regional [26] scales. Westervelt et al. [27] discovered a robust positive relationship between surface temperature and PM 2.5 throughout the continents and oceans except East and South Asia. This positive correlation may be related to temperature-dependent reaction rates and changes in oxidant abundance, that is, the oxidation of aerosol precursors (such as sulfur dioxide) could be enhanced with higher temperature [27]. However, in China, Yang et al. [33] revealed that surface temperature was negatively related to PM 2.5 in autumn but positivity correlated with PM 2.5 in winter. Dawson et al. [45] found an averaged negative effect of temperature on PM 2.5 east of the USA, primarily due to the volatilization of nitrate at high temperature. Since nitrate aerosols were not assimilated in the MERRA-2 PM 2.5 reanalysis, there was little negative relationship between PM 2.5 and temperature over the YRB. In addition, Figure A6 and Table A2 showed an average increase of 0.03 K per year in temperature from 1980 to 2017. Based on the MLR model between temperature and the PM 2.5 (Figure 13a), the average sensitivity of 0.11 µg m −3 K −1 could lead to PM 2.5 growth of up to 0.003 µg m −3 per year. Overall, the temperature effect on PM 2.5 (0.11 µg m −3 K −1 ) was weaker than the influence of wind speed (−2.42 µg m −3 m −1 s) over the YRB (Table A2).  Figure 14 shows the slope coefficients between each PM2.5 component and precipitation. Each PM2.5 component was significantly anti-correlated with precipitation in most areas of the YRB, primarily due to wet scavenging [26][27][28][29][30][31][32][33][34][35]. However, a positive relationship was observed in the SYR, where dust particle was the main air pollutant (see Section 3.2) and precipitation did not occur frequently [46][47][48]. Even if there was a small amount of precipitation over the SYR, it may have only increased the relative humidity, resulting in the growth of hygroscopic particles [27]. Precipitation changes were typically more positive than negative from 1980 to 2017, along with an annual trend of 0.02 mm day -1 ( Figure A6 and Table A2). Overall, based on the MLR model of PM2.5 and precipitation (Figure 14(a)), the average sensitivity of −0.34 µg m −3 mm −1 day could lead to a decrease of 0.007 µg m −3 per year for PM2.5, which was comparable to the effect of temperature on PM2.5 (Table A2).  Figure 14 shows the slope coefficients between each PM 2.5 component and precipitation. Each PM 2.5 component was significantly anti-correlated with precipitation in most areas of the YRB, primarily due to wet scavenging [26][27][28][29][30][31][32][33][34][35]. However, a positive relationship was observed in the SYR, where dust particle was the main air pollutant (see Section 3.2) and precipitation did not occur frequently [46][47][48]. Even if there was a small amount of precipitation over the SYR, it may have only increased the relative humidity, resulting in the growth of hygroscopic particles [27]. Precipitation changes were typically more positive than negative from 1980 to 2017, along with an annual trend of 0.02 mm day −1 (Figure A6 and Table A2). Overall, based on the MLR model of PM 2.5 and precipitation (Figure 14a), the average sensitivity of −0.34 µg m −3 mm −1 day could lead to a decrease of 0.007 µg m −3 per year for PM 2.5 , which was comparable to the effect of temperature on PM 2.5 (Table A2).  Figure 14 shows the slope coefficients between each PM2.5 component and precipitation. Each PM2.5 component was significantly anti-correlated with precipitation in most areas of the YRB, primarily due to wet scavenging [26][27][28][29][30][31][32][33][34][35]. However, a positive relationship was observed in the SYR, where dust particle was the main air pollutant (see Section 3.2) and precipitation did not occur frequently [46][47][48]. Even if there was a small amount of precipitation over the SYR, it may have only increased the relative humidity, resulting in the growth of hygroscopic particles [27]. Precipitation changes were typically more positive than negative from 1980 to 2017, along with an annual trend of 0.02 mm day -1 ( Figure A6 and Table A2). Overall, based on the MLR model of PM2.5 and precipitation (Figure 14(a)), the average sensitivity of −0.34 µg m −3 mm −1 day could lead to a decrease of 0.007 µg m −3 per year for PM2.5, which was comparable to the effect of temperature on PM2.5 (Table A2).  Figure 15 shows the slope coefficients for each PM 2.5 component and relative humidity (RH). As can be seen in Figure 15a, PM 2.5 was positively correlated with RH in most areas of the YRB, with the exception of the YRD. Throughout the YRB, significant positive correlations with RH were also found in SO 4 , BC and OC, likely due to the dependence of sulfate and carbonaceous PM 2.5 formation on RH [26,27]. In contrast, Dust 2.5 had an anti-correlation with RH, which probably explained the negative association of PM 2.5 with RH in the YRD. Yang et al. [33] found a positive correlation between PM 2.5 and RH in north China and a negative correlation in south China. A negative PM 2.5 -RH regression was also estimated over the Pearl River Delta Region [34]. They considered that when RH was greater than 80%, rainfall often occurred, leading to a decrease in PM 2.5 in the atmosphere [34]. In this study, YRD had a temperate monsoon climate, which made precipitation events occur in most days with high RH (>80%), especially in summer. Thus, there was a negative PM 2.5 -RH regression over the YRD. However, with the relatively low slope coefficient for total PM 2.5 and RH (0.05 µg m −3 % −1 ), the sensitivity of PM 2.5 to RH was small (Table A2).  Figure 15 shows the slope coefficients for each PM2.5 component and relative humidity (RH). As can be seen in Figure 15(a), PM2.5 was positively correlated with RH in most areas of the YRB, with the exception of the YRD. Throughout the YRB, significant positive correlations with RH were also found in SO4, BC and OC, likely due to the dependence of sulfate and carbonaceous PM2.5 formation on RH [26,27]. In contrast, Dust2.5 had an anti-correlation with RH, which probably explained the negative association of PM2.5 with RH in the YRD. Yang et al. [33] found a positive correlation between PM2.5 and RH in north China and a negative correlation in south China. A negative PM2.5-RH regression was also estimated over the Pearl River Delta Region [34]. They considered that when RH was greater than 80%, rainfall often occurred, leading to a decrease in PM2.5 in the atmosphere [34]. In this study, YRD had a temperate monsoon climate, which made precipitation events occur in most days with high RH (> 80%), especially in summer. Thus, there was a negative PM2.5-RH regression over the YRD. However, with the relatively low slope coefficient for total PM2.5 and RH (0.05 µg m -3 % -1 ), the sensitivity of PM2.5 to RH was small (Table  A2).

Sensitivity of PM2.5 Components to Total Cloud Cover
In Figure 16, anti-correlations between each PM2.5 component and the total cloud cover were observed mostly over the middle and lower reaches of the YRB. In those regions, clouds could partially prevent aerosol precursors (such as sulfur dioxide) from oxidizing, and thus, reduced PM2.5 concentrations in the atmosphere [27]. However, positive regressions between each PM2.5 component and the total cloud cover occurred in the SYR, possibly due to the enhancements in in-cloud production of sulfate aerosols. Overall, the average sensitivity of PM2.5 to the total cloud cover was approximately −0.07 µg m −3 % −1 (Figure 16(a) and Table A2).

Sensitivity of PM 2.5 Components to Total Cloud Cover
In Figure 16, anti-correlations between each PM 2.5 component and the total cloud cover were observed mostly over the middle and lower reaches of the YRB. In those regions, clouds could partially prevent aerosol precursors (such as sulfur dioxide) from oxidizing, and thus, reduced PM 2.5 concentrations in the atmosphere [27]. However, positive regressions between each PM 2.5 component and the total cloud cover occurred in the SYR, possibly due to the enhancements in in-cloud production of sulfate aerosols. Overall, the average sensitivity of PM 2.5 to the total cloud cover was approximately −0.07 µg m −3 % −1 (Figure 16a and Table A2).

Sensitivity of PM2.5 Components to Boundary Layer Height
In Figure 17, the relationship of each PM2.5 component with boundary layer height (BLH) was robustly negative throughout the YRB. The reason may be that a well-developed BLH was conducive to the diffusion and dilution of PM2.5 particles [31]. Recently, Wang et al. [31] reported that BLH was lower than the normal during heavy pollution in Beijing, China. During 1980-2017, the YRB experienced a significant decrease of −0.98 m per year in BLH, especially over the YRD, CC and SB ( Figure A6). Based on the MLR model between BLH and PM2.5 (Figure 17

Sensitivity of PM 2.5 Components to Boundary Layer Height
In Figure 17, the relationship of each PM 2.5 component with boundary layer height (BLH) was robustly negative throughout the YRB. The reason may be that a well-developed BLH was conducive to the diffusion and dilution of PM 2.5 particles [31]. Recently, Wang et al. [31] reported that BLH was lower than the normal during heavy pollution in Beijing, China. During 1980-2017, the YRB experienced a significant decrease of −0.98 m per year in BLH, especially over the YRD, CC and SB ( Figure A6). Based on the MLR model between BLH and PM 2.5 (Figure 17a and Table A2

Sensitivity of PM2.5 Components to Boundary Layer Height
In Figure 17, the relationship of each PM2.5 component with boundary layer height (BLH) was robustly negative throughout the YRB. The reason may be that a well-developed BLH was conducive to the diffusion and dilution of PM2.5 particles [31]. Recently, Wang et al. [31] reported that BLH was lower than the normal during heavy pollution in Beijing, China. During 1980-2017, the YRB experienced a significant decrease of −0.98 m per year in BLH, especially over the YRD, CC and SB ( Figure A6). Based on the MLR model between BLH and PM2.5 (Figure 17

Conclusions
Based on the MLR model, the sensitivities of PM 2.5 and its major components to meteorological variables were quantified using daily mean MERRA-2 PM 2.5 and meteorology reanalysis data over the YRB from 1980 to 2017. First of all, in order to verify the accuracy of the MERRA-2 PM 2.5 reanalysis, it was compared with ground-based measurements derived from 476 air quality monitoring sites in the YRB. The results suggested that MERRA-2 PM 2.5 was underestimated relative to their corresponding ground observations throughout the YRB. In addition to the lack of nitrate aerosols in GOCART, the bias in MERRA-2 AOD assimilation could also partly explain the underestimation of MERRA-2 PM 2.5 .
The YRB witnessed severe air pollution, with its annual mean PM 2.5 concentration of 23.43 µg m −3 . In particular, approximately 67.49% of the pixels in the YRB exceeded the annual mean threshold of 10 µg m −3 recommended by the WHO. The high PM 2.5 concentrations were observed over the YRD, CC and SB in winter, while the low values occurred in the SYR in summer. For each PM 2.5 component, SO 4 , BC and OC showed similar spatiotemporal distributions with PM 2.5 concentration. The temperate monsoon climate could account for the seasonal and spatial variations of PM 2.5 concentrations in the YRB. However, Dust 2.5 had the highest concentrations of 8.63 µg m −3 over the SYR and the northern areas of the YRB in spring, which came from local arid sources. Most of the SS 2.5 were concentrated in the eastern coastal areas and there was no significant seasonal variation. Over-increasing trends (1980 onward) in all PM 2.5 components were confirmed in the YRB; however, they were curbed after 2007.
Based on the MLR model, we found the highest adjusted-R 2 (0.67) in the SYR, suggesting that the six meteorological variables could explain up to 67% of the daily PM 2.5 changes in that region. Furthermore, each PM 2.5 component was significantly anti-correlated with surface wind speed and precipitation throughout the YRB. High wind speeds and precipitations facilitated the circulation and dilution of PM 2.5 from the atmosphere. Nevertheless, a positive wind-PM 2.5 regression was observed in the SYR, probably due to the dependence of Dust 2.5 on wind. Furthermore, SYR witnessed a positive relationship between precipitation and PM 2.5 , which was likely associated with an increase in relative humidity (RH). Additionally, boundary layer height (BLH) was robustly anti-correlated with each PM 2.5 component. However, the impact of BLH on PM 2.5 (−0.006 µg m −3 m −1 ) was far less than those of wind speed (−2.42 µg m −3 m −1 s) and precipitation (−0.34 µg m −3 mm −1 day). In contrast, surface temperature was positively correlated with each PM 2.5 component across the YRB, and the sensitivity of PM 2.5 to temperature was approximately 0.11 µg m −3 K −1 . This was possibly due to the enhancement in oxidation of aerosol precursors at high temperatures. The RH-PM 2.5 regression showed regional dependencies, with negative correlation in the YRD and positive correlation in the other areas. A negative relationship between total cloud cover and PM 2.5 occurred in the middle and lower reaches of the YRB. However, the sensitivities of PM 2.5 to RH (0.05 µg m −3 % −1 ) and total cloud cover (0.07 µg m −3 % −1 ) were relatively small. Overall, this study highlighted the impact of meteorological conditions on PM 2.5 , even though it was relatively minor in comparison with the anthropogenic emissions effect [27]. When assessing the effects of emission reductions and further formulating mitigation policies, the unfavorable atmospheric diffusion conditions should not be ignored in China.             Figure A5. Coefficients of determination (R 2 ) of the MLR model between PM2.5 and the six meteorological variables in spring (a), summer (b), autumn (c) and winter (d). Figure A4. Spatial distribution of PM 2.5 annual trend derived form MERRA-2 (a) and GWR (b) over the YRB from 1998 to 2016. Grids marked by grey points represented PM 2.5 trends that did not exceed a 95% significant level.  Figure A4. Spatial distribution of PM2.5 annual trend derived form MERRA-2 (a) and GWR (b) over the YRB from 1998 to 2016. Grids marked by grey points represented PM2.5 trends that did not exceed a 95% significant level.    .   Table A2. Summary of regression coefficients between surface PM 2.5 and meteorological variables, annual trends in meteorological variables and theoretical annual trends in PM 2.5 based on MLR model. β 1 . . . β 6 refer to regression coefficients between PM 2.5 and wind speed (Wind), temperature (T), precipitation (Precip), relative humidity (RH), total cloud cover (Cloud) and Boundary layer height (BLH), respectively.