Aerosol concentrations in eastern China are very high due to both direct emissions and secondary aerosol formation from precursor gases such as NO2, SO2 and volatile organic compounds (VOCs). They are produced by anthropogenic activities such as industry, transportation and heating, black carbon and other carbonaceous aerosols produced by biomass burning, and dust aerosols produced from the deserts. Aerosol particles influence the local climate such as monsoon intensity and the distribution of precipitation. In eastern China, the monsoon in turn plays an important role in the wet deposition and transport of aerosol particles (Li et al., 2016). The Asian monsoon system plays an important role in the precipitation across the country (Kourtidis et al., 2015). In early April, the pre-monsoonal rain period starts over southern China and the summer monsoon rain belt moves northward to the Yangtze River basin in June. Further, the rain belt arrives in northern China in July and the monsoon rain belt propagates back to southern China in August. The length of the rain season differs between southern and northern China with the migration of the monsoon across China (Song et al., 2011). Based on these characteristics, four regions with different aerosol emission levels and climate characteristics were selected to study the indirect effects of aerosol particles on cloud micro- and macrophysical properties. The Beijing–Tianjin–Hebei (BTH), Yangtze River Delta (YRD) and Pearl River Delta (PRD) urban clusters are characterized as a temperate monsoon climate region, a subtropical monsoon climate region and a tropical monsoon climate region, respectively. The BTH domain (35.5–40.5∘ N, 113.5–120.5∘ E) is an area with high AOD levels due to rapid industrial and economic development (Fig. 1). The YRD domain (28–33∘ N, 117–122∘ E) is a major source region of black carbon (Streets et al., 2001; Bond et al., 2004) and sulfate (Lu et al., 2010). The PRD domain (21.5–24.5∘ N, 111.5–115.5∘ E) is an area within the Intertropical Convergence Zone (ITCZ) migration belt, with high anthropogenic aerosol emissions (Streets et al., 2003, 2008; Lei et al., 2011). In addition, one domain (20–25∘ N and 125–130∘ E), which is located in the East China Sea (ECS for short), has been selected as the study area for comparison. The ECS domain is relatively clean, but it is often impacted by aerosol particles transported from the highly industrialized eastern China (Wang et al., 2014). The study period is 10 years, i.e., 2008–2017.

The aerosol and cloud properties used in this study were derived from the MODIS instruments on the Terra and Aqua satellites. Since these instruments are of the same design, errors due to instrument differences are minimal although some differences have been reported due to degradation of MODIS/Terra (Xiong et al., 2008; Levy et al., 2010). The MODIS L3 collection 6.1 data (which were downloaded from https://ladsweb.modaps.eosdis.nasa.gov/, last access: 12 December 2018) provide daily aerosol and cloud parameters on a 1∘ by 1∘ spatial grid. The time difference between the Terra and Aqua overpasses is about 3 h, with variations due to swath width. In the following, the time difference between the MODIS/Terra and Aqua observations is referred to as the time step. The application of daily MODIS satellite data on a 1∘ by 1∘ spatial grid in this study on aerosol–cloud interaction ensures that the aerosol and cloud retrievals are coincident. The MODIS instruments have 36 spectral bands – the first seven of these (0.47–2.13 µm) are used for the retrieval of aerosol properties (Remer et al., 2005) while cloud properties are retrieved using additional wavelengths in other parts of the spectrum (Platnick et al., 2003). More detailed information on algorithms for the retrieval of aerosol and cloud properties is provided at http://modis-atmos.gsfc.nasa.gov (last access: 12 December 2018). In this study on ACI we use the AOD at 550 nm (referred to as AOD throughout this manuscript), CDR, cloud liquid water path (CWP), cloud optical thickness (COT), cloud fraction (CF), cloud top pressure (CTP) and cloud top temperature (CTT) from both instruments. AOD is used as a proxy for the amount of aerosol particles in the atmospheric column to investigate ACI (Andreae, 2009; Kourtidis et al., 2015). To reduce a possible overestimation of the AOD, cases with AOD greater than 0.8 were excluded from further analysis. The focus of this study is on warm clouds with CTP larger than 700 hPa, CTT larger than 273 K and CWP lower than 200 g m-2, as most aerosols exist in the lower troposphere (Michibata et al., 2014).

In addition, to explore the effect of meteorological conditions on ACI, we use the daily temperature at the 1000 and 700 hPa levels, relative humidity (RH) at the 750 hPa level and pressure vertical velocity (PVV) at the 750 hPa level. LTS is defined as the difference in potential temperature between the free troposphere (700 hPa) and the surface, which can be regarded as a measure of the strength of the inversion that caps the planetary boundary layer (Klein and Hartmann, 1993; Wood and Bretherton, 2006). These meteorological data were obtained from daily ERA-Interim reanalysis data which contain global meteorological conditions on a grid of 1∘×1∘ with 37 levels in the vertical (1000–1 hPa) every 6 h (00:00, 06:00, 12:00, 18:00 UTC) (http://apps.ecmwf.int/datasets/data/interim-full-daily/, last access: 12 December 2018). The meteorological properties were resampled to 10:30 (local time) by taking a weighted average of the properties at the two closest times (00:00 and 06:00 UTC) provided by ERA-Interim.

In this study, high and low AOD are defined as the highest and lowest quartile for each 1∘×1∘ location to reduce climatological spatial gradients in aerosol and cloud parameters. As a result, the difference between high and low AOD varies by location. So, for each 1∘×1∘ grid cell, 3642 data samples are available for the 10-year study period.

For the comparison of the difference in cloud properties in high- and in low-AOD conditions and the change in this difference during the time step, we need to ensure that the initial conditions are similar; i.e., the probability distributions of a cloud parameter Cloud_X at the start of the time step for the low- and high-AOD cases should be similar. Any change in this distribution at the end of the time step can then be attributed to changes in cloud properties due to aerosol and/or meteorological effects. To reduce the difference between the initial probability distribution of Cloud_X in high- and low-AOD conditions at the start of the time step, normalized histograms of cloud properties and meteorological parameters are made for high- and low-AOD conditions following the method described by Gryspeerdt et al. (2014).

In Fig. 2 we illustrate the process to remove possible effects linking, as an example, CF and AOD. Normalized histograms of CF are made for the high- and low-AOD conditions following Gryspeerdt et al. (2014), with the difference that in the current study AOD is used instead of AI (Andreae, 2009; Kourtidis et al., 2015). The CF probability density functions for low- and high-AOD conditions at the start time are different as illustrated in Fig. 2a. This difference indicates a link between CF and AOD at the start of the time step which needs to be removed to detect the effect of changes during the time step. This is achieved following the process described in more detail by Gryspeerdt et al. (2014). In brief, for each bin datapoints are drawn out randomly from the conditions with the larger probability density frequency until both distributions match. This is performed independently for each bin and the entire process is repeated until the normalized histograms in both AOD conditions are similar. As a result of this normalization process, the CF distributions at the start of the time step are nearly identical for both AOD conditions; i.e., the non-aerosol effect linking CF and AOD has been removed. This technique has also been applied to ensure that the high- and low-AOD conditions have the same probability distributions for CDR, COT, CWP and CTP at the start time. Among those cloud properties, this process of normalization has the greatest effect on the cloud fraction and its dependence on aerosol–cloud interaction. Throughout the study, we only take a subset of original data by removing random samples until the histograms are similar.

Note that here and in the following sections, normalized histograms of cloud properties for the high- and low-AOD populations are made for the whole region (Sect. 3.1), because the data volume based on each 1∘×1∘ location is relatively small. However, the difference between the cloud properties for low and high AOD at the start time is based on each 1∘×1∘ location (Sect. 4.1). So the difference in the cloud properties between the low and high AOD at the start time is not zero.

After removal of the potential relationships between AOD and cloud parameters at the time of the Terra (morning) overpass, as described in Sect. 3.1, effects of aerosol particles on cloud properties are investigated from the change in the relationship between AOD and cloud parameters over the time step. For cloud property Cloud_X (where X = CF, COT, CWP, CDR or CTP), the change during the time step is indicated by ΔCloud_X. The mean ΔCloud_X for high AOD is then indicated by ΔCloud_X[High AOD]‾ and by ΔCloud_X[Low AOD]‾ for low AOD. The difference between the mean change in Cloud_X during the time step in high- and low-AOD conditions is then indicated by d(Cloud_X): d(Cloud_X)=ΔCloud_X[High AOD]‾-ΔCloud_X[Low AOD]‾. The high AOD is representative of polluted atmospheric conditions, and the low AOD is representative of clean atmospheric conditions. The difference (d(Cloud_X)) between the mean values of the cloud property Cloud_X during clean (low AOD) and polluted (high AOD) conditions indicates the effect of these two aerosol cases on the cloud property Cloud_X. For example, d(CWP) would be the difference between the mean change in CWP in high-AOD conditions minus that in low-AOD conditions.

A Student's t test is used to determine whether two data sets are significantly different from each other. The marker * at the top right corner of symbol “+” (or “-”) denotes that the difference between a change in cloud property and zero is significant (at a 95 % confidence level).

The difference in the mean cloud properties (CDR, CF, COT, CWP and CTP) during high- and low-AOD conditions at the start time for each 1∘×1∘ grid cell; i.e., Cloud_X[High AOD]‾-Cloud_X[Low AOD]‾t=0 represents the change in cloud properties due to the higher AOD. Figure 3 shows the spatial distributions of these differences (left column) and sample series of the difference (right column) for the four regions of interest. The selection of samples for each region is according to the pixels in the region. Figure 3a1–a2 show that, over the ECS, CDR is smaller at high AOD than at low AOD, which is consistent with Twomey's effect. In contrast, over the three urban clusters, CDR is larger at high AOD. This behavior has been observed before for warm clouds in conditions with high AOD (Liu et al., 2017) and may result from the intense competition for the available water vapor and the evaporation of smaller droplets as a consequence of the high aerosol abundance over these regions (Yuan et al., 2008; Tang et al., 2014; Wang et al., 2014; Liu et al., 2017). For COT (Fig. 3b1–b2) the values are significantly higher at high AOD over the ECS and the BTH; however, COT does not show a significant difference between the situations at low and high AOD over the YRD and PRD. These results indicate that there is no clear dependence of COT on aerosol load and also the aerosol type may influence the aerosol effect on COT. Figure 3c1–c2 show that CWP is lower at high AOD over the ECS, which is in clear contrast with the so-called “lifetime effect” proposed by Albrecht in 1989. In contrast, over the BTH, CWP behaves similar to COT and is higher at high AOD. Furthermore, CWP is also higher at high AOD. Ackerman et al. (2004) reported that CWP is not generally observed to be larger, but significantly smaller in high-AOD conditions. They reported that CWP response to the increasing AOD is determined by the balance of two competitive factors: moistening from precipitation decrease and drying from increasing entrainment of dry overlaying air. With increasing AOD, CF does not show any significant correlation between changes in AOD and CDR variations over the BTH and YRD. However, CF is larger at high AOD over the PRD and ECS. Wang et al. (2014) also found that, when aerosol loading is relatively small, cloud cover is found to increase over the YRD and ECS in response to aerosol enhancement regardless of RH conditions. Meanwhile, over the YRD urban cluster CTP is higher at high AOD, as suggested by Liu et al. (2017). In contrast, CTP is lower at high AOD over the BTH and ECS. Many studies have also reported that with higher cloud altitude CTP decreases in most of the places as AOD increases except for some regions at low AOD (Myhre et al., 2007; Kaufman et al., 2005; Alam et al., 2010). This might have resulted from the suppression of precipitation by increasing cloud lifetime and thus also affecting the cloud albedo and cloud top pressure.

To better characterize the variation in cloud properties between high and low AOD, Table 1 summarizes the difference in cloud properties between high and low AOD at start time for the four study areas. We find that different regions with various aerosol emission levels and different climate characteristics show different ACI patterns. Some links between aerosol and cloud in the four regions are different from those of previous studies over China (Wang et al., 2014; Tang et al., 2014; Kourtidis et al., 2015; Liu et al., 2017), which might be due to the use of different data sets (MODIS C6.1 versus older versions), hypothesis, and target areas characterized by complex aerosol composition and varying meteorological conditions. Overall, the result implies that the interaction between aerosol particles and clouds is more complex and of greater uncertainty over land (BTH, YRD and PRD) than over ocean (ECS). Jin and Shepherd (2008) also noted that aerosols affect clouds more significantly over ocean than over land. They suggested that dynamic processes related to factors like urban land cover may play at least an equally critical role in cloud formation.

The meteorological and aerosol effects on clouds are reported to be tightly connected, and this connection must be accounted for in any study of aerosol–cloud interactions (Stevens and Feingold, 2009; Koren et al., 2010). Although normalized histograms of meteorological parameters are made for high- and low-AOD conditions at the start time, the normalization described in Sect. 3.1 is based on the whole region. Differences in meteorological conditions may still occur between each 1∘ by 1∘ grid cell. In this study, we analyze the meteorology of the different regions, in support of the interpretation of the regional variation of the relationships between aerosols and clouds.

The spatial variations of meteorological parameters over the four regions, averaged over the years 2008–2017, are shown in Fig. 4. Over the urban clusters, we can see an increasing north–south pattern in RH and LTS, with the highest values found in the PRD. For the positive PVV, the spatial distributions for the low- and high-AOD situations are remarkably similar, with the highest values over the BTH and decreasing toward the south to near zero over the PRD. In contrast, the negative PVV is highest over the BTH, with little variation over the study area. Overall, the meteorological parameters over the YRD and PRD are similar to those over the ECS, irrespective of the AOD. Furthermore, the LTS is significantly larger in the high-AOD conditions for all four regions. Zhao et al. (2006) proposed that the enhancement in atmospheric stability tends to depress upward motion and precipitation, leading to an increase in aerosol particles. The spatial distributions of both positive and negative PVV in the low-AOD conditions are similar to those in high-AOD conditions.

The differences between the mean afternoon and morning values of cloud properties in each 1∘×1∘ grid cell in either low- or high-AOD conditions show the variation of cloud properties during 3 h of cloud evolution at low and high aerosol concentrations. Figure 5 presents the spatial distributions (left, a1–e1) and the sample series (right, a2–e2) of differences in cloud properties (CDR, COT, CWP, CF and CTP) after this 3 h period for the lowest MODIS/Terra AOD quartiles. Figure 6 shows the spatial distributions (left, a1–e1) and sample series (right, a2–e2) of these differences for the highest MODIS/Terra AOD quartiles.

Overall, we look at statistics for a large data set of 10 years. Concerning the effect of aerosol loading on cloud parameters in each urban cluster, a decrease in CF occurs over the BTH for low-AOD conditions, which is opposite to the CTP variation for both AOD conditions. For the variations of CDR over the YRD urban cluster, a significant increase occurs under high-AOD conditions, which may be attributed to the higher RH (see Fig. 4a1, a2). As regards the variation of CF and CTP, a significant decrease occurs under low-AOD conditions. Likewise, an increase in the CDR was observed for high-AOD conditions over the PRD urban cluster. Furthermore, decreases in CF and CTP were observed for low-AOD conditions and increases in CF and CTP were observed for high-AOD conditions. From the perspective of considering all urban clusters (BTH, YRD and PRD), both COT and CWP increase over land during the 3 h time step for both low and high AOD. Overall, the variation in cloud properties after the time step over BTH is less significant than over the YRD and PRD for both low- and high-AOD conditions. This may result from less humid and more unstable atmospheric environments over the BTH than over the other two urban clusters (as shown in Sect. 4.2). Over the ECS, in both low- and high-AOD conditions, CDR, CF and CTP decrease during the time step while COT and CWP increase (see Figs. 5 and 6).

In general, the variations over 3 h in COT and CWP over land are similar to those over ocean for both low- and high-AOD conditions. Another similarity is that CF decreases for low-AOD conditions over land and ocean during the 3 h time step. Having a closer look at the CF variation over the YRD and PRD, we see that CF increases in high-AOD conditions during the 3 h time step. This implies that the variation of CF may depend on the initial AOD conditions. The decrease in afternoon cloud cover over ocean confirms that the largest cover for marine clouds is reached early in the morning as was also concluded by Meskhidze et al. (2009). Meanwhile, a significant difference is found between land and ocean areas, i.e., in high-AOD conditions CDR increases over land but decreases over ocean during the 3 h time step. Table 2 summaries the differences in cloud properties between the Aqua and Terra overpasses for high- and low-AOD conditions over land and ocean during the time period 2008–2017.

The differences between the mean changes in cloud properties (CF, COT, CWP, CDR and CTP) between the Terra and Aqua overpasses in high- and in low-AOD conditions (d(Cloud_X) as defined in Sect. 3.2) are investigated to identify the effect of aerosol particles on the cloud properties. Figure 7 shows the differences between the mean change in cloud properties at low- and high-AOD conditions during the two observations at 10:30 and 13:30.

Figure 7 shows that the values of d(CDR) over the three urban clusters are not mostly positive or negative, which indicates that in two AOD conditions over land the variation in CDR during the 3 h between the MODIS/Terra and Aqua overpasses is similar. Over the ECS the values of d(CDR) are positive, which indicates that the CDR in high-AOD conditions decreases much more than during low-AOD conditions over ocean. Wang et al. (2014) also reported a negative correlation between CDR and AOD over the ECS, in accordance with the Twomey effect. Furthermore, CDR tends to be smallest in polluted and strong-inversion environments, an outcome in good agreement with the findings of Matsui et al. (2006). Most of the d(COT) values are negative over the four regions, especially for the YRD, PRD and ECS. This shows that the COT increases less in high-AOD conditions than in low-AOD conditions, over both land and ocean, which is in contrast with the findings of Meskhidze et al. (2009). Likewise, the values of d(CWP) are almost all negative over the four regions although over the BTH urban cluster the values are not clear. This indicates that in high-AOD conditions the CWP increases less during the time step than in low-AOD conditions, a result in accordance with the conclusion that higher LTS is linked with a slightly lower CWP (Matsui et al., 2006). We can conclude that the variation trend of COT and CWP after 3 h depends little on the initial AOD, but the initial AOD conditions can affect the amplitude of variation of COT and CWP. Meanwhile, the values of d(CF) are smaller than zero over the ECS. This shows that the cloud fraction in high-AOD conditions over the ECS decreases less than that in low-AOD conditions. However, Meskhidze et al. (2009) found that an increase in the aerosol concentration may lead to enhanced reduction of afternoon cloud coverage and optical thickness for marine stratocumulus regions off the coast of California, Peru, and southern Africa. Therefore, the connection between AOD and variation of cloud cover could be a response to regional-scale changes in aerosol covarying with meteorological conditions. The value of d(CF) is overall positive over the PRD, which indicates that over the PRD in high-AOD conditions the cloud cover increases much more than the cloud cover decreases in low-AOD conditions. Mauger and Norris (2007) have shown that scenes with large AOD and large cloud fraction experienced greater LTS. As regards CTP, we find that the values of d(CTP) are positive over the BTH and PRD urban cluster, but the values of d(CTP) over the other two regions do not show a clear pattern. This indicates that in high-AOD conditions over the PRD region the CTP increases much more than the CTP decreases in low-AOD conditions. We can conclude that the variation in d(Cloud_X) is different for continental and oceanic clouds. This applies to CDR, cloud fraction and CTP, but not to COT and CWP. Table 2 summarizes the differences between the mean changes in cloud properties for low and high AOD over the time step of 3 h.

Based on the above findings, we conclude that over the ECS the values of CDR, CWP and CTP are smaller but the values of COT and CF are larger in high-AOD conditions. After the 3 h time step, CDR, CF and CTP become smaller, irrespective of the AOD. Furthermore, CDR decreases much more in high-AOD conditions but CF and CTP decreases much more in high-AOD conditions. In contrast, COT and CWP become larger in both AOD conditions, and more significantly in low-AOD conditions. Over the urban clusters, COT and CWP also increase over the time step in both AOD conditions, especially for the low-AOD condition. For CF the values in low-AOD conditions decrease over the time step. The CTP change behaves different among the three urban clusters during the 3 h.

In order to explore the initial meteorological effects on the correlations between AOD and the cloud fraction, we determine the difference in mean cloud parameters between the high- and low-AOD conditions at the end of the time step (d(Cloud_X)) in meteorological variable space rather than in longitude–latitude space. Therefore, we define high and low AOD as the highest and lowest quartile for each bin of the meteorological parameters, respectively. Figure 8 shows the effect of meteorological factors (PVV, RH, LTS and initial cloud fraction) on the d(CF) when the cloud cover increases (ΔCloud_X > 0) under both low- and high-AOD conditions over land after the 3 h time step. Figure 9 shows the effect of meteorological factors on the d(CF) when the cloud cover deceases (ΔCloud_X < 0) under both low- and high-AOD conditions over land after the 3 h time step. From both figures we find that almost all d(CF) values are positive, indicating that the variations of CF are larger in high AOD than that in low AOD after the 3 h time step.

The PVV, a measure of dynamic convection strength, is very important for cloud formation. Negative PVV is indicative of upward air motion; adiabatic expansion and cooling; and hence, if cooling is sufficient, cloud formation (Jones et al., 2009). Figure 8a shows that the d(CF) decreases with the PVV over the range from -0.05 to 0.05 Pa s-1 as cloud cover increases in both conditions over the 3 h time step. This indicates that the weaker downward motion and stronger upward motion of air parcels makes the difference between the increment of cloud cover in high- and low-AOD conditions larger: in other words, the increase rate of cloud cover is larger for high AOD under stronger upward motion of air parcels. Jones et al. (2009) showed that stronger upward motion of air parcels can promote the cloud formation in both high- and low-AOD conditions, but they did not report the increase rate of cloud formation in both AOD conditions. While cloud cover decreases in both conditions over the 3 h time step, Fig. 9a shows that the d(CF) increases with the PVV over the range from -0.05 to 0 Pa s-1 and decreases with the PVV over the range from 0 to 0.05 Pa s-1. This indicates that the decrease rate of cloud cover is smaller for high AOD both under stronger upward motion of air parcels and stronger downward motion of air parcels. Outside this range of PVV values the relationship becomes harder to determine due to the reduced data volume in both cases.

Figure 8b shows that the d(CF) decreases with increasing RH when RH is lower than 20 %. This implies that the increase rate of cloud cover is smaller for high AOD with increasing RH. However, when RH is larger than 20 %, the increase rate of cloud cover is larger for high AOD with increasing RH. An increase in d(CF) occurs due to activation of CCN and formation of clouds (Feingold et al., 2003; Liu et al., 2017). It should be noted that the variation of d(CF) with increasing RH above around 80 % is uncertain as the sample sizes of high- and low-AOD conditions are small. In contrast, the d(CF) values become smaller with increasing RH over the whole RH range (see Fig. 9b), indicating that the decrease rate of cloud cover is smaller for high AOD than that for low AOD with increasing RH.

The LTS is an indicator for the mixing state of the atmospheric layer adjacent to the surface. It describes to some extent the atmosphere's tendency to promote or suppress vertical motion (Medeiros and Stevens, 2011), which in turn affects cloud properties (Klein and Hartmann, 1993). Low LTS represents a relatively unstable atmosphere and high LTS represents a more stable atmosphere. Both Figs. 8c and 9c show that the d(CF) increases and then decreases with increasing LTS when LTS is lower than 20 K, but increases with increasing LTS for higher values (LTS > 20 K). However, the sample sizes of high- and low-AOD conditions are extremely disproportionate when LTS is larger than 20 K. Therefore, it is difficult to reach a conclusion from the relationship between d(CF) and LTS when LTS is larger than 20 K.

Figure 8d shows a strong negative relationship between d(CF) and initial cloud fraction. The d(CF) increases with increasing initial cloud cover, even though the data volume becomes smaller over the range from 0 to 1.0. This implies that the increase rate of cloud cover becomes smaller for high AOD with an increase in the initial cloud cover. Likewise, Fig. 9d also shows that d(CF) decreases with increasing initial cloud cover, indicating that the decrease rate of cloud cover becomes larger for high AOD with increasing initial cloud cover. This phenomenon is different from the observed weak relationship between d(CF) and initial cloud fraction in the oceanic shallow cumulus regime (Gryspeerdt et al., 2014). It may result from the combination of the above two cases.