In order to examine the cloud cover trend in both historical and future periods, the cloud fraction data from 36 CMIP6 models are used in this work (Eyring et al., 2016). Considering the time span of all the models, the analysis is performed using the historical experiment driven by all forcings for the period from 1850 to 2014 and two future scenarios, the ssp245 (i.e., Shared Socio-Economic Pathway 2 and 2100 climate forcing level of 4.5 Wm-2) and ssp585 (i.e., Shared Socio-Economic Pathway 5 and 2100 climate forcing level of 8.5 Wm-2) experiments, for the period from 2015 to 2100 (O'Neill et al., 2016). ssp245 assumes a central pathway with continued historical tendencies, while ssp585 envisions optimistic but fossil-fueled development trends. Different future emission scenarios may provide further insights into the impacts of global climate change on cloud cover. Direct comparisons between the total cloud cover in models and satellite observations may be hampered by uncertainties due to the differences in cloud cover definitions and determination algorithms (Engström et al., 2015). Therefore, this investigation mainly employs the total and layered cloud fraction produced by the CALIPSO simulator, and the direct cloud cover simulations are used to verify the representativeness of the limited CALISPO simulator results. Three layered clouds are categorized according to the pressure thresholds of 440 and 680 hPa, i.e., high clouds (<440 hPa), middle clouds (680–440 hPa) and low clouds (>680 hPa). Note that clouds that straddle two (three) pressure layers are counted as two (three) cloud layers at the same time. Table 1 provides a list of CMIP6 models used in this study.

CVS can be fairly complex with numerous conceivable configurations; therefore reducing its complexity into a handful of manageable classes requires simplification. In accordance with Oreopoulos et al. (2017), the two atmospheric boundaries of the ISCCP cloud categories, as specified in Sect. 2.1.2, are adopted as the basis for the three standard layers of CVS classifications in each CCCM cloud group profile. Since the CCCM product also gives the cloud base location, vertically separated cloud layers can be identified. When multiple overlapping clouds coexist inside the same standard layer or they contiguously span two or three standard layers, we treat them as one single-layer cloud. Under the above presumptions, 12 combinations illustrated in Fig. 1 are conceivable, including six single-layer clouds: isolated high clouds with base pressures <440 hPa (H), middle clouds with cloud base pressures <680 hPa and top pressures >440 hPa (M), and low clouds with top pressures >680 hPa (L), as well as contiguous clouds of H and M (H×M); M and L (M×L); and H, M and L (H×M×L) and six multi-layer clouds: non-contiguous clouds of H and M (HM); H and L (HL); M and L (ML); H, M and L (HML); H and M×L (HM×L); and H×M and L (H×ML). After classification, the cloud fraction of a certain CVS in each CCCM group profile can be derived.

The radiative impacts of the total clouds and different CVS categories are investigated using the CCCM dataset. Since the satellites of the A-Train constellation used to generate the CCCM product are polar orbiting with fixed crossing times of approximately 01:30 and 13:30 local solar time (LST), the instantaneous solar irradiance is initially adjusted with the daily average solar insolation, Fs↓, as applied by Haynes et al. (2013). The CRE at the TOA or surface is then calculated by 1CRE=(Fx↓-Fx↑)cloudy-(Fx↓-Fx↑)clear, where F is the irradiance at the TOA or surface; the subscript x is either shortwave (SW) or longwave (LW); and the cloudy and clear indices denote cloudy-sky and clear-sky conditions, respectively. The superscripts ↓ and ↑ indicate downward and upward fluxes, respectively. The sum of SW and LW CRE gives the net CRE. The difference between the CRE at the TOA and the surface is the CRE within the atmosphere.

The vertical irradiance profile provided by the CCCM product is used to obtain the heating rate (HR) profile, and the HR at a certain layer is computed by 2HR=dTdt=(Fx↓-Fx↑)upper-(Fx↓-Fx↑)lowercp(plower-pupper)/g, where T is the layer temperature, t is time, F is the irradiance, p is the pressure, cp=1004 Jkg-1K-1 is the specific heat capacity of air at constant pressure, and g=9.81 ms-1 is the gravitational constant. The subscripts upper and lower denote, respectively, the upper and lower boundary of a layer, and x is either SW or LW. The unit of HR is converted to Kd-1.

After calculating the HR, the cloud heating rate (CHR), which denotes the HR between cloudy-sky and clear-sky conditions, is represented by 3CHR=HRxcloudy-HRxclear, where the subscript x is either SW or LW, and the superscripts cloudy and clear denote cloudy-sky and clear-sky conditions, respectively. The sum of SW and LW CHR indicates the net CHR.

Note that when we examine the CRE and CHR for a specific CVS class, the cloud group containing only one CVS class with a cloud fraction of 100 % is considered. In terms of the spatial distributions, the cloud fraction and CRE calculated in a CERES footprint are further monthly averaged to a grid of 2∘×2∘.

Before delving into the CCCM product, we first assess its cloud fraction in comparison to the data from the Level-2B GEOPROF-LIDAR. Although Ham et al. (2017) have conducted a 4-month thorough comparison between these two products, the findings are only applicable to the previous version of the CCCM data (RelB1). Overall, the 4-year assessments in Fig. S1 in the Supplement indicate that the CCCM and GEOPROF-LIDAR products capture quite comparable features, and the temporal correlations are extremely strong for both total and different types of cloud cover. However, biases between these two products cannot be ignored, notably for the middle cloud, which has a global average cloud fraction bias of 5.74 %. These biases are mainly induced by differences in cloud mask algorithms between the CCCM and GEOPROF-LIDAR discussed in Sect. 2.1.2, despite their employing of the same satellite products (Ham et al., 2017).

The spatial distributions of the 12 CVS categories are shown in Fig. 2. Although there exist slight differences in daytime and nighttime cloud fractions of the 12 CVSs, the magnitudes are consistent (Fig. 1); thus we merely show the average values here. The statistical results demonstrate that single-layer clouds of L, H and, to lesser extent, H×M×L, as well as the multi-layer cloud of HL, occur frequently, whereas the other eight CVSs all have relatively lower cloud fractions, with area-weighted averages of less than 5 %. Basically, H distributes with latitudes, with its high values in the Tropics across the west-central Pacific Ocean warm pool, Indonesia, western Africa and central South America. The Tibetan Plateau, which is dominated by high topography, is another region with a high H fraction. In contrast, L is distributed throughout the low-value zones of H and shows a clear land–ocean difference with very low values over land, except for some regions in the Northern Hemisphere mid-to-high latitudes. The L class has very high values over oceans globally, except where the H class is prevalent. The distributions of HL and H×M×L generally follow a similar pattern as H, apart from the low values over the Tibetan Plateau due to the absence of low-level clouds. For the eight infrequent CVSs (average cloud fraction <5 %), their spatial patterns resemble the feature of H×M×L, excluding the high values of H×M over the Tibetan Plateau and Antarctica caused by high terrain. In conclusion, the spatial patterns of these 12 CVSs reveal consistent distributions with well-known cloud top height/pressure characteristics (Marchand et al., 2010; King et al., 2013) and, for the single-layer clouds, documented cloud regimes from approaches (Tselioudis et al., 2013, 2021; Unglaub et al., 2020). However, the quantitative cloud fractions of the various CVS types provided here have not previously been achievable with passive satellite sensors.

For the six overlapping cloud types, it is more challenging to derive the cloud radiative effect. However, their fractions are much lower than the ones of the single-layer clouds. Figure 3 shows the spatial distributions of the single-layer and multi-layer clouds, as well as their ratio. In the global average, single-layer clouds are 3.74 times more frequent than multi-layer clouds. This ratio exhibits considerable regional variations that are associated with clearly distinguishable regimes. In the time average, nowhere are multi-layer clouds more frequent than single-layer clouds. Over tropical convective zones, multi-layer clouds are almost as frequent as single-layer clouds. A reason for this is that cirrus clouds from either large-scale ascents or dissipating deep convections are ubiquitous in these regions, both in the absence (Fig. 2a) and in the presence (Fig. 3c) of low-level clouds below the cirrus. In contrast, near the descending branch of the Hadley cell in both hemispheres, single-layer clouds are often an order of magnitude more frequent than multi-layer clouds. There, the prevalent subsidence is unfavorable for the formation of mid- or upper-level clouds (Yuan and Oreopoulos, 2013). The large ratio values over Antarctica and Greenland are influenced by the descending branch of the polar circulation. In brief, the single-layer clouds prefer regions with a stable troposphere, while the multi-layer clouds favor the regions with strong ascents.

In addition to the global horizontal distributions, another concern is the vertical extent of each type of CVS. These, provided by the top and base heights, are shown as global averages in Fig. 4. In the presence of high-level clouds, different CVSs present distinctive top heights. When the high-level clouds occur alone or overlap thinner low-level clouds (average geometric thickness ≤1 km), the cloud tops exceed 12 km (H, HM and HL), primarily in the Tropics throughout the west-central Pacific Ocean warm pool, Indonesia, western Africa and central South America. However, the cloud tops drop to approximately 10–11 km when thicker low-level clouds (average geometric thickness ≥2 km) are overlapped by high clouds (e.g., HML and HM×L), mainly across equatorial and mid–high latitudes. The average cloud top of deep convective clouds (H×M×L) distributed mostly over equatorial and mid–high latitudes is >11 km but lower than the isolated H. The cloud tops are much lower (<10 km) when the high-level clouds and mid-level clouds are contiguous (e.g., H×M and H×ML), and they are generally spread over high altitudes and mid–high latitudes.

Morphological differences in low-level clouds among the CVSs are also observed. The cloud base heights are higher (>1 km) when the low-level clouds connect with the mid-level clouds (e.g., M×L and HM×L), mostly occurring across equatorial and mid–high latitudes. Contrarily, deep convective clouds (H×M×L), which prevail over the same regions as M×L and HM×L, have the lowest base height.

The CCCM product makes use of the combination of cloud profiles from active satellites, cloud optical properties from MODIS and broadband radiation fluxes from CERES to compute radiative flux profiles using radiative transfer modeling. Figure 5 shows the total SW, LW and net CREs at the TOA and surface as well as within the atmosphere as provided by 4 years of CCCM data from 2007 to 2010. Overall, this matches the global patterns of CREs examined in previous studies using other datasets (Allan, 2011; Dolinar et al., 2019), although values deviate somewhat. In total, clouds act to cool the Earth atmosphere system with a global average net TOA CRE of -18.7 Wm-2, which is due to the cloud albedo effect. Over northern Africa and other bright surfaces (e.g., Greenland, the Arctic, Antarctica), there is a slight warming at the TOA, and the cloud greenhouse effect dominates. The SW CRE manifests primarily as surface cooling, with an average residual heating of 4.69 Wm-2 within the atmosphere. This heating effect is partly related to an enhanced SW absorption by water vapor in the atmosphere in cloudy compared to clear skies (Sohn et al., 2006; Allan, 2011) and partly to SW absorption by the clouds themselves (Slingo and Schrecker, 1982). The SW CRE within the atmosphere is homogenous globally, so the spatial pattern of net CRE within the atmosphere is driven by LW. At the surface, net CRE is dominated by SW cooling, except over the poles, which are warmed by LW heating.

Given the systematic differences in the total opacities and thermal emissions caused by vertical extent and temperature-dependent cloud phase, it is evident that different CVSs influence the radiative flux for both SW and LW within the atmosphere in distinct ways. The geographic variations in prevalent CVSs over different regions cause the spatial pattern of CRE in Fig. 5. Therefore, quantifying the global average CREs induced by various types of CVSs is valuable. Figure 6 shows the global average SW, LW and net CREs at the TOA, within the atmosphere and at the surface for the 12 classified CVSs.

In terms of SW CRE, the magnitude of each CVS is similar at the surface and TOA, with a relatively small value within the atmosphere. The minor positive values of SW CRE within the atmosphere are due to the increased atmospheric path length for radiation reflected by clouds that caused an enhanced absorption by water vapor, but for some high-level clouds (e.g., H and H×ML), the reflection at high altitudes instead decreases the SW absorption by the atmosphere that induces tiny negative values. At the TOA and surface, all types of CVSs exert cooling effects in the SW. In general, low-level clouds have stronger SW CREs than high-level clouds due to the generally vertically decreasing profile in cloud water content and due to cloud-phase differences. When they are overlapped or connected with upper-level clouds, the SW CREs further increase as the vertically integrated water content and thus optical thickness increases. The deep convective clouds of H×M×L with an average geometric thickness >10 km have, as expected, the strongest albedo effect (SW CRE<-85 Wm-2).

Compared to SW, LW CRE exhibits more complex features. At the TOA and surface, all types of CVSs act as warming effects in LW, but the magnitude differs a lot. At the TOA, since the LW CRE highly depends on the temperature difference between the surface and cloud top, CVSs containing high-level clouds all have a strong LW CRE (>40 Wm-2), notably for the H×M×L with its value >90 Wm-2. On the contrary, LW CRE at the surface greatly depends on the cloud base thermal emission. Therefore, the CVSs containing low-level clouds all have a strong LW CRE at the surface (>38 Wm-2), whereas the H with the highest cloud base presents the weakest surface LW CRE (<10 Wm-2). As the LW CREs at the TOA and surface are quite dissimilar, the LW CREs within the atmosphere display a wide range among all the CVSs. The clear distinction is that the H causes LW radiative warming within the atmosphere, while the L causes LW radiative cooling. The LW CREs generated by the clouds between the locations of L and H or the combination of the two can be offset to some extent.

The net CRE, which combines both SW and LW, indicates whether a specific type of CVS has an overall warming or cooling effect. At the TOA, four CVSs have warming effects, all of which include high-level clouds, while eight CVSs have cooling effects, and most of them contain low-level clouds. For M or HL, SW and LW cancel out almost entirely. Within the atmosphere, there are seven CVSs that in the net warm the atmosphere, while five CVSs cool it. At the surface, all types of CVSs have net cooling effects, as the reduction in SW reaching the surface is larger than the increase in downwelling LW. However, in terms of spatial distribution, there are positive values of net CRE observed at the surface over bright areas (e.g., Greenland, the Arctic, Antarctica) (Fig. 5i), where the cloud greenhouse effect prevails. Nevertheless, when examining the global average, the net positive CREs of cloud types that dominate over these bright regions are rather small in magnitude compared to the average albedo effects of the same cloud types over most other regions, ultimately resulting in net cooling at the surface. We conclude that these intriguing discrepancies in the CREs of all kinds of CVSs contribute to large uncertainties in estimating changes in the radiative budget when the spatial patterns of CVS change.

Apart from the integral CREs within the atmosphere, the CHR profiles of 12 CVS types are further illustrated in Fig. 7, which can provide detailed profiles of how clouds vertically affect radiative heating as provided by the CCCM dataset. Overall, the net CHR profiles are driven by the LW component, and the CHR profiles of multi-layer clouds are more curved and complex than those of single-layer clouds. Regarding SW, all the CVSs exert similar characteristics, which are shown as heating near the cloud layers and cooling beneath the clouds. Due to the SW absorption within the upper parts of some optically thick clouds (e.g., M×L and H×M×L), the SW cooling starts to appear in the middle and lower portions of the clouds. Concerning LW, the heating below the cloud layers is due to the absorption of LW radiation emitted from the surface or the lower clouds below, while the LW cooling near and above the cloud layers is the result of radiative emissions by the clouds. Strong greenhouse effects are produced by ample ice particles inside the H, exhibiting inconsistent characteristics distinguished from the other CVSs and even heating all levels below the cloud top. In conclusion, the SW albedo effects, LW greenhouse effects and the interactions between cloud layers result in rather complex radiative profiles, contributing to manifold atmospheric thermal stratifications. The precise assessment of these stratifications is inextricably linked to the accurate observation of cloud properties, especially the detection of vertically overlapping clouds.

In light of the widely disparate radiative effects of different kinds of CVSs, the response of CVS to a warming climate appears to be particularly important. However, how the CVS changes during the historical period and future projections remains poorly constrained. Some aspects have been documented in Norris et al. (2016) from passive satellite sensors. In this section, the trends in total, high-, middle- and low-level cloud fractions from 1850 to the end of this century are analyzed based on CMIP6 models. The historical experiment driven by all sorts of forcing for the period from 1850 to 2014 and two future scenarios (the ssp245 and ssp585 experiments) for the period from 2015 to 2100 are used. Furthermore, we investigate whether the global spatial distribution of the dominating cloud type will change as a result of climate warming.

Because GCMs only parameterize cloud fraction, the performance of the CMIP6 models is initially assessed with CCCM observations. Figure 8 presents scatterplots of the monthly average cloud fraction between the MME using the CALIPSO simulator diagnostics available from eight CMIP6 models and the CCCM datasets, including total, high, middle and low clouds between 2007 and 2010. Land and ocean regions are separated. The results indicate that CMIP6 models in general capture the monthly mean cloud fraction for both total and layered clouds rather well, especially over the ocean with correlation coefficients ≥0.9. Cloud fraction in CMIP6 is systematically underestimated due to the ability of two active satellites (CALIPSO and CloudSat) in CCCM to retrieve more clouds than CALIPSO alone. Although this underestimation mainly occurs over the tropical regions, the correlations against CCCM are even stronger than in high latitudes (Fig. S2). It is concluded that the CMIP6 models perform well in simulating the cloud fraction for both the total and the layered clouds, implying that estimating the historical and projected cloud fraction using CMIP6 is reliable. Additionally, since there are only two models (GFDL-CM4 and IPSL-CM6A-LR) available for the future period, it is crucial to assess whether these two models have representation comparably good to the MME mean. Figure S3 compares the total, high-, middle- and low-cloud fractions of the two CMIP6 CALIPSO simulator MMEs with the total eight CALIPSO simulator MMEs for the historical period from 1850 to 2014. Figure S4 further analyzes the total cloud fraction correlations between the two CALIPSO simulator MMEs and 32 model MMEs for four different periods from the past to the future. The results demonstrate that for both historical and future periods, the two models, for which CALIPSO simulator output is also available for future scenarios, have a fair representation of the simulated cloud fractions over both land and ocean regions. Although the cloud fraction from the direct simulation of the 32 GCMs is defined differently from that in CALIPSO simulators, they are still highly correlated. Besides the intercomparison between the models, a similar assessment result as in Fig. 8, but the relationship between the average of two models (GFDL-CM4 and IPSL-CM6A-LR) and CCCM during 2007–2100, is depicted in Fig. S5.

Time series of annual average cloud fraction based on the two CMIP6 models from 1850 to 2100 are presented in Fig. 9, including two scenarios of ssp245 and ssp585 for the future. Here, four time periods are specifically focused on to understand the temporal variations, which include the baseline (1994–2014), near term (2021–2040), midterm (2051–2070) and long term (2081–2100). The spatial differences between the future periods under ssp585 (ssp245) and the historical baseline are illustrated in Fig. 10 (Fig. S6). The results show that the projected total cloud fraction decreases faster over land than over the ocean. High clouds over oceans increase dramatically, while other types of clouds over land and the ocean all continue to decrease, which helps to almost offset the reduction in the oceanic total cloud fraction. Though the global average middle- and low-cloud fractions both decrease over land and the ocean in the future, the low-cloud fractions over the tropical ocean and the Arctic show a significant increase. The increase in oceanic high clouds is spatially concentrated across the tropical Pacific Ocean and the high latitudes, emphasizing the ensuing positive cloud feedback generated by the increased ice clouds. The decreasing trend difference between continental and oceanic total cloud cover is also influenced by spatial patterns; cloud fraction over land (except the polar regions) consistently decreases, whereas the opposing tendency between high and low latitudes over the ocean offsets the global average.

For the near term, there are only slight differences in the total cloud fraction over land between ssp245 and ssp585, but for the middle and long term, the cloud response becomes more sensitive to different anthropogenic forcing. The total cloud fraction over the ocean, however, only shows a discernible difference between the two scenarios in the long term. The different layered cloud fractions exhibit the same time series feature as the total cloud fraction over land. Over the oceans, the high cloud shows small differences between the two scenarios over all the projected periods, while the low cloud only shows a discernible difference in the long term. Although the oceanic middle-cloud fraction changes significantly over the middle and long terms, its small value means that it contributes little to the total cloud cover. The combined changes in the high and low clouds over the ocean mainly result in the time series features of the total cloud fraction.

As the above analysis demonstrates, the changes in total cloud cover and distinct types of cloud cover are noticeable in the context of climate change. The interesting question that follows is could climate warming and anthropogenic forcing change the driving cloud type in a certain area? Here, we analyze the correlations between the cloud types and the total cloud cover using historical forcing and future scenarios. Two correlation coefficient thresholds of 0.66 and 0.9 with p value <0.05 are employed, which means likely positive correlation and very likely positive correlation, respectively (Chen et al., 2021). When only one cloud type is correlated with the total cloud fraction, we presume that only this cloud drives the total cloud cover. When two or three cloud types are simultaneously correlated with the total cloud fraction, we assume that the total cloud cover is driven by these several cloud types together.

Figure 11 depicts the results with a correlation coefficient >0.66, and the results with a correlation coefficient >0.9 are displayed in Fig. S7. Over the ocean, the changes in total cloud fraction are mainly governed by low cloud cover, with the exception of the tropical Pacific Midwest and the Indian Ocean, where high clouds or a combination of high and middle clouds dominate the total cloud cover change. Over land, the regional differences are more pronounced. At high northern latitudes, middle and low clouds together drive the total cloud cover. In low-latitude regions, high clouds have a greater influence on the total cloud cover, and some regions, such as South America, South and Central Africa, and Indonesia, are synchronously affected by middle and low clouds. In the Antarctic region, the total cloud fraction is primarily affected by the middle cloud, which has a clearer signal when the correlation coefficient threshold is increased to 0.9, as seen in Fig. S7. Moreover, by comparing the results of different periods and scenarios, we can conclude that climate change and human activities have little impact on this spatial pattern.