Contrasting temporal trends in low-level clouds and mid- and high-level clouds over the Indian Ocean in the last four decades (1979–2018)

Understanding the climatic trends in cloud fraction (f c ) and its drivers is critical in climate science. Here, we analyzed 40 years (1979–2018) of hourly f c data at 0.25° × 0.25° spatial scale from ERA5 to examine the trends in the 3D distribution of f c over the oceanic region adjacent to the Indian Subcontinent in view of its drivers. We found that the mid-level f c (MCF) and high-level f c (HCF) have increased by 0.05 and 0.05–0.12 fraction in the last four decades in this region. On the contrary, the low-level f c (LCF) decreased by 0.04 fraction, resulting in a net marginal increase in total f c . The observed contrasting trends in LCF, MCF and HCF are manifestations of the changes in sea surface temperature and meteorological conditions. Although LCF showed a regionally averaged declining trend, it has increased over the south-southwestern part of the domain. The increasing trend of MCF and HCF can be attributed to the increase in natural convection due to surface warming. Our results suggest that the observed contrasting trends might have resulted in positive radiative feedback on the Indian Ocean warming.


Introduction
Clouds play an important role in modulating the Earth's energy budget and the hydrological cycle (Stephens 2005, Kato et al 2008).Clouds at low, mid, and high levels exhibit unique radiative responses and hydrological properties, thus, any change in f c and the frequency of their occurrences can significantly influence the global climate (Graf 2004).Clouds are also a major source of uncertainty in climate models as it is difficult to parameterize cloud-related processes and their interaction with different components of the Earth system (IPCC 2021).The ability of the climate models to simulate long-term changes in cloud characteristics (that are governed by greenhouse gas and aerosol forcing) is critical in reducing uncertainty in climate projections.For this, understanding the long-term trends of f c at regional scales from the observational datasets is key.
Various efforts have been made to document the long-term variation in f c over different parts of the world.Eastman et al (2011) compiled data from ship observations and developed the Extended Edited Cloud Reports Archive (EECRA) cloud atlas over the oceans for the period 1954-2008.An increase in total f c was reported over the central Pacific, and a decrease in low-level stratiform clouds over the regions of persistent marine stratiform clouds.Eastman and Warren (2013) also compiled a cloud dataset over land utilizing ground observations.While these datasets provide insights into the long-term trends of the global f c , the data has human observation error and is available at a very coarse resolution (10°× 10°over the ocean and 5°× 5°over land).The other potential data for this analysis is the International Satellite Cloud Climatology Project (ISCCP) f c data available for the period 1983-2009(Rossow and Schiffer 1999).Analyzing this dataset, Norris et al (2016) found a poleward retreat in mid-latitude storm tracks with the expansion of subtropical dry zones and an increase in the height of the high-level cloud tops.The ISCCP data has been reported to be affected by the view-angle and resolution effect, and it also suffers from spurious trends due to sensory degradation and deviation of the satellite orbits (Evan et al 2007, Norris andEvan 2015).
A slew of new-generation sensors, both passive and active, were launched at the end of the 20th and the beginning of the 21st century for Earth observations.Comparative analysis of f c reported by the sensors revealed large discrepancies among themselves (Stubenrauch et al 2013).The Global Energy and Water Cycle Experiment (GEWEX) program attributed these discrepancies to five major factors in view of the strengths and weaknesses of the sensors (Stubenrauch et al 2013).Furthermore, these sensors only provide data for the last two decades, that too at satellite overpass times, limiting a comprehensive analysis of the climatic trends that are also affected by the diurnal cycle of clouds.
Amongst the global oceans, the Indian Ocean (including the Arabian Sea and Bay of Bengal) surrounding the Indian subcontinent is crucial due to its significant influence on regional weather patterns, monsoon dynamics, and global climate systems.Given the importance of the Indian Ocean in regulating the South Asian climate, it is important to elucidate the changes in LCF, MCF, and HCF in the region under a warming climate.
To date, two global studies (Eastman et al 2011, Norris et al 2016) have examined long-term changes in f c over the Indian Ocean.Both these studies processed the data till 2008-2009 at a coarse spatial scale, and the changes in more recent years were not reported.Moreover, the studies did not focus exclusively on the clouds at low, mid and high levels.Here we filled this knowledge gap by analysing and reporting the 3D variations of f c and its trend over the Indian Ocean in the last four decades .We explained the temporal trends in LCF, MCF, and HCF in terms of their driving factors.

Methodology
We processed three types of cloud datasets.Our main cloud dataset is from the European Centre for Medium-Range Weather Forecasts (ECMWF) fifth-generation reanalysis (ERA5).Since ERA5 f c was not used earlier for climate studies in this region, we evaluated the data against the resolution-corrected Multi-angle Imaging SpectroRadiometer (MISR) f c data (Jones et al 2012).We also assessed the vertical profiles of ERA5 against GCM-Oriented Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observation (CALIPSO) Cloud Product (GOCCP) product.We further processed variables such as sea surface temperature (SST), vertical wind at 700 hPa (w700), and convective available potential energy (CAPE) from ERA5 data to explain the observed trends.
We also compared cloud radiative forcing from ERA5 and Clouds and the Earth's Radiant Energy System (CERES) and a temporally consistent bias of 5 Wm −2 to 10 Wm −2 was found between the two datasets.We later quantified the change in cloud radiative forcing from 1979-2018.A forcing adjustment feedback framework was used (Sherwood et al 2015, Klein et al 2018), where, the change in cloud radiative forcing (R) is assumed to be linearly related to global mean surface air temperature (T g ) as follows- Here, a is cloud fraction over the basin and dC Here x represents the cloud-controlling factor.We used multilinear regression to quantify the respective contribution to cloud-controlling factors on f c .

Cloud datasets
ERA5 provides global three-dimensional f c data at 0.25°× 0.25°spatial and hourly temporal resolution at 37 pressure levels from 1000 hPa to 1 hPa (Silber et al 2019, Hersbach et al 2020).In ERA5, an advanced parameterization scheme developed by Tiedtke (1993) was used for large-scale precipitation and stratiform clouds.The f c , water vapor, and cloud liquid water were predicted with a model that considered autoconversion, accretion, snow riming, and the Wegener-Bergeron-Findeisen processes parameterized to represent the interactions between the various water species via a bulk mass flux scheme (Bechtold et al 2014 Cloud datasets derived by all modern-day passive sensors are affected by resolution and/or view-angle effects (Stubenrauch et al 2013).Recently, the resolution effect was corrected in the MISR f c product by Jones et al (2012) following an earlier work by Di Girolamo and Davies (1997).Global analysis by Dutta et al (2020) revealed a large (by 0.3-0.5)reduction in absolute f c values in the tropics frequented by cumulus clouds.They also found that the new dataset agrees within ±0.05-0.08 of f c derived from ASTER data and interpreted that this dataset can be considered as closest to being 'true.'For the comparison with ERA5, we processed the MISR data for the period 2000 to 2018.
The GOCCP cloud dataset was developed from the CALIOP L1 data for consistent comparison between CALIOP observations and 'GCM + lidar simulator' outputs (Chepfer et al 2010).Here, we processed the GCMoriented GOCCP data for the period (2007-2018) and compared it against the vertical profile of ERA5 f c data over the Indian Ocean (figure S2).An earlier study by Binder et al (2020) combined CloudSat radar and CALIPSO lidar data and found that ERA5 data is capable of capturing the vertical structure of warm conveyor belt clouds related to extra-tropical cyclones (figure S2).

Meteorological variables
We used the estimated inversion strength (EIS) defined by Wood and Bretherton (2006) as a measure of the strength of the inversion layer at the top of the boundary layer: where θ 700 and θ sfc denote potential temperatures at 700 hPa and surface, respectively, and Z 700 and Z LCL denote local altitude of 700 hPa and lifting condensation level, respectively.m 850 G is the moist adiabatic lapse rate at 850 hPa level.We analyzed these meteorological variables from the ERA5 data and estimated hourly EIS for the study period.
ERA5 hourly dataset of vertical wind at 700 hPa (w 700 ) was also processed because this pressure level is close to the upper limit of low-level f c .Vertical velocity is useful for understanding the large-scale dynamics of the atmosphere, including areas of ascent and subsidence (Myers and Norris 2013).Convective available potential energy (CAPE) is an indicator of the instability (or stability) of the atmosphere and was used to assess the potential for the development of convection, which can lead to heavy rainfall, thunderstorms, and other severe weather.In the ECMWF Integrated Forecasting System, CAPE is calculated by considering parcels of air departing at different model levels below the 350 hPa level.ERA5 reports CAPE for most unstable air parcels (Bechtold et al 2014).The 2m air temperature (t2m) from ERA5 was used as surface air temperature in the analysis.

Sea surface temperature (SST)
SST was found to influence f c variations (Gadgil et al 1984).ERA5 data assimilation system used the HadISST2 dataset before Sep 2007 and the OSTIA dataset afterwards to estimate SST.In a study over the Pacific and Atlantic Oceans (Yao et al 2021), the ERA5 SST was found to be in good agreement with the observation when wind speeds were above 6 ms −1 and below 6 ms −1 a cold bias appeared, indicating ERA5 SST remained unaffected by solar radiation heating.

Analysis
The analysis is divided into two parts.First, we processed various cloud datasets to compare ERA5 f c with MISR and GOCCP.Since no 'true' cloud datasets exist, we restricted ourselves mostly to inter-comparison and inferred whether the broad patterns are matching.We further generated the cloud climatology and analysed the trends in view of the changing meteorological conditions.
For the climatology, we converted the monthly f c data to the respective seasons-winter (Dec-Feb, DJF), premonsoon (Mar-May, MAM), monsoon (Jun-Sep, JJA) and post-monsoon (Sep-Nov, SON).For the meteorological variables, climatology was established following the same method.The temporal correlation between the f c and meteorological variables was performed to analyse the spatial correlation.We estimated temporal trends after removing the seasonality of monthly data using a linear regression model and statistical significance using Durbin-Watson statistics (Durbin and Watson 1950).The Pearson Correlation Coefficient was used to calculate the temporal correlation and statistical significance between two variables.In proximity to coastlines, ERA5-fc values are higher than MISR-fc (figure 1(S)), likely attributable to the existence of thin cirrus clouds over and near the landmass, which may go undetected by MISR.We derived the vertical cross-sections of f c from ERA5 and GOCCP at the available pressure levels for the same period across the latitudes (figure S2).In each season, the broad zonal patterns resemble each other quite well despite a mismatch in sampling density.The comparative statistics suggested that ERA5-f c is quite reasonable in capturing the known spatiotemporal distributions of f c in this region, and therefore, we proceeded with the data for subsequent analysis.

Seasonal climatology of f c over the Indian Ocean
The f c at all levels shows a strong seasonal cycle over different parts of the Indian Ocean (figure 2).LCF dominate over the southern part of the Indian Ocean throughout the year, while over the northern part of the domain, LCF are scant and scattered yet cavort to seasonal changes.During DJF, LCF shows a distinct maximum on the West coast of Australia at around 105 ˚E and rapidly declines toward the west, but LCF is more zonally uniform during JJA and SON, which agrees with Miyamoto et al (2018).Seasonal maxima of LCF can be noted over the Arabian Sea and the Bay of Bengal during DJF and JJA.At the low level, the northern part of the South China Sea remains cloudy throughout the year.
MCF over the Indian Ocean have the lowest climatological magnitude among the clouds at three levels.Throughout the year, high values of MCF were predominantly found over the Equatorial Indian Ocean except during JJA when they were found spreading from North to South.The spatial pattern of MCF matches HCF, yet with less spatial coverage.The HCF dominate throughout the year over the Equatorial Indian Ocean between 15°N and 15°S due to the favourable meteorology (Cetrone and Houze 2009, Nair et al 2011), where the eastern part remains more cloudy as compared to the western part.Intense high-level clouds can be seen over the Bay of Bengal and south-eastern parts of the Arabian Sea during JJA and SON.The LCF showed a negative correlation with SST over the Indian Ocean (figure S3).Strong negative correlation persists over the low SST waters, but positive correlations are prevalent over the warm SST regions.The cool SST increases the lower tropospheric stability (LTS, θ 700 -θ sfc ) by reducing the air temperature close to the surface and thereby enhancing the LCF.The cumulus-type low-level clouds are positively correlated to SST due to trade-offs between SST and LTS resulting in a transition from stratiform to cumuliform clouds (Clement et al 2009, Klein andHartmann 1993).
The MCF showed a dual correlation pattern with SST (figure S3), with coexisting seasonal and spatial variations of positive and negative correlation.In a study conducted over Darwin, Australia, Riihimaki et al (2012) observed that the precipitating high-level clouds are succeeded by the formation of thin mid-level clouds, which are attributed to increased humidity at mid-level altitudes.Other studies also confirm the presence of thin-mid-level clouds over the tropical oceans (Seifert et al 2010).If some mid-level clouds over the Indian Ocean are hypothesised to be formed after rainfall, the rainfall can relax the SST.It can be reflected as a negative correlation between SST and MCF.
The HCF showed mostly a positive correlation with SST during all seasons.A negative correlation was also observed over the equatorial region.The presence of convective clouds in the intertropical convergence zone (ITCZ) region can lead to the relaxation of SST after precipitation.It hence can be reflected as a negative correlation in the monthly average of HCF and SST (Mace et al 2006).

Temporal trends of 3D cloud structure over the Indian Ocean
We found a decreasing trend in LCF across the Indian Ocean, similar to the previous observations by Eastman et al (2011) and McCoy et al (2017).However, the MCF and HCF show an increasing trend (figure 3) in the last four decades.The highest declining trend in LCF was observed in SON, while the highest increasing trends in MCF and HCF were observed in DJF and MAM, respectively.
Over the last four decades, the MCF and HCF have increased by 0.051 fraction and 0.05-0.12fraction respectively.However, the LCF decreased by 0.04 fraction over the same period (figure 3).
The temporal trends of f c at various heights along zonal and meridional cross-sections shown in figure 4 revealed the distinct polarity more clearly.Our findings reveal a dual pattern in LCF over the Indian Ocean.Clouds near the surface exhibit a decreasing trend across the entire region, while those above the marine boundary layer show an increasing trend, particularly over the south-southwestern part.Oceanic LCF is influenced by EIS andSST (Eastman et al 2011, Miyamoto et al 2018).In our study region, the EIS has decreased and this has led to a decrease in LCF (figures S5 and S6).However, in the regions where EIS show an increasing trend, LCF also show an increasing trend (figure S5).The MCF demonstrated an increasing trend over the Indian Ocean.For MCF, w700 shows a significantly positive correlation through all seasons (figure S4).However, the temporal trend of w700 shows an insignificant trend (figure S6).The HCF also show an increasing trend over the northern part (north of 10°S) and a decreasing trend over the southwestern part of the basin.Our findings are consistent with the earlier studies reporting an increase in cloud top height and HCF (Norris et al 2016, Zelinka et al 2016).The CAPE shows a strong positive correlation with HCF (figure S4) but a decreasing temporal trend (figure S5 and S6).The decreasing trend of CAPE is consistent with DeMott and Randall (2004).

Cloud radiative effect and its trend
The mean cloud radiative forcing calculation shows an increase of 8 Wm −2 in the long-wave cloud radiative forcing and a decrease of 10 Wm −2 in the short-wave cloud radiative forcing over the last four decades.The effective top-of-the-atmosphere change in the cloud radiative cooling effect of 2 Wm −2 over the last 4 decades.Therefore, total cooling of 0.50 Wm −2 decade −1 is due to changes in clouds.Loeb et al (2021) found a total radiative forcing of 0.50 ± 0.47 W m −2 decade −1 attributable to decreases in clouds and sea ice and increases in trace gases and water vapour.
Multilinear regression was used to quantify the feedback of individual cloud types using cloud-controlling variables with a confidence interval above 95% (Supplementary section S1).The multilinear regression was performed on the spatially averaged temporal variables.The LCF feedback of −0.0071K −1 , MCF feedback of −0.05914K −1 and HCF feedback of −0.002844K −1 were observed over the Indian Ocean.
Although the correlation between the f c and atmospheric variables was discussed earlier (sections 3.1 & 3.2), the controlling factors can be highly inter-correlated.The multilinear regression model suggests that SST, EIS, and w700 are significant predictors of LCF.The SST, W700, and CAPE are significant predictors of MCF, while the EIS, w700, and CAPE are significant predictors of HCF.
Equation (1) was utilized to measure how changes in f c over the past forty years have affected the change in cloud radiative feedback.Our analysis also found the LCF and HCF linearly correlated with the top of the atmosphere long-wave radiative forcing (figures S9 and S10).The cloud radiative feedback increased by 0.000284 Wm −2 K −1 due to a change in the LCF, it decreased by 0.00502 Wm −2 K −1 due to an increase in MCF and it reduced by −0.0003412 Wm −2 K −1 due to increase in the HCF.

Discussion and conclusions
The lack of highly accurate and stable observations over the decades has been a major challenge in studying the long-term trends of f c .Here, we explored the ERA5 f c record to understand the trends in LCF, MCF, and HCF over the Indian Ocean in four decades.We want to highlight that ERA5 is a reanalysis dataset, and therefore the reported f c can be influenced by the choice of parameterization scheme in the ECMWF modelling framework.Since ERA5 data was not used earlier for cloud studies in this region, we compared ERA5 total f c data with resolution-effect corrected MISR and the vertical distribution of f c from ERA5 with that from GOCCP datasets.However, we note that our interpretation of the robustness of ERA5 f c data is based on comparison during overpass times of MISR and CALIPSO-CloudSat.Unfortunately, there are no 'true' cloud datasets available for further assessment.
Our results have important climatic implications.A decreasing trend in LCF implies a gradual reduction in daytime cooling via a decrease in short-wave reflection and night-time cooling via the long-wave radiative effect (figures S9 and S10).Simultaneously, an increasing trend in HCF implies enhancement in night-time warming over the years.We hypothesize that the overall consequence of the observed changes in f c would have net positive feedback in this region, which, in turn, could also play a key factor in the observed increase of the MCF and HCF.Further modelling study is recommended to examine this aspect.
The key conclusions of this study are as follows.
• The LCF exhibit an overall decreasing trend driven by the interplay of SST, EIS and w700.Regionally averaged, LCF has decreased by 0.04 fraction in the last four decades over the Indian Ocean.However, an opposite trend was observed over the southern part of the domain due to an increase in EIS under global warming.
• From 1979 to 2018, MCF increased by 0.05 fraction, and the increasing trend is prominent over the entire Indian Ocean.The MCF have an ambiguous correlation with SST but a very strong positive correlation with the vertical wind at 700 hPa.Yet the interplay between SST, w700 and CAPE plays a vital role in the observed trend of MCF.
• The HCF have increased by 0.05 to 0.12 fraction from 1979 to 2018.The interplay between EIS, w700 and CAPE controls the variability of HCF in the region.Overall, the reduction in LCF and the rise in HCF could enhance net warming and influence the rise of SST through positive climate feedback in this region.
dT g shows cloud feedback due to local cloud-controlling factors.Here C represents the fraction of individual cloud types.The local cloud feedback resulting from the local cloud controlling factors can be expanded as follows- 3.1.Inter-comparison of ERA5 data with MISR and GOCCP For comparison between ERA5 f c and the new MISR-f c dataset, we processed the ERA5 f c during the MISR overpass hour (10 to 11 AM local time) for a nineteen-year (2000-2018) period.We then constructed monthly climatology over the study region (30S-30N and 40E-120E) for the two datasets and found (figure 1) the ERA5-f c values to be within ±0.05 of MISR-f c .The spatial pattern of the difference between the climatological f c of the two datasets (see figure S1 in the Supplementary Information, SI) indicates a close match over a large part of the study region.The overestimation of ERA5 data over the Equator (figure S1), could result from MISR's inability to detect thin cirrus clouds dominating in this region (Dey et al 2015).In a study based on CALIPSO and CloudSat data, Sassen et al (2009) identified the highest frequency of thin cirrus clouds over equatorial landmasses.

Figure 1 .
Figure 1.Box plots of differences in MISR and ERA5 total f c for each month over the Indian Ocean (40°E-120°E and 30°N-30°S) from 2000-2017.

Figure 2 .
Figure 2. Seasonal climatology of ERA5 f c (unitless) for low-level (left column), mid-level (middle column), and high-level (right column) clouds for the period 1979-2018 in DJF, MAM, JJA, and SON, respectively (from top to bottom).

Figure 3 .
Figure 3. (a) Time series of de-seasoned and normalised f c (vertical axis), (b) quantile statistics of changes in seasonal f c over the Indian Ocean in the last four decades using the monthly ERA5 reanalysis dataset (1979-2018).The changes in the cloud fraction are calculated by multiplying the trend with the number of time steps.

Figure 4 .
Figure 4. Vertical profiles of seasonal trends (in year −1 ) of f c over the Indian Ocean using ERA5 reanalysis dataset (1979-2018).Statistically significant (at 95% CI) trends are marked by dots.
Danso et al 2019, Lei et al 2020, Wright et al 2020, Yao et al 2020, Wu et al 2023), the ERA5 f c was found best among the other modern reanalysis products like ECMWF's Interim Reanalysis (ERA-Interim) and Modern-Era Retrospective Analysis for Research and Applications, Version 2 (MERRA-2) and National Centre for Environmental Prediction (NCEP).It was found temporally consistent with Moderate Resolution Imaging Spectroradiometer (MODIS) data.The diurnal cycle of LCF from ERA5 also matches with the Extended Edited Synoptic Cloud Reports Archive (EECRA) data (Dommo et al 2022).
).The type of f c in ERA5 is calculated based on pressure levels.Low clouds are those found below 80% of surface pressure (1000 hPa), mid-level clouds are between 45% and 80% of surface pressure, and high clouds are above 45% of surface pressure.Assumptions are made about the degree of overlap/randomness between clouds in different model levels.More details about the ERA5 cloud data can be found in Binder et al (2020).According to earlier studies (