Observation of ENSO linked changes in the tropical Atlantic cloud vertical distribution using 14 years of MODIS observations

vertical distribution using 14 years of MODIS observations Nils Madenach1, Cintia Carbajal Henken1, René Preusker1, Odran Sourdeval2, and Jürgen Fischer1 1Institute for Space Sciences, Freie Universität Berlin, Carl-Heinrich-Becker-Weg 6–10, 12165 Berlin, Germany 2Laboratoire d’Optique Atmosphérique, Université de Lille, Villeneuve d’Ascq, France Correspondence: N. Madenach (nils.madenach@wew.fu-berlin.de)

For this work the focus lies on the Tropical Atlantic Ocean (TAO). The tropical Atlantic ocean-atmosphere system is largely influenced by the Hadley-Walker circulation. Large deep convective systems with elevated Cirrus clouds at the convective outflow predominate at the Intertropical convergence zone and shallow cumulus clouds at the trade wind regions. The interannual variability of climate variables at the tropical Atlantic is coupled with the El Niño Southern Oscillation (ENSO). An overview 5 of ENSO linked influences on Earth's climate system can be found in e.g. (Timmermann et al., 2018). In order to better understand questions as e.g. how clouds, circulation and climate interact (Bony et al., 2015), this region is of major interest of recent research efforts. At the Barbados Cloud Observatory long-term ground based measurements are done (e.g., Stevens et al., 2016). In December 2013 and August 2016 the Narval campaigns within the High Definition Clouds and Precipiation for advancing Climate Prediction (HD(CP) 2 ) project (Klepp et al., 2014) produced a large amount of airborne based measure-10 ments for this region. Furthermore, the study area is one simulation domain of the cloud resolving ICON-LEM (ICOsahedral Nonhydrostatic-Large Eddy Simulation) model developed within the HD(CP) 2 project (Dipankar et al., 2015;Heinze et al., 2017).
In this work 14 years of the MODerate-resolution Imaging Spectroradiometer (MODIS) level 2 cloud product (Platnick et al., 2015) were analyzed using Multiple Linear Regression Model (MLRM) techniques in order to get insights of the long-term 15 variability and possible trends within the CTH and cloud vertical distributions due to a warming climate. The passive imager MODIS aboard the polar orbiting NASA satellites Terra and Aqua provide the possibility to obtain observational information about the cloud vertical distribution on high temporal (on climate scales) and spatial resolution. MODIS aboard Aqua delivers reliably high resolution data of cloud properties since 2002 (Platnick et al., 2015). Every point on Earth is seen every one to two days. For the evaluation of the MODIS observations, the DARDAR (CloudSat RADAR and CALIPSO LIDAR) data 20 set which relies on information of active instruments were used (Delanoë and Hogan, 2010). Further more we analyzed the level 3 monthly mean Sea Surface Temperature (SST) and Total Column of Water Vapor (TCWV) from Advanced Microwave Scanning Radiometer -Earth Observing System (AMSR-E) aboard Aqua (Wentz, 2004). For the interpretation of the results we examined Era-Interim reanalysis data of the vertical velocity ω at 500 hPa (Dee et al., 2011).
In Sec. 2 we will give an overview of the data used for the analysis. Section 3 introduces the methodology used. In Sec. 4 the 25 results are presented and discussed, and in Sec. 5 the work is summarized and some conclusions are drawn.

Data
In this section an overview of the used data and its processing is given. The study area is defined by a latitude-longitude box of 30 • S-30 • N and 70 • W-20 • E and is hereinafter referred as tropical Atlantic. The study period ranges from September 2002 to September 2016.
Both satellites are in a polar sun-synchronous orbit at an altitude of 705 km. In this work solely data from MODIS aboard Aqua (Platnick et al., 2015), which is part of the Afternoon Train (A-train) constellation, is considered. With a swath of 5 2330 km (cross track) by 10 km (along track at nadir) and a scan rate of 20.3 rpm (cross track) global coverage is acquired almost daily. The level 2 data of collection 6 (Platnick et al., 2015) provides cloud optical and microphysical property data at a spatial resolution of 1 km and cloud top property data at 5 km as well as at 1 km resolution (Menzel and Strabala, 2015;Platnick et al., 2017). For the analysis only data from overpasses during day time (ascending node) were used. The vertically resolved cloud fractions were calculated using the 1 km cloud mask and the International Satellite Cloud Climatology Project (ISCCP) 10 cloud classification scheme explained in Sec. 3.1.
In order to get daily composites from the MODIS overpasses the segments were regridded on a regular 0.1 • × 0.1 • grid. For every grid cell the mean, minimum, maximum and the variance of the containing pixels were computed. Due to the proximity to the equator of the analyzed region for the majority of the grid cells, the pixels arise from a single overpass. To account for possible miss-classification of the MODIS cloud mask (clear sky conservative) due to e.g. cloud edges, broken clouds, smoke, 15 dust or sun-glint, pixels that were poor retrieval candidates were excluded using the clear sky restoral flag (Hubanks, 2015).
After computing the daily composites for every day of the investigated period, grid cell based monthly means were computed and afterwards regridded to a 0.2 • × 0.2 • grid. Based on the monthly means, climatologies were calculated by averaging every month of the year over the 14 years of MODIS data. Monthly anomalies were produced by subtracting the climatology from every single monthly mean. to the much lower sampling of the active lidar and radar instruments the monthly means were computed on a 2 • × 2 • grid.

30
Furthermore, data from the AMSR-E instrument aboard Aqua was used to acquire information about the SST and TCWV.
AMSR-E measures the brightness temperature at 6 different wavelengths in the microwave range between between 0.34 and 4.35 cm. For every wavelength, the horizontal and vertical polarized radiation is measured leading to 12 channels in total. The spatial resolution depends on the the channel and varies from 5.4 km to 56 km . The radiometer has a viewing swath width of 1445 km and an incidence angle of 55 • . For the analysis the SST and the TCWV from version 2 of the monthly Level-3 product (AE_MoOcn) with a spatial resolution of 0.25 • × 0.25 • were used (Wentz, 2004). The AMSR-E data is available from June 8th 2002 through October 4th 2011.

ERA-Interim
To get insights about the predominant large scale dynamics, the vertical velocity ω at 500 hPa was acquired from the Era-Interim reanalysis (Dee et al., 2011). In order to roughly match the MODIS overflight time, the daily Era-Interim analyses data at 12:00 was used. From the daily data (0.2 • × 0.2 • ), monthly means were computed the same way as explained in Sec. 2.1.

10
Following, the ISCCP cloud classification scheme is introduced. Furthermore, the MLRM used for the time-series analysis is described.

Assessment of the cloud vertical distribution
The ISCCP cloud classification scheme is widely used to simply distinguish between different cloud types using remote sensed information. The ISCCP is a project, started 1982 as part of the World Climate Research Programme (WCRP), with the goal 15 to collect and analyze satellite radiance measurements to infer the global distribution of clouds, the cloud properties, and the cloud diurnal, seasonal, and interannual variations. As Fig. 1 illustrates the clouds are distinguished by its Cloud Top Pressure (CTP) in hPa and its optical thickness. There are three categories for the vertical distribution namely high (CTP < 440 hPa ), middle (≤ 440 CTP <680 hPa ) and low (CTP ≥ 680 hPa ). For the MODIS data the cloud height category flag from the MODIS 5 km quality assurance (Hubanks, 2015), that is based on the explained ISCCP thresholds was used to discriminate 20 between the height categories.

Multiple linear regression
To obtain information of linear changes within the data-sets a simple MLRM was developed. For the model development, the time series of the analyzed parameters (y) were assumed to be composed of a constant µ, a seasonal S, a trend τ and a noise η component. These assumptions lead to the following model equation.
Where the seasonal component was specified by the sum of sine and cosine with a period of 12 months.
Due to bi-annual variations in the time series the first harmonic (i=2) of sine and cosine was added to the model.
In a general form Eq. 1 can be written as: Where x T t represents a vector including all explanatory parameters required to describe the parameter y at time step t and b a vector containing all regression coefficients. In Eq. 4 all possible T linear combinations of the n parameters are summarized in 5 matrix notation.
Here y stands for the T ×1 vector with the measured parameter for all T month, X for the T × n model matrix including all predictor variables for the estimation for all T months and b for the n×1 vector containing all the regression coefficients. The variations that are not explained by the model are stored in the T ×1 noise vector η. 10 To find the vector, which best fits to the data (denoted asb), the ordinary least square method is used. This method tries to predict the expected parameterŷ by minimizing the sum of squares of the distances between measured and estimated values.
The remaining noise between prediction and measurement is termed residual: It is important to remember that y is the vector with the measured values andŷ is the vector containing the values estimated by the model. Consequently, the following expression has to be minimized, in order to minimize the sum of squared residuals: With and some simplifications Eq. 6 results in: To find the values of the vectorb, including all regression coefficients, which minimizes the sum of squared residuals, the derivation of M with respect tob must be zero: After transposing Eq. 8, the following linear equation system in matrix notation must be solved in order to obtainb: 10 Multiplying both sides with ([X] T [X]) −1 leads to the final equation to be solved: To get an idea of the the linear correlation between two variables x and y the Pearson Correlation Coefficient (PCC) was used.
It is defined as the ratio of the co-variance of x and y and the product of the two standard deviations s x and s y : The PCC is bounded by -1 and +1. For PCC = -1 there is a perfect, negative linear association between x and y and for PCC = +1 there is a perfect positive linear association between the two variables. For PCC = 0 there is no linear correlation, but there might be non linear correlations between x and y.
A more detailed explanation of the used MLRM can be found in e.g. Wilks (2011).

Monthly means and climatologies
In Fig. 2

10
To get insights of the interannual variability and the temporal evolution of the variables, the monthly tropical Atlantic three month running mean anomalies (see Sec. 2) were analyzed. As illustrated in Fig. 4   to the tropical Atlantic ocean-atmosphere system is well known even though the exact mechanism of teleconnection is still not fully understood. Positive ENSO events were found to be associated with higher SSTs in the tropical Atlantic (Klein et al., 1999), a weakening of the Walker circulation and the Atlantic Hadley circulation (Klein et al., 1999;Wang, 2004, e.g.) and a stronger vertical wind shear over the tropical Atlantic (Zhu et al., 2014). The lower panel of Fig. 4 displays the three month running mean of the HCF anomaly and additionally the Nino3.4-Index (obtained from the NOAA climate prediction center). 10 The interannual variability of the mean HCF at the tropical Atlantic and the Nino3.4-Index are negatively linked (PCC=-0.53). Where negative ENSO events are associated with more high clouds and vice versa. As shown in (Marchand, 2013) a strong negative correlation between ENSO and high cloud amount is also present in MODIS (and Multi-angle Imaging

Multiple linear regression analysis
In order to get insights of possible linear changes in the investigated variables associated with a warming ocean and changes in atmospheric dynamics due to climate change, a multiple linear regression analysis, explained in more detail in Sec. 3.2, was  In addition, the Nino3.4-Index is displayed in red. 10 The CTH in Fig. 5 (a) and the HCF in Fig. 5    and RADAR instruments was analyzed. Due to the low sampling, monthly means were computed and compared to MODIS with a 2 • × 2 • resolution. As Fig. 6 illustrates the seasonal and interannual variability of the HCF, as well as its decrease in P2, is despite of the low sampling visible in the DARDAR data as well. As already seen in the climatologies (Fig. 3) DARDAR has a positive offset.
To get spatial information of the anomalies (trends) the MLRM was applied to every 0.2 • × 0.2 • grid-box. Fig. 7

Link between TAO and ENSO
During a positive ENSO phase the anomalous high SST in the Pacific induces an average warming of the tropical Atlantic troposphere, e.g. trough Kelvin waves causing an increase in the meridional tropospheric temperature gradient (e.g., Horel and Wallace, 1981;Yulaeva and Wallace, 1994;Chiang and Sobel, 2002;Zhu et al., 2014). This leads to an increase of the vertical 15 wind shear (Aiyyer and Thorncroft, 2006;Shaman et al., 2009;Zhu et al., 2014, e.g.) and tends to increase the static stability over the tropical Atlantic (e.g., Tang, 2004;Larson et al., 2012). Moreover, a weakening of the Atlantic Hadley and Walker circulation and an east-ward shift of the latter (Klein et al., 1999;Wang, 2006) has been associated with higher Pacific SSTs.
Furthermore, a reduction in equatorial Atlantic rainfall (Saravanan and Chang, 2000;Chiang and Sobel, 2002), a lagged SST increase caused by a warmer troposphere and reduced latent and sensible heat losses due to weaker trade winds and a reduction 20 in cloudiness (Curtis and Hastenrath, 1995;Enfield and Mayer, 1997;Klein et al., 1999;Saravanan and Chang, 2000;Huang, 11 Atmos. Chem. Phys. Discuss., https://doi.org/10.5194/acp-2018-1136 Manuscript under review for journal Atmos. Chem. Phys.  2002) were found during El Niño. The reduction in cloudiness leads to changes in radiation fluxes due to higher absorption of solar radiation by the ocean, which in turn increases the mean tropical Atlantic SST (Curtis and Hastenrath, 1995;Klein et al., 1999) strengthening the positive SST feedback between the Pacific and the Atlantic. In general the described effects are observed to be inverted during La Niña.
As shown in Sec. 4.2 & Sec. 4.3 a link between ENSO and TAO cloudiness, primarily in high cloud amount, was observed in 5 our analysis as well. In addition, the analysis of AMSR-E retrieved SST and TCWV data show a consistent (Klein et al., 1999, e.g.) three months delayed TAO SST and TCWV increase after an El Niño event (Fig. 8). Further more, a significant positive trend of 0.3 • C dec −1 was found for the regional mean SST.
To investigate whether the observed anomalies in HCF are linked to ENSO induced changes in the large scale circulation over the TAO, the large scale vertical velocity ω at 500 hPa acquired from ERA-Interim data was analyzed. The results are illustrated 10 in Fig. 5 (c) for the TAO mean and in Fig. 7 (e) for the grid-box based analysis. ω is negatively defined so that negative velocities imply upward movements and vice versa. For the TAO mean a decrease in P1 and an increase in P2 was found. The anomalies are mostly due to strengthening/weakening in upward movement rather than weakening/strengthening in downward movement. This is coherent with the results of the grid-box based analysis displayed in Fig. 7 (e) where in both phases the strongest and most of the significant anomalies are found at the ITCZ region where upward movement predominates.

15
The comparison of the patterns of ∆HCF (Fig. 7 (d)) and ∆ω 500 (Fig. 7 (e)) reveal similar patterns for both phases. The increase/decrease in the high cloud amount in P1/P2 seems to be associated with an increase/decrease in large scale upward motion in equatorial Atlantic ocean. A stratification of the HCF into different ω 500 -bins as listed in Table 1 confirms that the decrease of the high cloud amount in P2 occurs in regions with strong upward movements, Fig. 9 (d) and Fig. 9 (e). This supports the idea that the observed anomalies, particularly the decrease of CTH/HCF in P2, are caused by an ENSO induced

Summary and Conclusion
In this study the MODIS (aboard Aqua) cloud products were used to analyze the interannual variability of the cloud vertical distribution at the TAO for a period of 14 years (2002 to 2016). From the level 2 data, daily composites on a 0.1 • × 0.1 • regular grid were generated. For the analysis 0.2 • × 0.2 • monthly means, calculated on the basis of the daily composites, were used.
The data well represented large circulation patterns as the band with high cloud tops and cloud amount of the ITCZ and its 5 seasonal shift. Also the trade wind inversion regions with its characteristic low cloud tops and the broad stratocumulus region at the west coast of Namibia were represented in the data.
The analysis of the time-series revealed strong interannual variability of the vertical cloud distribution, which was found to be mainly caused by changes in the large scale circulation due to ENSO associated teleconnective causes. The changes might be associated with the east-ward shift and the weakening of the Walker circulation and a weakening of the Hadley cell during 10 El Niño conditions. The largest ENSO linked anomalies were found for the HCF, which also drives the observed anomalies in the mean CTH. With an El Niño event a decrease in HCF mainly at the equatorial ITCZ influenced region was observed and was linked to a decreasing large scale vertical upward movement by including Era-Interim reanalysis data. The HCF and LCF showed opposite behaviour, which might mask much of a possible signal if using solely TCF. This supports the efforts to consider the vertically resolved cloud fraction rather than the cloud fraction as a whole. The decrease in cloudiness is consistent 15 with findings from e.g. Klein et al. (1999).
Furthermore, a three month lagged increase in SST and humidity during El Niño was found examining data from the AMSR-E, which is consistent with literature (e.g., Curtis and Hastenrath, 1995;Klein et al., 1999). Besides other factors as alleviated surface wind speed and a warmer troposphere the change in cloud cover is associated with positive SST anomalies e.g. Klein et al. (1999). Subsequently, a sensitivity study of the CRF associated with ENSO linked changes could be done to quantify the  influence of radiation changes.
As climate projection show similar changes in the tropical Atlantic circulation in a warming climate (e.g., Vecchi and Soden,