The AIRS instrument (Chahine et al., 2006) provides very high-spectral-resolution measurements of Earth-emitted radiation in 2378 spectral bands in the thermal infrared (3.74–15.40 µm). The spatial resolution of these measurements varies from 13.5 km × 13.5 km at nadir to 41 km × 21 km at the scan extremes. The polar-orbiting Aqua satellite provides observations at 01:30 and 13:30 LT (local Equator crossing time). Nine AIRS measurements (3 × 3) correspond to one footprint of the Advanced Microwave Sounder Unit (AMSU), grouped as a “golf ball”.

The CIRS cloud retrieval uses measured radiances along the wing of the 15 µm CO2 absorption band. We have chosen AIRS channels closely corresponding to the five channels used in the TIROS-N Operational Vertical Sounder (TOVS) Path-B cloud retrieval, at wavelengths of 14.19, 14.00, 13.93, 13.28 and 10.90 µm, and three additional channels at 14.30, 14.09 and 13.24 µm (with peaks in the weighting function at 285, 415, 565, 755 hPa and surface as well as at 235, 375 and 855 hPa, respectively). The multi-spectral cloud detection, based on the spectral coherence of retrieved cloud emissivities, decides whether the AIRS footprint is cloudy (Sect. 2.5.3). For the latter, radiances in the atmospheric window between 9 and 12 µm are used, at six wavelengths of 11.85, 10.90, 10.69, 10.40, 10.16 and 9.12 µm.

Ancillary data necessary for the cloud retrieval, which include atmospheric temperature and water vapour profiles as well as surface skin temperature, are provided by the NASA Science Team L2 standard products (version 6 (V6); AIRS Science Team/Texeira, 2013). They were retrieved from cloud-cleared AIRS radiances within each AMSU footprint. The methodology remains essentially the same as described in Susskind et al. (2003). Compared to version 5 (V5), the most significant changes are as follows: (i) V6 uses an IR–microwave neural network solution (Blackwell et al., 2014) as a first guess for the retrieval of atmospheric temperature and water vapour profiles as well as for surface skin temperature, instead of the previously used regression approach (Susskind et al., 2014). This leads to physical solutions for many more cases than in V5. (ii) The retrieval of surface skin temperature only uses shortwave IR window channels (Susskind et al., 2014). These modifications resulted in significant improvement of accurate temperature profiles and surface skin temperatures under partially cloudy conditions (Van T. Dang et al., 2012): Compared to V5, the surface skin temperature is larger over land in the afternoon (especially over desert) and over maritime stratocumulus regions.

In addition, we use the microwave identification of snow- or ice-covered surfaces, also provided by the NASA L2 data.

Since the retrieved cloud pressure should be within the troposphere/lower stratosphere, we have determined the tropopause pressure from the atmospheric profiles, using the concept described in Reichler et al. (2003) and in Feofilov and Stubenrauch (2017). The CIRS cloud retrieval allows cloud levels up to 30 hPa above the tropopause.

IASI, developed by CNES in collaboration with EUMETSAT, is a Fourier transform spectrometer based on a Michelson interferometer, which covers the IR spectral domain from 3.62 to 15.5 µm. As a cross-track scanner, the swath corresponds to 30 ground fields per scan, and each of these measures a 2 × 2 array of footprints. The latter have a 12 km diameter at nadir. IASI raw measurements are interferograms that are processed to radiometrically calibrated spectra on board the satellite. Two instruments were launched so far on board the European platforms Metop-A and Metop-B, with measurements in October 2006 and September 2012, respectively, at 09:30 and 21:30 LT (Metop-A) and 10:30 and 22:30 LT (Metop-B). IASI has been providing water vapour and temperature sounding profiles for operational meteorology (accuracy requirements of, respectively, 1 K and 10 % in the troposphere) as well as trace gas concentrations and surface and atmospheric properties, including those of aerosols and clouds (Hilton et al., 2012). For the cloud retrieval, we use radiances at the wavelengths 14.30, 14.20, 14.06, 14.00, 13.93, 13.40, 13.24 and 10.90 µm and for the multi-spectral cloud detection the radiances at 11.85, 10.90, 10.70, 10.41, 10.16 and 9.13 µm.

At the time we started incorporating IASI data to the CIRS cloud retrieval, two data sets of IASI-retrieved atmospheric profiles and surface temperature were available: one provided by EUMETSAT (version 5) and one by NOAA. EUMETSAT L2 temperature and water vapour version 5 products were only available for clear and partly cloudy scenes, leaving atmospheric and surface retrievals in only 9 % of all cases. Therefore we first used IASI L2 ancillary data provided by NOAA. The comparison with collocated temperature profiles of the Analyzed RadioSoundings Archive (ARSA, available at the French data centre AERIS) has shown that, while AIRS-NASA and ERA-Interim (Sect. 2.3) temperature profiles do agree in general with the ARSA profiles within 1 K, differences between IASI-NOAA and ARSA profiles were often larger than 1 K in the lower troposphere (not shown). In addition, a study of the influence of the different ancillary data on the CIRS cloud amount (CA) has demonstrated that the amount of low-level clouds over ocean was underestimated when using those deduced from IASI-NOAA (Feofilov et al., 2015a). This might be explained by an underestimation of the sea surface temperature (SST) linked to cloud contamination. From this we concluded that the AIRS–IASI synergy to explore cloud diurnal variability in a coherent way needs ancillary data from similar retrievals or from the same source. Thus we also implemented ancillary data from the European Centre for Medium-Range Weather Forecasts (ECMWF) meteorological reanalyses into the CIRS cloud retrieval.

ECMWF provides the meteorological reanalyses ERA-Interim, covering the period from 1989 onwards. Dee et al. (2011) give a detailed description of the model approach and the assimilation of data. The data assimilation scheme is sequential: at each time step, it assimilates available observations to constrain the model, which then provides a short-range forecast for the next assimilation time step. Gridded data products (at a spatial resolution of 0.75∘ latitude × 0.75∘ longitude) include 6-hourly surface temperature, atmospheric temperature and water vapour profiles, as well as dynamical parameters such as horizontal and vertical large-scale winds. These data are given at universal time of 00:00, 06:00, 12:00 and 18:00. To match these data with the AIRS and IASI observations, we interpolate them to the corresponding local time, using a cubic spline function, as in Aires et al. (2004).

All satellites of the A-Train follow each other within a few minutes. We use the same collocation procedure as in Feofilov et al. (2015b): first, each AIRS footprint is collocated with NASA CALIPSO L2 cloud data averaged over 5 km (version 3; Winker et al., 2009) in such a way that for each AIRS golf ball, three CALIPSO samples are matched to the centres of three AIRS footprints. These data are then collocated with the NASA L2 CloudSat-lidar geometrical profiling (GEOPROF) data (version R04; Mace and Zhang, 2014). Each of these AIRS footprints thus includes cloud top and cloud base for each of the cloud layers, detected by lidar or radar, at the spatial resolution of the radar footprints (1.4 km × 2.3 km) from the GEOPROF data. Cloud optical depth (COD), cloud top, ztop and apparent cloud base (corresponding to the real cloud base or to the height at which the cloud reaches opacity), zappbase, are given at the spatial resolution of the CALIPSO cloud data (5 km × 0.09 km). A cloud feature flag indicates whether the cloud is opaque. The CALIPSO cloud data also indicate at which horizontal averaging along the track the cloud was detected (1, 5 or 20 km), which is a measure of the COD. As in Stubenrauch et al. (2010), for a direct comparison with AIRS cloud data, we use clouds detected at horizontal averaging over 5 km or less. This corresponds to clouds with visible COD larger than about 0.05 to 0.1 (Winker et al., 2008).

The scene type of an AIRS footprint is estimated as cloudy when the CALIPSO sample as well as the GEOPROF sample include at least one cloud layer. Clear sky is defined by cloud-free CALIPSO and GEOPROF samples within the AIRS footprint.

For the evaluation of cloud height, we identify the GEOPROF cloud layer which is closest to zcld from AIRS and estimate the height at which the cloud reaches a COD of 0.5, zCOD0.5, from CALIPSO. zCOD0.5 is required to be located within the corresponding GEOPROF cloud layer. zCOD0.5 is deduced from the CALIPSO L2 COD, assuming a constant increase of COD from cloud top towards cloud base, except for high-level clouds, for which the shape of the ice water content profile as a function of cloud emissivity is taken into account (Feofilov et al., 2015b). As the COD of CALIPSO might be slightly underestimated (Lamquin et al., 2008), especially for larger COD, we reduce the ratio 0.5 / COD to 0.4 / COD, used in the estimation of zCOD0.5.

The cloud property retrieval is based on a weighted χ2 method using channels along the wing of the 15 µm CO2 absorption band (Stubenrauch et al., 1999). Cloud pressure and effective emissivity are determined by minimizing χ2(pk), computed at different atmospheric pressure levels by summation over N wavelengths λi: χ2pk=∑i=1NIcldpk,λi-Iclrλi⋅εcldpk-Imλi-Iclrλi2∗W2pk,λi. Im corresponds to the measured radiance. Iclr is the simulated radiance the IR sounder would measure in the case of clear sky, and Icld(pk) is the radiance emitted by a homogeneous opaque single cloud layer at pressure level pk. Icld is calculated for 42 pk levels (from 984 to 86 hPa) for the viewing zenith angle of the observation. A sensitivity study has shown that five (for HIRS) to eight channels (AIRS and IASI) are sufficient, as doubling the number of channels in the retrieval did not change the results.

By introducing empirical weights W(pk,λi), the method takes into account (i) the vertical contribution of the different channels, (ii) the growing uncertainty in the computation of εcld with increasing pk and (iii) uncertainties in atmospheric profiles. These weights are determined for each of five typical air mass classes (tropical, midlatitude summer and winter, polar summer and winter) as in Stubenrauch et al. (1999) and in Feofilov and Stubenrauch (2017), using the spread of clear-sky radiances within these air mass classes. The clear-sky radiances have been simulated for each of the atmospheric profiles of these five air mass classes, using the 4A radiative transfer model (Scott and Chédin, 1981), and stored in the Thermodynamic Initial Guess Retrieval (TIGR) database (Chédin et al., 1985, 2003; Chevallier et al., 1998). Minimizing χ2 in Eq. (1) is equivalent to dχ2/dεcld=0, from which one can extract εcld as εcldpk=∑i=1NImλi-Iclrλi⋅Icldpk,λi-Iclrλi⋅W2pk,λi∑i=1NIcldpk,λi-Iclrλi2⋅W2pk,λi. In general, the χ2(p) profiles have a more pronounced minimum for high-level clouds than for low-level clouds. We stress here that for the identification of low-level clouds it is important to allow values larger than 1 for εcld, because at larger pressure Iclr and Icld become very similar and their uncertainties may lead to values larger than 1 (Stubenrauch et al., 1999). Thus only pressure levels leading to εcld >  1.5 are excluded from the solution. Typical pcld uncertainties have been estimated from a statistical analysis of the χ2(p) profiles: they range from 30 hPa for high-level clouds to 120 hPa for low-level clouds, corresponding to about 1.2 km in altitude, zcld.

In the case of atmospheric temperature inversions in the lower troposphere, the cloud height is moved to the inversion level, zinv, defined as the highest level with T(zinv) > Tsurf. To detect these cases, the inversion strength, defined by T(zinv)-Tsurf, has to be larger than 2 K. Depending on the ancillary data, these cases occur in about 7 to 15 % of all cloudy cases. εcld as defined in Eq. (2) does not have a physical meaning in the case of an inversion, since Icld(pcld) will be greater than Iclr. Therefore, we scale εcld and the spectral emissivities in accordance with the ratio pinv / pcld.

Cloud temperature, Tcld, is determined from pcld, using the ancillary temperature profile similar to the observed situation (see Sect. 2.5.1). Cloud types are distinguished according to pcld and εcld. High-level clouds are defined by pcld < 440 hPa, mid-level clouds by 440 hPa < pcld < 680 hPa and low-level clouds by pcld > 680 hPa. High-level clouds may be further distinguished into opaque (εcld >  0.95), cirrus (0.95 > εcld >  0.50) and thin cirrus (εcld < 0.50). pcld is transformed to cloud altitude, zcld, using a standard hydrostatic conversion.

For the computation of Iclr and Icld in Eq. (1), we need (i) surface type (ocean, land, ice/snow), surface temperature and spectral emissivities, (ii) atmospheric temperature and water vapour profiles as well as spectral transmissivity profiles for the atmospheric situation of the measurements. The latter have been calculated using the 4A radiative transfer model, separately for each satellite viewing zenith angle (up to 50∘) and for about 2300 representative clear-sky atmospheric temperature and humidity profiles of the TIGR database.

In the cloud retrieval, the TIGR database is searched for the atmospheric profile corresponding best to the observational conditions by applying a proximity recognition which compares the atmospheric temperature and water vapour profiles from the ancillary data with those from TIGR as in Stubenrauch et al. (2008). The preparation and evaluation of these ancillary data is presented in Sect. 2.5.1.

The hit rates (fraction of agreeing cloudy and clear cases) between the AIRS-CIRS cloud detection and the lidar–radar cloud detection (Sect. 2.4) are 85 % (84 %) over ocean, 82 % (79 %) over land and 70 % (73 %) over ice/snow. Values in parentheses correspond to ERA-Interim ancillary data. Table 1 presents separate comparisons for the three latitude bands. In general, the hit rates are quite high, considering that CALIPSO and GEOPROF data only sample a small area of the AIRS footprints. They are slightly higher over ocean than over land. Compared to the AIRS-LMD cloud retrieval presented in Stubenrauch et al. (2010), the agreement with CALIPSO–CloudSat has improved both over ocean and land but slightly decreased over sea ice. The latter can be explained by applying only one test over all surface types. In the earlier version we used an additional brightness temperature difference test related to temperature inversions. A detailed analysis (not shown) indicated that it also introduced noise.

To further illustrate CA uncertainties linked to ancillary data, we investigate, in Fig. 2, geographical maps of differences in CA and Tsurf, using ancillary data from AIRS-NASA and from ERA-Interim. With AIRS-NASA ancillary data, CA over land is often smaller during night and larger in the afternoon, with Tsurf also smaller during night and larger in the afternoon over large parts of the continents. Considering the Tsurf comparison with ARSA (Sect. 2.5), this means that over land CA is slightly underestimated during night with AIRS-NASA ancillary data, while slightly underestimated in the afternoon with ERA-Interim ancillary data. Patterns of differences in atmospheric water vapour are less reflected in those of CA (not shown), but slightly more atmospheric water vapour in the ancillary data (as in the tropics for AIRS-NASA compared to ARSA and ERA-Interim) might lead to a slight underestimation of CA.

Figure 3 presents normalized distributions of the difference between zCOD0.5 from CALIPSO (Sect. 2.4) and zcld, from AIRS for the three latitude bands. We compare results for pcld < 440 hPa and pcld ≥ 440 hPa separately for AIRS-NASA and ERA-Interim ancillary data. In general, all distributions peak around 0 km and are slightly narrower for lower-level clouds than for high-level clouds. Results are similar for both ancillary data, with a slight cloud height overestimation of lower-level clouds over tropical ocean for ERA-Interim (not shown) and a height overestimation of some clouds over polar ocean for AIRS-NASA ancillary data (not shown). The latter can be explained by the fact that in some of these regions Tsurfand atmospheric profiles of good quality are only available 10 % of the time. When comparing distributions of ztop-zcld, the peaks for lower clouds are still around 0 km, whereas for high-level clouds zcld lies on average 1.5 km below the cloud top (not shown), very similar to results in Stubenrauch et al. (2010). This means that Tcld is about 10 K warmer than the cloud top (Fig. S2 of the Supplement). The broader distributions for high-level clouds compared to low-level clouds may be explained by the fact that high-level clouds often have diffuse cloud tops (e. g. Liao et al., 1995), especially in the tropics (ztop-zcld is slightly larger for the same εcld, as shown in Fig. 5). To summarize, zcld can be approximated by (i) the height of maximum lidar backscatter (Stubenrauch et al., 2010), (ii) zCOD0.5 (Fig. 3) or (iii) the mean layer height (for optically thin clouds) or the mean between cloud top and the height at which the cloud reaches opacity), as shown in Fig. S2 in the Supplement (considering mid-pcld) .

For a more detailed investigation of the different height approximations, Fig. 4 compares median values of zcld-zCOD0.5, ztop-zcld and (ztop-zcld)/(ztop-zappbase) as functions of εcld for high-level clouds. For this analysis we have selected cases for which zcld lies between top and base of the closest GEOPROF cloud layer. This leaves about 82 %, 73 % and 57 % of the statistics in tropics, midlatitudes and polar regions, respectively. zcld varies from 1 km above for εcld = 0.1 to 1 km below zCOD0.5 for εcld = 1, assuming that zCOD0.5 is accurately estimated for all εcld (Sect. 2.4). In that case, zcld of thin cirrus should be approximated by a height with COD < 0.5 and zcld of opaque high clouds by a height with COD > 0.5. In contrast, zcld lies about 1 km to 2 km below ztop, the difference to cloud top increasing with εcld (except for εcld close to 1). Since ztop-zappbase also increases with εcld (not shown), (ztop-zcld)/(ztop-zappbase) does not depend on εcld and is about 0.5. We deduce that it probably needs less vertical extent for opaque clouds than for semi-transparent cirrus to reach a COD of 0.5, while the χ2 method determines a height within the cloud, which corresponds well to the mean between cloud top and base or the height at which the cloud reaches opacity, independent of εcld. This is important to take into account for the determination of radiative fluxes and heating rates of UT clouds, when using CIRS cloud heights. We want to stress that (ztop-zcld)/(ztop-zappbase) is about 0.5 (0.4 to 0.6) for low-level clouds as well, while zcld lies only about 0.1 to 0.4 km below zCOD0.5 and about 0.5 km below ztop (Fig. S3 of the Supplement).

Finally, Fig. 5 presents normalized frequency distributions of zcld, using both sets of ancillary data, and zCOD0.5, whenever clouds are detected (excluding subvisible cirrus; see Sect. 2.4). The CALIPSO zCOD0.5 distributions have a slightly larger part of high-level clouds, especially in the tropics, and the AIRS zcld distributions show a slightly larger part of low-level clouds over land. The latter disappear if one considers only cases with all three CALIPSO samples cloudy within an AIRS golf ball. Thus these low-level clouds are part of partly cloudy fields for which it is difficult to compare results from samples of very different spatial resolution. The distributions compare better when only mostly covered cloud fields are considered (three CALIPSO samples cloudy within an AIRS golf ball). In the tropics, the peak of the AIRS zcld distributions for high-level clouds is still slightly broader towards lower heights than for CALIPSO (not shown). Additional filtering, excluding multi-layer clouds, ultimately leads to very similar distributions, also presented in Fig. 5. A plausible interpretation is that in cases of multiple cloud layers with the upper cloud layer not fully covering the large AIRS footprint, instrument received radiation is mixed from different cloud layers, and thus zcld is slightly lower than the one of the uppermost cloud layer. The distributions in the midlatitudes still peak at slightly lower heights, because high-level clouds in these latitudes are on average optically thicker (storm tracks) than in the tropics. In these cases zcld lies below zCOD0.5, as we have seen in Fig. 4. The choice of ancillary data influences only mildly the zcld distributions, with a slightly larger contribution of low-level clouds over land for ERA-Interim. This difference disappears if we consider only mostly covered cloud fields, as the contribution of low-level clouds strongly decreases over land. Over ocean, the effect is much smaller. This indicates that low-level clouds over ocean appear more often as stratus decks whereas those over land appear more frequently as cumulus, as expected.

While the NH and the SH reflect the same amount of sunlight within 0.2 Wm-2 (Stephens et al., 2015), there is a small energy imbalance between both hemispheres of our planet, with slightly more energy absorbed by the SH (0.9 Wm-2). This yields more frequent precipitation in the SH and more intense precipitation in the NH (Stephens et al., 2016). The latter might be linked to the characteristics of the ITCZ, a zone of strong convection, which itself produces large cirrus anvils. As the size of these anvils is on average positively related to convective strength (e. g. Protopapadaki et al., 2017), we explore the annual mean and seasonal hemispheric difference of high CA and try to relate it to the characteristics of the ITCZ, such as its peak strength, the latitudinal position of the peak and its width.

The ITCZ characteristics have been determined by fitting a Gaussian around the tropical peak of the latitudinal CAH distributions (Fig. 8), per month and year. This yields the latitude of the peak position, the value of the peak itself and the width of the tropical CAH distribution. From Fig. 11 we deduce that the annual NH–SH difference in CAH is 0.05, with a pronounced seasonal cycle of about 0.3 in amplitude. Results from the three CIRS cloud climatologies (AIRS with two ancillary data sets and IASI), AIRS-LMD, CALIPSO-GOCCP, ISCCP and MODIS-CE are similar. This seasonal cycle is well related to the one of the ITCZ peak latitude, which moves up to 12∘ N in July. It is interesting to note that the width of the ITCZ is smaller in July and August (10.5–12.5∘) than in January (17∘) and the CAH peak is about 10 % larger in August than in January. This might suggest a more intense ITCZ (and hence more intense precipitation) when it is located in the NH than when it is located in the SH.

All data sets agree well on the ITCZ peak latitude. The smaller maximum CAH values of MODIS-CE and ISCCP are due to smaller sensitivity to thin cirrus, and the reduced seasonal cycle of maximum CAH and of ITCZ width for CALIPSO-GOCCP is due to the inclusion of ubiquitous thinner cirrus, leading to less-well-pronounced CAH minima in the subtropics. The CIRS climatologies reveal the seasonal behaviour of the ITCZ characteristics clearly. Figure 11 confirms and extends the interpretation of the results of Stephens et al. (2016) by displaying a relation between the hemispheric difference of CAH and characteristics of the ITCZ, which seems to be more intense when its peak is situated in the NH (smaller width and larger maximum).

Since the observational period of AIRS and IASI is too short to directly study long-term cloud variability related to climate warming, an alternative approach is to analyse cloud variability in response to interannual climate variability. Though interannual global mean surface temperature anomalies might not directly relate to patterns of anthropogenic climate warming, Zhou et al. (2015) have shown that interannual cloud feedback may be used to directly constrain the long-term cloud feedback. Changes in tropical UT clouds lead to variations in atmospheric heating and cooling, which then may influence the large-scale circulation, as has already been shown by Slingo and Slingo (1991).

Since the radiative effects of high opaque clouds and thin cirrus are quite different, we investigate the geographical patterns of UT cloud amount anomalies with respect to tropical and global mean surface temperature anomalies, by separating them into opaque, cirrus and thin cirrus (εcld > 0.95, 0.4–0.95 and < 0.4, corresponding to visible COD > 6, 1–6 and < 1, respectively). By making use of the whole period between 2003 and 2015 (covering 156 months), we estimate a change in UT cloud amount as a function of change in mean surface temperature by a linear regression of their deseasonalized monthly anomalies, at a spatial resolution of 1∘ latitude × 1∘ longitude. Similar techniques were already utilized in other studies related to El Niño–Southern Oscillation (ENSO) and cloud feedback (e.g. Lloyd et al., 2012; Zhou et al., 2013; Liu et al., 2017). Figure 12 presents the change in amount of high opaque cloud (mostly of convective origin), in thick cirrus (often formed from convective outflow as anvils) and in thin cirrus (which might be formed as anvil or via in situ freezing) per kelvin of global surface warming, obtained as the linear slopes of these deseasonalized monthly anomaly relationships. The cloud amounts are from AIRS-CIRS, while the surface temperatures are from the ERA-Interim ancillary data. Results are very similar when using Tsurf anomalies from AIRS-NASA (not shown). Zhou et al. (2013) have shown that ERA-Interim Tsurf anomalies give similar results in their short-term cloud feedback analysis compared to other Tsurf data sets. In our study, we concentrate on the change of UT clouds of different height (pcld < 440 hPa and pcld < 330 hPa), and we compare changes in absolute UT cloud amounts and in UT cloud amounts relative to total cloud amount. The geographical patterns of the relative slope uncertainty are shown in Fig. S5 in the Supplement. In general, large changes in cloud amount per K of warming have smaller uncertainty than small ones, indicating robust patterns.

During this period, global mean Tsurf anomalies and tropical mean Tsurf anomalies are strongly correlated (not shown), and the spatial patterns in Fig. 12 are compatible with ENSO-like patterns. The left panels of Fig. 12 agree quite well with Fig. 8 of Liu et al. (2017), based on MODIS cloud amount and HadCRUT4 Tsurf anomalies, even though our cloud types categories differ slightly. In particular, we have separated thin cirrus. Therefore the analyses suggest that the change patterns address ENSO variability rather than long-term trends. When considering relative cloud type changes (middle panels in Fig. 12), the signals are stronger. An interesting feature appears when considering changes in the relative amounts of higher clouds (pcld < 330 hPa, left panels of Fig. 12): while the high opaque clouds, linked to strong precipitation (Protopapadaki et al., 2017), relative to all clouds, increase in a narrow band in the tropics, there is a large increase in relative thin cirrus amount around these regions; the latter might directly affect the atmospheric circulation through their radiative heating (e.g. Sohn, 1999; Lebsock et al., 2010).

As in Liu et al. (2017), we have also examined linear regression slopes from anomaly averages over the tropics and other latitudinal bands. Although in general the relationships are very noisy, on the interannual scale tropical cirrus amount slightly decreases with warming (-0.76 ± 0.21 % K-1), while thin cirrus amount seems not affected (-0.09 ± 0.20 % K-1), in agreement with Liu et al. (2017). However, when considering changes in tropical cirrus and thin cirrus amount relative to total cloud amount, at higher altitude (pcld < 330 hPa), both increase with warming (1.87 ± 0.52 and 1.70 ± 0.54 % K-1), which means that these clouds are more frequent among all clouds when Tsurf gets warmer.

Even though the changes in mean Tsurf are mostly linked to interannual variability over the studied period and it is still uncertain how to relate these to long-term patterns due to anthropogenic climate warming, it is very interesting to note that changes in amounts of high opaque clouds and thin cirrus, relative to all clouds, show very different geographical patterns. To get a better understanding on the underlying feedback processes one has to consider the heating rates of these UT cloud systems and link them to the dynamics, which is foreseen in future work.