Evaluation of tropical water vapour from CMIP6 GCMs using the ESA CCI "Water Vapour" climate data records

The tropospheric water vapour data record generated within the ESA Climate Change Initiative "Water Vapour" project (ESA TCWV-COMBI) is used to evaluate the interannual variability of global climate models (CMIP6 framework under AMIP scenarios) and reanalysis (ECMWF ERA5). The study focuses on the tropical belt, with a separation of oceanic and continental situations. The intercomparison is performed according to the probability density function (PDF) of the total column water vapour (TCWV) defined yearly from the daily scale, as well as the evolution of the large-scale overturning 5 circulation. The observational diagnostic relies on the decomposition of the tropical atmosphere into percentile of the PDF and into dynamical regimes defined from the atmospheric vertical velocity. Large variations are observed in the patterns among the data records, especially over tropical-land, while oceanic situations show more similarities in both interannual variations and percentile extremes. The signatures of El Nino/La Nina events, driven by the sea surface temperatures, are obvious over the oceans. Differences also occur over land for both trends (a strong moistening is observed in the ESA TCWV-COMBI data 10 record which is absent of CMIP6 models and ERA5) and extremes years. The discrepancies are probably associated with the scene selection applied in the data process. Other sources of differences, linked to the models and their parametrizations, are highlighted.

the evaluation period is restricted to before 2014 for consistency with the available period of the CMIP6 experiment. Such cut in the ESA TCWV-COMBI excludes the OLCI observations.

CMIP6 Models
A subset of seven GCMs participating to CMIP6 is evaluated here, limited by the availability of the required geophysical 95 variables at daily resolution (at least) that is comparable with the CCI_WV CDRs. However, there was no TCWV field at the daily frequency that was available from the Earth System Grid Federation (ESFG) (node of Institut Pierre Simon Laplace, IPSL). Therefore we recomputed TCWV from the vertical profiles of specific humidity q (in g/kg) that were provided at the model vertical resolution. High vertical resolution of specific humidity (more than 19 vertical levels in the troposphere) were then necessary to be certain to capture the full tropospheric water vapour (using the extraction on a selection of pressure levels 100 would bias the computation of TCWV).
The TCWV (in kg/m 2 ) from each model is thus calculated using: where g is the gravitational acceleration constant, and dp is the difference between adjacent pressure levels (hPa).
We focus on the AMIP (Atmospheric Model Inter-comparison Project) (Ackerley et al., 2018) Table 2. In addition to the CMIP6 models, the ensemble mean of the seven models is also included in the following analysis to represent the mean state of the CMIP6 models.
Since the TCWV-COMBI data is cloud screened over land area, it is important to carefully analyze the cloud conditions 110 for CMIP6 models before making the quantitative comparison. A series of tests have been done (not shown) using different screening thresholds from the cloud fraction and precipitation rate to determine the best cloud conditions for the modelobservation comparison. The key aspect here is to have a filter on clouds and/or precipitation that is stringent enough for an apple-to-apple comparison but also that leaves enough point from a statistical point of view. Then, over land the thresholds are set with maximum cloud fraction of 50% for each vertical level and, over ocean with precipitation rate less than 0.001 115 kg/m 2 /s 2 (Sohn and Bennartz, 2008). These thresholds result in a reduction of the size of the datasets that differs according to the region and the percentages of remaining data for land and for ocean are indicated in Table 2. Over land, the percentage of data remained for the CMIP6 models after the cloud screening are globally in the range 46.14% -76.10%. Over tropical oceans the percentage of data remained is higher and range from 99.79% to 99.98%. Although the scene selection is more stringent over land, this indicates that the CMIP6 data used in the following analysis are comparable in terms of size of sample to the 120 data from TCWV-COMBI and ERA5. It is worth mentioning that although the screening thresholds for the models are set to meet the criteria of the TCWV-COMBI product, the number of data retained for the comparison are not exactly the same for all models. Therefore, differences among data sets may be observed in the analysis. This is particularly true over tropical land.

ERA5
The reanalysis data are widely analyzed in atmospheric sciences to assess the impact of changes in observation system, to 125 scale progress in model simulations, and to calculate climatology for forecast-error evaluation (Hersbach et al., 2020). The ECMWF's ERA5 TCWV data are based on the integrated forecasting system (IFS) Cy41r2, with considerable enhanced horizontal resolution of 31 km compared to 80 km for ERA-Interim. Here the ERA5 TCWV with hourly frequency are averaged into daily data. To compare the data under the same conditions, the ERA-5 land-sea mask is employed for land and ocean separation, and a scene selection is performed and is similar to the process of the CMIP6 data. Hence data with total cloud 130 cover less than 95% and total column cloud liquid water less than 0.005 kg/m 2 over land (Sohn and Bennartz, 2008), and data with total precipitation less than 0.001 kg/m 2 /s 2 over ocean are retained.

Methods
The time series of the daily means of the CMIP6, ERA5 and ESA CCI_WV TCWV-COMBI are analyzed with tropical-land and tropical-ocean separation over the common observation period that covers July 2003 to December 2014.

135
The intercomparisons are conducted according to two approches: 1) The first approach evaluate the interannual variation of TCWV based on the probability distribution function (PDF) established from the daily records for each year of the period. The percentiles of the TCWV are defined from the yearly distributions and the data is sorted by intervals of 10 percentiles. Finally, the mean TCWV of each interval is computed and normalized by the corresponding mean TCWV of the whole observation period for this given percentile.This is generalized for every per-140 centile. This approach is meant to highlight the tropical anomalies with respect to the mean and trace back to the inter-annual variability of the tropical atmosphere.
2) The second approach is based on the fact that the water vapour distribution is strongly controlled by the large-scale vertical motion of the atmosphere. Therefore, we can use the mid-tropospheric atmospheric vertical velocity at 500 hPa (noted ω500 in hPa/day) as a proxy for the vertical motions in the tropics (Bony et al., 2004). While such framework has been greatly 145 used to study tropical clouds and their distribution (e.g., Konsta et al., 2012;Höjgård-Olsen et al., 2020), this link between vertical motion and TCWV is documented (e.g., Brogniez and Pierrehumbert, 2007) and further illustrated on Figure 1. Figure  1 presents the TCWV-COMBI averaged over the whole 2003-2014 period as well as the mean Winter (December, January, and February -DJF) and mean Summer (June, July, and August -JJA), together with the corresponding ω500 taken from ERA5 at a monthly scale. As expected, a moist troposphere is associated with large-scale ascending motion (ω500 < 0 hPa/day) while 150 a dry troposphere is associated with large-scale subsidence (ω500 > 0 hPa/day). The TCWV data are sorted upon 10hPa/daybins of monthly values of ω500. The dynamical decomposition is performed for all TCWV data records at each year of the time period. Moreover the TCWV data averaged over the whole 2003-2014 period is also sorted into the corresponding ω500 bins of the period and this value is considered as the reference to normalize the results. This second approach allows to study the trends in TCWV for a given state of the large-scale dynamics, and thus overcome issues associated with variations (such as 155 shifts or expansion) of the atmospheric circulation (Vallis et al., 2015;Mbengue and Schneider, 2017).

Results and Discussions
This section aims to assess the degree of agreement in the TCWV climatology and interannnual variations between the ESA CCI_WV TCWV-COMBI, CMIP6 models and ERA5 reanalysis data over the tropics (30 • S -30 • N). The distribution of the water vapour over tropics and its link to large-scale circulation (ω500) are discussed in detail. signal. More specifically, the ESA TCWV-COMBI data is globally moister compared to the CMIP6 models and to ERA5 data, and this moist bias is even more pronounced over tropical land (Fig 2a, ∼ 10kg/m 2 over land vs ∼ 2kg/m 2 over ocean). On the other hand, the observed ESA TCWV-COMBI over ocean reaches globally higher values than over land. This difference 170 can be explained because the ESA TCWV-COMBI dataset is composed with clear-sky-only data, which are likely drier than the nearby cloud area for a given location and thus translates into a dry bias associated to moistening processes by convective clouds (Sohn et al., 2006). Besides, the boundary layer is drier in the continental subtropics while the maritime stratocumulus zones are wetter at low levels and very dry at high altitude.

175
The normalized PDFs of the daily TCWV obtained from all data records over both land and ocean are displayed in Figure 3. because of the data screening. Over land, all datasets reach a first maximum at around 10-13 kg/m 2 and present a secondary maximum near 40-50 kg/m 2 . Moreover, more lower values are observed in the CMIP6 models and ERA5 than in the ESA TCWV-COMBI dataset and this cannot be explained by the cloud-screening method alone. Indeed, the ESA TCWV-COMBI 180 dataset is strictly cloud-free while some cloudy scenes remain in both ERA5 and CMIP6 (see Section 2), which could translate in a moister TCWV of the latter data records.
Over oceans, most of the TCWV data are located around 20-60 kg/m 2 . The main peak is around 30 kg/m 2 , and a secondary peak appears near 50 kg/m 2 . While the frequency of situations of the main peak is nearly identical for TCWV-COMBI, ERA5 and CMIP6, there is a divergence for the secondary peak. This secondary peak even dominates the PDF of the CMIP6 models 185 while ERA5 and ESA TCWV-COMBI are still quite similar. The bimodal distributions can be explained by the presence of more humid columns in the Intertropical Convergence Zone (ITCZ) and relatively drier ones in subtropical regions.

Extremes of the distributions
The data records are then evaluated following the approach (1) described in Section 3: the percentiles of the annual distributions of TCWV (at daily resolutions) are sorted into bins of 10% intervals, and this is done for each year of the period.  in the ESA TCWV-COMBI or CMIP6 ensemble mean. The very good agreement among the various data sets is largely due to the fact that the CMIP6 models that are evaluated under the AMIP scenario, so with the same prescribed SST for all models and ERA5, and that the relationship between SST and TCWV is largely explained by the Clausius-Clapeyron law (Stephens, 1990). Hence this explains that anomalous years are the same, most notably those concerned by El Nino Southern Oscillation (Trenberth et al., 2005)

General assessment
The interannual variability of TCWV is then analyzed from its links with the large-scale atmospheric circulation, and follows the approach (2) described in Section 3. The monthly ω500 of individual data records are decomposed into 10 hPa/day intervals 215 in the range of -120 to 120 hPa/day. Figure 6 displays the normalized PDFs of the ω500 of the CMIP6 models and ERA5. As mentioned earlier there is no atmospheric circulation data from the ESA TCWV-COMBI data record, so the ω500 from ERA5

Trends over lands
This global assessment is further discussed by applying the TCWV-ω500 approach for every year of each data record to delineate the trends in TCWV. As shown in Figure 7, all the data records (except for ERA5) agree that the most positive ω500 is clearly the moistest model and CanESM5 is the driest. The moist bias of IPSL-CM6A-LR is already documented (Boucher et al., 2020) and is explained by the (too) efficient parametrization scheme of the transport of evaporated air from the surface to the top of the boundary layer. The CanESM5 behavior is opposite and this shall be partly explained by its strong effective climate efficiency (Virgin et al., 2021) which yields to adjust too strongly the atmospheric response to a perturbation. ERA5 displays also a very dry troposphere whatever the dynamical regime. Moreover, all models show that the transition dry/moist 245 occurs around 60hPa/day and this transition is similar also for ERA5 and ESA TCWV-COMBI.

Trends over oceans
Oceanic situations as considered similarly with respect to the large-scale circulation. The results are shown in Figure 9 Figure 9. Mean of TCWV over tropical ocean areas at each dynamical intervals (ω500) in 10 hPa/day computed from each data record.

275
Despite the importance of water vapour in the study of climate variability, our ability to evaluate the water vapour feedback is constrained by its measurements at ranges of scales that are adapted for local, regional and global studies. This deficiency is attributable in part to the fact that it is difficult to quantitatively and accurately measure the distribution of water vapour.
To work towards the requirement of GCOS on satellite-based water vapour observation as ECV, the ESA Climate Change Initiative "Water Vapour" project (ESA CCI_WV) tackled this challenge by generating gridded products on stratospheric and 280 tropospheric water vapour from multiple satellite observations suitable to climate and process studies.
We have conducted a comprehensive evaluation of the tropical water vapour (30 • N-30 • S) of seven GCMs (CMIP6 models, AMIP scenario) and ERA5 using the TCWV-COMBI climate data record developed within the ESA CCI_WV project as a reference. The study focused over tropical-land and tropical-ocean areas at the daily frequency and over the 2003-2014 period.
The variability of TCWV was analyzed according to (i) its probability density function (PDF) defined at a yearly scale over 285 the period and (ii) to the large-scale circulation using the atmospheric vertical velocity at 500hPa (ω500) as a proxy of the tropospheric overturning circulation.
Different patterns of variability are observed among the various datasets, the largest discrepancies being noticed over land areas, while over the oceans the datasets are closer to each other: over land, the PDFs of the ESA TCWV-COMBI present a clear moistening trend of their driest percentiles, with a 290 tiping point in 2011, probably associated to the addition of the MODIS observation in the climate data record. The projection of the TCWV onto regimes of ω500 shows the same behavior, the drying trend being present for all regimes of ω500. Interestingly, the CMIP6 ensemble mean and the ERA5 reanalysis are in good agreement in terms of interannual anomalies, although the ERA5 TCWV is clearly too dry.
over ocean, the PDFs of all datasets present the same interannual variability. The extreme dry and moist years, associated 295 to El Nino and La Nina events, are the same. This similarities hold when using the large-scale circulation as an evaluation tool, with the same transition between the dry/subsiding regimes and the moister/ascending regimes.
The results show that the ESA TCWV-COMBI data, ERA5 data vary within the ensemble spread of CMIP6 models, indicating that the mean models could correctly represent the evolution of water vapour with respect to large-scale circulation. The humid area is related with the ascending motion (negative value in ω500) and dry area is related with the subsiding motion (positive 300 value in ω500) over both tropical land and tropical ocean area. There are discrepancies observed among the data records, because of the lateral mixing, outflows from clouds, and the precipitation efficiencies of the convective schemes. It is difficult to track entirely the reasons of the differences, however, the differences and similarities can be explained by several factors : 1) the use of different satellites with different accuracies and resolutions within the ESA TCWV-COMBI may explain part of the moistening trend observed for this dataset over land.

305
2) the cloud-masks applied to the GCMs and ERA5 and defined to mimic the cloud-mask of the observation can also explain the differencies.
3) the parametrization of the moisture fluxes at the surface and of convection, as well as the climate efficiency of the GCMs also contribute to the observed differences.
4) the use of AMIP scenarios, defined from prescribed sea surface temperatures, as well as a scene selection that is much more 310 conservative than over land, explain almost entirely the very good agreement between the ESA TCWV-COMBI, ERA5 and the GCMs.
It is really necessary to underline the role of the cloud mask in the assessment of water vapour fields in climate models using observations, even though water vapour seems to be an easier parameter than clouds. Climate models provide water vapour profiles (and sometimes the integrated values) at the scale of their mesh which is usually a lot larger (see Table 2) than 315 the observed water vapour, and whatever the cloud distribution within the mesh. However, the water vapour estimated from observations by satellite sensors is strictly analyzed with respect to cloud contamination. This clearly poses the question about having access to the simulated water vapour (full profiles as well as integrated values) for the clear sky part of the meshes of the climate models.