The GRUAN (Global Climate Observing System Reference Upper-Air Network) measurement programme consists of a network of about 30 quality-controlled radiosonde measurement sites around the world to detect trends in essential climate variables such as temperature and humidity . Here we pick out the GRUAN site on Manus Island, where radiosondes were launched from January 2011 to July 2014, run by the Atmospheric Radiation Measurement programme . This is a particularly suited reference dataset for the scope of our work for two reasons. Firstly, Manus Island is located at about 2∘ S in the western Pacific warm pool, a region where EMLs are expected to occur frequently due to their link to deep convective events. Secondly, the standard radiosonde launch times at 00:00 and 12:00 UTC with a local time shift of UTC+10 h turn out to coincide well with IASI overpasses at the fixed Equator crossing time (ECT) of the Metop satellites at around 09:30 local time.

The GRUAN sounding data used in this work are obtained from the RS92-GDP.2 data archive. Uncertainty estimates are 6 % for relative humidity (RH) and between 0.15 and 0.6 K for temperature depending on daytime and altitude . When binning the launch times of the full sounding dataset into hourly intervals, about 60 % of the soundings occur around the 00:00 and 12:00 UTC launch times. A significant anomaly in radiosonde launch times occurred from 24 September 2011 to 31 March 2012 with launches every 3 h as part of the DYNAMO campaign .

As a first step of preparing the GRUAN sounding data for our processing, relative humidity values are transformed from being defined with respect to the saturation vapour pressure above water (GRUAN standard) to a mixed phase approach as described by . We then linearly interpolate the sounding dataset to a fixed altitude grid ranging from 0 m at the surface to 15 km altitude at 10 m intervals. In the event of missing values in the original data, we interpolate over intervals of up to 100 m and leave the missing values for larger intervals. We then deduce H2O volume mixing ratios (VMRs) from RH, temperature and pressure.

We use the ECMWF Reanalysis v5 (ERA5) high-resolution atmospheric data on a 31 km spaced horizontal grid, on 137 vertical levels and in hourly intervals. Detailed descriptions of spatial and temporal discretisation of ERA5 are provided in the overview paper and in the IFS (version Cy41r2) documentation .

We use a total of 21 ERA5 pixels around Manus Island as depicted in Fig. . The data are originally stored on a T639 spectral grid or a reduced Gaussian grid depending on the variable. We transform the grids of all variables to a 0.25∘ evenly spaced latitude–longitude grid using bilinear interpolation. We deduce H2O VMR as our main humidity quantity from the specific humidity that is originally provided in ERA5. We deduce altitudes for each ERA5 profile by assuming a hydrostatic atmosphere and using the fixed pressure grid and the temperature profiles as input.

The IASI Level 2 retrieval dataset used in this work is called the “IASI All Sky Temperature and Humidity Profiles – Climate Data Record Release 1.1 – Metop-A and -B” and is provided by . We use only data from Metop-A. We refer to this dataset as the IASI L2 CDR in the frame of this study. The dataset is aimed to be a consistently reprocessed long-term dataset based on the most recent version of the statistical piecewise linear regression (PWLR) EUMETSAT retrieval algorithm. Data of the near-real-time operational IASI L2 retrieval are subject to significant jumps over the years due to algorithm updates . Since the algorithm of the period between 2011 to 2014 is not representative of today's standard, we use the reprocessed IASI L2 CDR.

Details about the IASI L2 CDR are provided in the product user guide . Here we summarise some of its main properties. The retrieval algorithm makes use of IASI spectra and radiances observed by the microwave sounders AMSU-A (Advanced Microwave Sounding Unit-A) and MHS (Microwave Humidity Sounder) aboard the same satellite to also retrieve information about atmospheric temperature and humidity in the presence of clouds. The retrieval is conducted on the native pixel resolution of IASI with a pixel diameter of about 12 km at nadir. Only AMSU-A information is available on 2×2 arrays of IASI pixels. To train the PWLR retrieval algorithm, global sensing data of 4 d of each month of the years 2015 and 2016 are matched with ERA5 temperature, humidity and ozone profiles on 137 vertical levels. Cloudy scenes are included in the training step of the algorithm to allow for the retrieval of atmospheric quantities in all-sky scenes. The retrieval is conducted on 137 atmospheric levels and an additional surface level. All-sky retrievals are conducted for atmospheric temperature and specific humidity profiles as well as for surface temperature and total column water vapour. A cloud fraction estimate is also provided based on AVHRR (Advanced Very High Resolution Radiometer) data that are integrated over the retrieval's field of view. The dataset also comes with uncertainty estimates for temperature and humidity profile retrievals that reflect the mean uncertainty of the surface level and the mid-troposphere . These uncertainties are provided in units of kelvin in temperature and dew point temperature. As recommended in the user guide, we filter cases considered highly defective with uncertainties > 4 K. This filtering only removes about 1 % of data. We want to highlight that we repeated our analysis based on more strict filtering criteria with uncertainties < 1 K but found no significant changes to our results.

The only variable we add in our own processing is the height associated with the retrieval's vertical levels. For this purpose we assume a hydrostatic atmosphere and use profiles of pressure and temperature as input.

The Community Long-term Infrared Microwave Combined Atmospheric Product System (CLIMCAPS)-Aqua Level 2 product is based on AIRS spectra and AMSU-A radiances. The processing uses a sophisticated stepwise optimal estimation procedure following the formalism of of various atmospheric quantities such as temperature, moisture, cloud heights and fractions, and concentrations of trace gas species O3, CO, CH4, CO2, HNO3 and SO2. The retrieval is conducted on about 50 km spatial pixels at nadir (150 km at scan edge). One pixel is referred to as field of regard (FOR) and is made up of 9 (3×3) AIRS field of views (FOVs). The retrieval procedure and a characterisation of retrieval errors are described by . In an evaluation of the CLIMCAPS observing capability, it is found that CLIMCAPS has sensitivity to multiple narrow tropospheric layers in temperature and humidity – a promising premise for our study .

We limit our use of available CLIMCAPS variables to the retrieved surface temperature, temperature and humidity profiles, the total cloud fraction, the geopotential height, and the respective quality control flags and error estimates. Temperature profiles are provided on 100 fixed vertical pressure levels from the surface to the top of atmosphere, and specific humidity profiles are provided on 66 levels from the surface to about 50 hPa. Since surface pressure is not a retrieval quantity and instead MERRA2 (Modern-Era Retrospective analysis for Research and Applications, Version 2) surface pressures are used as input to the retrieval, we calculate surface values of humidity following the boundary layer adjustment procedure that is described in the CLIMCAPS science application guide . Surface values of humidity are important for our method of analysing moisture anomaly characteristics that is described in Sect. .

The quality control flags are provided for each variable on all vertical levels. They subdivide the retrieval into “best”, “good” and “rejected” quality. We filter cases where the specific humidity quality control flag of the level closest to MERRA2 surface pressure is labelled “rejected” and cases with more than 10 “rejected” vertical levels in humidity between 900 and 100 hPa. These criteria are quite stringent as they filter about 90 % of the data. However, we do not aim to analyse data that are already flagged as being of deficient quality.

A significant difference between the IASI L2 product and the CLIMCAPS product lies in the estimation of the total cloud fraction and the way cloudy scenes are handled. While for the IASI L2 product cloud fraction is estimated based on an independent instrument (AVHRR), CLIMCAPS estimates cloud fraction based on a subset of cloud-sensitive AIRS channels. CLIMCAPS does so for each AIRS FOV and provides a derived FOR-integrated total cloud fraction, i.e. over 3×3 FOVs. To retrieve atmospheric quantities in cloudy conditions, the CLIMCAPS and IASI retrieval products deploy conceptually different methods. While the IASI product attempts retrieval through the cloud on the native IASI pixel size of 12 km, CLIMCAPS deploys a cloud-clearing technique where information from the 3×3 AIRS FOV spectra are combined to represent the atmospheric state around the clouds throughout the total retrieval FOR. For the IASI L2 CDR, a retrieval through the cloud can only be achieved through the water vapour and temperature information available from the microwave sounders since IR sounders do not yield information through clouds. We specifically compare the retrievals' capabilities to resolve vertical moisture structures in all-sky and clear-sky conditions in Sect.  and .

We collocate the datasets pairwise in space and time to seek high direct comparability of the investigated scenes. This is done using a collocation toolkit that is freely available as part of the “typhon” collection of Python functions for atmospheric science (https://www.radiativetransfer.org/tools/, last access: 3 January 2023). We want to highlight that direct comparability of the datasets is inherently limited due to different horizontal resolutions of the datasets in the presence of inhomogeneous atmospheric fields. By still comparing the datasets' vertical humidity structures directly, we implicitly assume that differences in the datasets’ effective vertical resolution are significantly higher than the averaging effect of different horizontal resolutions on the vertical moisture structure. We would argue that, if for any of the compared datasets, this assumption may break for some cases of point source GRUAN radiosonde comparisons to ERA5 or IASI/AIRS retrievals. However, as shown in Sect. , we find ERA5 to match EML characteristics of GRUAN quite well, indicating that horizontal resolution mismatches are not the significant error source in the conducted comparisons.

We conduct the collocation for four dataset pairs, namely ERA5/GRUAN, IASI/GRUAN, IASI/ERA5 and CLIMCAPS/ERA5. With GRUAN being the gold standard reference dataset, we use it as reference where sufficient collocations are available. The standard launch times at 00:00 UTC and 12:00 UTC in conjunction with a local time difference on Manus Island of UTC+10 h yield launches at local times of about 10:00 and 22:00, matching up well with the IASI Equator crossing time of about 09:30 and 21:30. Unfortunately for the AIRS-based CLIMCAPS retrieval, there is a systematic offset of about 4 h in GRUAN radiosonde launch time and the Equator crossing time of the Aqua satellite at around 01:30 and 13:30. We decide to choose rather conservative collocation criteria of 50 km and 30 min. The temporal criterion of 30 min is chosen due to the expected 30 min offset of IASI overpasses and regular radiosonde launches. In addition, 30 min assures temporal collocation with ERA5, which has hourly sampling. These collocation criteria still yield a high number of matches between IASI, GRUAN and ERA5 while disabling matches between CLIMCAPS and GRUAN. Although previous case studies of EMLs suggest EML lifetimes of about a day e.g., we prefer to use quite strict collocation criteria since atmospheric variability over the course of hours can still be significant . In addition, while retaining tight collocation criteria, we find that an evaluation of CLIMCAPS retrievals is still possible with respect to ERA5, which we find to represent EMLs reasonably well when compared to GRUAN (see Sect. ).

For collocation pairs where the secondary dataset has higher spatial resolution than our spatial collocation radius of 50 km, several pixels of the secondary dataset can match with one pixel of the reference dataset. This is frequently the case for all collocation pairs because the lowest-resolution dataset CLIMCAPS has a 50 km pixel diameter while the collocation diameter is 100 km. In these cases, we randomly select one of the matching pixels. In cases where the spatial resolution of the reference dataset is higher than our spatial collocation radius, it can occur that the same data points of the collocated secondary dataset collocate with several data points in the reference dataset. This is the case for the collocations with reference to ERA5. We reject cases where the same data point of the secondary dataset would be used more than once to assure that all collocations are truly independent data pairs.

Applying the described collocation criteria and the dataset-specific filtering criteria described above, we obtain 1921 ERA5/GRUAN collocations, 648 IASI/GRUAN collocations, 37 491 IASI/ERA5 collocations and 2500 AIRS/ERA5 collocations. The strong discrepancy in the number of collocations has several reasons. First, matches with ERA5 within our collocation criteria are available for all data points of the other datasets, yielding a high number of collocations for every dataset with reference to ERA5. The reason for the number of IASI/ERA5 and AIRS/ERA5 collocations deviating significantly lies in the different reduction of data points in IASI and AIRS retrievals when applying the quality criteria described in Sects.  and .

The moisture anomaly identification and characterisation method introduced in Sect.  is applied to the humidity profiles of the four collocation datasets. Figure  shows the resulting distributions of moisture anomaly characteristics for the four collocation pairs. In the following, we discuss what these results tell us about the different datasets' ability to capture EMLs. We start off with ERA5, then go to IASI and finally to AIRS.

As a first indicator of a dataset's ability to capture moisture anomalies, we compare the number of detected moisture anomalies to the reference dataset, i.e. the areas under the distributions depicted in Fig. . ERA5 captures about 99 % as many anomalies as collocated GRUAN data, indicating a good amount of vertical water vapour variability in ERA5 (Fig. , first row). Moisture anomalies in ERA5 are about 50 % weaker and 28 % thicker than moisture anomalies in the collocated GRUAN dataset. Moist layers that are less than 2 km in thickness are particularly underrepresented by ERA5, while moist layers with thickness >3 km occur more often. These biases suggest that ERA5 is subject to some degree of smoothing due to limited vertical resolution, which we quantify in the following.

To investigate to what extent smoothing alone can explain the biases between the moisture anomaly characteristics of ERA5 and GRUAN, we apply a running mean with vertical window size of 1 km and constant weighting to the GRUAN profiles. The resulting distributions are shown as thick lines in the first row of Fig. . They show that biases in all three moisture anomaly characteristics can mostly be eliminated through the artificial smoothing. This indicates an effective vertical resolution of the ERA5 humidity profiles in the free troposphere of about 1 km. We conclude that ERA5 captures vertical humidity structures on scales of 1 km and greater well as no systematic deviations from the GRUAN distributions are apparent. Hence, we argue that ERA5 is a suitable reference for assessing the satellite retrieval datasets.

We assess the IASI L2 CDR by comparing it to GRUAN data (Fig. , row 2) and ERA5 data (Fig. , row 3). The IASI L2 CDR captures about 75 % as many moisture anomalies as in collocated GRUAN data and about 79 % as many moisture anomalies as in collocated ERA5 data. This is a first indicator that the IASI L2 CDR captures less vertical water vapour variability than ERA5. In addition, the maximum in moisture anomaly thickness at around 2 km altitude detected in both GRUAN and ERA5 data is missing in the IASI L2 CDR. Instead, the anomaly thickness distribution is shifted towards significantly higher values with differences in the means of 85 % against GRUAN and 28 % against ERA5. Moisture anomalies are also significantly weaker in the IASI L2 CDR, with mean differences of 53 % and 10 % against GRUAN and ERA5 data, respectively. At this point we want to highlight the added value of assessing the vertical moisture structures of a dataset through moisture characteristics opposed to just comparing the mean profiles (Fig. ). While we find a strikingly good agreement of the IASI, ERA5 and GRUAN humidity profiles on the mean, quite significant biases become apparent when applying the moist layer characterisation method and then taking a statistical look at how the resulting metrics compare.

As for ERA5, we investigate whether the found biases in anomaly strength and thickness against GRUAN can be explained by smoothing. We apply a 1 km moving average to the GRUAN profiles collocated with the IASI L2 CDR and obtain the moisture anomaly distributions represented by the thick lines (Fig. , row 2). While the biases in anomaly strength and height against the IASI dataset are significantly reduced, a strong bias remains in the anomaly thickness. We also attempted to adapt the smoothing window width to different values between 1 and 5 km but do not find the anomaly thickness distribution to approach the one of the IASI dataset much more (not shown). Hence, the bias in anomaly thickness originates from some other source of error in the IASI dataset than smoothing. We come back to this in the next subsection when concentrating on the clear sky.

To assess the AIRS CLIMCAPS retrieval, we rely only on ERA5 as a reference as outlined in Sect. . The AIRS CLIMCAPS retrieval captures about 92 % as many moisture anomalies as collocated ERA5 data, which is significantly more than the IASI L2 CDR. Moisture anomalies in the AIRS CLIMCAPS retrieval are on average 26 % stronger and 5 % less thick than those in collocated ERA5 data. Also, moist layers in the AIRS CLIMCAPS retrieval are typically found significantly lower in the troposphere compared to the three other datasets; in particular, there are many more moist layer cases below 5 km compared to ERA5. The mean moist layer height is about 1.3 km lower in the AIRS CLIMCAPS retrieval compared to ERA5. We already saw this bias in terms of a shift of the mid-tropospheric humidity peak towards lower altitudes when comparing the dataset mean profiles in Fig. . Moisture anomaly strength is a somewhat height-dependent quantity with generally stronger anomalies in the lower troposphere than further up . Hence, the increased strength of moist layers in the AIRS CLIMCAPS retrieval is to some degree also caused by a bias in moisture anomaly height. Nonetheless, the number of moisture anomalies in the AIRS CLIMCAPS retrieval speaks towards a good capability of the dataset to capture vertical moisture variability – more so than the IASI L2 CDR.

These findings are coherent with the notion of previous case studies that optimal-estimation-based retrievals are more capable of capturing vertical moisture structures than regression-based retrievals . A plausible explanation for the superiority of the optimal-estimation-based AIRS CLIMCAPS retrieval is that capturing EMLs is not sufficiently emphasised in the training of the regression-based IASI retrieval. The AIRS CLIMCAPS retrieval is constrained by a priori assumptions about mean and variability of the atmospheric state, but if the optimal estimation setup is tweaked well, deviations from the mean state can be captured well with this method.

The satellite retrieval products do operate in the presence of clouds, but information content is limited with increased cloudiness and cloud depth, in particular from the infrared instruments. Hence, we are interested in whether our analysis of moisture anomaly characteristics yields different results when limited to clear-sky scenes compared to the previously investigated all-sky scenes. Possible differences could then potentially be linked to the different cloud handling schemes deployed by the retrieval products (Sect. ).

The AIRS CLIMCAPS and IASI L2 retrievals come with an estimate of total cloud fraction for each retrieval pixel, which are obtained based on quite different methods as outlined in Sect. . ERA5 also provides a total cloud fraction variable, which we show in addition but do not base our further analysis on since it appears quite biased against the satellite-derived cloud fractions. As suggested in the CLIMCAPS science application guide, we use a cloud fraction threshold of 0.2 to distinguish clear-sky from cloudy scenes . For the two collocation datasets IASI/ERA5 and AIRS/ERA5, this leaves about 22 % of the all-sky amount of data. For the IASI/GRUAN comparison, sampling becomes too limited, which is why we limit this analysis to the satellite collocations with ERA5.

Figure  shows the resulting cloud fraction distributions of the two collocation datasets. It is striking that when limiting satellite-based cloud fractions to 0.2, ERA5 cloud fraction estimates show maxima near cloud fractions of 1. Without any applied thresholds, both satellite datasets also have their global maxima near cloud fractions of 1 (not shown). However, the secondary maximum near 0 found in both satellite datasets is not at all present in ERA5. Compared to CLIMCAPS, we would have expected a stronger bimodality between high and low cloud fractions in ERA5 due to the higher spatial resolution of ERA5 of 31 km compared to about 50 km in the nadir view of the CLIMCAPS product. However, since cloud fraction requires subgrid-scale knowledge, it is difficult to define this variable in a model framework. Hence, finding significant differences to satellite-derived estimates is not completely surprising.

Figure  shows the resulting moisture anomaly characteristics after application of the clear-sky filter. All datasets consistently show an increase in mean anomaly strength of about 20 % compared to the all-sky results. Note that our method for quantifying anomaly strength is designed to capture the magnitude of vertical moisture variability rather than absolute amount of humidity, which would be highest in the event of clouds . The found increase in anomaly strength in the clear sky is in line with our expectations because in cloudy conditions vertical humidity variability is limited by the saturation humidity, leading to weaker moisture anomalies.

We also see a significant change in the shape of the anomaly height distributions when comparing clear sky to all sky. IASI and ERA5 both show a clear bimodal structure in anomaly height in the clear sky, which was not the case in the all-sky data. Physically, we explain the position of the maxima near 5 and 12 km by levels of preferred detrainment of moist air from mid-level or deep convective plumes into the clear-sky environment . The mid-level detrainment is thought to be driven by enhanced stability near the melting level, and the upper-tropospheric detrainment is associated with increased stability towards the tropopause as the atmosphere goes into pure radiative equilibrium aloft. We hypothesise that the mid-tropospheric peak is more pronounced than the upper-tropospheric peak in ERA5 and IASI anomaly height distributions because both deep (cumulonimbus) and mid-level (cumulus congestus) convection causes mid-level detrainment, while only deep convection causes upper level detrainment. AIRS also shows peaks in anomaly height near 5 and 12 km and another peak in between at around 7 km that we can not link to a physical mechanism in this height. However, when interpreting the detailed shape of the distributions to this extent, we advice caution due to the limited number of AIRS/ERA5 collocations, which is only about 10 % of the number of IASI/ERA5 collocations.

We do not find significant changes in biases between satellite retrievals and ERA5 in anomaly strength or thickness when limiting our data to clear sky. While we do see changes in the means of the distributions as described above, biases remain similar. Although biases do not change much, we see that the all-sky secondary maximum at large anomaly thickness values of IASI, which is not present in ERA5 (Fig. ), vanishes in the clear sky, indicating better vertical resolution. However, going to clear sky does not reduce the gap between satellite retrievals and ERA5 at anomaly thickness values below 3 km. We conclude that the retrievals' observing capability of moist layers is not significantly limited by clouds.