The RPG-HATPRO Generation 5 multi-channel microwave radiometer BCOHAT has been continuously operating on top of a container at 25 ma.s.l. in proximity to the island shore (see Fig. a and ). An absolute calibration with liquid nitrogen was performed before the start of the EUREC4A operations on 14 January 2020. BCOHAT measured according to the following regularly occurring scan strategy: azimuth scans at 30∘ elevation angle were performed for the duration of 5 min every 40 min, followed by an elevation scan covering 10 elevation angles (90, 30, 19.2, 14.4, 11.4, 8.4, 6.6, 5.4, 4.8, 4.2∘) at 0∘ azimuth position (later referred to as elevation scan; ). Zenith measurements are performed for 15 min at a temporal resolution of 2 s. Due to technical difficulties with the scanning unit, the scanning strategy changed after 1 February 2020: azimuth scans were not performed, and operations were limited to zenith mode with elevation scans available every 15 min. These technical difficulties also affected the associated BCOHAT weather station. From 26 January 2020 onwards, data from the adjacent BCO weather station were used instead to flag measurements for precipitation (10.25326/54, ). No measurements were performed between 29 and 31 January 2020, due to maintenance on the instrument. A blowing unit was operational to mitigate the deposition of rain on the radomes during and after precipitation events.

Aboard Meteor, the Leipzig Institute for Meteorology (LIM) of Leipzig University operated an MWR-type RPG-HATPRO Generation 5 (here referred to as LIMHAT) and a radar–radiometer system of type RPG-FMCW-94 dual polarization (DP), operating actively in the W-band (94 GHz) and containing a passive radiometer channel at 89 GHz (; here referred to as LIMRAD). Both instruments were placed 4.5 m apart on the navigation deck of the ship at 15.8 ma.s.l. to avoid sea spray. LIMHAT operated at a temporal resolution of 1 s in zenith mode. Elevation scans, as done by BCOHAT, were performed by LIMHAT every full hour. An absolute calibration with liquid nitrogen was performed on 15 January 2020.

LIMRAD was operated with two different radar settings as specified in . Between 17 and 29 January 2020 as well as between 31 January and 28 February 2020, the temporal resolution of LIMRAD was 2.9 and 1.6 s, at a vertical resolution between 22 and 42 m, respectively. Radar absolute calibration was performed on 16 January 2020. Data gaps exist between 27 and 31 January 2020, when different radar chirp table settings were tested, and on 3 February 2020, when all instruments had to be turned off while Meteor was near Trinidad. As explained in , LIMRAD was operated in a novel passive horizontal stabilization system (two-axle Cardan mount) to assure zenith-pointing of the instrument. Stabilization is required to eliminate the effect of horizontal wind on the radar Doppler velocities. Means and standard deviations of absolute values of radar attitude measurements amounted to 0.36∘ ± 0.31∘. It should be noted that since LIMRAD was operated in a horizontal stabilization platform while LIMHAT was not, the exact (near-zenith) viewing direction of both instruments was not always the same. This effect should be negligible for retrieved IWV and LWP, however, as the larger opening angle of the LIMHAT (half-power beamwidth HPBW = 3.5∘) covered the LIMRAD column (HPBW = 0.5∘) even in events of slight mis-pointing.

Aboard Merian, the Institute for Geophysics and Meteorology of the University of Cologne operated a radar–radiometer system of the type RPG-FMCW-94 dual polarization (DP), which measures in the W-band (94 GHz) and includes a passive radiometer channel at 89 GHz (; here referred to as MSMRAD). MSMRAD is of the same type as LIMRAD. The system was positioned on an active stabilization platform from the US Atmospheric Radiation Measurement (ARM) program Mobile Facility 2, which keeps the radar in zenith position by adapting the table surface position to compensate for ship motions (for more information, see ). As for LIMRAD, stabilization helps eliminate the effect of horizontal wind and ship roll and pitch tilting from the radar Doppler velocities. MSMRAD was operated with three chirp programs, established after initial testing, and worked for the entire campaign. The chirp programs had 0.846, 0.786, and 1.124 s integration times, respectively, with vertical resolutions of 7 m in the boundary layer and 30 m in the free troposphere, resulting in brightness temperatures at 3 s temporal resolution (see Table 2 in ). The data browser (https://bit.ly/3ZcAusN, last access: 16 January 2024) displays availability and observational quality for every day of the entire campaign.

IWV, LWP, and coarse temperature and humidity profiles are retrieved from all 14 TB measurements by applying the statistical quadratic regression retrieval equation (see Eq.  with k indicating the number of channel and “var” indicating the retrieved variable). The seven K-band channels (22–31 GHz, channels 1–7) provide information for IWV, LWP, and the absolute humidity profiles, while the seven V-band channels (51–58 GHz, channels 8–14) are used for temperature profiling. The IDL software MWR PRO was used to process the data . 1var=c0+∑k=0N=6c1,k⋅TB,k+c2,k⋅TB,k2

The coefficients c0, c1, and c2 are derived from a climatological training data set comprising 10 871 daily radiosoundings launched from 1990 until 2018 from Grantley Adams International Airport (GAIA, station ID 78954 TBPB) in close vicinity to BCO. Sounding measurements were obtained from http://weather.uwyo.edu/upperair/sounding.html (last access: 16 January 2024). During EUREC4A, radiosoundings of the type GRAW DFM-09 were used .

Following the approach by and, more recently, by , we use a radiative transfer model to link atmospheric conditions with TB. In the model, gas absorption is calculated according to with modifications in the water vapor continuum and 22 GHz line . Liquid cloud absorption is calculated following . A liquid water cloud was modeled using a modified adiabatic liquid water content approach following in vertical levels where radiosounding relative humidity exceeded 95 %. To imitate the instrument noise, a random noise factor was added to the simulated TB, taken as a random sample from a Gaussian distribution with standard deviation of 0.4 K . For the temperature retrieval, only a linear regression was used as in . To derive temperature profiles from the elevation scans, coefficients c1 and c2 were calculated by adjusting the angle for which radiative transfer was performed. Theoretical LWP uncertainty scales with retrieved LWP and is further discussed in Sect. .

To further improve the standalone LWP retrieval, a clear-sky offset correction method is applied to the retrieved LWP . The correction scheme identifies a liquid-free condition if the standard deviation of LWP in a running 2 min window, as well as the previous and subsequent 2 min window, is below 2.5 gm-2. The median LWP during the identified 2 min clear-sky period is subsequently subtracted from all following LWP measurements until the next clear-sky period. Note that, due to the statistical retrieval approach, negative (unphysical) LWP values can occur. The remaining few negative LWP values are not set to zero to keep for statistical noise evaluation and to avoid biasing the overall statistical distribution of LWP. That way, the clear-sky LWP noise can be estimated by analyzing the LWP distribution in independently identified clear-sky periods as presented in Sect. .

As opposed to a multi-channel LWP retrieval, the retrieval of LWP from a single channel is underdetermined as both water vapor and liquid water contribute to the measured TB (e.g., ). In order to extract the LWP signal in TB at 89 GHz, we present a novel retrieval approach based on the difference in brightness temperature, ΔTB, between cloudy-sky TB and the closest clear-sky TB,0. Parameter ΔTB is used in a third-order regression (Eq. ) to estimate LWP. 2LWP=a⋅ΔTB+b⋅ΔTB2+c⋅ΔTB3 with ΔTB=TB-TB,0

Instrument biases are reduced by using the difference in brightness temperatures, so the unbiased portion of the signal from LWP remains. The clear-sky brightness temperature is obtained by selecting profiles not showing any radar reflectivity through the cloud mask, excluding measurements up to 5 min after rain events to avoid biases due to wet radome conditions. The unknown coefficients of the regression (a, b, and c) are derived from a training data set compiled from artificial LWPs and simulated brightness temperatures calculated with the forward model operator Passive and Active Microwave TRAnsfer model (PAMTRA; ). Atmospheric profiles were constructed from 401 radiosondes launched on the respective research vessels (Merian: 182; Meteor: 219) and artificial clouds between 0 and 5 km with LWPs up to 1 kgm-2. To retrieve LWP from the measured ΔTB in non-precipitating conditions, the coefficients derived for the closest radiosounding were applied following Eq. () to ΔTB which was in turn adjusted with noise by a random number of a Gaussian distribution with width of 0.5 K.

IWV is retrieved from the single-channel TB measurements only in clear-sky conditions as emissions are then dominated by water vapor. A quadratic regression is applied as in Eq. (), weighed by the variability of TB around the radiosonde launch. By applying a weight to the regression, misidentified clear-sky radiosoundings are excluded from the training. Thus, 120 and 65 clear-sky radiosoundings were identified aboard Meteor and Merian, respectively, by applying a 98 % relative humidity threshold, and these were used to derive the coefficients linking TB and IWV. The coefficients were then applied to the measured TB in clear-sky conditions as detected by the cloud-masking algorithm presented in the following section.

Ground-based passive microwave radiometer measurements are not reliable during precipitation events due to additional liquid water emissions on the radome contributing to the column emissions. Flagging precipitation is, thus, crucial to guarantee high-quality retrievals. The HATPRO precipitation mask is set to “True” when precipitation was detected by the internal HATPRO or an adjacent weather station. Cloud radar measurements are added to the standalone precipitation flagging to improve the precipitation detection. At BCO, Ka-band (35 GHz) zenith-pointing radar measurements are used. Aboard the ships, measurements of the LIMRAD and MSMRAD cloud radar operating in the W-band (94 GHz) are added. Precipitation is flagged if any reflectivity above -50 dBZ was recorded below 350 m. This reflectivity threshold was chosen according to to exclude sea salt aerosols from being mis-flagged as precipitation. Aboard the ships, precipitation was also flagged if reflectivity exceeded 0 dBZ anywhere in the column or if a rain rate was derived by the radar.

Independent cloud masking was performed using the adjacent radar and, at BCO and aboard Meteor, ceilometer measurements from a Jenoptik/Lufft CHM15k Nimbus ceilometer, respectively. Ceilometer measurements are identified as cloudy if a cloud base height above 100 m is derived by the internal instrument software. If no valid cloud base height is derived, the scene is treated as clear. At BCO and aboard Merian, radar measurements indicate cloudy conditions if a reflectivity of more than -50 dBZ is recorded in more than two range gates above 300 m. The reflectivity threshold was carefully chosen to exclude occurring sea-spray from being flagged as cloudy . Due to the different radar chirp settings (see Sect. ) and resulting radar sensitivities, a threshold of -40 dBZ was applied to the LIMRAD measurements to optimally exclude sea spray and clutter. An additional liquid cloud mask is derived by enforcing that reflectivity above the respective threshold only occurred between 300 and 4000 m. Clear sky is identified if reflectivity is not a number (nan) in all range bins. Due to MSMRAD's maximal range of 10 km, high-occurring cirrus clouds might not be detected and could be mis-flagged as clear conditions aboard Merian.

In the presented analyses, the individual cloud masks are combined to a joint cloud mask as follows: clear conditions prevail if both ceilometer and radar flags are clear; probably cloudy conditions prevail if either ceilometer or radar sensed a cloud; and confidently cloudy scenes refer to measurements in which both radar and ceilometer sensed a cloud. Making use of the additional liquid cloud flag allows us to additionally derive probably liquid cloudy and confidently liquid cloudy conditions to exclude scattering from ice in the LWP statistics. Probably cloudy occurrences are mainly due to sensor beam mismatch, platform motions, or sensitivity differences between the ceilometer and radar as outlined in . Aboard Merian, scenes were classified as clear or confidently cloudy based solely on MSMRAD radar observations.

Missing data of ceilometer or radar led to discarding of 3.7 %, 20 %, and 8.2 % of all measurements as a cloud mask could not be determined at BCO, Meteor, or Merian, respectively. The comparatively higher percentage aboard Meteor is dominated by data availability of LIMRAD. For presented analyses, we additionally demand a valid IWV and LWP, as well as a valid cloud mask for a measurement to be considered, thus excluding scenes affected by precipitation or instrument measurement or retrieval quality. This reduces the availability of valid measurements to 50.5 %, 66.8 %, 69.5 %, and 83.1 % of all 3 s measurements in the core period, dominated by instrument availability as shown in Fig. . Table  summarizes the respective cloud cover fractions of clear, probably (liquid) cloudy, and (liquid) cloudy scenes relative to this subsample for all instruments. The clear-sky fraction is highest aboard Merian and lowest at BCO. We relate the highest clear-sky fraction of 75.2 % aboard Merian to the missing ceilometer and reduced sensitivity of the radar to optically thin and geometrically small clouds . Compared to the airborne cloud cover products presented in , the here-presented ground-based derived confidently cloudy cloud cover estimates are closest to the airborne lidar-derived cloud cover of 34 %. Differences arise due to the fact that airborne operation was limited to selected days and daytime and that airborne horizontal resolution is lower than when measured from ground. Here-presented cloud cover matches the cloud cover observed at BCO from 2 years of measurements . More than 80 % of detected clouds are classified as liquid.

Liquid cloud LWP is analyzed by applying the joint cloud mask (Sect. ) to retrieved LWP. Figure  illustrates retrieved LWP distributions observed by BCOHAT, LIMHAT, LIMRAD, and MSMRAD in confidently liquid cloudy scenes. Corresponding distribution parameters are summarized in Table . Mean LWP conditions in confidently liquid cloudy conditions at BCO and aboard Meteor and Merian were 63.1, 62.5, 52.4, and 46.8 gm-2. The mean conditions at BCO and Meteor align well with the mean airborne LWP of 63 gm-2 observed during NARVAL-1 in similarly dry winter trade conditions across the same region . BCOHAT and LIMHAT retrieved mean LWP of 63.1 and 62.5 gm-2 agree well within their associated LWP uncertainties (see Sect. ).

Even though similar mean LWP conditions were observed, more detailed trajectory analyses are necessary to investigate the effect of ocean surface and island impact on the cloud evolution between Meteor and BCO. Median and mean LWP differ as the mean LWP is influenced by single events of high LWP, e.g., through un-flagged precipitation or sea-spray, while the median is driven by the large amount of small LWP below the instruments' detection limit.

So, 90 % of observed confidently liquid cloudy columns were associated with a LWP of around 160, 120, and 110 gm-2 at BCO, and aboard Meteor and Merian, respectively. The comparatively higher LWP 90th percentile and standard deviation at BCO are most probably related to wet radome conditions as the blower unit of BCOHAT was broken throughout some of the core period (as opposed to the other instruments). We also suspect that the sea-spray-altered and aged radome was less hygroscopic compared to the newer LIMHAT radome, leading to additional moisture on the radome and longer drying times. An additional island impact triggering deeper convection in prevailing non-trade-wind conditions is in ongoing analysis. The close-to-zero 10th percentile reflects the fact that the statistical regression covers negative values to avoid biasing the overall distribution (see Sect. ), and this indicates that the cloud mask did not perform perfectly well in all conditions. Likely, the wider field of view of the MWR compared to the active remote sensing instruments used for the cloud mask led to misidentification of scenes.

Expanding the analysis to include probably liquid cloudy conditions slightly reduces mean and median LWP, as well as all other parameters of the distribution given in Table , likely due to the fact that more mis-flagged clear-sky conditions impact the LWP distribution. This shift in distribution parameters illustrates the sensitivity of the derived LWP properties to the cloud mask performance.

At BCO and Meteor, retrieved LWP is compared to the ERA5 estimates by resampling BCOHAT and LIMHAT's LWP, respectively, to every hour (see inlet in Fig. ). Instrument-derived mean LWP of 33.2 and 39.5 gm-2 agree well with ERA5 mean LWP of 34.5 and 36.0 gm-2 at BCO and Meteor, respectively. The measured LWP variabilities, here quantified as standard deviation, of 58.0 and 84.7 gm-2 are higher than the ERA5 variabilities of 27.9 and 26.4 gm-2, respectively, which we attribute to the horizontal resolution of ERA and the small cloud sizes.

Characterizing the uncertainty of the retrieved LWP by independent measurements is not straight forward as LWP retrieved from measurements by visible or infrared remote sensing techniques is not sensitive to the same column as the microwave measurements e.g.,. Therefore, a clear-sky LWP noise can be derived by analyzing retrieved LWP in independently classified clear-sky cases as a generally accepted strategy . Retrieval offsets to zero are due to the statistical nature of the retrieval approach, due to calibration artifacts and radiometric noise. The lowest detectable LWP is then calculated from the clear-sky LWP noise for different water vapor conditions. Cloudy-sky LWP uncertainty can be estimated as a function of LWP by calculating a root-mean-square difference (RMSD) of true versus retrieved LWP. True LWP here refers to the LWP used to forward-model TB in the radiative transfer calculations (see Sect. ), while retrieved LWP is the result of applying the respective retrieval equation to the same TB.

The retrieved clear-sky LWP distribution at BCO is illustrated in Fig. a, and Table  summarizes the distribution characteristics for all sites. Percentages of 49.0 %, 59.0 %, 61.1 %, and 75.2 % of all valid LWP BCOHAT, LIMHAT, LIMRAD, and MSMRAD measurements, respectively, are identified as clear-sky. Note that the fractions disagree for LIMRAD and LIMHAT aboard Meteor due to different observational gaps in the measurements. Applying a Gaussian fit to the distribution yields a mean and standard deviation, which is interpreted as clear-sky LWP bias and clear-sky LWP noise, respectively. The Gaussian fit widths of 9.9 and 12.0 gm-2 for BCOHAT and LIMHAT, respectively, quantify the clear-sky LWP noise and match clear-sky noises previously identified for retrievals based on the similar channels . The single-channel clear-sky LWP noises are smaller (3.4 and 4.5 gm-2, respectively), as IWV is fixed due to retrieving from the TB difference of cloudy and clear-sky and as water vapor absorption is stronger at 89 GHz compared to the lower frequencies used in BCOHAT and LIMHAT. The lowest detectable LWP depends on the vertical water vapor distribution which, in cloudy conditions, is not available at any of the sites. Therefore, we estimate the smallest detectable LWP as the clear-sky LWP noise which, in turn, depends on the performance of the independent cloud-masking algorithm.

Quantifying the detection limit allows us to analyze which clouds are missed by the different radiometers. Percentages of 79.2 %, 92.8 %, 88.5 %, and 81.8 % of all confidently liquid cloudy flagged measurements contain LWP above the respective detection limits of BCOHAT, LIMHAT, LIMRAD, and MSMRAD (see Table ). The remaining undetected LWP compared to the ceilometer-radar cloud mask is most likely associated to optically thin clouds with low water contents e.g., and to cloud mask performance. This reduction in cloud cover when derived from passive microwave sensors is also observed by the airborne cloud masks . One-third of detected LWP is seen between the detection limit and 30 gm-2, as well as between 30 and 100 gm-2, averaging to mean LWP conditions of 19 to 22 gm-2 and around 55 gm-2, respectively. Only 11 % to 16 % of detected LWP at BCO and aboard the ships, respectively, are associated with thicker clouds of higher than 100 gm-2.

Cloudy LWP uncertainty varies as a function of retrieved LWP as illustrated in Fig. b binned to logarithmic bins of LWP. The mean RMSD for HATPRO-derived LWPs below 20 gm-2 varies below 5 gm-2, corresponding to a relative RMSD between 75 % and 50 %. For LWP between 20 and 100 gm-2, the RMSD moderately reduces from 50 % to 15 % of retrieved LWP (15.8 at 50 gm-2). Above LWP of 100 gm-2, the relative uncertainty is better than 15 % (e.g., 29.9 gm-2 at LWP of 200 gm-2). Higher LWP values are in reality often affected by precipitation and, thus, not sensed by ground- or ship-based MWR measurements. find, on average, higher RMSDs, which we relate to the additional uncertainty given by the background emission characterization for airborne LWP retrieval.

The single-channel retrieval, different in retrieval design and training data set compared to the multi-channel retrieval, is characterized by lower uncertainties and detection limit. Higher liquid water emissions in the 89 GHz channel compared to the 31.4 GHz channel used in the multi-frequency HATPRO retrieval leads to a higher sensitivity of the retrieval to smaller clouds with less liquid. This retrieval, however, strongly depends on the knowledge of IWV conditions and accurate clear-sky flagging. Relative LWP uncertainty for LWP between 10 and 100 gm-2 increases by few percentage points if closest clear-sky and cloudy IWV differ by 1 to 2 kgm-2.

The availability of both multi- and single-channel retrievals aboard Meteor allows for a direct comparison of the two different retrieval approaches. A direct intercomparison in confidently liquid cloudy conditions reveals a LWP RMSD of 31.3 gm-2, a bias of -5.9 gm-2 (LIMHAT LWP higher than LIMRAD), and a high correlation of 0.92 (not shown). As LWP varies strongly in time and space and sensors fields of view are different, comparing the liquid cloudy LWP distributions through percentiles is a preferable method. The percentiles of the liquid cloudy LWP distribution of LIMRAD and LIMHAT, illustrated in Fig. , show that LIMRAD-retrieved LWP is skewed to lower values compared to LIMHAT's LWP. The different clear-sky correction approaches in the two retrievals constitute themselves in the fact that LIMRAD LWP approaches zero when LIMHAT LWP ranges between 5 and 17 gm-2. Above this range, the negative bias towards LIMRAD, showing less LWP than LIMHAT, moderately decreases towards higher LWP values.

Inter-platform evaluation of single-channel-retrieved LWP and clear-sky IWV is performed for the two periods of ship collocation (see Sect. ). The ships's visiting times at BCO could not be used as BCOHAT was not operational at those times. LWPs obtained from LIMHAT, LIMRAD, and MSMRAD are intercompared in a statistical way rather than directly as clouds might overpass with an unknown time shift. Both LIMRAD and MSMRAD exhibit larger LWPs (median of 50.4 and 42.6 gm-2, respectively) than LIMHAT (28.4 gm-2), confirming the percentile-based comparison of LIMRAD and LIMHAT. Cloudy profiles of above 100 gm-2 were mostly seen by MSMRAD, which, however, might be related to single events that did not overpass Meteor given the small sample size. Additionally, both radars were operated with different chirp table settings, leading to different sensitivity to boundary layer clouds which, in turn, might affect the performance of the cloud mask. Given the uncertainties of each LWP product identified in the previous section and the uncertainty related to the applied cloud mask, the distributions match well and are suitable for site intercomparison. Clear-sky IWV, less variable in space and time, is compared point to point and exhibits a RMSD of 1.0 and a bias of -0.2 kgm-2 (MSMRAD slightly drier). Both single-channel retrievals agree within the expected uncertainties.

Assuming that BCO and Meteor were exposed to similar conditions and given the fact that multi-channel derived LWP is generally more reliable, we conclude that the LIMHAT measurements should be used as truth for the Meteor site compared to the single-channel LIMRAD LWP. While LIMRAD single-channel LWP is biased by -5.9 gm-2 compared to the multi-channel LWP estimates, presumably due to a higher sensitivity towards smaller clouds, this bias cannot be directly translated to MSMRAD LWP due to absolute calibration differences of the two cloud radars. Clear-sky TB data are affected by a RMSD of 2.6 K and a bias of 5.6 K (Merian warmer), but a correlation of 0.66 is low due to temporal spatial mismatch. Given this TB bias and assuming all other instrument characteristics being the same between LIMRAD and MSMRAD, Merian single-channel LWPs might in reality be lower. An extended analysis can help to quantify this bias, e.g., by comparing similar-looking clouds as seen in the active cloud radar part.

The obtained temperature RMSD and bias are illustrated in Fig. . At both sites, zenith mode RMSD increases throughout the moist layer from less than 0.5 K below the lifting condensation level (LCL) to 1.5 K at the trade inversion around 2 km. As zenith, HATPRO measurements generally contain 2 degrees of freedom (independent pieces of information) for retrieving the temperature profile ; the retrieval information content is too low to resolve the trade temperature inversion. Rather, the MWR profiles smooth the inversion, resulting in on average warmer MWR conditions at the base of the inversion and colder conditions at inversion top, similar to conditions found in the Arctic . Temperature information content is highest below 4 km , which makes the MWR insensitive to the conditions in the middle troposphere as seen by further increasing RMSD.

Elevation scans have been shown to improve the derived temperature profile in the lowest kilometer of the boundary layer . As illustrated in Fig. , however, the BCOHAT and LIMHAT scans increase RMSD and bias in the layers below 1 km. We suspect that the GAIA sounding data set used for training is impacted by the island surface, leading to warmer temperatures in the moist layer compared to the zenith column at BCO or over the ocean. Typically, when trade winds prevail, radiosoundings launched at GAIA or BCO drift westwards over the island when ascending through the sub-cloud layer. Paired with small seasonal temperature variations in the tropics and, thus, little variability in the temperature training data set, this systematic training error translates into warm temperature biases of the retrieved temperature profiles compared to the launched radiosoundings. A physical, iterative retrieval approach such as Optimal Estimation would help to constrain the covariances of the prior temperature profile data set. Elevation scans aboard Meteor, in particular the low-elevation-angle measurements, are additionally affected by ship motion as LIMHAT was not stabilized. The functionality of the HATPRO-attached weather station might additionally impact the quality of the temperature retrieval.

Comparing radiosoundings and MWR yields to the RMSD and bias illustrated in Fig. a. At both sites, the RMSD from ground to LCL is 1.3 gm-3, and it increases to 2.5 gm-3 in the area of the hydrolapse associated with the trade inversion. The tendencies of the bias can be further understood when analyzing the mean profiles as illustrated in Fig. b. From ground towards hydrolapse, MWR underestimates the humidity, resulting in a negative bias of -1 gm-3. Throughout the hydrolapse, MWR and sounding profiles converge, which is due to the smoothing of the MWR profile. Depending on the strength of the hydrolapse, MWR overestimates the humidity in the dry layer, balancing the overall profile to match overall IWV conditions. Above the hydrolapse in the free troposphere, dry conditions prevail, and MWR is not sensitive to elevated moist layers. While the MWR covers the variability of moist layer water vapor well as seen by similar standard deviations of sounding and MWR profile, it does not resolve the variability in the hydrolapse or free troposphere. The overall negative bias in the absolute humidity profile translates into a dry bias in the IWV estimate (compared to the radiosoundings), which confirms the findings in Sect. .