The HATPRO-G2 microwave radiometer and the Cimel sun photometer used in this study were located at the Romanian Atmospheric 3D Observatory (RADO, 44.82∘ N, 26.82∘ E, 93 m a.s.l.), part of the National Institute of Research and Development for Optoelectronics (INOE2000). The observatory is located in the city of Măgurele, Ilfov, at the central part of the Romanian plain, approximately 10 km southwest of Bucharest, the capital city of Romania, and is surrounded by research facilities, residence buildings and a small forest. The central Romanian plain has a temperate climate influenced by the western circulation, the east European anticyclone, the Mediterranean cyclones and the tropical advection . The relative humidity at Măgurele, as calculated from observations from the RADO weather station between 2007 and 2016, has high values (>80 %) during November–February and low values (<60 %) between May and September (Fig. ). Since the Cimel sun photometer performs measurements only when the solar disk is free of clouds, two critical parameters for the availability of the Cimel sun photometer data are the sunshine duration and the cloud fraction. For the calculation of the climatology of the sunshine duration the Surface Radiation Data Set – Heliosat (SARAH) – Edition 2 of the EUMETSAT's Satellite Application Facility on Climate Monitoring (CM SAF) was used. The sunshine duration (SDU) product is the daily sunshine duration per day at which direct normal irradiance (DNI) exceeds the WMO threshold of 120 W m-2. SDU is derived by the ratio of sunny slots to all slots during daylight multiplied by the length of day. The length of day is calculated depending on the date, longitude and latitude. The length of day is restricted by a threshold of the solar elevation angle (SEA) of 2.5∘. The SDU product is provided on a regular latitude–longitude grid with a spatial resolution of 0.05∘ × 0.05∘. In this study, for the calculation of the climatological sunshine duration, the daily SDU at the closest pixel over Măgurele during the period 2005–2015 was used. A full description of the SDU product can be found at . For the calculation of the cloud fraction climatology, the CM SAF cloud property dataset using SEVIRI – edition 2 (CLAAS-2; ) was used. The cloud fractional cover (CFC) is defined as the fraction of cloudy pixels per grid cell compared to the total number of analysed pixels in the grid cell and is expressed as a percentage. In this study the daytime CFC during the period 2005–2015 was used. The daily CFC product is provided on a regular latitude–longitude grid with a spatial resolution of 0.05∘ × 0.05∘. A full description and evaluation of the CFC product is given in . High cloud coverage (>70 %) affects the RADO site from November to February (Fig. ), while the lowest cloud fraction (<40 %) is during July–August. The rest of the months the cloud fraction ranges between 50 % and 60 %. The high percentage of clouds, in combination with the small sunshine duration (Fig. ), during late autumn and winter affects the availability of the Cimel sun photometer data during these months. Thus calculations of multi-year annual mean TPW values from Cimel sun photometer observations are biased from the highest number of data points during summer. The sunshine duration exhibits a clearly annual cycle with a minimum during winter and maximum during summer and ranges from ∼2.3 h during January (minimum) to up to more than 10 h during July (maximum) (Fig. ).

A Cimel Electronique 318A sun photometer (serial number 359) was installed at the RADO facilities in July 2007 and was operated until May 2016, when it was reallocated to Poland. As a replacement a Cimel lunar photometer has been operating since 2016, but data from this instrument have not been used in this study due to the limited availability of level 2 data. The Cimel sun photometer is the standard instrument of AERONET used for the study of the aerosol total column load. It performs spectral measurements of the direct sun irradiance and sky radiance at six discrete wavelengths using interference filters. The filters are centred at the wavelengths of 340, 380, 440, 500, 675, 870 and 1020 nm. An additional channel at 935 nm is used for the retrieval of the TPW. The instrument is calibrated almost annually following the procedures and the guidelines of AERONET. TPW is calculated based on a modified expression of the Beer–Bouguer–Lambert law. Since provide a full description of the TPW retrieval algorithm (see Sect. 2 of that paper), in this section just the major differences between V2 and V3 and some other factors that may influence the TPW retrieval are discussed. For the computation of TPW a necessary preliminary step is the subtraction of the AOD and Rayleigh optical depths from the total optical depth at 935 nm. Since AOD is not calculated directly for the 935 nm channel due to the strong effect of water vapour, the AOD at 870 nm is extrapolated at the 935 nm using the Ångström exponent (AE) at 440–870 nm. The main differences in the computation of TPW in V3 are that the new algorithm accounts for an updated continuum look-up table , using total internal partition sums and using the extraterrestrial spectral solar irradiance from . In this study all available data from July 2007 to May 2016 for level 2 from versions 2 and 3 of AERONET algorithms were used. Level 2 data are screened for clouds, quality controlled, and pre-field and post-field calibrations are applied. The newest released version 3 incorporates improvements for the direct sun measurements (1) related to the screening of clouds, (2) the automated data quality assurance, (3) inclusion of data with higher air masses (i.e. from 1 to 7, in contrast with V2 that ranges from 1 to 5) and (4) implementation of spectral temperature corrections based on laboratory measurements (i.e. unlike version 2 that was based on the manufacture specifications). Details about all the improvements implemented in V3 of AERONET can be found at . The AERONET TPW measurement uncertainty is estimated to be <10 % , which is consistent with the one-sigma uncertainty for AERONET V2 provided by of 7 %–9 % after evaluating the TPW from AERONET at the U.S. Department of Energy Atmospheric Radiation Measurement Program (ARM) sites against microwave radiometers, GPS and radiosondes.

The HATPRO-G2 microwave radiometer used in this study was produced by Radiometer Physics GmbH. It is a passive instrument working in the microwave regime. It consists of two working bands at 22–31 and 51–58 GHz, each with seven channels. The relevant receiving optics, the ambient load, the internal scanning mechanism, the electronics and the data acquisition system of the radiometer are described in . For humidity profiling only the first band is used. The vertical resolution for profiling is variable, ranging from 200 below 2000 to 800 m for altitudes higher than 5000 m. Water vapour emission dominates the signal in the 23.8 GHz channel, which is on the wing of the 22.2 GHz water vapour absorption line, whereas liquid water emission constitutes the primary portion of the signal at 31.4 GHz . From these two observations, both integrated water vapour (IWV) and liquid water path (LWP) can be retrieved. The retrievals are performed in the zenith direction. In this study, the IWV was used, which presents, according to the manufacturer (RPG-HATPRO-G4 series microwave radiometers for continuous atmospheric profiling, available at https://www.radiometer-physics.de/download/PDF/Radiometers/HATPRO/RPG_MWR_PRO_TN.pdf, last access: 13 July 2018), an accuracy of ±0.2 kg m-2 RMS and noise of 0.05 kg m-2. Considering the density of liquid water, the IWV expressed in kilograms per square metre is equivalent with the TPW expressed in millimetres of liquid water . In this study measurements are performed each 2 s. To ensure the high quality of measurements, the instrument is absolutely calibrated with liquid nitrogen every 6 months following the instructions of the manufacturer. All the available data between 16 December 2009 and 31 December 2017 were used. The internal data quality has three options for filtering level 2 data (retrieved atmospheric data). The “Flag Data Quality” (level 2) option does not filter the level 2 data according to the quality level but flags each data sample in the rain flag byte. With the option “Remove Medium/Low Q.”, medium- and low-quality samples are not transmitted by the radiometer. In this case, the sample sent to the personal computer that controls the instrument is the repeated latest high-quality sample. The filter “Remove Low Quality” only removes the worst-quality data and transmits high- and medium-quality data. In the present study, the first option was used for the creation of the level 2 data; thus only data that have been flagged as rain from the internal sensor of the instrument have been removed. In addition, all days with data have been visually inspected for identification of instrumental malfunctions, which can include periods when there are no changes in the TPW values due to bad transmission of data or periods with low-quality data (i.e. when the TPW remained high after rain, until returning to its previous levels after some time).

The radiosonde measurements were obtained from the sounding database maintained by the University of Wyoming (http://weather.uwyo.edu/upperair/sounding.html, last access: 13 July 2018). Between July 2007 and December 2017, 3760 radiosonde measurements for 00:00 UTC and 3759 for 12:00 UTC were available from the Bucharest site, situated at approximately 30 km northeast from the RADO facility and operated by the Romanian National Meteorological Administration. The radiosondes used during the study period were of the Vaisala RS92 type. For this type of radiosondes, showed that the accuracy of the humidity sensor during daytime depends on the calibration error and the dry bias due to the solar heating effect and during the night-time just from the calibration error. The overall uncertainty of TPW from radiosonde measurements has been estimated to be ±5 % . TPW over the entire sounding was calculated as 1TPW=1ρg∫p1p2xdp, where xp is the water vapour mixing ratio at the pressure level p, ρ is the density of water and g is the acceleration of gravity.

For the computation of the daily mean values of TPW all available measurements that qualify the quality criteria were used. A preliminary step was the averaging of TPW from the microwave (MWV) radiometer into 1 min intervals. Table  gives an overview of the total number of observations, along with their total number of corresponding days that have been analysed for each instrument and for the different versions of the AERONET algorithms in order to compute the daily averages. Due to the different schedule of each instrument and the gaps in each database, the computed averages cannot be directly compared between them. For a direct comparison we extracted the common measurements between V2 and V3 and they were averaged for ±20 min around the launch time of the noon radiosonde. The same averaging was applied to the MWV radiometer data, so as to extract a dataset of simultaneous or nearly simultaneous measurements from all instruments. Since the exact hour of the radiosonde launch is not explicitly known, this 40 min interval has been selected in order to ensure that the instruments detect the same air masses and to limit the atmospheric variability that takes place on timescales larger than 1 h . If the GPS information of the radiosondes is available, a further improvement in the coincidence criteria would be to average the Cimel sun photometer data for ±20 min since the time the balloon reaches the altitude of the 4 km, following . However, in our case access to the raw data is not available; thus the averaging was performed ±20 min around the launch time. For the comparison of the MWV radiometer and Cimel sun photometer data with the radiosondes, the same coincidence criteria as described above were used. The comparison of the two different algorithms of AERONET is based just on their common measurements. This way the comparison provides insight into the TPW calculation differences between the two algorithm versions rather than impacts due to cloud screening and instrument quality controls. For the comparison of the Cimel sun photometer data with the MWV radiometer, the exact time matched measurements were selected. For the evaluation of the MWV radiometer and Cimel sun photometer data, the radiosonde TPW was used as the reference measurements because they are considered more representative of the actual atmospheric conditions. However, since the radiosonde site is at a distance of ∼30 km from the RADO facilities, there is the possibility that the different instruments detect air masses with different characteristics, especially when the radiosondes are affected by southwest winds. Thus, the calculated uncertainty expressed as the 1σ of the mean difference among the different datasets is expected to be little overestimated when compared to the radiosondes. The absolute and relative differences between two sets of measurements were defined as 2X-Xref and 3100⋅(X-Xref)Xref, respectively, where (Xref) is the reference measurement (i.e. the radiosonde measurement, except for the comparison between the Cimel sun photometer and the microwave radiometer).

The times series of the daily mean values for the TPW from the different instruments employed in this study are shown in Fig. . In general, the radiosonde measurements are available twice per day (i.e. 00:00 and 12:00 UTC). The Cimel sun photometer measurements are restricted only during daytime and under conditions that require the solar disc to be clear of clouds, while the microwave radiometer performs measurements during daytime and night-time under all weather conditions. Although there are differences in the measurement schedule, all three instruments depict the same annual cycle, demonstrating their capability of performing long-term measurements for climatological applications (Fig. ). The gaps in Cimel sun photometer time series are due to the calibration of the instrument, which requires the reallocation of the instrument. Data gaps of the microwave radiometer are due to malfunction of the instrument or controlling personal computer (usually solved with a restart after a maximum of couple of days) or due to the relocation of the instrument during different measurement campaigns (data not included in this study). Furthermore, in the beginning of 2016 the instrument was sent to the manufacturer for testing and replacement of several components.

The observed differences in the mean values calculated from all instruments (Table ) can be mostly attributed to the different operating period of each instrument and their different sampling rates. However, even though the overall mean from Cimel sun photometer measurements is not significantly different from the radiosondes and the microwave radiometer estimates, Cimel sun photometer measurements are actually biased towards the higher TPW values observed during the summer. Since the cloud fraction during the winter months at Măgurele is pretty high, more than 70 % from November to January (Fig. ) when TPW also attains its minimum values (Fig. ), the number of Cimel sun photometer observations is substantially reduced, leading to the inclusion of a reduced number of low-TPW days in the Cimel sun photometer dataset. This observed summer (wet) bias is partly compensated for by the inherent Cimel sun photometer dry bias e.g. due to restrictions of measurements when the solar disc is cloud free and thus the overall TPW mean from the Cimel sun photometer is similar to the other methods (Table ). This dry bias for the mid-latitudes is more pronounced during winter and can range from 25 % to 50 %, while in summer it ranges from 5 % to 25 % . The clear-sky monthly bias can be clearly seen in the mean monthly values of TPW (Table ), for which the Cimel sun photometer measurements during January can be lower by ∼25 % compared to the radiosondes while the summer mean monthly values are lower by only a few percent (e.g. ∼4 % for August) (Fig. a). Such behaviour is not observed for the MWV radiometer, with the differences in their mean monthly values ranging within ±10 % for all months (Fig. b). The minimum values daily values can be as low as 2 mm, while the maximum values exceed 44 mm (Table ). The peak-to-peak range during the year (i.e. from minimum to maximum) can be up to 20 mm.

The annual cycle of the TPW as depicted by all three instruments has a minimum during winter months (DJF) and a maximum during summer months (JJA) (Fig. ). Higher air temperature during the summer implies a larger capacity to store water vapour without saturation . The small differences in the monthly median values for all instruments are due to their different sampling rates. For example, the increased number of outliers in the radiosonde box plots, compared to the other instruments, can be attributed to the limited number of measurements (i.e. a maximum of two per day). Thus, some high or low values are not smoothed by averaging all measurements during the day (Fig. a). In any case, the main aim of the analysis presented here is to show that the annual cycle of TPW can be depicted fairly well by all instruments and demonstrate their capabilities for long-term monitoring for climatological applications. A direct comparison of the daily values from each instrument is not valid due to the very different sampling rates and the diurnal variation in TPW as shown in Fig. . An overview of the statistical values based on all available measurements for all three instruments is shown in Table . A dataset was constructed, as described in Sect. 2.4, containing the common measurements and thus allowing for a direct comparison among the three instruments. This dataset consists of a total of 234 days during the measurement period, which is limited by the Cimel sun photometer observations during conditions in which the solar disc was free of clouds. For this reason, the comparison of the different instruments is not affected by the clear-sky dry bias. An overview of the long-term averages of the common measurements from all instruments can be seen in Table . The MWV radiometer has the higher mean TPW (18.57 mm), followed by the radiosondes (17.96 mm), Cimel V2 (17.80 mm) and finally Cimel V3 (17.65 mm). Although this dataset consists of nearly time-matched measurements, the small differences in the long-term averages may occur from differences in the geometry of the measurements and subsequently the sounding of air masses with different characteristics. For example, the Cimel sun photometer measures the direct sunlight and can track the sun between clouds, while the MWV radiometer measures the zenith sky radiance and it may not be completely cloud free for the same sky. The radiosondes are also launched from a different area, which could possibly track different air masses. The mean monthly TPW values (Table ) appear to have very good agreement (within ±5 %) among the different instruments when using common data periods (Fig. b). The Cimel sun photometer clear-sky dry bias that had been observed, especially during the winter months in the long-term averages when computed from all measurements (Table ), has been cancelled out, as can be clearly seen in Fig. b.

As mentioned previously, the temporal resolution of the microwave radiometer on the order of a few seconds in combination with its capacity to operate under all weather conditions allow the detection of the TPW diurnal variations. In addition, under clear-sky conditions, the Cimel sun photometer performs measurements at about every 15 min. To verify if both instruments depict the same daily cycle, the diurnal variability for 6 selected days was examined. The days were selected to meet the following conditions: the Cimel sun photometer measurements cover most of the day and in particular for high solar zenith angles (SZAs, >70∘), no discontinuation due to clouds in Cimel sun photometer measurements was observed from sunrise to sunset and the measurements cover all seasons. The Cimel instruments and MWV radiometer depict the same diurnal variation during daytime (Fig. ), with some small differences in their absolute values that are further investigated in the following sections. For some of the selected days (i.e. 8 June 2012, 26 October 2013) there are differences in the diurnal variation during the early morning or late afternoon hours, which are most likely artefacts associated with direct sun measurements at high air masses (e.g. SZA > 70∘). These artefacts are due to Cimel clock deviations that result in some minor deviation in the optical air mass calculation and thus slightly impact AOD but within uncertainty expectations (see Sect. 3.3.1 of ).

To account for spatial and temporal differences between the radiosonde and the microwave radiometer, all the microwave radiometer data were averaged over an interval of 40 min centred on the radiosonde launching time. A total number of 2820 common measurements, out of which 1416 during daytime (i.e. at 12:00 UTC) and 1404 during night-time (i.e. at 00:00 UTC), were extracted for the comparison. The relative difference between the two datasets is in general within ±25 % (Fig. ). The MWV radiometer slightly overestimates TPW with the overall difference from the radiosondes to be 1.82±9.61 % (0.17±1.66 mm). This overestimation is more evident during daytime (i.e. 3.12±9.93 % or 0.35±1.71 mm) due to the radiation dry bias effect that affects the radiosondes e.g., which is more pronounced for TPW values less than 10 mm (Fig. b). During night-time the differences are almost negligible (i.e. -0.50±9.10 % or -0.01±1.57 mm).

The two datasets are highly correlated (Fig. a; R2=0.97), with the majority of the points over the y=x line. However, for the higher values of TPW (i.e. TPW > 30 mm) an increased scatter of the data is observed, without being significantly high. The histogram of the relative differences between the two instruments, which peaks at about 1 % (Fig. b), does not follow a normal distribution, as indicated by the Shapiro–Wilk test for normality (p value < 2.2e–16). About 96 % of the data are within ±20 %, while ∼78 % lie in the range of ±10 %. The difference between the two datasets has a small dependence from the TPW amount of -0.169 % mm-1 (Fig. a). This dependence is more evident for the daytime measurements (i.e. for radiosondes launched at 12:00 UTC; Fig. b), while for the night-time measurements the dependence is almost negligible (i.e. -0.092±0.052 % mm-1; Fig. c). The best agreement between the two datasets is achieved for TPW values ranging between 15 and 35 mm. The increased difference for TPW values higher than 40 mm cannot be fully evaluated due to the very small number of observations (i.e. just 19 measurements).

To assess the differences of the TPW derived from the newly released version 3 from AERONET and the previous version 2, only their common measurements were used. The difference in the number of observations between the two versions (see Table ) arises from the fact that they have different quality control and cloud screening procedures . A total of 27 707 common observations between the two versions were extracted for comparison. In general, the differences between the two versions are small, ranging within ±2 % and rarely exceeding 5 %, with V2 having higher values than V3 (Fig. a and b). The overall difference between the two datasets for the period 2007–2016 is 0.60±0.91 % (0.08±0.14 mm). The differences at the AOD at 870 nm between the two different algorithm versions (Fig. c) are generally pretty low and rarely exceed the ±0.01 AOD units. The cyclic nature of the AOD differences (Fig. c) suggests the variation in the AOD with temperature for version 2. The V2 data are not temperature-corrected for the 870 nm filter and this produces a difference in AOD between temperature-corrected (V3) and not-corrected (V2) data due to this specific filter before 2009. The 870 nm filter was changed in 2009 in this specific instrument and its dependence on temperature was a magnitude lower than the initial filter. As a result, the filter used in the instrument from 2009 and onward shows less deviation from V2 since the temperature correction needed for the filter is minimal. This is a clear example of how implementation of temperature correction in version 3 significantly improved the AOD and TPW, before 2009.

To further evaluate the differences between TPW from the two different versions of the AERONET algorithm, a series of factors that could affect the measurements (i.e. the total amount of TPW, the SZA, the sensor temperature and the differences at the AOD at 870 nm) were examined. No significant dependence was found with SZA when comparing the two versions. The relative difference between V2 and V3 show a dependence on TPW (Fig. a). The biggest differences (i.e. ∼2.5 %) are observed for TPW values lower than 10 mm, while the agreement between the two datasets improves with increased TPW values. However, the decrease in the relative difference of TPW between V2 and V3 is due to the different treatment of the temperature correction in the versions. As shown in Fig.  the lowest TPW values appear during wintertime, when the temperature is low as well. Corresponding to these low temperature values the differences between V2 and V3 shows a mean maximum value of ∼2.5 % (Fig. b). A very pronounced dependence is also seen by the temperature of the internal sensor of the instrument. This dependence is due to the different temperature coefficients in the two versions of the retrieval algorithm. For V2 the temperature coefficients are based on the manufacturer specifications, while in V3 the temperature characterisation is based on laboratory measurements during the calibration of the instrument. The highest positive differences, on the order of ∼5 %, appear for low temperatures (<10 ∘C). For the whole range of temperatures that are recorded in the instrument (i.e. ∼50 ∘C) a total difference of up to 5 % is observed (Fig. b).

To have a better overview about how the differences between the two versions affect the agreement with the other instruments, the evaluation of Cimel sun photometer measurements with radiosondes and the microwave radiometer was based on the common dataset between the two different algorithm versions (Sect. ). Since this Cimel model is not capable of night-time measurements, the comparison is limited to daytime measurements only (i.e. radiosondes launched at 12:00 UTC). To account for spatial and temporal differences, the same procedure with the one described for the comparison between microwave radiometer and radiosondes was used (i.e. averaging all Cimel sun photometer points over an interval of 40 min centred on the radiosonde launching time). Thus, a total of 682 common measurements were identified.

The differences between the Cimel sun photometer and radiosondes range within ±20 % (Fig. ), while the overall mean difference is -1.95±10.97 % (or -0.39±2.1 mm) and -2.74±10.56 % (or -0.50±2.05 mm), for V2 and V3. These results are in agreement with previous studies that showed that AERONET sun photometers generally underestimate TPW in comparison with other instruments . Version 3 shows an increased underestimation of TPW in comparison with the radiosondes; however the standard deviation is slightly better than in the previous version (Fig. b).

The TPW from both versions is highly correlated with the TPW from the radiosondes (i.e. R2 is 0.95 for both Cimel V2 and Cimel V3; Fig. a and c), with the slope of the least-square regression line being very close to unity. The histogram of the relative differences between the two datasets has a very small kurtosis towards negative values, for both Cimel V2 and Cimel V3 (Fig. b and d). According to the Shapiro–Wilk test for normality , it does not follow a Gaussian distribution (p value = 0.01427 and 0.004603, for V2 and V3, respectively). For Cimel V2 about 65 % of the differences are within ±10 %, while ∼93 % are within ±20 %. For Cimel V3 the respective numbers are 67 % and 93 %. The low number of the coincidence measurements, and their big scatter among different SZAs, TPWs and temperatures of the sensor, does not allow a further evaluation of the influences from these factors.

The comparison between the Cimel sun photometer and the microwave radiometer is based on their coincident measurements, with the microwave radiometer observations averaged over a 1 min interval. This common dataset consists of 8505 observations for the period December 2009–May 2016. The differences between the TPW from both versions of AERONET algorithms are in general within ±10 % (Fig. ). The Cimel sun photometer underestimates the TPW by 2.75±5.85 % (or 0.70±1.22 mm) and 3.57±5.54 % (or 0.81±1.17 mm), for V2 and V3, respectively. The comparison of the Cimel sun photometer with the MWV reveals a lower overall uncertainty of ∼6 % estimated as the one sigma of the mean difference, compared to the one (∼10 %) that was calculated from the comparison of the Cimel sun photometer with the radiosondes. This lower uncertainty can be attributed to the collocation of the Cimel sun photometer and MWV and subsequently the sounding of the same air masses from both instruments. The distance between the RADO site and the radiosonde launching site increases the estimated uncertainty of the retrieved TPW from Cimel sun photometer; however it still remains within the limits that have been estimated by other studies in the past e.g..

The TPW values from the Cimel sun photometer (both Cimel V2 and Cimel V3) and the microwave radiometer are highly correlated (Fig. a and c; R2=0.99). Taking into consideration that the microwave radiometer and the Cimel sun photometer have the same diurnal variations (Sect. ), a very high correlation of the two datasets was expected. For higher values of TPW there is a deviation from the identity line.

The histogram of the relative differences between the two datasets has a very small flattening towards negative values, for both Cimel V2 and Cimel V3 (Fig. b and d). According to the Shapiro–Wilk test for normality (p value < 2.2e–16 for both Cimel V2 and Cimel V3) it does not follow a Gaussian distribution. For Cimel V2 about 88 % of the differences lie within ±10 %, while differences for almost the entire dataset are within ±20 % (>99 %). For Cimel V3 the respective values are similar. These results show a very good agreement between the two different methods for the retrieval of TPW.

The difference of the TPW between the Cimel sun photometer and the microwave radiometer does not show a pronounced dependence on the SZA (Fig. a and b), for both versions of AERONET algorithms. However, there is an increased scatter for SZAs higher than 70∘. This is due to the clock shift effect (see Sect. 3.2) that can affect the direct sun measurements from the Cimel sun photometer at high air masses, resulting in an increased uncertainty on the retrieved TPW.

The difference of the TPW between the Cimel sun photometer and MWV radiometer has a small dependence on the total amount of TPW of -1.97 % per 10 mm for Cimel V2 and -1.38 % per 10 mm for Cimel V3 (Fig. c and d). The lower dependence of TPW from Cimel V3 on the total amount of TPW in comparison with Cimel V2 is an indication that the changes applied in the newer version of the algorithm are more correct. Both versions show a higher variability for TPW values lower than 10 mm due to the increased uncertainty of both instruments for dry conditions. However, this variability is based on a relatively low number of observations and is highly affected by some outliers (i.e. differences >20 %) observed for extremely low TPW values (i.e. 1.5–2 mm). When the TPW values lower than 10 mm are excluded from the analysis, the dependence of the difference between the Cimel sun photometer and MWV radiometer becomes -1.69 % per 10 mm and -1.19 % per 10 mm, for V2 and V3, respectively. In addition the very low variability for TPW values higher than 40 mm cannot be evaluated because they are based on a very limited number of observations (i.e. six observations).

The new laboratory-based temperature coefficients for the sun photometer filters improve the quality of the retrieved TPW from the Cimel sun photometer, as can be depicted from the comparison with the MWV (Fig. f). The dependence of the difference between Cimel V3 and the MWV from the temperature recorded in the sensor of the Cimel sun photometer is substantially improved in comparison with the one of Cimel V2 (the order of -0.61 % per 10 ∘C and -1.07 % per 10 ∘C for Cimel V3 and Cimel V2, respectively; Fig. e and f). Thus the corrections from the application of the new temperature coefficients are important, since they significantly improve the quality of the retrieved TPW for all the operating temperatures.