Completeness of radiosonde humidity observations based on the IGRA

Radiosonde measurements from the 1930s to present give unique information on the distribution and variability of water vapor in the troposphere. The sounding data compiled in the Integrated Global Radiosonde Archive (IGRA) Version 2 10 (released by the NOAA's National Centers for Environmental Information) are examined here until the end of 2016, aiming to describe the completeness of humidity observations from radiosondes in different times and locations. The IGRA stations reporting radiosonde data in at least 5 % of the annual soundings for at least one year are evaluated according to specified completeness parameters for every year in their period of record. The selection of source data essentially removes pilot-balloon sites, retaining a set of 1723 stations (designated ‘IGRA-RS’), including 1300 WMO upper-air stations, of which 178 belong 15 to the current GUAN network. Completeness of humidity observations (either relative humidity or dewpoint-depression) for a radiosonde station and a full year is defined by: the number of humidity soundings; the fraction of days having humidity data; the mean vertical resolution of humidity data; the mean atmospheric pressure and altitude at the highest measuring level; and the maximum number of consecutive days without humidity data. The completeness of the observations qualified for calculating precipitable water vapor – i.e., having adequate vertical sampling between the surface and 500 hPa – is particularly 20 studied. Individual soundings are described by the (vertically averaged) vertical resolution and the pressure level and altitude of the top of humidity measurements. For illustration, the study presents a global picture of the completeness of radiosonde humidity observations over the years, including their latitudinal coverage. This overview shows that the number of radiosonde stations having a long enough record length for studies on the climatic variability and trends of humidity-related quantities depends critically on the temporal continuity, regularity and vertical sampling of the humidity time-series. It is hoped that the 25 derived metadata will help climate and environmental scientists to find the most appropriate radiosonde data for humidity studies by selecting upper-air stations, observing years or individual soundings according to various completeness criteria – even if differences in instrumentation and observing practices require extra attention. A dataset is presented for that purpose, consisting of two main sub-sets: 1) humidity metadata for each of the IGRA-RS stations and year within the period of record (yearly metadata); and 2) humidity metadata for individual observations from the same stations (ascent metadata). These are 30 complemented by 3) a list of the stations represented in the dataset, along with the observing periods for humidity and the corresponding counts of observations. The dataset is to be updated on a two-year basis, starting in 2019, and is available at https://doi.org/10.5281/zenodo.1332686. Earth Syst. Sci. Data Discuss., https://doi.org/10.5194/essd-2018-95 O pe n A cc es s Earth System Science Data D icu ssio n s Manuscript under review for journal Earth Syst. Sci. Data Discussion started: 15 October 2018 c © Author(s) 2018. CC BY 4.0 License.

Satellite-based remote sensing of atmospheric water vapor is part of modern weather forecasting and climate monitoring (Kley et al., 2000;Andersson et al., 2007). In the present state-of-the-art, some satellite retrievals of moisture-related quantities are used as a reference to compare humidity measurements from different radiosonde types, aiming to monitor radiosonde stations and improve satellite calibration (Kuo et al., 2005;John and Buehler, 2005;Sun et al., 2010;Moradi et al. 2013). However, limb-sounding satellite techniques with high vertical resolution (albeit very course in the horizontal), using GPS radio-5 occultation, are a recent acquisition, of main interest to access water vapor in the upper troposphere and lower stratosphere so far (Kishore et al., 2010;Shangguan et al., 2016;Rieckh et al., 2018;Vergados et al., 2018). Reanalysis outputs based on past radiosonde data, assimilating also satellite data when available, offer multiple-level, global gridded, synoptic-scale moisture fields up to 4 times daily from a beginning year (e.g.: 1948 in NCEP/NCAR Reanalysis 1; 1979 in NCEP/NCAR Reanalysis 2 and ECWMF's ERA Interim) to present timeeven though radiosonde observations are scarce over the ocean, unevenly 10 spaced over land, and taken normally twice a day, with significant differences in vertical coverage. Naturally, since air moisture is highly variable in time and space, humidity data from different reanalysis models show discrepancies and can differ significantly from the collocated radiosonde data (e.g., Noh et al., 2016). Therefore, the radiosonde archives represent the primary source of information on the short and long-term distribution of moisture in the troposphere, despite various data inhomogeneities. Namely, geographical-temporal sampling differences (Wallis, 1998), uncertainties related to observation 15 time and balloon drift (Kitchen, 1989;McGrath et al., 2006;Seidel et al., 2011;Laroche and Sarrazin, 2013), differences in vertical coverage and data gaps related to reporting practices of humidity (Dai et al. 2011 and references therein) and differences in humidity data accuracywhich depend on humidity sensors and vary with measured conditions (WMO, 1995;Nash, 2002;Sappuci et al., 2005;Moradi et al., 2013;Dirksen et al., 2014).
The growth interest in climate change motivated a renewed attention to radiosonde data since the 1990s. Soon it was 20 realized that the usefulness of radiosonde data archives to investigate climate trends relies on homogenization procedures to overcome biases and sudden shifts arising from instrument changes, reporting practices and sampling differences Schwartz and Doswell, 1991;Parker and Cox, 1995;Luers and Eskridge, 1998;Lanzante et al. 2003).
Subsequent climate studies based on radiosondes have mostly focused on the detection of climate change in temperature trends (Free and Seidel, 2005;Thorne et al., 2005;Haimberger et al. 2008). Concerning humidity, radiosonde-based climatic studies 25 are for now confined to the lower and middle troposphere, because of the large uncertainty of measurements and biases in the upper troposphere and lower stratosphere (Elliot and Soden and Lanzante, 1996;Wang et al., 2003) and the extremely large relative biases and insufficient data in the lower stratosphere (Miloshevich et al., 2006;Nash et al. 2011).
Radiosonde data have been used for studying the long-term trends and the annual cycle of several humidity parameters (precipitable water vapor, specific humidity and relative humidity), at least in well-sampled regions of the globe and if data 30 inhomogeneities are removed (Elliot et al., 1991;Gaffen et al., 1992;Ross and Elliott, 1996;Ross and Elliot, 2001;McCarthy el al., 2009;Durre et al., 2009;Dai et al., 2011). On a rather different scale, radiosonde measurements with high vertical resolution near the ground are suitable for studying water evaporation over land and the structure of the planetary boundary layer, provided that the time-lag of humidity sensors as they move through a rapidly changing environment is accounted for (Sugita and Brutsaert, 1991;Connell and Miller, 1995;Seidel et al., 2010).
Since its creation in 2004, the Integrated Global Radiosonde Archive (IGRA) was meant to be the largest data set of up-to-date weather-balloon observations freely available, by collecting quality-controlled data from upper-air stations across all continents. The first version of IGRA (a successor of the Comprehensive Aerological Reference Data Set (CARDS; 5 Eskridge et al., 1995) contained practically data after 1945 (Durre et al., 2006). The IGRA Version 2 used in this paper, released by the NOAA's National Centers for Environmental Information  and recently described in Durre et al. (2018), has enhanced data coverage and extends back in time as early as 1905, although (for historical reasons) humidity data begin in 1930 with a sole location in Europe. The extension to observations prior to 1946 resulted mainly from the addition of data from the Comprehensive Historical Upper-Air Network (CHUAN), which is the most important collection of upper-air 10 observations taken before 1958 (Stickler et al., 2009). In view of the huge amount of data collected in IGRA (which is a combination of radiosonde and pilot-balloon observations) and the differences in the observing period, temporal regularity and continuity, vertical resolution and vertical extension of humidity data among different stations, finding the most suitable humidity-reporting stations (or humidity soundings from different stations) for a specific purpose can be difficult to put into practice . 15 Radiosonde humidity measurements involve the simultaneous measurements of pressure, temperature and relative humidity or dew-point depression. Therefore, except for horizontal wind, which is indirectly measured with the aid of a remote tracking device, humidity represents the most accomplished of the radiosonde observations. The purpose of this paper is to study the completeness of humidity observations collected in IGRA according to various needsnumber and latitudinal distribution of observing stations, fraction of observing days in a year, resolution and range of vertical levels, length and 20 continuity of the time-series, minimal sampling between the surface and the 500-hPa levelaiming to facilitate the use of radiosonde humidity data by atmospheric and environmental scientists. The task is two-fold: first, to elucidate the completeness of the humidity observations from IGRA for each year in global terms, including the latitudinal coverage of stations and the length of regular time-series; second, to provide metadata describing the completeness of humidity observations from each station. The observing periods without missing years in humidity data must be clarified. Latitudinal and regional differences 25 should be easily derived from the geographic coordinates of stations.
The remainder of this section is intended to clarify the term "completeness of observations" concerning the use of radiosonde data and to present an historical account of the main factors that limit the completeness of humidity observations from radiosondes: vertical levels available in radiosonde reports; missing observations associated with humidity sensor limitations. The next sections are organized as follows. Section 2 indicates the IGRA data set used in the study; selects the 30 IGRA stations reporting a minimum of radiosonde data (coined as 'IGRA-RS') by discharging stations with practically windonly data in their period of record; and explains the data analysis. Section 3 presents a global picture of the completeness of humidity observations over the years, as derived from the IGRA-RS stations. Section 4 provides the definition of the metadata parameters describing the completeness of humidity observations from each IGRA-RS stationeither as annual statistics or for individual soundingsand the format description of the corresponding data sets. The availability of the resulting dataset supplied by the study (Ferreira et al., 2018) is reported in Sect. 5. A summary of results and some suggestions for future application are given in Sect. 6.

Completeness of observations for radiosonde humidity studies 5
Data completeness in a data set refers to the extent to which the data set collects the expected elements: quality-assured data are not left-out; missing or invalid values are properly indicated. This is a basic requirement for data quality, and it is assured in IGRA. In a different way, and uncommon, data completeness may refer to whether the required data for a specific purpose are available or not. That meaning is not new in the field of meteorology. For instance, the WMO recommendations on "data completeness" (sic) required for calculating monthly means and climate normals from meteorological surface data refer to the 10 temporal continuity and regularity of observations for different climate elements (WMO, 1989). Broadly, Bellamy (1970) discussed the acceptability of meteorological observations in terms of their degree of completeness, considering that the goal of meteorological observations is to "depict the space-time distributions of everything-atmospheric everywhere always, ever more completely in ever-increasing detail"; appropriately, he used the expression completeness of observations. This is the terminology used in the present paper. 15 Concerning the completeness of radiosonde humidity observations, the vertical coverage and vertical resolution of sounding data is of first concern, chiefly between the surface and the middle troposphere (500 hPa) regarding the precipitable water vapor content; furthermore, the period of record and the regularity and continuity of radiosonde data is a relevant issue for long-term monitoring of the climate system (Karl et al., 1995), as exemplified by temporal sampling requirements used in trend and seasonal analysis of temperature, humidity and integrated water vapor Gaffen and Elliot, 1992;20 Karl et al., 1995;Ross and Elliot, 1996;Zhai andEskridge 1997, Lanzante et al. 2003;McCarthy et al., 2009). Although the vertical and temporal completeness of station-based humidity time-series can be treated separately from the geographical coverage of stations, studying the completeness of observations in a global, historical data set of radiosonde observations should address both issues simultaneously. This is particularly true concerning the subsampling of radiosonde stations for studies of atmospheric temperature or water vapor trends on a regional or global scale (Wallis, 1997). 25 Several factors contribute to differences in the completeness of humidity observations among radiosonde stations and individual soundings: i) The geographical coverage of radiosonde stations evolved over time, and so, the period of usage varies among stations; ii) A lack of equipment maintenance may result in interruptions of observations; iii) The number of vertical levels and vertical extent in radiosonde reports depend on the standard pressure levels in use, as well as on the reported significant levels (assuming that the balloon bursts at the proper altitude); iv) Missing humidity observations arise from 30 difficulties associated with the performance of humidity sensors and the observing practices related to their working range. While (i) and (ii) are of a random nature, points (iii) and (iv) deserve an explanation because of historical changes with implications in the vertical coverage and resolution of radiosonde humidity profiles.

Vertical levels in radiosonde observations
In radiosonde soundings, temperature, relative humidity (and/or dewpoint depression) and wind speed and direction are measured together with atmospheric pressure, while geopotential height is indirectly measured from hypsometric calculations 1 5 (but may be missing in radiosonde reports). As a common practice, only standard pressure levels and significant levels are stored and reported. Currently, the standard levels are 1000, 925, 850, 700, 500, 400, 300, 250, 200, 150, 100, 70, 50, 30, 20, and10 hPa (WMO, 1996). But historical changes deserve due attention. An inspection of the earliest soundings collected in IGRAmade in 1905 at Lindenberg, Germany, a quarter of a century before radiosondes were availablereveals temperature data reaching sometimes 100 hPa, with the reported levels being 1000, 925, 850, 700, 600, 500, 400, 300, 250, 200, 150 and 10 100 hPa (although most of those soundings did not reach beyond 700 hPa). Radiosonde humidity measurements at the same station, as collected in IGRA, began in 1950. Nevertheless, the 150-and 100-hPa levels were first recommended by the WMO in 1953, while the levels 70, 50, 30, 20, and 10 hPa were proposed in 1957, the International Geophysical Year. Even so, the levels above 200 hPa were still referred as non-standard by the WMO in 1958, until the 100-hPa level was finally adopted that year (WMO, 1957;WMO 1958). In the years that followed, the pressure levels  150 hPa (representing roughly the 15 stratosphere) became common worldwide. As to the lower levels  200 hPa (representing roughly the troposphere), they were in general use since the early 1940s, with two exceptions: first, the 250-hPa level was only adopted in 1970, to satisfy aviation demands (WMO, 1970); second, the 925-hPa (within the planetary-boundary layer above low-altitude stations), although planned since 1977, was first required in WMO Antarctic stations in 1987, given the low surface pressure over the Antarctic Plateau, until it was adopted worldwide by the end of 1991 (WMO, 1977, p. 15;WMO, 1987, pp. 57-58;Oakley, 1993, p. 23). 20 Note, however, that these two levels were in use to in some stations before international agreement, and for a long time as exemplified by the Lindenberg station. Besides the standard levels, some intermediate fixed levels within the troposphere (e.g., 800, 750, 650, 600 hPa) are regularly used in some stations (Shea et al., 1994). The Lindenberg station also indicates the early use of 600 hPa. The additional (high-stratospheric) levels 7, 5, 3, 2 and 1 hPa have been used in agreement with WMO recommendations (WMO, 1970), depending on regional and national practices. E.g., they form part of 'upper-level ' 25 observations in the U.S. National Weather Service (OFCM, 1997).
The number of significant levelsnon-standard levels needed to reproduce the vertical temperature and dew-point temperature profiles, capturing turning points or abrupt changes (such as thermal inversions and the tropopause)depend on atmospheric conditions, manual rules and, before automation, on the observers' skills. By the late 1950s, the rules for choosing significative levels were still under discussion (WMO, 1957), being established over time by WMO regulations (WMO, 1988). 30 1 Except in some Soviet/Russian radiosonde-radar systems and the last generation of GPS radiosondes -in which pressure is deduced from the (radar or GPS, respectively) profile of geometric height and the radiosonde profiles of temperature and humidity (Zaitseva, 1993;Nash et al., 2011). Interestingly, the almost linear increase in the average number of non-standard levels in weather-balloon sounding reports (radiosondes + pilot balloons) from about zero by 1945 to about 30 by 2000as revealed from IGRA v1, inferring from Fig.   7 in Durre et al. (2006) -, can hardly be attributed to an increased attention to significant levels alone. It suggests that a significant number of stations have reported additional levels apart from the standard and significant levels (both "mandatory" in WMO's nomenclature). 5 The surface level, which is treated separately in upper-air sounding reports, was reported in most of the radiosonde stations since the mid-1940s. However, it has been reported systematically only since around 2000 (as shown later in Sect.

2.3).
The current migration of radiosonde reports from alphanumeric (TEMP) to the binary universal form for the representation of meteorological data (BUFR), together with the conversion of radiosondes to generate native BUFR messages, 10 allows the transmission of high-resolution data (2 to 10 s sampling rate, i.e., ~ 5 to 50 m resolution in a typical balloon ascent) along with the balloon drift position, the observation time for each level and other metadata . Currently, 20% of the radiosonde stations send high-resolution BUFR reports through the Global Telecommunication System (GTS), many coming from Europe; however, such data are not yet available in an open archive.

Missing humidity observations 15
Combining adequate spatial and temporal resolution with enough accuracy for synoptic use, modern radiosonde measurements reach the upper troposphere and lower stratosphere, much beyond the layers where most of the atmospheric water vapor resides. That has not always been so. While the vertical sampling of temperature soundings is limited by the burst altitude and the mandatory levels (standard and significant), the maximum height and the vertical resolution of humidity soundings are further restricted by sensor limitations. Upper-air humidity measurements began in the 1930s but became substantial only in 20 the 1940s. Despite radiosonde hygrometers (measuring relative humidity (RH)) have improved over time, humidity has been always difficult to measure in very cold or dry air due to the poor response of many instruments at very small vapor concentrations (by lowering saturation vapor pressure, cold temperatures are associated with low water vapor pressures). As it was once pointed out, "humidity measurements in the free atmosphere are probably the least satisfactory of the regular aerological observations" (Hawson, 1970). Balloon-based chilled mirror hygrometers, designed to measure water-vapor 25 mixing ratios in the stratosphere (an extremely cold and dry environment), has been used for more than half a century but are exclusive to scientific research or comparison with humidity measurements from operational radiosondes (Mastenbrook and Daniels, 1980;Vömel et al. 2007;Hurst et al., 2011;Hall et al., 2016). Since a long time ago, whether services need to rely on meteorological radiosondes consisting of expandable balloons carrying relatively low-cost and light instrument packages (Brettle and Galvin, 2003). 30 Here is a brief review of the main humidity sensor types and their limitations, since the time when registering balloons were abandoned by national weather services and electric hygrometers began to be incorporated in radiosondes (circa 1940;DuBois (2002)). The lithium chloride humidity sensors, which were widely used in radiosondes between the mid-1940s and the mid1960s, did not respond to temperatures below around −40º C. From the early 1960s onwards, the new carbon hygristor allowed measurements at lower temperaturesdown to −65°C in the early 1990s, however with a time lag in the sensor´s response as large as 10 minutes (Garand et al., 1992). In practice, humidity measurements at temperatures below −40°C were discontinued in many countries before the 1990s, limiting the vertical extent of routine humidity observations to about 400 5 hPa ( 7 km altitude) (Gutnick, 1962;Gaffen, 1993). Besides, the radiosondes using lithium chloride hygrometers suffered from a low-frequency limitation in the transmission of RH less than 15-20 %, known as motorboating (Wade, 1994). The radiosondes using the carbon hygristor enabled, in principle, measurements in that low RH regionhowever, the accuracy and reproducibility of low-RH values was little known and suspected to be poor for many years, giving the wrong impression that relative humidity lower than about 20 % did not occur in the lower troposphere (Wade, 1994;Nash, 2015). Therefore, 10 values of RH below 20 % were usually cut off in humidity reports; in the radiosonde network of the U.S.A. this happened between 1973 and 1992 . Note that changes in instrument and reporting practices in different countries took place at different times: the threshold value of RH varied in the range 10-20%; the lowest temperature of −40⁰C for reporting humidity, and the shift to lower temperatures, was applied in different periods depending on country; humidity could be reported up to a specified pressure level. Moreover, mechanical sensors are not exclusive to pre-1940s radiosondes: 15 hair hygrometers were only abandoned in the mid-1950s and rolled hair hygrometers were used in a few places until about 1980; the goldbeater's skin sensors introduced in the 1950s became particularly important in the Soviet Union. [For historical details on these changes, see Gaffen (1993).] The new capacitive thin-film sensors introduced in 1981, with the RS80 radiosonde, have improved the response time at low temperatures and the capability of measuring very low humidity. Two important enhancements occurred in the late 1990s. First, the protection from chemical contamination arising from outgassing 20 of RS80 radiosonde packages, thus making the dielectric polymer more selective to water vapor molecules and reducing dry bias (Wang et al., 2002). Second, the dual sensors introduced in the RS90 radiosonde, in which two sensors were alternately heated to remove condensation from the measuring sensor, thus preventing wet biases after measurements in saturated conditions. In the RS92 radiosonde, in use since 2004, the lowest temperature of the heating cycle extended down from -40 ⁰C to -60 ⁰C. The smaller size of and the better ventilation of the RS90 and RS92 sensors compared to RS80 improved the 25 response time. However, RS80 sondes were less affected by dry biases in daytime measurements because of the protective rain cap which also prevented direct sunlight (Smit et al., 2013). However, RH reports at temperatures lower than −40 °C did not develop significantly until about 2000; in recent years, for temperatures of −50 °C to −70 ° C only the newest humidity sensors respond quickly enough to make useful measurements; moreover, the best ones had an uncertainty of around 16 % RH at temperatures as low −70 ºC (occurring over Antarctica and around the tropical tropopause), which is barely acceptable for 30 numerical weather prevision but not suitable for climatic studies (Nash, 2015). Improvements over time were not restricted to sensor type but also to data reduction and calibration. E.g.: measurements from the carbon hygristor in VIZ radiosondes were improved in the 1990s by correcting the low-humidity algorithm; some modern radiosonde systems apply corrections for slow time constant of response and for daytime heating of the humidity sensor; calibration at low temperatures was perfectioned (Dirksen et al., (2014) and references therein). While radiosonde humidity measurements are now generally reliable in the troposphere, uncertainties remain concerning the upper-stratosphere, with temperatures below −50ºC, in addition to dry conditions found above the lower troposphere and wet conditions that occur in thick clouds (Miloshevich et al., 2006).
Although the capacitive thin-film sensors have been widespread (with Vaisala radiosondes RS80 and RS92), two older sensor 5 types continued in use for many years: the carbon hygristor (in VIZ/Sippican radiosondes, currently in disuse, and in the GTS1 radiosonde, in use in China) and the goldbeater's skin sensor used in some radiosonde types made in Russia and China until a few years ago; this peculiar sensor responded too slowly to be useful at temperatures lower than -20 °C and suffered from hysteresis following exposure to low humidity (Nash, 2015;Moradi et al., 2013). For the current radiosonde types, see Ingleby (2017). 10 The trouble in measuring upper-air humidity affects the completeness of observations in several ways: the vertical extent of humidity soundings varies much among radiosonde stations and over time owing to sensor limitations in very cold air; likewise, vertical gaps in low humidity regions are expected, due to cut-off of RH below sensors' measuring capability; lastly, missing days in radiosonde humidity records may originate from adverse conditions (dry days, wet days, cold days) at individual stations (Garand et al., 1992;Ross and Elliott, 1996;McCarthy et al., 2009;Dai et al., 2011). As explained above, 15 the actual extent of missing data depends on the observing practices combined with sensor limitations. In addition, failures in some part of the radiosonde system can compromise soundings. Faulty ground-equipment used for control checks (sensors' calibration before balloon release), data reduction and data recording or telecommunication of coded reports may cause long inoperative periods; poor signal reception from the radiosonde make sometimes data processing impossible. Radiosonde operations in remote environments, particularly performed from ships, present their own challenges; Hartten et al. (2018) give 20 a vivid illustration. In sum, the vertical extent, vertical resolution, temporal regularity and continuity of humidity reports are quite heterogeneous.

Input data and methods
We have examined the IGRA 2 main dataset until the end of 2016. Section 2.1 presents briefly that dataset, including the 25 quality assurance of humidity data. Section 2.2 provides a first look on the data, to find out how many and which of the IGRA stations have a non-negligible amount of radiosonde observations (RAOB), and at the same time to give a hint of the amount of humidity data. Section 2.3 describes the data analysis, aiming to explore the completeness of humidity observations in the sense introduced in Sect. 1.

IGRA 2 -sounding data
The IGRA 2 consists primarily of radiosonde 2 and pilot-balloon 3 observations from over 2700 globally and temporally distributed stations, even though the coverage over oceans is limited to ships, buoys and remote islands. This paper concerns with sounding data , comprising over 45 million soundings from 2761 (2662 fixed and 99 mobile) stations [based on data accessed in September 2017]. 5 The main difference between IGRA 2 and IGRA 1 is the amount of sounding data: 33 data sources instead of the initial 11, implying about 80 % more stations; new data from hundreds of stations before 1946 and the addition of floating stations (fixed weather ships and buoys, mobile ships, and Russian ice islands); furthermore, humidity data prior to 1969 were added. The latter change is related to how humidity values were stored in radiosonde reports. Until 1969, humidity observations were given only as RH; from then on, RH measurements have normally been converted to dewpoint depression (DPD) and 10 reported mostly in that form. Different assumptions in the conversion code can lead to inconsistencies of data (Garand el al., 1992). The former IGRA contained only DPD, while IGRA 2 contains humidity data in either form, as available in original reports, provided they pass the following conditions: i) Data completeness: valid temperature accompanies humidity data; ii) Valid range: 0-100 % for RH; 0 to 70 ºC for DPD; 15 iii) Internal consistency: DPD-derived RH differs from reported RH by 10 % at most; iv) Plausibility: (derived water-vapor pressure)  0.1  (atmospheric pressure).
Quality checks (i)-(ii), save for the RH range, are integral to IGRA from its creation (Durre et al., 2006); (ii) for RH (with the later introduction of this variable in the archive) and (iii)-(iv) were added in IGRA 2 (Durre, 2016;Durre et al., 2018). Note that a RAOB message must have at least temperature data at several pressure levels, while humidity or wind data may be 20 missing, and geopotential height is not always given. The recording of pressure levels, and consistence between pressure and geopotential height whenever the latter is reported in source data, has been assured in IGRA since its first version. IGRA uses a consistent data format, irrespective of the provenience of the data (PIBAL or RAOB). Therefore, RAOBs in IGRA can be simply identified by the presence of temperature data. Wind observations from pilot-balloons (PIBAL) have only wind data at several geopotential heights (adjusted from geometrical height measurements and the gravitational field).. Concerning 25 humidity data, the precision and accuracy of RH and DPD data vary substantially as a function of RH and temperature, degrading in dry or cold conditions to a greater or lesser extent depending on the radiosonde type (for a review on the subject, 2 In modern usage, the term radiosonde refers not only to the early radiosondes but also to the rawindsondes (in use since the 1950s), which, besides measuring thermodynamic parameters, provide wind information with the aid of a radio-theodolite, a radar device, a radio navigation system or, more recently, GPS (Dabberdt et al. 2002;Nash et al., 2007). Observations from either radiosonde type are often abbreviated as 'raob' in meteorological jargon. 3 Free balloon tracked by optical theodolites or radar to measure upper-air winds (Wenstrom, 1937, Hickman, 2015. Often abbreviated as 'pibal'. The common single theodolite technique requires the approximate ascent rate to obtain position, while the double-theodolite method allows a pure trigonometric calculation. In visual tracking, rarely used today but still important where radar tracking or wind measurements from a rawindsonde are not possible, a flashlight is used during night or twilight hours. see Smit et al. (2013)). The information about instrument changes (stations' history), whenever available, is provided in a separate metadata file in IGRA 2 [update of the metadata given in the first version of IGRA, which were mostly taken from Gaffen (1996)].
The most frequent nominal observation times are 0300 and 1500 UT until 1957 and 0000 and 1200 UT afterwards, which reflects the shift in observing time that occurred in 1957 in major WMO radiosonde networks. In the beginning of 1958 5 the primary standard hours of WMO upper-air observations were already 0000 and 1200 GMT (WMO, 1958). However, in some countries, different synoptic hours were practiced over the years; sometimes stations have performed up to four soundings per day for certain demands (see Gaffen, 1993).
For a description of data coverage and data sources of IGRA 2, a full description of quality assurance of data, and further detail on the differences between IGRA 1 and IGRA 2, the reader should see Durre et al. (2018) published after the 10 initial draft of this paper was written. NB -Hereafter, IGRA 2 is simply referred to as IGRA, unless stated otherwise.

Identification of radiosonde stations
The examination of the IGRA reveals that 958 stations have wind-only observations in their full period of record, i.e., 34.7 % of the stations represented in the entire archive. These stations form part of the global pilot-balloon (PIBAL) station network, which evolved over time. As to the rest of the stations, some of them changed from a PIBAL launching site to a radiosonde 15 launching site at some point in their period of record,meaning that they are not, strictly, PIBAL stations nor radiosonde stations. Obviously, the number of PIBAL stations or radiosonde stations at a certain time depends on the stations opened and closed before that time, of either type. In the following, terms like "observations", "soundings" or "reports" refer to the upperair stations and balloon data of IGRA, which retains most of the source data. Figure 1a shows the yearly number of stations reporting RAOB any time of the yearmeaning they have at least 20 observations of temperature, regardless of simultaneous humidity/wind observationsand of stations reporting PIBAL observations alonei.e., reporting only wind throughout the year. For comparison, the number of stations reporting any humidity dataindicating that radiosondes are equipped with a hygrometerand with humidity observations in more than 95 % of the radio soundings, is also shown. Constituting the bulk of the IGRA stations until the early 1940s, the stations with only PIBAL data represent nowadays only 13 % of the total. The reason for the apparent discontinuity in the amount of stations 25 performing only PIBALobservations between 1972 and 1973 is this: beginning in 1973, IGRA data largely come from the GTS and include many more PIBAL data than prior data sources (Imke Durre 4 , personal communication, April 12, 2018).
The number of stations reporting RAOB increased rapidly since the mid-1940s, staying in the range 800-900 from around 1970 to present (2016). Note that, before the advent of the radiosonde, upper-air measurements of temperature and relative humidity were made using kites, registering balloons and aircrafts; these platforms were gradually abandoned until the 30 radiosonde era was established in the mid-1940s (DuBois, 2002Stickler et al., 2010). Since the first radiosonde prototypes were developed between 1929 and 1930, the early observations of temperature collected in IGRA, beginning with one station in 1905 (Lindenberg, Germany), were made by meteographs without radio-telemetry. The first upper-air humidity observations are from 1930, coming from a single station: Kjeller, Norway. According to IGRA, until 1942 there were less than 10 stations reporting humidity, with that number growing rapidly in the following years together with the total number of radiosonde 5 stations, reflecting the widespread use of radiosondes (see 'HUM' and 'TEMP' in Fig. 1a). The major relative change occurred between 1945 and 1946, coincident with the end of World War II, when the global count of radiosonde stations tripled. Note that the replacement of hair hygrometers by the lithium chloride humidity element began in some radiosonde networks shortly before; in the U.S.A., apparently that change took place between 1940 and 1943 DuBois, 2002), although IGRA does not contain data prior to 1946. Fig 1a shows also that the fraction of radiosonde stations measuring 10 humidity in more than 95 % of the soundings increased over the years, getting very close to 100% in the last decade (see relative difference between the curves 'TEMP' and 'HUM > 95 %'). Figure 1b shows the evolution of the global, annual mean number of soundings performed per day, for the different atmospheric parameters apart from pressure: temperature, humidity and wind. Recall that pressure is always measured in RAOB soundings, while in PIBAL soundings wind is measured as a function of altitude 5 . For clarity, the PIBAL wind 15 soundings are depicted separately from all wind measurements, which also comes from RAOB soundings since the mid-1940s.
Although, as a rule, radiosonde launches are carried out twice a day, in fact there is a significant number of missing days in temperature and humidity data, i.e. days without any RAOB data: roughly 1 in 5 days during the year, on average for the years after the mid-1950s [as concluded by comparing the yearly number of observing stations (TEMP, HUM in Fig. 1a) with half the global number of daily observations (TEMP, HUM in Fig. 1b)]. 20 Aiming to study humidity completeness, the IGRA stations having a negligible amount of temperature data in every year of their period of record were excluded, because temperature is required to measure RH or DPD and so all humidity data in IGRA are accompanied by temperature data. Specifically, we have selected the stations with RAOB soundings in 5 % or more of the annual soundings in at least one calendar year within their full period of record until the end of 2016. These will be hereafter referred to as IGRA-RS stations (RS stands for radiosonde), even if some of them contribute with relatively few 25 RAOB data. The above selection reduces the number of IGRA stations by 38 %, whereas the number of soundings is only reduced by 13 %, amounting to 39.5 million, out of which there are 30.2 million radio soundings including 29.8 million humidity soundings (see Table 1). Wind-only soundings are still present in 23.6 % of the soundings from the selected stations.
Note, however, that 92.2 % of the removed stations are strictly PIBAL stations and the remaining 7.8 % have RAOB data in less than 0.6 % of the corresponding soundings, apart from two cases with a period of record shorter than one year. In sum, 30 the IGRA-RS subset retains practically all the RAOB soundings (99.999%), particularly the humidity soundings, as shown in Table 1.   The IGRA-RS stations and their locations are listed in supplementary Table S1, along with the full periods of record (Full POR), the periods of record for humidity (Hum POR) and the corresponding numbers of humidity observations (Hum Obs), i.e., the number of individual soundings reporting either DPD or RH data. Since humidity time series can be interrupted 5 for long periods of time, the full period of record of one station may be segmented into two or more periods for humidity (both are rounded to years). Table S1 comprises Table S2. '# Days' is the number of days in Hum POR, excluding the months before Begin Date. The last three columns 20 give the corresponding count of humidity observations around the principal nominal hours, 0000 UT and 1200 UT (± 1 h), and at any other times (0200 UT through 1000 UT and 1400 UT through 2200 UT). Stations are identified by the WMO region and WMO number, followed by the station name and country. (To find out the corresponding IGRA ID codes in Table S1 it suffices to observe that the last nine characters must be 'M000' followed by the WMO number.) Note that most of the GUAN stations have humidity data at or around 0000 UT and 1200 UT almost every day; however, the exceptions to the rule, and 25 even gap years, are not negligible.
Moreover, the IGRA-RS contains 16 stations that form part of the GCOS Reference Upper-Air Network (GRUAN; Bodeker et al., 2016): half certified and half to be certified according to current GRUAN status, of which eight (half certified too) are also GUAN stations. Those specific GRUAN sites report default data (from radiosonde manufacturers) to the GTS; at present, most of them already send BUFR messages with high resolution (Michael Sommer 6 , personal communication, 30 September 18, 2018). The GRUAN aims to serve as reference network for climate applications, satellite validation and in 6 GRUAN Lead Centre, Lindenberg Meteorological Observatory -Richard Aßmann Observatory, Germany. support of other radiosonde networks, by providing long-term high-quality records of vertical profiles of selected essential climate variables, accompanied by traceable estimates of measurement uncertainties (WMO, 2011a;Dirksen et al., 2014).
Naturally, real-time meteorological data transmitted from GRUAN sites to the GTS may differ from GRUAN internal data regarding raw data processing. For reference, the IGRA ID codes of the GRUAN sites appearing in the IGRA-RS station list are underlined on Table S1; of course, other GRUAN sites performing only research measurements are not part of IGRA. 5 Likewise, the WMO numbers of the GUAN stations coincident with GRUAN sites are underlined on Table S2.

Analysis of humidity data
Overall, the analysis of data from IGRA-RS stations, selected as described in the previous section, aims to answer the following questions: 1. What is the spatial coverage of humidity-reporting stations in different years and latitudes? 10 2. What is the fraction of days in a year with humidity data and the number of consecutive missing days on average? 3. What is the typical vertical resolution and vertical extent of humidity observations? 4. How many stations have enough data in the vertical to allow the estimation of precipitable water? 5. How does the temporal and vertical completeness affect the availability of long-term humidity time-series?
Each question is explored as detailed below in Sects. 2.3.1-2.3.5, with the results presented later in Sect. 3. The description of 15 the related metadata parameters regarding each IGRA-RS station, is deferred to Sect. 4.

Global coverage
The geographical distribution of the radiosonde network measuring humidity evolved over time. Its size and distribution were studied in terms of the annual number of stations with any humidity observations during the year in different climatic zones, considering only the fixed IGRA-RS stations. The spatial coverage of observations was further studied by a parameter that is 20 closely related to the average spacing of stations, but it represents better the data coverage if stations are unevenly distributed.
The average separation between adjacent stations ( ) over a region of the Earth's surface can be estimated by √ ⁄ , where is the number of stations lying on a surface of area , ⁄ representing the average station density. This measure is, however, insensible to the spatial distribution of stations. The global radiosonde network has highly variable density since the observation stations are concentrated in continental regions, mostly in populated areas of developed countries. Sparse-data 25 areas occur on oceans and seas, near the poles and in certain parts of land continents. can be alternatively defined as the mean distance between each station and its nearest neighbor; but this definition ignores data-void areas. The average distance from a point on the surface to the nearest station (≈ 2 ⁄ for a uniform network) is more informative because it depends on the distribution of concentrated-and sparse-data areas. Therefore, to study the global coverage of observations it is convenient to use the average distance to the nearest station, as measured from every point over the main landmasses or ocean/sea areas to study the spatial coverage of any subset of stations selected according to a given range for the metadata parameters presented in Sect. 4.

Annual frequency and temporal continuity
The frequency of humidity observations over time is studied in terms of the fraction of humidity observing days in the year.
Although this gives a sense of the regularity of observations, it says little about the continuity of data over the year. In this 25 respect, it is of interest to know the size of the maximum interval of consecutive days without humidity data in a yeardenoted hereafter as 'size of missing days'.
The above defined measures of temporal completeness are critical to study climatic trends (long-term changes in the annual mean or in the seasonal cycle of humidity-related quantities) on specific locations or areas of the globe, which otherwise requires merging procedures using radiosonde data from nearby locations to circumvent large data gaps. We have averaged 30 both quantities across all fixed stations reporting humidity within each major latitude region, year by year.

Vertical resolution and vertical extent
Since the vertical resolution varies with heightaccording to the height of the reported pressure levels (standard and significant) and depending on the number of levels with non-missing data for humidity -, the vertical resolution of an individual sounding must be defined by a vertical average. Since the vertical distance between consecutive levels, say , generally increases with height, with the lower levels being more populated than the upper layers, a geometric mean is more suitable 5 than an arithmetic mean. So, the mean vertical resolution of a single humidity sounding is here defined by the geometric mean of { } for all levels with humidity data in the sounding profile: where is the atmospheric pressure at level ( = 0 denoting the lowest level with humidity data), is the number of levels with humidity data above the lowest level, ̅ is the estimated mean temperature between level and its immediate, relevant lower level − 1 , is the specific gas constant for dry air and 0 is the standard gravity. (Note: IGRA's data-quality checks assures that vertical levels with valid humidity data have also valid temperature and pressure data.) Since geopotential altitude is only given in part of the RAOB reports, the vertical extent of an individual humidity 15 sounding, i.e. its maximum height above mean sea level reached by the humidity measurements, was estimated by adding the station's elevation to the height from the surface calculated upon pressure and temperature data from the surface level up to the top of the humidity sounding (highest level with a non-missing value for DPD or RH), whenever values of temperature and pressure at the surface are given; otherwise the height from the surface cannot be calculated. For mobile stations (ships and buoys), the elevation of the stations can be approximated to zero, unless the vertical extent of the sounding is too small, 20 requiring data for the balloon release height.. For the purpose, it suffices to neglect moisture in the hypsometric equation; given that the virtual temperature is typically within 4 K above the actual temperature, the error in calculating geopotential height amounts to less than 1 %.
Following the above definitions, we have studied the statistical distributions of the vertical extent and the vertical resolution in humidity soundings from all IGRA-RS stations (including mobile) over time by grouping individual values of 25 both parameters in annual bins. To assess the shortness of humidity observations in RAOB, we have also calculated the vertical extent of temperature observations and their vertical resolution up to the top of the co-located humidity observations.

Soundings eligible to estimate precipitable water vapor
Usually, the precipitable water vapor (column integrated water vapor mass per unit surface area) is estimated from the profile of water vapor mixing ratio between the surface to the 500-hPa leveli.e. the layer where  95 % of the columnar mass of 30 water vapor is and where humidity data from radiosondes are more often available and generally more accurate (Elliot et al., 1991;Gaffen et al., 1992;Ross and Elliott, 1996;Durre et al., 2009). In this paper, a humidity profile is considered eligible to estimate precipitable water vapor under the following conditions: i. Humidity data are given at the station's surface and at all standard levels laying between the surface and the 500-hPa level, except for the 925-hPa level.
ii. If humidity data is missing at a standard level apart from 925-hPa, a nearby significant level is acceptable if its height 5 from the surface differs from the height of the missing standard level by less than 5 %.
iii. The distance between any consecutive levels with humidity data between the surface and the lowest level located more than 1 km away from the surface should not exceed 1 km, unless the station elevation is larger than 500 m.
The 925-hPa level is not required here because this level was not standard until 1991. However, condition (iii) assures a minimal resolution in the planetary boundary layer, by including near-surface significant levels as well as the 925-hPa level 10 when it is given; this is required because water vapor is highly variable and abundant in this region; the exception for very elevated stations contemplates the case when the first upper-air humidity record is at 850-hPa, i.e.  1.5 km above the mean sea level, but the height from the surface is less than 1 km. The IGRA-RS soundings fulfilling the above conditions will be hereafter referred to as Sfc-to-500hPa humidity soundings. Typically, the first standard level higher than 1 km from the ground is 850-hPa. Thus, by including enough data at 15 significant levels below the 850-hPa level, instead of requiring data at the current standard levels in the same layersorted out of 925-hPa and 1000-hPa, depending on the surface pressurethe definition given above accommodates much more soundings, particularly before 1992 when the 925-hPa level was not mandatory. To be sure, the relative amount of humidity data at near-surface levels is now examined, excluding 1000-hPa (around 0.1 km altitude) since this is frequently placed below the stations' elevation (339 m on average for the IGRA-RS fixed stations). 20 Figure 2 shows the evolution of the global percentage of humidity observations at the 925-hPa level and at any significant level between the surface and 850-hPa, out of the IGRA-RS soundings having humidity data at the surface on condition that the surface level pressure is higher than 925 hPa and 850 hPa, respectively (referred in the following as surfaceupper-air soundings). The percentage of surface observations in all humidity soundings is also shown. First, note that the surface observations began in 1943 (in fact, not only of humidity but also of temperature), rising rapidly in the next 5 years to 25  95 % of the humidity soundings, decreasing then to a minimum of 60% in 1965 and broadly increasing again until 2000, staying above 95 % since then. In short, the humidity measurements at the surface level are mostly available since 1945 but were in widespread use only since 2000. Secondly, the percentage of the surface-upper-air soundings having humidity data at the 925-Pa increased almost as a step-function around 1992. It increases from only 2 % in 1991 to 60, 86 and 98 % in the following three years. This change is coincident with the introduction of the 925-hPa level as an additional standard level in 30 radiosonde messages in November 1991, and consistent with the fraction of stations already reporting that level in mid-1993 (Oakley, 1993, p. 24). Lastly, the percentage of the surface-upper-air soundings that have humidity data at any significant level below the 850-hPa level, beginning in 1948, generally increased with time, mainly in the 1960s, with a value larger than 80 % in recent years.
Using the definition given at the beginning of this section, we have studied the number and the percentage of stations (fixed and mobile) whose Sfc-to-500hPa humidity soundings exceed a given percentage out of the humidity soundings made in each year. The distance between missing standard levels and nearby significant levels was calculated from pressure and 5 temperature data, neglecting moisture. A stricter definition of Sfc-to-500hPa humidity soundings, specifically, having humidity data at the surface and all upper-air, current standard levels up to 500-hPa was also studied for comparison.

Current record length of time-series
The 'current record length' of a humidity time-series in a given station and year, is herein defined as the number of elapsed years in the time-series with no gap years in the interim. To simplify, a calendar year with any amount of humidity 10 data is counted as one. For past years, the current record length generally differs from the span of the entire time-series, which may continue after the year under consideration. Also, note that one station can have more than one time-series for humidity, as the humidity observations can be interrupted by one or more gap years. So, the full 'period of record' of a radiosonde station may be divided into several sub-periods of record for humidity. Recall that the years with humidity observations at each IGRA-RS station are indicated in Table S1. 15 We have studied the evolution of the average current record length of the humidity time-series, from all stations, year by year. The same was done for the time-series with humidity soundings 90 % of the days in the year, at least, and particularly consisting of Sfc-to-500hPa soundings, with their own record lengths. The calculation of the distribution of the current record length of the time-series with Sfc-to-500hPa soundings 90 % of the days in the year or more was repeated by restricting the corresponding size of missing days to less than 10 days. 20

Overview on the completeness of radiosonde humidity observations
This section gives a general picture of the completeness of humidity observations over the years, using the data from the IGRA-RS stations defined in Sect. 2.2 and following the data analysis described in Sect. 2.3. The results of Sects. 3.1 and 3.2 refer to fixed stations, i.e., over 1600 stations on continents and islands, 14 ocean fixed weather-ships and 2 environmental buoys. In 25 the remaining Sects. 3.3-3.5 mobile stations (99 ships) are equally included (see Table S1; moving stations are denoted by unspecified geographical coordinates). Figure 3 shows the geographical distribution of the IGRA-RS fixed stations at specific years illustrating the growth of the global radiosonde network (cf. Fig. 1a); stations reporting humidity observations are highlighted and counted. Recall that a single station can change from PIBAL observations to RAOB, or even the reverse, during its period of activity. The IGRA-RS retains practically all RAOB data of IGRA; however, some stations have years with only PIBAL observations. Since almost 5 all of the IGRA stations measuring temperature in a given year do also measure humidity at least part of the time (as seen by comparing black and solid blue lines in Fig. 1a), it is clear that most of the IGRA-RS station-years without humidity data (red crosses in Fig. 3) correspond to years of PIBAL observations alone (no RAOB). Concerning humidity observations, Fig. 3 shows that, by 1945, 2/3 of the stations were set in South Asia (British India) and Australia. Most of the data coverage over North-America, Greenland, Europe and North Asia, including the Artic region, as well as over the surrounding oceans took 10 place between 1945 and 1955. By 1975, Central and South-East Asia, Africa, South-America, Antarctica (along its coastline, except for the Amundsen-Scott South Pole Station) and the surrounding oceans were already coveredalthough not as well as farther north regarding the continental regions, with the noteworthy exception of China territory. While the global number of fixed stations measuring humidity has remained practically unchanged since then (see Fig. 4 ), their geographical distribution changed significantly. At present (2015) there are more observation sites in South America, much less in Central and East 15 Africa, much more in Western Asia, and a more even distribution in the rest of the world.

Geographical coverage of humidity observations
Although oceans have been mainly covered by island stations, ocean weather-ships were important for more than 30 years, between the late 1940s and the early 1970s. Judging from IGRA data, coverage from these ships was optimal between 1967 and 1972, with 7 to 12 fixed ocean weather-ships transmitting radiosonde observations simultaneously, almost all located in the North Atlantic except for one ship in the Norwegian Sea and another one in the North Pacific. Apparently, upper-air 20 observations from these platforms ended a few years later (save for the station "M" in the Norwegian Sea, which continued until 1990), coincident with the growing use of satellite retrievals in weather forecasting. However, in-situ data coverage on oceans was improved by using balloon sounding systems on board of merchant shipswhich obviously are not shown in Fig.   3. About 10 to 20 ships of opportunity have launched radiosondes concurrently along their routes from the 1980s to present.  One can see that before 1937 there was only one observing station and three years without data (1933-34 and 1936). The few humidity-reporting stations existing by 1940 were placed in the Northern Hemisphere (in fact, most of them were in the Artic).
From 1945 on they were located predominantly in the northern extratropics, even though they raised other latitudes, at the fastest rate in the subsequent two decades. From 1970 onwards, the number of stations in the southern extratropics did not follow the observed growth observed elsewhere. This is not surprising since regions south of parallel 30⁰ S have significantly 5 more ocean and less land. Note that the absence of mobile stations affects very little Fig. 4, since mobile stations are short numbered and have a short period of record (a few years), despite being important in covering the oceans. The northern extratropics have accounted for about half of the world stations for decades, although its relative weight has been decreasing over the years; the decrease in number of stations situated in that region after 1990 is counterbalanced by an increase in the Tropics since then. Using Eq. (1) and the method outlined in Sec. 2.3.1, Fig. 5a 500 km over land during the same period, the little change observed despite considerable changes in the distribution of continental stations can be explained as follows: the better spatial coverage over South America at present is offset by the poorer coverage over a large part of sub-Saharan Africa, save for the country of South Africa (see Fig. 3, panels c and d). In contrast, in ocean/sea regions ̅ not only is two to three times larger than over land regionsexcept in the southern polar region, i.e. in the Southern Ocean when compared to Antarcticabut has also degraded slightly over time in the Northern 20 Hemisphere oceans and seas from the subtropics (23.5⁰-35⁰ N) to the Arctic (66.5⁰-90⁰ N). From a global perspective, the hemispheric differences in ̅ due to the distribution of oceans and continents can be appreciated in Fig. 5c. Note that 2 ̅ gives an estimate of the average separation between adjacent stations in regions where the radiosonde network is relatively regular.
For example, Fig. 5a indicates a typical separation of about 400 km in the northern temperate continental regions (coincident with the wealthier countries of the North Hemispherewhich is acceptable for synoptic weather forecasting. While the same 25 is impracticable in many other parts of the world and over the oceans, distances up to two or three times larger than ideal need to be filled by satellite-based data and supplemented by surface observations, which are generally much denser than radiosonde stations. Nevertheless, on a scale suitable for climate monitoring, the WMO recommends that upper-air stations should have a maximum average separation of 1000 km (WMO, 2011b). This would require ̅ ≤ 500 km.  Figure 6 represents the fraction of days in a year having humidity observations, averaged across all humidity-reporting fixed stations in each of the three major latitude regions of both hemispheres, along with the standard deviation. The plot begins in 1945, about when upper-air humidity measurements became routine in radiosonde soundings (cf. Fig. 1). In all regions, the average fraction of the days in a year with humidity observations increased rapidly until around 1960, stabilizing 5 since 1965 to values in the range 70-80 % in low latitudes and 80-95 % in mid-and high latitudes. The corresponding standard deviations indicate that a non-negligible percentage of the stations have observations on every day in the year since the 1950s. The fraction of days with humidity observations, the size of missing days and the count of observations, detailed by station and year, is part of the first metadata set presented in this paper (see Sect. 4.1). Note that gap years are also evaluated in the metadata. Similar information is given for the Sfc-to-500hPa humidity soundings alone. 20 Figure 8a shows the distribution of the vertical-mean resolution in the annual humidity soundings since 1945, along with the homologous distribution in the simultaneous temperature soundings, limited to the highest level with humidity measurements for comparison. The vertical-mean resolution of each soundings was calculated by Eq. (2), rounded to the nearest decameter. 25

Global vertical extent and resolution of humidity observations
For clarity, the curves displaying the mean and the quartiles are smoothed by a 5-year running mean. The differences between the distributions for temperature and humidity indicate missing humidity data in a few percent of the vertical levels with temperature data; although statistically irrelevant, such vertical gaps may be quite significant in individual soundings. The vertical resolution was relatively poor on average and highly variable in the early decades: until around 1965 the mean and median were coincident and varied in the range 1.1-1.4 km, with a midspread (interquartile range) of almost 1 km. Both the 30 average resolution and the midspread improved consistently from 1965 to between 2000 and 2005. Since 2005, ¾ of the soundings have a vertical-mean resolution better than 0.5 km, with half of the values ranging from 0.3 to 0.5 km. Figure 8b shows how the maximum height above the mean sea level reached by either temperature or humidity measurements is distributed among the corresponding soundings on each year. Curves are smoothed by a 5-year running mean, as in Fig. 8a. However, the distribution is restricted to temperature and humidity soundings with surface data, these 5 representing, respectively, 86.7 % and 86.0 % of the total RAOB soundings in the period 1945-2016. In addition, the soundings from mobile stations with missing geopotential height at the surface level are excluded, for consistency with the hypsometric calculations used in the dataset presented in this paper. According to Fig. 8b, by 2003, 75% of the temperature soundings with surface data reached an altitude of 22 km, i.e. ~ 50 hPa. Note that Durre et al. (2006) reported that, by the same year, 74% of all IGRA 1 soundings reached at least the 100-hPa level, i.e. ~ 16 km; this lower height is due to the inclusion of PIBAL data 10 in their analysis of IGRA 1, whereas our analysis is restricted to RAOB in IGRA 2.
Contrary to temperature, which can be measured up to the maximum height achieved by the sounding balloon (burst altitude, although the highest reported level is usually limited by the standard levels in use), the vertical reach of humidity measurements depends on the working range of humidity sensors (and, to some degree, on reporting practices). In Fig 8b we can observe that the top of the humidity observations is at present situated almost 5 km below the maximum altitude of the 15 temperature measurements (close to the burst altitude) on global average (either mean or median). But the difference in vertical extent between temperature and humidity observations has changed greatly over the years, with a maximum mean value of 12.5 km by 1980. Until the late 1960s the burst altitude increased at a much faster rate that the top of humidity observations, except for a few years after 1965 when the reverse happened. This last feature is coincident with the introduction of the carbon hygristor. Likewise, the isolated peak of the global mean maximum height of humidity observations around 1970, as seen in 20 Fig. 8b ('HUM'), seems to indicate an exploratory period with the new instrument. Noteworthy, in 1970 the WMO stated that "no routine observations of humidity are made in the stratosphere and no practical use is envisaged for such current observations" (Hawson, 1970). Figure 8b undoubtedly shows that humidity was mostly measured in the troposphere until the mid-1960s. In contrast, from the mid-1980s onwards, the vertical extent of humidity observations has increased consistently and faster than the RAOB-toproughly by 4 km per decade on global average -, denoting improvements of humidity sensors 25 (see Sect. 1.3). In 2015, three quarters of the humidity-top values extend to over 13 km, and half to over 26 km, indicating that many humidity reports extend into the lower stratosphere. The accuracy of radiosonde humidity measurements in the stratosphere is beyond the scope of this paper. However, only 15 years ago, international experts pointed out that the accuracy of current operational measurements of humidity in the upper-troposphere was inadequate for addressing climate variability and change (despite the usefulness of some sensors) and a challenge for future operational radiosondes (Durre et al., 2005). It 30 is interesting to note that the interquartile range of the top height of humidity soundings has increased much in the period 1985-2000, i.e., individual values became widely dispersed around the median. This is likely related to the proliferation of humidity sensors of different kinds, thus increasing the instrument variations among stations as new instruments coexist with older ones. For instance, by 1989 the WMO had identified 20 major radiosonde types in use worldwide (Kitchen, 1989a).
While this number has fallen to 13 by 2002 (Elms, 2003), Fig. 8b shows that the dispersion in the vertical range of different humidity sensors has increased in the meantime; changes in instrument-dependent observing practices may have a role.
The average vertical resolution of the humidity observations, by year and station, as well as the individual values by station, date and time, are given in the metadata sets provided in this paper (see Sect. 4). Similar metadata is provided for the 5 Sfc-to-500hPa humidity soundings, as defined in Sect. 2.3.4, in which case the vertical resolution is calculated for the levels between the surface and the 500-hPa level and is normally finer than depicted in Fig. 8a.
Since IGRA-RS observations do not have surface humidity data prior to 1945, and surface data is not always given after that year, to account for all soundings the vertical extent of humidity observations must be alternatively represented by the lowest pressure corresponding to humidity data. The same applies to soundings from mobile stations if geopotential height 10 is not given at the surface level (it is missing in 30 % of the corresponding RAOB). Although the local sea level is normally within ± 10 m from mean sea level, taking zero as the baseline height can lead to large relative errors if humidity is only measured very close to the radiosonde station elevation (balloon release height). Therefore, the average pressure of the top of humidity soundings, by station and year, are given in the first metadata set presented in this paper; average values are represented by a geometric mean (see Sect. 4.1). As to the second set for individual observations -metadata by station, date 15 and timeboth the top pressure and the corresponding altitude, whenever this can be calculated, are provided (see Sect. 4.2).

Global relative amount of Sfc-to-500hPa humidity soundings
Recall that our definition of Sfc-to-500hPa humidity soundings (Sect. 2.3.4) is intended to represent the soundings with enough vertical level, and almost evenly distributed near the surface, such that the water vapor profiles can be properly described and the precipitable water can be estimated. In this respect, the completeness of such humidity observations based on current 20 standard pressure levels alone is unsatisfactory for two reasons: first, the level 925-hPa was barely used before 1992 (see Fig. 2); second, the sounding data at significant levels, often related to features of temperature rather than RH, or at other additional levels, are equally good provided that the vertical resolution (habitually increasing towards the surface) is not too different.
Next, we will compare the distribution of Sfc-to-500hPa soundings among IGRA-RS stations in each year using alternative definitions: 25 A) Humidity data at the surface and all current standard levels above the surface up to 500 hPa, i.e. at pressure levels The soundings meeting either of the above two definitions are coined as Hum-A or Hum-B in the following analysis. Note, however, that only definition B was used to prepare the metadata sets supplied in this paper. 30 Let P be the percentage of Sfc-to-500hPa soundings (HUM-A or HUM-B, at our choice), out of all humidity soundings from an arbitrary station in a given year. Figure 9a shows the evolution of the number of stations with P for Hum-A soundings exceeding given values, since nearly the time radiosonde humidity data at the surface level are first available. Comparing Fig.   9a with Fig. 1a, we can see that between 1945 and 1991 only a small fraction of the radiosonde stations carried out Hum-A observations in most of the soundings, say, with P within the range P > 95 %. Things changed drastically between 1992 and 1994, when the number of stations with a significant percentage of Hum-A soundings, e.g. 20 % < P < 80 %, increased by an order of magnitude. This change reflects the change in the observing practice shortly after 925-hPa was internationally adopted 5 as a standard pressure level (see Fig. 2). In recent years there are over 400 stations with P > 95 % and almost 800 with P > 80 %. Figure 10a shows how the probability of finding a IGRA-RS station with a percentage of Hum-A soundings exceeding a given value was reversed over the last five decades: by 1965, only 12 % of the stations had at least 20 % of Hum-A soundings, whereas by 2015  91 % of the stations had at least 80 % of Hum-A soundings. Figure 9b is the counterpart of Fig. 9a, for the less stringent Hum-B soundings. Recall that 925-hPa is now treated in 10 the same way as any near-surface non-standard level, and, for the rest, significant levels close to standard levels are allowed.  The number of Hum-B Sfc-to-500hPa humidity soundings in each station and year, as well as the fraction of days in 25 a year having such soundings, are both given in the first metadata set presented in this paper (see Sect. 4.1). Information on other parameters describing humidity completeness but focusing on these soundings is also provided. The metadata set regarding individual observations identifies the Sfc-to-500hPa humidity soundings and provides information on their vertical resolution between the surface and 500 hPa (see Sect. 4.2).

Amount of long-term time-series 30
Figure 11a (black lines) illustrates the number of humidity time-series, one per station, fixed or mobile, with a 'current record length' (elapsed years until the year in abscissa) exceeding a given number of decades: 1, 3 and 5. The time-series refer to periods of consecutive years with any observations in the year. For comparison (color lines) the same is shown for the timeseries with data 90 % of the days in the year or more. It should be noted that the concurrent series with more than 10 years of back data begin in 1948, even though the first upper-air humidity measurements began in 1930 and by 1949 there were already around 300 stations measuring humidity (cf. Fig. 1a). The initial slope of the curves in Fig. 11a denotes the rapid growth of the global radiosonde network after the second world war, with many stations measuring humidity regularly. Nevertheless, the 5 curves are not monotonous, due to the closing of stations in the past and to the existence of gap years in many stations. E.g., if we want to collect the largest amount of parallel humidity time-series extending back in time to over 10 years, Fig. 11a   more; secondly, it further limits the time-series to those with less than 10 consecutive missing days in every year, meaning that all months are evenly represented. We can see that the number of parallel time-series with regular (fraction of days in a year  90 %) Sfc-to-500hPa soundings and extending back in time to more than 10 (30) years has increased rapidly since 2000, after four (two) decades with a value of less than 100 (20), only then increasing steadily to about 330 (50) by 2016. However, the Sfc-to-500hPa time-series with the same length but being almost continuous (fraction of days in a year  90 %; size of missing 20 days < 10 days) are much less numerous. Figure 11 is only illustrative. There is no simple answer to the question of when we have enough data to (in theory) perform climate studies from radiosonde humidity data: it depends on the strictness of completeness criteria. (In practice, it depends also on the accuracy and homogeneity of measurements by different instruments.) Note that the length of the timeseries under user-specified conditions for any of the metadata parameters defined in the next section, either backward or 25 forward in time, can be derived from the information contained in the related metadata set (Sect. 4.1). In addition, the years with humidity observations can be found in Table S1 (as well as in a data file accompanying the two main metadata sets): one or more periods per station, depending on whether there are gap years or not in the station's full period of record.

Metadata sets of completeness of radiosonde humidity observations based on the IGRA
The metadata sets describing the completeness of radiosonde humidity records collected in the IGRA are outlined next. They are constructed upon the metadata parameters introduced in Sect. 2.3 and examined in Sect. 3, except for minor adaptions described in the subsections below. The combination of selection criteria, by simply specifying value ranges for the metadata parameters, offers a plethora of choices to the user. Both sets refer to sounding data from the IGRA-RS sub-set of IGRA 2 5 stations, as defined in Sect. 2.2 and listed in Table S1, which essentially excludes pilot-balloon stations.

Metadata by station and year
The metadata parameters regarding each IGRA-RS station and year (annual statistics) are defined in Table 2. Since pressure is the vertical coordinate that is always present in humidity radiosonde data, and the calculation of height above mean sea level is impossible when the surface level is missing, the average vertical extent of humidity data at a given station and year is 10 represented by the geometric mean value of the lowest pressure with humidity data. (Given the huge number of values involved, the arithmetic mean of the logarithm of pressure was calculated first and then exponentiated.) Note that the geometric mean pressure of the highest level with humidity data provides a natural measure of the corresponding arithmetic mean altitude: insomuch as pressure decrease almost exponentially with height, the arithmetic mean height above sea level is roughly proportional to the logarithm of the geometric mean pressure. The statistics for the vertical resolution of humidity data consists 15 in the annual mean of the 'mean vertical resolution' of individual soundings at each station, calculated by Eq. (2). The IGRA data meeting the definition of Sfc-to-500hPa humidity soundings as stated in Sect. 2.3.4 have their own metadata: number of observations, average vertical resolution, size of missing days, and fraction of days in a year having humidity data.
The metadata set is as a plain-text file, each data record containing the metadata values, station-by-station and yearby-year. For each station, the yearly variables defined in Table 2 are displayed chronologically. To put humidity metadata into 20 context, the yearly number of soundings and of RAOB soundings are also given in the dataset. The years within a station's period of record with observations of any meteorological variable (wind, temperature, humidity) are kept, using appropriate missing values when humidity data is missing throughout a whole year. While humidity data begin in 1930, pre-radiosonde years are included to preserve the stations history.
NB -The metadata set is complemented by a list of the IGRA-RS stations with information on geographic coordinates 25 (if fixed), name, and country, along with original metadata describing the periods of record for humidity and the corresponding amounts of observations. This is an ASCII version of Table S1, for computing purposes.

Metadata by station, date and time
The metadata parameters regarding individual soundings at each IGRA-RS station are defined in Table 3. The utility of the corresponding metadata set is to allow a fine selection of humidity data, as a complement to a first selection based on the 30 statistical parameters of Table 2. This metadata set is organized into one plain-text file per station, and one data record per sounding. Below the headline, each record contains: the date, nominal hour, latitude and longitude of the balloon launch, followed by the related humidity metadata.

Availability of metadata sets
A dataset combining the two metadata sets presented in Sect. 4 is available on Zenodo, DOI: 10.5281/zenodo.1332686. The 5 accompanying 'Readme' file gives the necessary information about the data format and file contents. The update is planned to take place each time two full years have been completed in IGRA, starting on January 2019 for the period 2017-2018. • The radiosonde network providing humidity observations in the northern temperate and polar latitudes was essentially established in the mid-1950s. In the mid-1970 the globe became covered by practically the same number of humidity-20 reporting stations (nearly the same as radiosonde stations) as it is today, although the distribution of stations has changed considerably over time. The averaged distance to the nearest station measured from points over continents and large islands (from Mindanao up to Greenland in size), ranging from ~ 200 km in the northern midlatitude countries to ~ 700 km in Antarctica, has little changed too. However, the spatial distribution of continental stations continue to show important regional disparities, such as the poor coverage of most of sub-Sharan Africa which has 25 worsened from 1970 to present. Concerning ocean/sea areas, and disregarding mobile stations (mostly 'ships of opportunity'), the oceans of the southern temperate latitudes always exhibit the poorest coverage, with an average distance to the nearest station exceeding 1000 km. Remarkably, since during the last half century the average distance to the nearest station has increased considerably in the North Hemisphere oceans and seas extending from the subtropics to the Arctic. 30 • The fraction of days in a year with humidity observations has greatly increased until 1960, on average calculated for each year across the stations on each of the main latitude regions. Since 1965, the mean values are in the range 70-80 % in low latitudes and 80-95 % in mid-and high latitudes annually, although individual values vary widely among stations. Between 1945 and 1955, the 'size of missing days' in humidity observations (largest gap in days) in each year having any humidity data decreased from about five months to one month on global average. After 1960, the 5 size of missing days exhibits averages around 20 days in middle and high latitudes; in low latitudes it has increased over time from around 30 to 40 days.
• Humidity was measured mostly in the lower to middle troposphere until the mid-1960s. Since the mid-1980s, the mean height achieved by humidity measurements has increased by  4 km per decade on global average; the gap distance to the burst altitude (denoted by the maximum height of the temperature measurements) more than halved. 10 At present, that gap is  5km and three quarters of the humidity soundings reach to at least 13 km altitude. However, the dispersion of the maximum height of humidity data around the mean has increased too, likely due to the coexistence of older humidity sensors with new sensors. The vertical-mean resolution of humidity observations (geometric mean of the distance between consecutive levels with data), has a median in the range of 1.3-1.4 km until 1965, among all humidity soundings in each year, with an interquartile range of  1 km. Thereafter, the median and 15 midspread have improved constantly before stabilizing in the early 2000s; since then, half of the annual soundings present a vertical-mean resolution between 0.3 and 0.5 km.
• The number of stations providing Sfc-to-500hPa profiles with adequate resolution for calculating precipitable water vapor in more than 80 % of the reported humidity soundings in a year has changed from a few dozen in the period 1945-1970 to around 800 in recent years (97 % of the active radiosonde stations by 2014/16). In general, the amount 20 of stations having a significant percentage of Sfc-to-500hPa observations shows a sudden increase around the years 1970, 1991/92 and 2000. These change points are associated with the availability of data at the surface level and at non-standard near-surface levels; the latter provide important information in the planetary boundary layer before 1992, when the 925-hPa level was not standard and was rarely used.
• The amount of humidity time-series with a given number of consecutive years of data until a given year depends not 25 only on the year and the length of record, but also on other completeness criteria. For example, the station-based timeseries extending back in time for at least 30 years and being 90 % complete (in terms of fraction of days in a year) begin with a few units in 1977 and represent around 200 stations in recent years; by further requiring Sfc-to-500hPa completeness, the other criteria being the same, the number of time-series drops to 50 by 2016. Evidently, the equivalent time-series until the same year but with a much shorter duration, e.g. 10 years as a minimum, are several 30 times more numerous and start much earlier. In short, the amount of humidity time-series that are potentially available to perform climate studies, i.e., discounting accuracy requirements and biases due to instrument changes, depends strongly on the strictness of the completeness criteria. Furthermore, this paper presents a dataset detailing the completeness of humidity observations (RH or DPD together with pressure and temperature) based on the data from the IGRA-RS stations. The dataset (Ferreira et al., 2018) consists of: 1) statistical metadata for each station and year, in a single file for the 1723 stations and their full period of record until 2016; 2) metadata specific to individual soundings, organized in one file per station and covering 39.5 million soundings for the same stations; and (3) list of stations along with the observing periods for humidity and the corresponding number of observations. 5 The metadata parameters were designed to facilitate the selection of upper-air humidity data from IGRA according to a plethora of choices, therefore being able to meet specific research needs in the areas of atmospheric and environmental sciences.
It is widely known that the usefulness of historical radiosonde data depends crucially on metadata information about instrumentation and observing practices (Schwartz and Doswell, 1991;Elliot and Gaffen, 1991;Gaffen et al., 1996;Parker and Cox, 1995). However, sampling differences among stations associated with geographical coverage, observing periods and 10 missing data, are no less important than differences in data precision and accuracy. Reporting practices related to limitations of humidity sensors affect particularly the humidity records (Garand et al., 1992;McCarthy et al. 2009). In this respect, the metadata presented in this paper, if used as a tool to find out the more complete humidity time seriesrelatively long, regular and continuous; vertically extensive and well resolved -, should be crossed with the coordinates of stations and the station history metadata available in IGRA. On the other hand, the present metadata might accelerate progress in the current research 15 on the homogenization of radiosonde humidity data (McCarthy et al., 2009;Dai et al., 2011), by sampling stations with coincident observing periods and satisfying reasonable completeness criteria.

Acknowledgments
We thank Imke Durre for clarifying some points about IGRA Version 2 during the preparation of this work, and David Carlson 20 and two anonymous reviewers for their helpful comments. We also thank to Orlando García for IT support at EPhysLab (Environmental Physics Laboratory at the Faculty of Science of the University of Vigo, Ourense Campus). This work was funded by the Spanish government within the EVOCAR project (CGL2015-65141-R), co-funded by the European Regional Development Fund, and partially supported by the Galician regional government through the program Consolidation and  Percentage of humidity observations: at the surface level out of all humidity soundings in each year; at 925-hPa, out of the soundings with surface data below that level; and at any non-standard levelexcluding the 925-hPa, irrespective of the yearbetween the surface and the 850-hPa level, whenever this is above the surface.       ( 2 ) Arithmetic mean of the 'mean vertical resolution' given by Eq. (2).
( 3 ) Calculated as RESa, but between the surface and 500 hPa.
15 ( 3 ) Calculated as RESa, but between the surface and 500 hPa.