Automation of upper-air sounding data processing has been making steady progress since the early 1970s and is now widespread (Kostamo, 1992). The Vaisala Autosonde project was started in late 1992 and a working prototype was presented at CIMO, Vienna, in 1993. The prototype was tested in Norway and Sweden in 1993 and 1994. This coincided with the replacement of manual balloon-tracking systems by the Omega and Loran networks. It was provided by Vaisala Oy (Finland) and was permanently installed at the Landvetter station in Sweden in 1994. As of today, about 80 Vaisala ARLs have been installed worldwide and the number of soundings performed has exceeded 800 000, while the annual number of new soundings will soon exceed 70 000 (Lilja et al., 2018). With the newest Autosonde model, it is possible to perform 60 soundings without replenishment, while the earlier models allowed up to 24 soundings.

The first radiosonde type used for an automatic launch was the RS80-15N (during 1994–2006). The RS80 radiosonde was followed by the RS92 (manufactured 2005–2017) and RS41 (available since late 2013) models. The RS92 radiosonde (Dirksen et al., 2014) performs measurements with a nominal measurement uncertainty (provided by the manufacturer) of 0.5 ∘C for temperature, 1.0 hPa for pressure below 100 and 0.6 hPa above, 0.15 m s-1 for wind speed and 5 % RH for relative humidity (https://www.vaisala.com/sites/default/files/documents/RS92SGP-Datasheet-B210358EN-F-LOW.pdf, last access: 3 July 2020). RS41 sonde specifications for nominal measurement uncertainties (provided by the manufacturer) are 0.3 ∘C for temperatures below 16 km and 0.4 ∘C above, 0.01 hPa for pressure sensor, 0.15 m s-1 for wind speed and 4 % RH for relative humidity (https://www.vaisala.com/sites/default/files/documents/RS41-SGP-Datasheet-B211444EN.pdf, last access: 3 July 2020). Note that the Vaisala RS41 radiosondes are of two different types: RS41-SG which is equipped with a pressure sensor and using the GNSS-based method to infer pressure (Lehtinen et al., 2014), and RS41-SGP which uses a pressure sensor as the default. More stations use the RS41-SGP than the RS41-SG; in November 2019, 158 stations were using type RS41-SGP versus 66 stations using type RS41-SG.

To launch the RS41 sondes, the Autosonde ground-check (GC) procedure has been updated. The GC device of the RS41 sondes consists of a wall-mounted box and an activator that contains a wireless reader for the radiosonde. The device is designed to automatically activate the radiosonde and to enable wireless data transfer. An activator is connected to the reader box with a coaxial cable. The ground-check device also includes a barometer, while the surface pressure used as a reference for the launch is obtained from a separate co-located automatic weather station. However, the ground-check pressure device can be used as a backup for the weather station sensor. The GC performs a temperature check where the actual temperature sensor is compared with the one integrated on the humidity sensor chip. In contrast to the RS92 GC, a pre-flight fine tuning of the temperature measurement is no longer applied to the RS41 because the manufacturer found that the performance of the RS41 temperature measurement is practically unchanged during storage.

Humidity is also checked in the GC. The RS41 humidity check consists of two main steps – the sensor reconditioning phase and the 0 % RH check. In the reconditioning phase, the sensor is heated to remove possible contaminants that might affect the measurement results and cause a slight degradation of the sensitivity of the humidity sensor. Then, the humidity sensor is checked and then corrected against a dry humidity condition. Specifically, the dry reference condition of the new zero humidity check is generated in open air by heating the sensor using the integrated heating element on the sensor chip. The procedure is based on the decrease of relative humidity towards zero as the temperature rises high enough (Vaisala, 2013, 2015). This method differs from the RS92 GC where the correction was based on a dry condition generated with desiccants, whose drying capacity gradually fades with time.

The radiosonde's humidity sensor is reconditioned and ground check is performed during the automated launch preparation in order to ensure similar performance as in manual stations (Lilja et al., 2018). Figure 2a provides a schematic picture of the most recent Vaisala AS41 Autosonde system configuration, while Fig. 2b shows a photograph of the Autosonde system operational at the Finnish Meteorological Institute GRUAN site in Sodankylä (WMO Integrated Global Observing System (WIGOS) station identifier 0-20000-0-02836; 67.34∘ N, 26.63∘ E; 179 m a.s.l.). In Table 1, the basic technical data of the AS41 Autosonde are reported. More details on the specifications of the Vaisala AS41 Autosonde can be found in the data sheet (B211636EN-A_2 pages.pdf) available on the Vaisala website (https://www.vaisala.com, last access: 3 July 2020).

The Meteomodem ARL is an automatic balloon launcher system that can perform up to 12 or 24 soundings without any manual control (http://www.Meteomodem.com/docs/en/Leaflet-robotsonde.pdf, last access: 3 July 2020). The system is compatible with M10 and M20 Meteomodem radiosonde types. It is built in a robust dry maritime container and composed of the following subsystems (Fig. 3):

operator room with electronic control unit and PC workstation, isolated from the launch tube by an air-tight safety door, and used only during radiosonde setup and restocking;

For each launch, there is a preparation phase which comprises the radiosonde GC and the loading of the balloon train (with the radiosonde, the unwinder, the parachute and the balloon) into individual bins before finally sounding parameters (e.g. launch time schedule, inflation volume) are set up.

During the launch phase, before powering on the sonde, the system performs a scan of the bandwidth in order to detect possible radio interference, then the radiosonde battery pack is powered on through an infrared link. According to the scan result, the system sets up the new frequency through an infrared link, and GNSS signal collection is initialized. Then, the system loads the calibration data of the relevant radiosonde stored during the preparation phase and checks consistency with PTU criteria. The Meteomodem ARL GC is a standard Meteomodem GC which consists of a sealed box enclosing a reference and a fan which homogenizes the inside temperature and relative humidity. It is recommended to return the Meteomodem GC every 3 years for calibration. The calibration is made with a certified Rotronic HC2A-S probe (https://www.rotronic.com/en/hc2a-s.html, last access: 3 July 2020).

Then, the ARL records the ground-check data and the metadata. Balloon inflation starts accordingly: the system monitors a flowmeter to inflate the balloon to the specified volume. The ARL may use either helium or hydrogen gas. Finally, the balloon is released at the specified launch time. In the event of launch failure before balloon release or during the flight, the procedure will restart for a new sounding immediately or can alternatively be manually launched according to a preset time schedule. At any time, an immediate start of the launch procedure can be initiated by an operator (locally or remotely).

For those stations operating an ARL and adopting a protocol based on GRUAN recommendations (Dirksen et al., 2014), as at Trappes station (WIGOS station identifier 0-20000-0-07145; 48.77∘ N, 2.02∘ E; 168 m a.s.l.; top panel of Fig. 3a), the GRUAN M10 ground-check procedure is performed in two steps: 5 min in a ventilated hut in ambient conditions together with calibrated T and RH sensors and, further, another 5 min to test the radiosonde performance in the SHC. Then each radiosonde is loaded in the ARL carousel (Fig. 3b).

A technical document describing the M10 sensor, corrections and uncertainties for both the temperature and relative humidity sensors will become available through the GRUAN community as soon as a Meteomodem M10 GRUAN data product is available.

The Meisei ARL, named “automated radiosonde system” is designed for fail-safe operation and high remote operability. Compared to the previous version developed in 2006, the new system, still under improvement, is able to load more radiosondes thanks to the development of the Meisei “canister type”. The operator can pre-load a maximum number of 40 sondes in the so-called “canister modules”. The canister has been recently implemented to reduce failures. Once the launch procedure has started, the respective canister fills a balloon independently. The right canister module and the left canister module are independent systems. It realizes high observation continuity by duplicating gas, air and electric systems. The canister module on one side can be moved to the preparation room to load the sonde and facilitate the operator's work. The new ARL version can also recover from balloon bursts without human intervention at the site by using a balloon from another canister. In the previous version, an operator had to visit the ARL to remove broken balloons and restart the ARL during the observation window in such cases.

The new system is also equipped with a new simplified wind shield for launches in strong wind conditions. All information and data are stored in a database available for each ARL. Various central monitoring/control functions are provided by using application software and a web browser to access the database on the workstation installed in the ARL. The Meisei ARL GC consists of a temperature and humidity reference sensor and an inspection box. The GC is performed before the sonde loading. The results from the GC are not used in the data processing but only to check if there are anomalies in the radiosondes.

In Table 3, the Meisei automated radiosonde system specifications are provided. Figure 4 shows a photo of the system along with a sketch of the interior of the system container. For more details on the Meisei ARL experimental setup, visit the Meisei website (http://www.meisei.jp/ars, last access: 3 July 2020). The Japan Meteorological Agency (JMA) has used Meisei ARL data since 2006. Parallel radiosoundings of auto launch and manual launch have not been done yet. This is the reason why this paper does not show additional datasets or comparisons involving Meisei ARL; therefore, the description of the Meisei ARL is the only information which can be shared with readers, according to recommendations provided by Meisei.

The performance of the Vaisala ARL has been evaluated through the analysis of a dataset collected at Sodankylä station. The Sodankylä Vaisala ARL was used to regularly launch RS92 radiosondes at 11:30 and 23:30 UTC over 2006 to 2012. Manual soundings were periodically performed in parallel using a similar Vaisala DigiCora-3 sounding system throughout this period. Parallel soundings have been selected with launch time difference between 2 and 20 min. A total of 283 parallel soundings have been considered: these are distributed evenly across the period, with the exception of 2006, which has more parallel soundings than other years, and most of these are daytime comparisons. In addition, two Vaisala ARL datasets from the Potenza GRUAN station (40.60∘ N, 15.72∘ E; 760 m a.s.l.) and the Minamidaitōjima station, run by JMA (WIGOS station identifier 0-20000-0-47945; 25.79∘ N, 131.22∘ E; 15 m a.s.l.), covering a similar time period, though much smaller sample sizes than in Sodankylä, have been used for comparison. Despite the less intensive sampling, Potenza and Minamidaitōjima data are useful data sources to compare with Sodankylä and, specifically, to check consistency of the GC correction across different stations and different batches of Vaisala sondes.

The availability of long time series of parallel sounding for the Sodankylä station permits investigation of the system performance also in the pre-launch phase. Two main aspects are evaluated: stability of the ground-check correction on temperature and potential effects related to the time periods the sondes were stored in before launch.

Figure 7 summarizes the temperature correction applied during the GC procedure for the RS92 sondes of the above-described datasets using the Vaisala GC25 ground-check device, with most of the launches performed since 2006. Figure 7 shows similar GC values at Sodankylä, Potenza and Minamidaitōjima stations despite the very different locations and launch scheduling, with a negative adjustment of between smaller than -0.5 K before 2010 and smaller than -0.3 K typically applied to most of the RS92 sondes with an improvement of the differences over the time in the batches launched after 2009. The results shown in Fig. 7 are based on the assumption that all the reported ARL GC temperature sensors were maintained according to recommendations described in the previous section.

Results similar to those from Sodankylä and Potenza GRUAN stations are reported by Payerne GRUAN station (Fig. 8) using the RS41 since April 2018 and operating the Vaisala AS15 ARL. Figure 8 shows that the distribution of temperature and relative humidity corrections have negative skewness with the GC adjustments within a few tenths of a degree, and the average adjustment is smaller than 0.1 K and 0.1 % RH, respectively. These results show an average negative GC corrections for the ARL in analogy to the results reported above for RS92 sondes at Sodankylä and Potenza, where also the old Vaisala ARL version was operated. Comparisons with the broader statistics collected by GRUAN stations launching manually (not shown) reveal results consistent with the GC time series shown in Figs. 7 and 8, thus excluding the presence of clear systematic effects in the GC corrections due to the use of ARLs. Nevertheless, the small differences observed between the ARL and manual GC corrections warrant further investigation to understand if performing the GC in a controlled temperature and humidity environment may generally improve or worsen the calibration in the long term.

In an operational station like Sodankylä, the time between balloon loading and ground check can vary from day to day. At Sodankylä, average loading time was 2–3 d prior to launch for regular soundings. The ARL software allows also longer times in the tray. Figure 9 shows, at different altitude ranges, the mean differences of simultaneous RH profiles (top panel) measured using the ARL and the manual soundings as a function of the number of days a sonde stays on a tray before launch, from 1 to more than 5 d. The corresponding mean standard deviations are also shown (bottom panel), while in brackets within the colour legend, the number of parallel soundings for each time period is reported. To calculate the statistics shown in Sects. 4 and 5, radiosounding temperature and RH from parallel soundings have been interpolated to a 100 m vertical grid. Figure 9 shows that there are no RH systematic differences when parallel launches are grouped according to the tray time, except for the launches with a tray time of 5 d or more at altitude levels above 6 km a.g.l., where a mean difference smaller than -2.0% RH is obtained up to 10–12 km a.g.l. Nevertheless, it must be noted that the size of the sample investigated for these tray time options (5 and > 5 d) is much smaller than for other tray times, and these launches also include parallel sounding with longer differences in the respective balloon release time.

To test if the estimated RH differences are meaningful, the Wilcoxon rank sum test has been applied. This test is a non-parametric test of the null hypothesis that it is equally likely that a randomly selected value from one population will be less than or greater than a randomly selected value from a second population. If the null hypothesis is rejected, then there is evidence that the medians of the two populations differ. In this study, the Wilcoxon rank sum test has been used instead of the z test because of its robustness in the event of a small observations sample (i.e. small number of parallel launches) and to avoid assumptions on the underlying data distribution (e.g. data distribution skewed or non-normal). For the RH profiles reported in Fig. 9, the probability computed using the Wilcoxon rank sum test ranges within 0.4–0.5, with smaller values only above 12 km a.g.l., where the probability becomes greater than 0.2. For the time-in-tray classes with a smaller sample of parallel soundings (1, 5 and > 5 d), the probability oscillates between 0.05 and 0.10. Therefore, it is possible to conclude that we cannot reject the hypothesis that the two data distributions (ARL and manual launches) have the same median value and the reported comparisons are consistent. Finally, the bottom panel of Fig. 9 shows that the standard deviations are substantially smaller than 5 % RH at all altitude levels without any evident correlation with tray time.

In Fig. 10, another way to study GC data is presented for the Payerne station. In this case, the average difference and the standard deviation of temperature and relative humidity found during the GC using Vaisala RS41 radiosondes into the Vaisala AS15 versus the ageing (up to 9 d into tray from the loading until launch) are shown. For both temperature and relative humidity, excluding only the launches which occurred within 24 h of the radiosonde loading, the bias is negative and independent of any further ageing. Until 1 d after loading, the bias is stable close to zero and thereafter it increases to about -0.14 K and -0.1 % over the following days. These results show how the use of ARLs also in remote places or where it is required to upload in advance a large number of radiosondes, to launch with a few days of delay, does not appreciably lead to changes in the Vaisala GC.

The performance of the Meteomodem ARL ground check has been evaluated through the analysis of a dataset collected at the Météo-France Trappes station, where M10 radiosondes have been launched regularly at 11:30 and 23:30 UTC since 2016. The availability of a long time series for the comparison between M10 temperature and humidity sensor and a reference temperature/humidity sensor (Vaisala HMP110; https://www.vaisala.com/sites/default/files/documents/HMP110-Datasheet-B210852EN_1.pdf, last access: 3 July 2020) at ambient conditions, inside a meteorological shelter for the Trappes station, permits the investigation of the system performance also in the pre-launch phase. Since June 2018, this comparison has been carried out during the 5 min before each automatic sounding. Figure 11 summarizes the time series and PDF of the difference between M10 and HMP110 sensor for temperature (black curve, upper panel) and relative humidity (blue curve, lower panel) recorded between June 2018 and June 2019. The relative humidity difference oscillates around 0 % and in more than 75 % of the cases the difference is smaller than 2 % RH in absolute value. For temperature, the observed residual difference around 0.5 ∘C requires further investigation.

Figure 12 provides a picture of the meteorological shelter and the position of the HMP110 and the M10 during the 5 min comparison shown in Fig. 11. These results need further investigation in order to determine if the systematic difference observed on temperature in the meteorological shelter is due to the Meteomodem M10 batches produced in 2018, though Meteomodem did not report similar systematic differences during the production checks, or if this could be due to the need for improvements in the experimental protocol. The meteorological shelter has been improved with the installation of a fan (Fig. 12), which should produce a better homogenization of the temperature and relative humidity around the two sensors. The development of a new experimental protocol is under consideration and should lead to the production of a tube ventilated by a laminar flow in which the Meteomodem M10 and a PTU reference could measure under the same environment, elucidating further upon the characterization of the spatial homogeneity of the temperature and relative humidity.

Finally, the M10 radiosonde is put inside a SHC chamber for 3 min before the sounding (with a relative humidity near 100 %); more than 95 % of the samplings are accepted after the test. For operational reasons, the Meteomodem probes used in the GRUAN protocol are tested in the meteorological shelter and in the 100 % RH test but not necessarily in this order each time. It is not known if the order of the checks makes any difference.

In Fig. 13, the statistics of the balloon vertical velocity and of the burst altitude for Sodankylä in the period from 2006 to 2012 are shown. In terms of vertical velocity (Fig. 13a), the ARL has a quasi-symmetric frequency distribution peaked around 5.3 m s-1 with a spread mainly between 4.7 and 5.9 m s-1. For the manual launches, the frequency distribution is quite wide, non-symmetric, peaked around 4.5 m s-1 with a larger spread of the values mainly between 3.5 and 5.7 m s-1. The comparison reveals the higher stability of the ARL compared to manual launches in controlling the balloon filling and therefore the sounding vertical velocity which is relevant for the quality of the measured profile. For the balloon burst altitude (Fig. 13b), a like-for-like comparison between the manual launches and the ARL is not feasible at Sodankylä due to the use of different balloon types (typically smaller for the ARL) which causes a strong difference in balloon altitude. Totex Tx800 or Tx600 types of balloons were used in winter and Totex Ta350 or Tx350 types of sounding balloons were flown during all other seasons. Due to smaller balloon volume, the summertime soundings had lower burst heights on average. The burst altitude for the ARL has also in this case a quasi-symmetric frequency distribution, which peaked around 25 km altitude a.g.l. with a spread of the values mainly between 17 km and 28 km a.g.l., while the distribution for manual launches is non-symmetric, with a maximum frequency around 33 km and most of values ranging within 21–35 km a.g.l. Differences between nighttime and daytime soundings were not significant, although nighttime soundings have on average lower burst heights during polar vortex overhead conditions in winter.

A more interesting comparison to show the eventual positive influence of automation on the burst altitude is those related to the dataset discussed in Sect. 3 and summarized in Table 5, shared by Météo-France for Trappes station (Fig. 14). In terms of vertical velocity (Fig. 14a), both the ARL and the manual launches have a quasi-symmetric frequency distribution peaked around 5.1 and 5.5 m s-1, respectively, with a similar spread of about 1.0 m s-1. For the burst altitude (Fig. 14b), we have for both the datasets a negatively skewed distribution with an evident peak around 33 km for the manual launches and 35 km for the ARL. The comparison reveals that the burst altitude (Fig. 14b) is generally higher for the ARL than for the manual launches, likely due to use of different balloons and the more limited human contact with the balloon which hence likely retains greater structural integrity. ARL frequency distribution has also a more peaked distribution that can be related to a more homogeneous balloon inflation (automatic inflation, same method, constant gas flow, more stable temperature). Furthermore, the vertical velocity of the balloon is stable (Fig. 14a). Overall, 40 % of the balloons burst before 30 km during the manual period, where only 20 % do during the automatic period. This result means that the Meteomodem ARL and/or the operational procedures, elaborated under a joint effort by Meteomodem and Météo-France, has increased by a factor of 2 compared to the number of balloons reaching an altitude higher than 30 km. The burst altitude for both periods (2012–2014 for the manual launches and 2016–2018 for the ARL) shows some seasonal signal. It appears that burst altitude is lower during the winter. A further study could evaluate burst altitude as a function of air temperature or potential vorticity in order to study the influence of polar vortex and its potential impact on the burst altitude.

In this section, two datasets are investigated to assess the differences in the vertical profiles of temperature and humidity: the set of RS92 parallel (automatic and manual) soundings performed with the automatic radiosonde launchers at Sodankylä, along with a second set of Meteomodem radiosoundings collected at Faa'a station, French Polynesia. These are near-coincident launches but the instruments are on physically distinct balloons which, as they ascend, likely at somewhat different rates if the balloons are not filled identically, will follow subtly distinct pathways leading to offsets in sampling. In the following analysis, given the latitude φ, the longitude λ, the Earth's radius R (mean radius of 6371 km), the distance between two balloons (1 and 2) has been calculated using the “haversine” formula (Sheppard and Soule, 1922) which provides the great-circle distance between two points (i.e. shortest distance over the Earth's surface): d=Rc, where c=2atan⁡2a,(1-a)a=sin⁡2Δλ2+cos⁡φ1cos⁡φ2sin⁡2Δλ2. The haversine formula remains particularly well conditioned for numerical computation even at small distances – unlike calculations based on the spherical law of cosines. The function “atan2” is described in Glisson (2011).

The two datasets are also investigated to show the correlation between the difference in the vertical profiles and the distance between the two flying sondes.

For the same 6-year dataset collected at Sodankylä discussed in Sect. 4, the vertical profiles of the average differences (automatic minus manual) and standard deviations of the temperature and RH measured during parallel soundings are shown in Fig. 15. Systematic differences in the temperature profile are negligible (on average smaller than 0.01 K) over the entire vertical range up to 25 km a.g.l., while the standard deviation increases with altitude from values smaller than ±0.5 K below 15 km to values larger than 1 K above. The result is in agreement with the increase in mean distance between near-simultaneous sonde paths at higher altitudes (Fig. 16). A subset of the parallel temperature soundings at Sodankylä has previously been analysed by Sofieva et al. (2008). Even though it is hard to separate difference components from non-colocation from those which may arise from instrument-to-instrument differences (e.g. arising from manufacture variations and differences in preparation, storage and launch at the uppermost altitudes), Sofieva et al. found differences in small scale structures in temperature profiles, when the horizontal separation was larger than 20 km. Moreover, to investigate whether the ARL and the manual radiosoundings datasets were selected from populations having the same distribution, i.e. if the calculated mean differences are statistically significant, the Wilcoxon rank sum test has been applied. The test result confirms that the two datasets are samples of the same population, showing a probability larger than 0.5 for temperature at all the altitude levels below 20 km and larger than 0.1 above, while for RH values the probability is larger than 0.3 over the entire range from the surface to 15 km a.g.l.

For the RH mean difference profile (Fig. 15b), there are no significant systematic differences up to 7 km and then again above 10 km a.g.l., while in between these altitudes a small negative mean difference lower than 1 % RH is found and may be related to the RH variability in the upper troposphere and the increased distance between the two sondes. The increase in standard deviation in the lower troposphere below 5 km a.g.l., with values generally smaller than 5 % RH, is due to the high RH variability which can be significant even for small horizontal distances between the two sondes. Above 5 km, and continuing through the profile to the upper troposphere/lower stratosphere (UT/LS) where the values of RH are on average smaller and less variable, RH difference decreases except when clouds or other uncommon events are detected (e.g. stratospheric–tropospheric exchanges).

In addition, the analysis was re-run after grouping the ARL flights according to the time a sonde had been loaded to the launcher system (see Sect. 4); variations of time period between sonde loading and actual launch time did not influence the comparison results.

Finally, the Wilcoxon rank sum test has been applied to the entire dataset, and the computed probability that the two samples belong to the same population is larger than 0.35 at all altitude levels.

A first evaluation of the performance of Meteomodem ARL is provided by the analysis of the datasets collected over 3–14 October 2018 at Faa'a station (French Polynesia, station identifier 0-20000-0-91938; 17.63∘ S, 149.84∘ W; 21 m a.s.l.), where 21 launches (9 daytime and 12 nighttime) of parallel radiosoundings have been undertaken (a picture is provided in Fig. 17) in order both to compare temperature, relative humidity, wind speed and direction, and to study further characteristics of the flights (burst altitude, ascent speed, for example). Météo-France has conducted the intensive operational period, while Institut Pierre Simon Laplace (IPSL) has produced the NetCDF files (data and metadata) for the analysis. Raw data without any correction for temperature and relative humidity have been considered in this paper. The GRUAN data processing, which remains under development at the present time for this data stream, has not been applied. The manufacturer Meteomodem IR2010 software was used for both manual and automatic launches.

The dataset collected by Météo-France at Faa'a station is not sufficiently large to draw robust statistical inferences. Nevertheless, this dataset is the first ever available to evaluate the performance of the Meteomodem ARL and can provide useful indications of any likely impact upon the data quality of ARL facilities.

Before comparing, the T and RH profiles of the parallel sounding dataset have been interpolated to a resolution of 100 m altitude. The difference between the launch time of the ARL and the manual balloons ranges between 1 and 12 s.

In Fig. 18, the horizontal distance between the pairs of parallel soundings at all the altitude levels up to 25 km a.g.l. is shown; the horizontal distance between the two balloons is typically within about 35 km.

In Fig. 19, the mean difference between the set of ARL and manual parallel soundings profiles of temperature and RH as a function of altitude regardless of time mismatch, along with the corresponding standard deviation is shown. Figure 19a shows the difference for temperature, while Fig. 19b shows it for RH. The mean temperature difference is smaller than ±0.2 K up to 12–13 km a.g.l. and typically smaller than ±0.5 K above. The difference is negative, up to -2.0 K, in the first 50–100 m, and this is probably due to the potential warming effect of the ARL environment on the radiosonde sensor.

For RH, the mean difference is instead always positive and smaller than 0.7 % RH up to 8 km a.g.l. with a standard deviation smaller than 3 %–4 % RH. Above 8 km, the mean difference becomes larger and less variable with a maximum of about 2 % RH and a standard deviation around 3 %. The Wilcoxon rank sum test has been applied to both temperature and RH. For temperature, the probability is higher than 0.3 until 17 km and higher than 0.2 above, while for RH is larger than 0.2 below 10 km and larger than 0.1 above. Only in the first 40 m for temperature and the first 20 m for RH, the Wilcoxon rank sum test fails with a probability lower than 0.05. The results of the test confirm the null hypothesis of the same median for the ARL and manual data distribution at all the height levels for both temperature and RH, with the only exception of a few decametres above the ground because of the ARL air-conditioned effect. The reason behind this bias could arise from GC effects or differences in the pre-launch procedures between the two systems affecting the performance of one of the two launches in a quasi-systematic manner throughout the vertical profile. This will be further investigated with the support of the manufacturer.

In terms of balloon burst altitude, the ARL proved to be reliable both during the daytime with a burst altitude ranging within 26 688–31 904 m a.g.l.) versus values within 24 970–30 621 m  a.g.l. calculated for the manual launches, while during nighttime the burst altitude ranges within 27 587–30 790 m a.g.l. for the automatic launcher versus values within 27 437–30 139 m a.g.l. for the manual launches. Applying the Wilcoxon rank sum test, the computed probability (0.05224) for the entire dataset is slightly greater than the 0.05 significance level, and therefore the two distributions of burst altitudes are not significantly different, indicating that ARL does not lead to significant improvements in the balloon burst altitude.