LUCA was designed to operate in mid-latitude and polar conditions. Therefore, the expected environmental constraints were evaluated from the radiosonde stations Neumayer in Antarctica and Lindenberg in Germany (Fig. ) episodically over 3 years (2016–2018). The temperature range covering 90 % of the operating conditions of the ascents at Lindenberg is between -60 and 20 ∘C and at Neumayer is between -75 and -2 ∘C. For the Lindenberg station, the median wind speed is up to 20 m s-1, and the wind speed that is encountered in 90 % of the cases (90th percentile) is up to 42 m s-1 for the altitude of the jet stream (7–15 km) . For the Neumayer station, the median wind speed is up to 14 m s-1, and the wind speed that can be expected in 90 % of the cases is up to 28 m s-1.

Operational challenges at Neumayer arise from the surface wind speed, which is generally higher than in Europe. The most frequent surface wind conditions are either from the east due to cyclonic activities near the polar front, with a typical wind speed of 20 m s-1, or from the south due to katabatic flow, with a typical wind speed of 10 m s-1 . High wind speed values are further reached at the altitude of the tropopause, around 9–13 km in the polar jet stream .

The LUCA system was designed for operation in the temperature range between -75 and +30 ∘C and for a wind speed of less than 28 m s-1 over the whole vertical profile to limit the maximum vector displacement of the drone from its launch site. Assuming the system is capable of operating in conditions exceeding the design temperature and nominal wind speed by 15 %, this would allow ascents in 87 % of the radiosonde days at Neumayer in the Antarctic and 72 % at Lindenberg. Restricting wind speed conditions such that the drone does not have to stay above the launch point but must be able to return to the base, the mean wind speed over the atmospheric profile to be observed by the drone should not exceed the nominal horizontal airspeed component of 28 m s-1. Despite the limited applicability of this approach when using the drone in a reserved airspace and ignoring possible environmental threats such as heavy precipitation or in-flight icing, the system is capable of covering 94 % of the measurements in Lindenberg and 98 % at the Neumayer station. Adding a margin of 15 %, the availability increases to 97 % for Lindenberg and 100 % for Neumayer. At the Neumayer station, on 96 % of the days, radiosondes can typically be launched. Lindenberg has a temporal coverage of 99 % of all days. In addition, the development of the drone addresses measurements during rainfall, snow and heavy turbulence and within clouds, but these capabilities have yet to be proven in upcoming measurement campaigns.

Particular attention was paid to takeoff and landing under high-surface-wind conditions, which are accompanied at the Neumayer station by low visibility due to drifting and blowing snow. The probability of 23 % to operate under conditions with a visibility below 500 m requires a highly automated takeoff and landing procedure that does not rely on visual contact of the operator with the system, similarly to operations shown by .

A possible threat for drone measurements comprises in-flight icing conditions, which depend on temperature, humidity and droplet size . An idea of the frequency of icing conditions might be available from icing forecast data using ADWICE (Advanced Diagnosis and Warning System for Aircraft Icing Environments; ) along with validation studies using PIREPs (pilot reports). Such comparisons exist for regions with dense air traffic, e.g. Europe , but no validation studies are found for the Antarctic. In addition, icing differs strongly between crewed aircraft (what ADWICE is made for) and drones . A report for Norway and its surrounding regions explicitly focused on icing for drones using meteorological reanalysis data (ECMWF ERA5; ) and an ice accretion model (ICE3D; ), and it found theoretical icing frequencies of 45 % at an altitude between 1–1.5 km between September and May and a lower risk with the highest frequency of 30 % peaking at 2.5 km altitude in June to August. These values are not directly applicable to the Antarctic, but the process to determine the likelihood of icing might be used to preliminarily estimate icing frequency and icing risk for drones in the Antarctic.

Drones are usually operated without sophisticated anti-icing and de-icing concepts like in crewed aviation, in particular as most drones are operated in visual line of sight. However, icing is a threat that may lead to a complete loss of the system. Icing protection for drones has been demonstrated but requires additional substantial energy for heating, even if combined with specific ice-phobic coatings or liquids .

For the demonstration flights shown here, the LUCA drone was prepared to be equipped with an icing sensor to measure in situations with a substantial risk of meeting icing conditions during the flight, but the sensor was not installed during the demonstration flights as zero risk of icing was present. Together with monitoring performance parameters, this allows for estimating the severity of icing and supports the decision of abandoning the mission in icing cases. More details about the all-weather strategy can be found in .

Before designing the physical aircraft system, the mission to be accomplished by the drone was defined. Although data assimilation is nowadays highly flexible regarding time, the mission was designed according to radiosonde observations to hypothetically surpass the 100 hPa surface at 12:00 UTC and ensure timeliness in anticipation of future in-flight reporting analogue to radiosonde reporting, where a first dataset is sent to the GTS when the sonde reaches the 100 hPa level . With a targeted climb rate of 10 m s-1 arbitrarily chosen from simple analytical estimates, the drone has to be launched around 11:33 UTC, and it reaches the design ceiling of 10 km at 11:50 UTC. During the flight, 1 Hz real-time data are available. After descending with a vertical rate similar to the climb rate, an approach procedure is flown in the vicinity of the landing site to determine wind direction and wind speed, and subsequently the approach trajectory is calculated automatically on the onboard computer. After the “splashdown” into a horizontal landing net, the drone can be recovered, and data processing including quality checks and transcoding begins. The observations of the complete flight are finally transferred into the GTS around 13:00 UTC. Thus, there is a target of making post-processed data from the descent profile available within an hour of the measurements being taken. Figure  illustrates the mission design.

As key design driving parameters, the ability to fly against high-speed winds, an efficient electric propulsion chain and a highly flight-state-independent position for the sensor package are essential. These requirements led to the development of a tailless fixed-wing configuration with a pusher propeller. As the result of a multi-variant optimisation for profiling the atmosphere vertically up to 10 km, the design weighs 5–6 kg, depending on the deployed sensor package; has a wingspan of less than 2 m; and operates at a constant airspeed of 30 m s-1 with a typical ascent rate of 10 m s-1 controlled by the autopilot system. Thus, the system has a minimum airspeed of 18 m s-1, exceeding the minimum airspeed of crewed ultralight aircraft. Subsequently, it is not feasible to launch the drone by hand, and a dedicated mechanical catapult was adapted and used during the measurement campaign (Fig. a). As automatic takeoff and landing are required for future operations in zero-visibility conditions, a horizontal landing net has been developed and an appropriate manoeuvre to get the drone safely into the net was implemented in the autopilot firmware (ArduPilot). The manoeuvre itself can be considered an automated vertical dive into the horizontally arranged net. While a belly landing is principally possible, the net-landing technique was preferred as it protects the drone and the sensors from the hostile surface in the Antarctic, and prevents the system from being blown away and lost in low-visibility conditions of drifting and blowing snow in the case of high near-surface wind speed, as occurs frequently in the Antarctic. For the launch and retrieval system, no specific operation limits are in place regarding wind speed, as long as the launch and the landing is performed against the wind direction. Preliminary operating limits are therefore bound to the nominal airspeed of the drone. To address the risk of disintegration in turbulence, the airframe was constructed to resist gusts according to EASA Certification Specifications 23.333 . On the avionics side, the systems on the drone were widely predefined by national regulations (e.g. the redundancy of the command and control link). reveal more technical details of the drone and its ground systems, which are not within the scope of this article.

An overview of the sensors and their accuracy according to the data sheet is provided in Table . The placement of the sensors is based on experience with similar drone-based systems e.g., and the sensor behaviour and possibilities of correcting, for example, sensor response time are well known. The performance of sensors and the data quality are assessed directly from the atmospheric measurements, without including wind tunnel tests for the overall setup.

To enable the autopilot sensors to measure dynamic and static pressure, a total pressure port (pitot tube) and static pressure ports on both sides of the drone were implemented in the nose section of LUCA. Below the pitot tube, an air inlet for the closed sensing path for the temperature–humidity probe (T/RH probe inlet) is installed, as sensor installation is known to influence the measurements . The combined resistance temperature and capacitive humidity sensor HMP110 (Vaisala, Finland), providing measurements of temperature and relative humidity in the closed sensing path, was mounted within the bulkhead, which separates the sensor chamber from the fuselage area. Subsequent to passing by the sensing elements, the air perfuses the measurement chamber and is vented out through apertures in the bottom of the nose section (venting). This ensures well-defined pressure conditions at the sensor location that are close to total pressure while maintaining ventilation of the sensing element. The ventilation rate, which is assumed to be still higher than the ventilation rate of radiosonde sensors, minimises the response time of the sensors as the sensor is exposed to an increased amount of air over time compared to radiosondes. The enclosure furthermore protects the sensors from damage due to ground contact, even during rough landings. An illustration of the sensor installation used for the measurement flights is provided in Fig. a, where the measurement nose of the drone is shown. As the body of the combined temperature and humidity probe is made of stainless steel and is mounted through the bulkhead between the sensor chamber and the fuselage area, thermal conduction into the sensor chamber which might affect the measurements is expected in the case of temperature differences between the ambient air temperature and the temperature inside the fuselage, as mentioned in, for example, . Additionally, the response times of the humidity measurements are expected to increase further, as the “wetted” area inside the sensor chamber is not well vented. Besides the sensors and pressure ports in the nose section, measurements of static and dynamic pressure, attitude angles, the Earth-related position, and velocities – all taken by the autopilot system (Cube Orange, HexAero, Singapore) – were recorded to derive atmospheric variables. As a drawback, the inertial navigation algorithm running on the autopilot system to calculate attitude angles, the Earth-related position and velocities relies on industrial-grade GNSS (global navigation satellite system), rotation rate, acceleration and magnetic field sensors, which limits the accuracy of the location and attitude estimation and subsequently limits the accuracy of the wind calculation, varying with the flight trajectory. Particularly, heading information might be adversely affected by electromagnetic interference originating from the power train, which is more pronounced during the ascent. Additionally, the estimation filter in the inertial navigation algorithm is not able to observe sensor (e.g. turn rate sensor) errors well during flight phases with low-trajectory dynamics. Measured parameters including data sheet uncertainties for both the LUCA drone and the radiosonde launched for data intercomparison are summarised in Table .

As an example for a flight trajectory, the measurement flight on 28 October 2021 at 07:20 UTC is shown in Fig. b. The trajectory during the measurement flight consists mainly of circular patterns and therefore provides trajectory dynamics to facilitate the calculation of aircraft attitude, velocity and position performed by the inertial navigation algorithm.

A cut-out in the left wing is foreseen to be used for an icing sensor as shown in , but other instruments can be fitted into it. As the risk of in-flight icing was negligible during the measurements, a camera was installed into the cut-out. The camera captured video and audio during the measurements, see Fig. b, and helped to analyse the behaviour of the flight controller and the motor controller of the drone.

The most important aspects for drone flights relate to safety. Operational requirements are different for each region of the world; however, in Europe, there are now unified regulations for drone operation . The requirements concerning redundancy and the level of integrity of the drone depend on the overall risk analysis. For relatively small and lightweight drones performing flights beyond visual line of sight, which is the case for any system operating up to an altitude of 10 km, the drone falls into the category “specific”, and precautions to avoid damage to a third party have to be met. An operational handbook includes, for example, regular checks and maintenance of vital parts like motors and propellers, redundancy in the electric system, regular training of the crew, independent control links, and many more aspects defined in the so-called operational safety objectives. The flight tests were therefore done in restricted military areas, in this case close to the Baltic Sea, Germany, and the internationally recognised high-quality portable radiosonde system of Graw, Germany , was deployed from the launching site for direct comparison. In the restricted areas, cooperation with the German Federal Armed Forces is required, and flight permission of the German Federal Supervisory Authority for Air Navigation Services (Bundesaufsichtsamt für Flugsicherung, BAF) is necessary.

For the operation in the Antarctic, a thorough risk assessment was performed with particular emphasis on safety and redundancy to avoid damage to third parties even in such a sparsely populated area. Further, for the deployment in the Antarctic, permission of the German Environmental Agency (Umweltbundesamt) is required, ensuring that no material stays in the pristine environment and that the penguin population near the Neumayer station is not disturbed.

For a comparison of data obtained by LUCA and radiosonde data, a procedure similar to what is presented in is applied to the data. In a first step, data within pressure bands of 2 hPa are found for drone observations and radiosonde observations. Assuming a constant wind situation at the particular altitude associated with the pressure band, the air parcel measured by the radiosonde is shifted with the wind according to the time difference between the drone measurement and the radiosonde measurement. For example, if the drone was measuring at 09:20 UTC within the pressure level of 500 hPa and the radiosonde was measuring at 09:30 UTC at the same pressure level, the radiosonde position would be shifted backwards by the distance an air parcel would have been travelling within this 10 min period in constant wind at this particular pressure level. This leads to a “source” position of the air parcel measured by the radiosonde at a specific time – the time when the drone was measuring at the same pressure level (altitude). These virtual positions are then taken into account to select measurements collocated in space and time. Collocation conditions identical to the conditions used in (spatial distance less than 50 km, temporal gap less than 30 min) were then applied, before measurements were intercompared. No explicit filter for outliers was applied, but the drone data were implicitly filtered for outliers by median-filtering the data measured in each particular pressure band of 5 hPa width. Differences were computed for all pressure bands on all six flights.

LUCA represents a new type of small fixed-wing drone. The combination of a relatively small wingspan below 2 m and a total weight of more than 6 kg leads to a high minimum flight speed of 18 m s-1 at sea level for obtaining the lift required to fly and therefore complicates the process of getting airborne and landing safely compared to aircraft with larger wing areas related to the total mass. During the flight test and measurement campaign, the design altitude of 10 km was reached. Ascending to such altitudes can be regarded as unique for an electrically powered fixed-wing drone without solar panels. To the authors' knowledge, it is the first time that a meteorological drone powered only by electrical batteries reached the altitude of 10 km. In addition, the landing manoeuvre has been proven to be repeatable during test flights without manual control. A list of flights performed can be accessed in . During flight tests at sea level, a maximum airspeed of more than 60 m s-1 was reached, indicating that the drone is able to operate in equally high wind speed . For operations, a safety margin has been applied, and the designated horizontal wind speed limit for normal operations was set to 28 m s-1, which equals the nominal horizontal component of the true airspeed during the ascent. Drifting away from the measurement location is prevented by forcing the minimum ground speed to 2 m s-1. The flight controller automatically increases the airspeed when the minimum ground speed falls below the threshold because of high wind speed – potentially trading the climb rate for airspeed. Resisting high wind speed and the corresponding turbulence was demonstrated during the flight on 25 October 2021 starting at 12:12 UTC, where the maximum measured wind speed was 28 m s-1 during ascent and descent at an altitude of 7800 m a.m.s.l. (above mean sea level). Operations in BVLOS (beyond visual line of sight) conditions and in the presence of closed cloud layers have been conducted successfully.

The drone measurements can take place starting from any location, up to the altitude of 10 km, if flight permission is obtained. Operations are completely independent of additional infrastructure, like an airport, or the availability of helium. The drone only requires a cylindrical airspace with a radius of about 11 km for the whole mission. The drone ascends and descends at a vertical speed of 10 m s-1 and a horizontal speed component of 100 km h-1. Therefore, the mission reaching up to 10 km altitude and returning to the landing site takes in total 33 min. The data are available by remote transfer with a temporal resolution of 1 Hz. The full dataset with a resolution of up to 25 Hz can be downloaded after landing and is preprocessed automatically for upload into the GTS. The subsequent launch after landing can be typically performed after a turnaround time of 20 min.

A measurement campaign with LUCA and radiosondes was performed over the Baltic Sea (54.4∘ N, 10.6∘ E) from 25 to 29 October 2021. An overview of the development of the meteorological conditions with altitude is shown in Fig. , based on hourly ERA5 reanalysis data . The distribution of relative humidity with time and altitude was highly variable during the measurement period. Also, wind speed and wind direction (indicated with wind barbs) varied with height and time. The temporal distribution of the LUCA flights and parallel additional radiosondes are indicated in the overview (Fig. ). LUCA was tested during a time period with high relative humidity, corresponding to cloudy conditions, and with high wind speed of up to 50 m s-1, but the drone was not operated at the time of the maximum wind speed at 9 km altitude. The overall wind direction was from the west with surface wind speed around 10 m s-1 and an absence of precipitation. Low and high scattered cloud layers were present, and a broad jet stream was forecasted just in the north of Denmark, with the jet core expected at 400 km distance to the north of the measurement campaign location.

Exemplarily, Fig.  shows measurements of temperature, relative humidity, wind speed and wind direction recorded by LUCA at up to 10 km. The measurements took place on 26 October 2021 at 08:45 UTC and on 29 October 2021 at 07:22 UTC. The data of the nearly simultaneous radiosonde are shown, as well as ERA5 reanalysis data.

Each profile provided by LUCA shows qualitatively the same atmospheric structures as the radiosonde measurements with a similar sensor package to that used on the radiosondes.

On 26 October 2021, the strong temperature inversion at 400 m as well as the transition to a different temperature gradient at an altitude of around 6500 m is equally captured by the drone and the radiosonde measurements (Fig. a). On 29 October 2021, the temperature inversion at the top of the boundary layer at 400 m altitude is captured by the drone and the radiosonde, as well as ERA5 analyses.

Humidity was generally moderately variable during the observation period. The drone and radiosonde measurements, as well as ERA5 data, agree that there was high relative humidity (small differences between dew point temperature and temperature) up to the temperature inversion at around 400 m altitude on 26 October 2021 (Fig. a). On both days, there were layers of lower and higher humidity. On 26 October 2021, in the lower troposphere up to an altitude of around 4 km, the individual layers of different humidities are identically resolved by the drone in terms of altitude and magnitude (Fig. a). On 29 October 2021, humidity measurements using the drone are in good agreement with radiosonde data up to 500 hPa (around 5 km altitude, Fig. b). Above 500 hPa, an increasing deviation between radiosonde and drone measurements can be observed in humidity for both days. This is likely caused by the dramatically increasing response time of the humidity sensor at lower temperatures in addition to the delaying effect of the implemented closed-path sensor setup. Moisture is a critical parameter to measure in the upper troposphere .

Wind measurements agree well between the drone and the radiosonde measurements. When comparing the measurements directly, one has to take into account that the radiosonde was launched 53 min after the drone. A systematic deviation can be expected for higher altitudes due to the drifting of the radiosonde with the wind, towards the north-east, which resulted in a spatial distance of 40 km from the launch point when reaching the altitude of 10 km.

Despite the simplistic sensor setup, data measured on board the platform LUCA have been shown to compare at least qualitatively with radiosonde data. To assess quantitative measures, the methods presented in Sect.  are applied to the data of all six vertical profiles (upwards and downwards) post-processed according to Appendix . These techniques were applied for ascents as well as descents, as data during the ascent are expected to suffer from electromagnetic interference as well as other factors such as thrust line misalignment. The collocated dataset for the pressure bands of 2 hPa width from 1000 to 250 hPa consists of 694 data points for the ascent and 1323 data points for the descent of the drone. Figure  shows histograms of the differences between the radiosonde observations and the drone observations for the variables temperature, relative humidity, specific humidity, wind speed, wind direction as well as the norm of the wind vector difference for the descending profiles. Within the histograms, average differences as well as the standard deviation of the intercompared observations are provided.

Despite the sparse database consisting of six vertical descent profiles, the distributions of the differences for the variables of temperature, relative humidity, specific humidity, wind speed and wind direction are similar to Gaussian distributions. The histogram for the norm of the wind vector difference follows a Rayleigh distribution, as is expected for the norm of a two-dimensional vector whose components are stochastically independent Gaussian processes. For the distribution of the observation differences, the average deviation and the standard deviation (root-mean-square deviation for the Rayleigh distribution) were calculated and are shown in the histogram plots (Fig. ). These values were retrieved for both the ascent and the descent and are shown in Table .

Comparing the statistical measures between ascent and descent data, increased differences between the radiosonde and the drone observations are revealed. This is most likely associated with a magnetic deterioration or a possibly induced sideslip angle when the thrust line is not aligned with the drone's body (see Sect.  and Appendix ). For the descent profile, the average differences as well as the standard deviation of the differences between drone observations and radiosonde observations are in the range of (or even below) statistical measures shown in the study of , which includes a comparison of radiosonde data with observation data from the AMDAR–TAMDAR programme. The statistics furthermore support the values of the theoretical error estimation in Appendix . For the humidity measurement, the average of the differences between radiosonde and LUCA measurements differs significantly between ascent and descent, indicating the possible need for further calibration and tuning of the post-processing parameters using a broader database in the future.

The overall uncertainties for the measured profile in Fig. b are shown in Fig. . As the contribution of the time lag correction and the wind calculation to the total uncertainty depends on the vertical profile itself, the total uncertainty for the variables of air temperature, relative humidity, wind speed and wind direction is shown as a grey area around the measurements.

For the air temperature Ts in Fig. a, the radiosonde observations deviate significantly in the lower boundary layer, likely caused by the temperature sensor being heated up by solar irradiance without venting on the ground before the launch. No particular uncertainty source dominates the total uncertainty for the measurements. For humidity observations in Fig. b, increased uncertainty around rapid changes in humidity is visible, originating in the uncertainty caused by the time lag correction. The data sheet uncertainty dominates the total uncertainty, in which the calibration drift (up to ±2 % per year according to the data sheet) is not yet included. Radiosonde observations lie partly outside the uncertainty area (1σ plus systematic errors) for the drone measurements, potentially pointing to the need for further measurements and investigations. Wind observations in Fig. c and d compared to the radiosonde measurements are within reason regarding the uncertainty and taking into account turbulence as well as the spatio-temporal distance between the radiosonde and the drone.