Hungarian Drone-Based Wind Measurements During the WMO UAS Demonstration Campaign—A Low-Level Jet Case Study

Highlights

What are the main findings?

• A purpose-built meteorological UAV can derive vertical wind profiles from orientation data with accuracy meeting WMO OSCAR operational thresholds when evaluated against radiosonde measurements.
• Low-level jet conditions provide a stringent real-world stress test and demonstrate that UAV profiling resolves sharp vertical wind gradients that are often smoothed in standard radiosonde products.

What are the implications of the main findings?

• UAV-based wind profiling offers a practical and mobile solution for observing the data-sparse boundary-layer region, where conventional measurements remain limited.
• Dedicated meteorological UAV platforms enable high-resolution, operational wind monitoring under demanding atmospheric conditions and can complement radiosondes in future multi-platform observing systems.

Abstract

This study presents an operational approach to atmospheric wind profiling using a purpose-built meteorological uncrewed aerial vehicle (UAV) and an orientation-based wind estimation method that does not rely on dedicated onboard anemometers. The quadrotor platform, designed and developed by our team, has a maximum take-off mass of 2.45 kg and is capable of acquiring vertical atmospheric profiles up to 3000 m under a wide range of weather conditions. Within the framework of the World Meteorological Organization’s (WMO) global demonstration campaign for evaluating the use of uncrewed aircraft systems in operational meteorology and associated field activities, twelve vertical wind profiles were collected in parallel with radiosonde observations. UAV-based wind estimates were evaluated against radiosonde data using the WMO OSCAR (Observing Systems Capability Analysis and Review) performance framework. Across most wind speed regimes, the central 50% of UAV–radiosonde wind speed differences remain within OSCAR threshold requirements, indicating operationally relevant accuracy. Systematic deviations are physically interpretable and arise primarily in strongly sheared boundary-layer flows. A representative low-level jet case is used as a stress test, demonstrating that the UAV system remains safe and that wind estimates remain reliable even under extreme wind conditions, supporting robust performance in less demanding regimes. These results establish UAV-based wind profiling as a viable and complementary observing technique in the lower atmosphere and provide a practical pathway toward high-resolution, operational boundary-layer wind measurements.

1. Introduction

There is a growing demand for highly accurate weather forecasts in surface and air transportation, energy production and precision agriculture [1,2,3,4]. However, numerical weather prediction models can only deliver such forecasts if they are constrained by sufficiently dense and high-quality observational data from both the surface and the atmosphere. Therefore, continuous sampling of the planetary boundary layer (PBL)—the lowest part of the atmosphere directly influenced by the Earth’s surface—is essential, as most weather-related processes originate in this zone. Despite its importance, the availability of spatially and temporally detailed observations that describe the vertical structure of the PBL remains limited [5,6,7].

Several established observing techniques aim to address this data gap, including radiosounding, remote sensing methods and aircraft-based measurements such as the Tropospheric Airborne Meteorological Data Reporting (TAMDAR) system. However, these approaches are often expensive, lack operational mobility and do not provide spatially collocated, high-resolution vertical profiles. For example, routine radiosonde operations are cost-intensive: launching at least four sondes per day at a single site results in annual expenditures of approximately 150,000–200,000 EUR. In addition, most sondes are not recovered, leading to environmental contamination by electronic waste. Remote-sensing systems, while valuable, are limited by high acquisition costs and fixed-site deployment. Aircraft-based measurements are constrained by their spatial focus on airport environments and by temporal availability governed by flight schedules [8,9,10,11].

To overcome these limitations, a growing number of solutions have been developed for PBL measurements using uncrewed aircraft systems (UASs). An overview of early small fixed-wing meteorological UAV platforms has been provided in [12]. Since then, several new UAS concepts capable of measuring temperature, humidity and wind parameters have been developed and deployed in both research and operational contexts [13,14,15,16].

In Hungary, our research group has been investigating the structure of the PBL using meteorological UAVs for more than a decade. In earlier studies, we primarily employed fixed-wing UAS platforms to perform atmospheric profiling along pre-programmed flight trajectories at multiple altitude levels and within a horizontal range of approximately 10 km [17]. These observations were assimilated in near real time into a numerical weather prediction model operated in a dedicated containerized computing environment [18].

Throughout this development project, particular emphasis was placed on meeting the prerequisites of the ultimate operational goal of UAS-based measurements, namely the provision of observational data suitable for nowcasting and short-range numerical weather prediction. As part of this effort, we developed a containerized, WRF-based modeling environment capable of performing data assimilation experiments using the three-dimensional variational (3DVAR) technique. This framework also supports Forecast Sensitivity to Observation Impact (FSOI) analyses, enabling the quantitative assessment of the potential contribution of UAV-derived observations to forecast skill.

Low-level jets (LLJs) are pronounced local maxima in wind speed occurring near the surface, typically within the 100–500 m above ground level (AGL) layer [19], resulting in non-logarithmic vertical wind speed profiles. Investigation of these jets began in the mid-20th century [20,21], and they remain an active research topic, as illustrated by the recent development of a global LLJ climatology based on ERA5 reanalysis [22]. LLJs significantly affect aviation [23], wind energy production [24,25,26] and the horizontal and vertical transport of atmospheric pollutants [27], and they can enhance extreme precipitation through vertically integrated moisture transport [28]. LLJs arise from multiple physical mechanisms, including nocturnal inertial oscillations following boundary-layer decoupling [20,29,30,31], frictional decoupling during warm-air advection over colder surfaces [32], synoptic-scale baroclinicity near extratropical cyclones [33,34] and shallow baroclinicity induced by differential heating over coastal or sloping terrain [35].

Despite decades of research, LLJ events are identified using a wide range of criteria, typically based on deviations from an idealized logarithmic wind speed profile. Detection algorithms employ absolute wind speed maxima and differences above and below the jet core [21], minimum thresholds for wind speed decrease [36,37], percentage-based criteria relative to the maximum, or combinations of these approaches [38,39]. In the present study, the selected case is evaluated using several of these established definitions.

In the present work, we focus on deriving vertical profiles of wind direction and speed using a UAV system designed, manufactured and operated by our team. The UAV-based wind estimates are evaluated through direct comparison with radiosonde observations. Rather than aiming at a climatological characterization of LLJs, we use them as a representative and dynamically demanding test case to demonstrate the accuracy and operational potential of the method. Low-level jets are characterized by strong winds and sharp vertical gradients; therefore, if the UAV system performs reliably under such extreme boundary-layer conditions, robust performance can reasonably be expected in less demanding atmospheric regimes as well. In particular, we highlight the capability of UAV-based wind profiling to resolve low-level jet structures, benefiting from its inherently higher vertical sampling resolution.

These UAV-based atmospheric measurements were conducted within the framework of the World Meteorological Organization’s (WMO) global demonstration campaign for evaluating the use of uncrewed aircraft systems in operational meteorology, a global initiative aimed at showcasing the potential of uncrewed aircraft systems for operational atmospheric data collection [40].

2. Preparation for WMO UAS Demonstration Campaign

2.1. Development of the Meteorological Measurement Prototype UAV

The custom meteorological UAV is a purpose-built quadrotor platform with a maximum take off mass of 2.45 kg designed to carry sensors for temperature and humidity measurements up to altitudes below 3000 m AGL (Figure 1). To extend its operational envelope, the platform was designed for flight under adverse environmental conditions, including wind speeds up to 20 m/s, ambient temperatures from −20 °C to 45 °C and exposure to precipitation and dust. Accordingly, the airframe and onboard systems were engineered to meet IP55 protection standards. Structural components, connectors and PA12 enclosure elements satisfy the corresponding water and dust resistance requirements, enabling repeated operation under harsh meteorological conditions.

During measurements, the UAV can carry up to eight environmental sensors connected via an I2C interface to an onboard computer. Meteorological variables are sampled at 1 Hz, while inertial and attitude data from the flight controller’s IMU (Inertial Measurement Unit) are recorded at higher internal rates and synchronized with the environmental measurements. Wind estimates are derived from ground calibration combined with inertial measurements. Vertical profiling can be performed at controlled ascent and descent rates, typically 6–10 m/s during climb and 2–8 m/s during descent, providing high vertical resolution while maintaining stable flight conditions.

Achieving the required altitude and endurance necessitated extensive optimization of the propulsion system, including the evaluation of multiple motor–controller–propeller combinations. The platform is operated using the ArduPilot (Copter-4.5.2) autopilot system, initially on an OrangeCube controller and later on an OrangeCube+ unit (CubePilot Global Pty. Ltd., Norlane, Australia).

Early prototypes employed an Newyenk 8000 mAh 4S1P lithium-polymer battery (Dongguan Xuanli Electronics Co. Ltd, Dongguan City, China). Subsequent iterations adopted a Yangda 22,000 mAh solid-state battery (Shenzhen Yangda Security Co. Ltd., Shenzen, China), increasing capacity by 150% with only a 50% mass penalty and approximately doubling flight endurance. In parallel, the propulsion system evolved from an 11-inch propeller with a KDE 2315-KV885 motor (KDE Direct LLC, Bend, OR, USA) to a 14-inch propeller with a Sunnysky 4006-KV600 motor (Zhongshan LangYu Model Co., Zhuhai City, China), yielding substantial efficiency gains.

The UAV structure was designed for additive manufacturing. Initial FDM-based PETG-CF components exhibited increased fatigue at low temperatures. Later iterations therefore adopted PA12 parts produced by Selective Laser Sintering, allowing thinner walls and reducing enclosure mass by 30–50 g. Additional geometric refinements minimized turbulence around the arms, further reducing mechanical fatigue. The resulting prototype fleet formed the technical basis of the UAV system deployed during the WMO’s demonstration campaign.

2.2. Data Gathering On-Board

2.2.1. Wind Estimation

Unlike radiosondes, UAV platforms allow wind speed to be measured directly using properly installed and calibrated onboard thermal [41] or sonic [42] anemometers. The main challenges of this approach are that the UAV’s own motions, its control responses and the flow disturbances generated by the propulsion system are all reflected in the sensor signal [43]. These effects must either be explicitly modeled during data evaluation or mitigated through dedicated filtering techniques. In order to obtain a first-order estimate of the ambient wind field from UAV telemetry data without relying on dedicated anemometric sensors, a set of simplifying assumptions is introduced below.

Let $\underline{GS}$ denote the velocity vector of the airframe in the Earth-fixed coordinate system (ground speed), $\underline{AS}$ be the velocity vector of the airframe relative to the surrounding air (true airspeed) and $\underline{WS}$ be the velocity vector of the ambient atmosphere, including possible vertical up- and downdrafts (wind speed). Under stationary flight conditions ($\underline{GS}=\mathrm{constant}$), $\underline{AS}$ is determined by the balance between (i) the aerodynamic forces acting on the vehicle (primarily drag), (ii) the thrust generated by the propulsion system and (iii) gravity. Drag depends on $\underline{AS}$; the horizontal component of thrust balances the horizontal component of drag, while the vertical component of thrust balances the vehicle’s weight together with the vertical component of drag. The kinematic relationship between these quantities follows directly from vector addition: the ground speed equals the true airspeed corrected for the ambient wind field,

$$\underline{GS}=\underline{AS}+\underline{WS}.$$

(1)

While $\underline{GS}$ and $\underline{WS}$ are defined in the Earth-fixed Cartesian coordinate system $(x,y,z)$, the vector $\underline{AS}$ is defined in the body-fixed coordinate system $(x^{\prime },y^{\prime },z^{\prime })$. In the Earth coordinate system, the x axis points eastward, the y axis points northward and the z axis is oriented opposite to the geopotential gradient vector (i.e., upward). In the body coordinate system, $x^{\prime }$ points along the longitudinal axis toward the nose of the UAV, $y^{\prime }$ points to the right along the lateral axis and $z^{\prime }$ points downward, orthogonal to the $x^{\prime }$–$y^{\prime }$ plane, defining the perpendicular axis of the airframe. See Figure 2 for a schematic representation of the reference systems.

Let t denote the tilt of the UAV, defined as the angle between the z and $z^{\prime }$ axes. In a general attitude, the airframe is rotated by an angle $\mathrm{\Theta }$ about its $x^{\prime }$ axis (roll) and by an angle $\mathrm{\Phi }$ about its $y^{\prime }$ axis (pitch). If the UAV’s $x^{\prime }$–$y^{\prime }$ plane is horizontal, i.e., parallel to the Earth-fixed x–y plane, then $t=0$ and the thrust vector is purely vertical. In this case, the horizontal component of the true airspeed $\underline{AS}$ is zero and the horizontal component of the ground speed $\underline{GS}$ coincides with the horizontal component of the ambient wind speed $\underline{WS}$, as the airframe is passively advected by the wind. For a tilted orientation, when the airframe’s horizontal plane is no longer parallel to the Earth’s geopotential surface, the thrust vector acquires a horizontal component and the UAV attains a non-zero horizontal airspeed. If $\underline{AS}$ is measured directly by an onboard anemometer, the wind vector can be obtained by vector subtraction,

$$\underline{WS}=\underline{GS}-\underline{AS}.$$

(2)

Air speed can also be inferred from UAV telemetry data—specifically from ground speed, attitude (orientation) and throttle—provided that the aerodynamic characteristics of the airframe are known. In practice, however, determining these characteristics and solving the associated force-balance equations is complex and computationally demanding. To obtain a tractable operational method, we therefore derived a simplified empirical relationship that estimates the horizontal component of the true airspeed from the vehicle’s orientation. This approach exploits the strong correlation between airframe tilt and the horizontal aerodynamic force required to maintain a stationary ground position during vertical profiling. The resulting formulation provides an efficient approximation of horizontal airspeed based solely on attitude information:

$$A{S}_h=A{S}_h\left(t\right).$$

(3)

In the absence of access to a suitable wind tunnel for controlled wind speed experiments, the initial calibration flights were conducted in the vicinity of a calibrated TriSonica Mini ultrasonic anemometer (Anemoment LLC, Louisville, CO, USA) mounted on a 6 m mast. During these tests, the UAV was flown in a quasi-stationary configuration close to the sensor. This arrangement provided only a rough calibration, because the vertical component of the drone’s $\underline{GS}$ had to be set to zero: no sufficiently elevated reference wind sensor was available to permit safe vertical ascent flights. Nevertheless, the experiments revealed a clear and approximately linear relationship between the horizontal airspeed $A{S}_h$ and the tilt angle of the airframe for wind speeds in the range of 1–5 m/s.

The second calibration step was based on processing tilt data obtained during flights along slanted trajectories with varying $\underline{GS}$, designed to emulate the conditions encountered during vertical profiling under windy environments. These calibration flights were conducted in still air or under weak wind conditions and in the absence of significant updrafts or downdrafts. Only flight segments for which the estimated wind speed remained below 5 m/s were retained for analysis, ensuring that the relationship between tilt angle and wind speed could be treated as approximately linear. The resulting flight pattern consisted of six characteristic stages, as illustrated in Figure 3.

First, the UAV traverses a fixed-altitude path between two predefined locations separated by at least 200 m (A), followed by the same trajectory in the opposite direction (B). In the third phase, the vehicle ascends along a slanted path between the two locations (C) using a climb rate comparable to that applied during operational profiling and then performs a vertical descent back to the initial altitude (D). Subsequently, the UAV ascends again along a slanted path in the opposite direction (E) with the same climb rate and finally returns to the initial location and altitude (F) before landing.

The calibration flights were conducted with fixed horizontal $\underline{GS}$ along all directions within each mission in order to eliminate the influence of horizontal $\underline{WS}$ on the vehicle’s tilt. The calibration was repeated for several prescribed horizontal $\underline{GS}$ values, providing an empirical basis for the relationship between tilt and horizontal wind speed. The selected calibration speeds were $GS=5$, 10 and 15 m/s, as well as the maximum attainable airspeed of the platform. Only quasi-stationary flight segments were retained for the analysis; data points were discarded whenever the actual horizontal $GS$ deviated by more than 5% from the prescribed value. For each calibration speed, tilt data were averaged separately for the horizontal legs A and B (with zero vertical velocity, $w=0$ m/s) and for the slanted ascent legs C and E (with non-zero vertical velocity, $w=6$ m/s), as illustrated in Figure 4.

For the purpose of calibrating tilt values against horizontal airspeed, a total of 15 calibration flights were conducted with different preset ground-speed values ranging from 5 to 20 m/s. Altogether, 1512 data pairs of tilt angle (t) and horizontal airspeed ($A{S}_h$) were obtained. Using the fit functionality of the GNUPLOT software (Version 5.4 patchlevel 1, modified 1 December 2020), the power-law function.

$$f\left(t\right)=a{t}^b,$$

(4)

was fitted to the dataset. The optimization converged after 15 iterations, yielding stable estimates for the parameters a and b. Fitting the function $A{S}_h\left(t\right)$ to the measured tilt values resulted in:

$$A{S}_h\left(t\right)=a{t}^b,$$

(5)

where $a=2.49$ and $b=0.61$. During the operational profile measurements performed with a climb rate of 6 m/s, this empirical relationship was used to derive horizontal wind speed from the orientation of the airframe. The horizontal airspeed was expressed as a function of the effective tilt angle t, computed from the Euler angles (pitch $\mathrm{\Phi }$ and roll $\mathrm{\Theta }$), as

$$t={cos}^{-1}(cos\left(\mathrm{\Phi }\right)\times cos\left(\mathrm{\Theta }\right)).$$

(6)

In addition, the horizontal projection of the thrust vector onto the Earth’s x–y plane defines the instantaneous flight direction of the UAV. This direction, denoted by d, is expressed relative to the body-frame longitudinal axis $x^{\prime }$. Assuming small angles and steady flight, d can be derived from the Euler angles as

$$d={tan}^{-1}\left({\displaystyle \frac{sin\left(\mathrm{\Theta }\right)}{tan\left(\mathrm{\Phi }\right)}}\right),$$

(7)

where $\mathrm{\Phi }$ and $\mathrm{\Theta }$ denote the pitch and roll angles, respectively.

The angle d represents the offset between the drone’s heading and its actual flight path relative to the surrounding airmass. Based on this relationship, the wind direction $wd$ in the Earth coordinate system can be expressed as

$$wd=H+d+\mathrm{\Delta },$$

(8)

where H denotes the heading of the UAV, i.e., the geographical direction of the body-frame longitudinal axis $x^{\prime }$ projected onto the Earth’s x–y plane, and $\mathrm{\Delta }$ is the local magnetic declination at the measurement location. This formulation yields the meteorological wind direction in geographic coordinates.

It should be noted that during real measurement flights, the vertical component of the wind speed may be significant in the turbulent planetary boundary layer and the horizontal component of the drag force may not be isotropic. Consequently, a more sophisticated estimation scheme is required, based on telemetry data from vertical ascending flights near elevated wind sensors on high towers and incorporating additional flight parameters such as throttle. For this reason, the calibration database was analyzed using a data-driven Artificial Intelligence (AI) framework, here understood as supervised machine-learning-based regression. Linear and logistic regression models were applied using the same input variables as in the classical least-squares approach and were found to reproduce the corresponding results. In the present study, the AI framework therefore serves primarily as a validation and baseline tool; its role is to provide a flexible structure that enables the systematic inclusion of additional telemetry parameters—such as throttle or dynamic attitude metrics—in future model extensions. Data-driven regression and machine-learning-based approaches are increasingly applied in wind estimation and forecasting, ranging from linear regression models to more complex nonlinear and state-estimation techniques [44,45,46].

At present, the calibration is based on low-altitude, near-surface flights; therefore, wind speeds inferred from pitch angle at higher altitudes are closer to the indicated airspeed than to the true airspeed. At the typical operational profiling altitude of 2500 m AGL, reduced air density introduces an estimated bias of about 14%. Addressing this altitude dependence, together with the integration of additional telemetry variables, is a key objective of future model development to obtain a more physically consistent and accurate wind estimation scheme.

2.2.2. Profiling Missions

As discussed earlier, the measuring UAV is specifically designed for meteorological profiling tasks. The nominal design altitude is 3000 m AGL; however, under favorable conditions, the platform can safely reach 4500 m AGL while maintaining continuous data transmission, demonstrating a substantial operational safety margin.

Because rotary-wing UAVs face inherent technological constraints during vertical descent, two alternative profile flight strategies were implemented:

• A vertical descent at 2.5 m/s while maintaining a fixed GPS position
• A 6 m/s descent along a spiral trajectory with a radius of 30 m and a tangential velocity of 8 m/s.

The selected parameter values, based on extensive flight experience (more than 200 missions), provide a favorable compromise between vehicle stability, mission endurance and power consumption, while fulfilling the requirements on temporal resolution and vertical sampling density for meteorological profiling.

With a climb rate of 6 m/s, a 2500 m AGL profile requires approximately 16 min when using spiral descent and about 25 min when using a vertical descent at 2.5 m/s. Owing to the reduced mission duration and the more efficient flight geometry, the spiral descent mode results in substantially lower power consumption than straight vertical positioning, making it the preferred operational configuration.

3. Results of the Parallel Wind Measurements Performed by UAV and Radiosounding

3.1. The Measurement Data and Method Used for Comparative Analysis

Within the framework of the WMO’s global demonstration campaign for evaluating the use of UAS in operational meteorology, opportunities arose to perform wind measurements in parallel with the Hungarian Defence Forces’ (HDF) Vaisala mobile radiosounding system. During the 42nd CISM (International Military Sports Council) Military Parachuting World Championship, comparative measurements were performed on two consecutive mornings (9–10 July 2024) at the Szolnok-Szandaszőlős airfield (47.1466° N, 20.1997° E; ICAO code: LHSS), Hungary, yielding a total of four atmospheric profiles. Additional profiling sessions were conducted on 30 July and 7 November 2024 near Hajmáskér and Bakonykúti (47.1652° N, 18.0404° E and 47.2483° N, 18.1768° E, respectively) as part of HDF training activities. Furthermore, during the 25th FAI (Fédération Aéronautique Internationale) World Hot Air Balloon Championship held in Szeged, three flights coincided radiosonde launches at the upper-air station of Szeged Airport (46.2517° N, 20.0965° E; ICAO code: LHUD). Altogether, this resulted in 12 wind profiles extending to altitudes of up to 4500 m AGL, which could be directly compared with radiosonde observations. Profiles were included in the analysis when UAV measurements were obtained within a ±30 min time window around the radiosonde launch.

The reference data were provided by Vaisala RS41-SGM radiosondes (Helsinki, Finland), processed using Vaisala MARWIN MW32 Sounding System (Helsinki, Finland) integrated into the TM-12 METEO HU system of the HDF and the Hungarian National Weather Service. These standard products offer a temporal resolution of 1 s throughout the ascent.

The comparison was conducted within the framework of the WMO OSCAR (Observing Systems Capability Analysis and Review) initiative [47], which provides a standardized, application-oriented reference system for evaluating the performance of meteorological observing technologies. OSCAR is designed to support the objective intercomparison of different measurement methods against operational requirements in a uniform and internationally consistent manner.

OSCAR defines specific performance benchmarks for observational data. The threshold represents the minimum acceptable standard to ensure data usability, while the goal specifies the level at which further improvements provide no additional value for a specific application. The breakthrough is an intermediate target between these two levels; achieving it significantly enhances application effectiveness and often represents an optimal cost–benefit balance for system design. Uncertainty is defined by the WMO as the estimated range of observation errors for a given variable, expressed as a 68% confidence interval (1$\sigma$). In the context of near-surface wind measurements, the OSCAR threshold, breakthrough and goal levels correspond to uncertainties of 2.0 m/s, 1.5 m/s and 0.5 m/s, respectively. For wind direction, the corresponding performance levels are defined as 10°, 5° and 1°. By adopting this framework, the present study evaluates UAV-based wind profiling not only in a relative sense, but against internationally recognized operational performance benchmarks. In some cases, uncertainty was calculated as the root mean square (RMS) of the differences between UAV and radiosonde measurements.

Following the methodology of [14], all profiles were interpolated onto a standardized vertical reference grid with 10 m spacing. This was achieved by averaging all original data points within ±5 m of each reference level.

3.2. The Comparative Analysis of Wind Measurements

To comprehensively evaluate the accuracy of wind speed measurements obtained by the UAV, we analyzed the absolute differences between wind speeds measured by the UAV and those recorded by the radiosonde, grouped into 1 m/s wind speed intervals (Figure 5). For each interval, the median and the interquartile range (first to third quartile) of the deviations are shown. The interquartile range therefore represents the central 50% of all paired observations within a given wind speed interval.

Our results indicate that, across most wind speed ranges, the interquartile range of the deviations remains within the WMO OSCAR threshold level for wind speed. This implies that half of all UAV-based wind speed estimates differ from the radiosonde reference by less than the threshold value. Consequently, most of the measurements can be regarded as sufficiently accurate for applications such as the initialization and constraint of numerical weather prediction models.

The differences in wind directions measured by the UAV and the radiosonde were analyzed in the same manner. At low wind speeds (<2 m/s), wind direction exhibits pronounced spatial and temporal variability, which explains the broader interquartile ranges observed in this regime (Figure 5). In addition, previous studies have shown that radiosonde-based wind measurements may not adequately resolve small-scale vertical variations in wind speed and direction due to several inherent error sources in their sensing and processing techniques [48].

In the wind speed range above 2 m/s, the spread of wind-direction differences between the first and third quartiles decreases. This reflects the reduction in spatial variability as the wind speed increases. At the same time, an additional source of discrepancy arises from the fundamentally different sampling geometries of the two platforms. While the UAV maintains a fixed GPS position during profiling, the radiosonde drifts with the flow, leading to an increasing horizontal separation between the instruments. This growing distance contributes to the observed differences in both wind speed and wind direction.

4. A Case Study: The Low-Level Jet

On 9 July 2024, the weather in eastern Hungary was influenced by the south–southeastern flow of an anticyclone located to the north–northeast. Overnight, the cloud band associated with a weak cold front that had produced evening storms moved away, leaving clear skies over Szolnok by 00 UTC. Despite 4–7 kt northeasterly winds, the temperature at 2 m dropped to 20.2 °C, resulting in a pronounced nocturnal inversion of approximately 4 °C within the lowest 400 m (Figure 6). At LHSS, two profile measurements were conducted: a 500 m test flight at 03:23 UTC and a 2500 m flight at 04:03 UTC, both compared with radiosonde data.

Figure 6 highlights the key structural features of this low-level jet event. In both profiles, a distinct wind speed maximum is observed at 210–250 m AGL, where wind speeds exceed 12 m/s (12.8 and 13.3 m/s, respectively), identifying the core of the LLJ. In contrast, near-surface winds remain weak (1.1 and 3.4 m/s), consistent with the stable stratification imposed by the nocturnal inversion. Above the jet core, wind speeds decrease markedly, delineating the upper boundary of the LLJ layer. At 500 m AGL the first profile records 7.6 m/s, representing the lowest value above the jet core, while in the second profile a pronounced minimum of 4.5 m/s is reached at 755 m AGL. The presence of a strong low-level maximum followed by a substantial reduction in wind speed above satisfies multiple commonly used LLJ definitions.

According to the first level of the classification system proposed by [21], an LLJ is identified if the maximum wind speed reaches at least 12 m/s and is followed by a decrease of at least 6 m/s. This criterion is fully satisfied in the second profile. In the first case, the observed structure also suggests compliance; however, the 500 m limitation of the profile height prevents direct confirmation of the required post-maximum decrease. In addition, the criterion of [38], which requires a reduction of at least 20% from the maximum wind speed, is met in both profiles. Furthermore, both measurements fall within the LLJ categories defined by [39]. In the first profile wind speed decreases by more than 40% from the maximum value (corresponding to at least 4 m/s), while in the second case a reduction exceeding 50% (at least 5 m/s) is observed, fulfilling even the strictest LLJ definitions.

Using this LLJ event as an example, the comparison of UAV-based and radiosonde-based wind profiles illustrates a key advantage of meteorological UAV measurements: their ability to resolve fine vertical structures. Taking the radiosonde as a reference, RMS deviations of wind speed and wind direction were calculated in 5 m/s wind speed intervals. The profiles were grouped according to the presence or absence of an LLJ.

Table 1. RMS deviations of wind speed (WS) and wind direction (WD) between UAV and radiosonde measurements in the presence and absence of a low-level jet (LLJ), grouped by wind speed ranges. The largest increase in RMS deviation in LLJ conditions occurs in the 5–10 m/s wind speed range, corresponding to the altitude interval of the jet core.

	0–5 m/s	5–10 m/s	10–15 m/s
RMS deviation (WS)—with LLJ	1.19 m/s	1.71 m/s	1.10 m/s.
RMS deviation (WD)—with LLJ	17.94°	15.12°	13.94°
RMS deviation (WS)—without LLJ	1.16 m/s	1.43 m/s	1.25 m/s
RMS deviation (WD)—without LLJ	23.34°	15.43°	11.31°

summarizes the RMS deviations between UAV- and radiosonde-based wind measurements calculated from all 12 flight missions conducted during the campaign. The two profiles obtained on 9 July 2024, during a well-developed low-level jet event, constitute the “LLJ” category. All remaining missions, during which no low-level jet was present, form the “non-LLJ” group.

Figure 7 shows the vertical wind speed profiles measured by the UAV and the radiosonde for the representative LLJ case of 9 July 2024, highlighting the differences within the lowest approximately 600 m, where the jet is present. The magnitude of such differences is quantified in a statistical sense by the RMS values summarized in Table 1, which are derived from all 12 flight missions.

Within the 5–10 m/s wind speed range, the average wind speed deviation between the UAV and the radiosonde is 1.28 m/s larger in LLJ cases than in non-LLJ situations. In contrast, within the 0–5 m/s range, the difference is only 0.03 m/s, while above 10 m/s, the deviation increases by 0.15 m/s when no LLJ is present.

This distribution reveals a systematic pattern. In the presence of an LLJ, the largest discrepancies occur precisely in the wind speed range associated with the jet core and its sharp vertical gradients. By contrast, at low wind speeds (0–5 m/s), where vertical shear is weak and spatial variability dominates, the RMS differences between LLJ and non-LLJ cases become negligible. These results indicate that the enhanced deviations are not random measurement noise, but arise from the physical structure of the flow and from the fundamentally different sampling geometries of the UAV and the drifting radiosonde.

As discussed in Section 3.2, the spatial variability of wind direction decreases with increasing wind speed. Consequently, within the high-wind core of the LLJ, the wind direction measured by the drifting radiosonde remains in good agreement with that observed by the meteorological UAV. However, during the ascent, the radiosonde gradually separates from the launch site. By the time it reaches the altitude of the first wind speed minimum above the LLJ core (755 m AGL), it has drifted approximately 890 m horizontally. This separation continues to increase, exceeding 1800 m at 2500 m AGL. Above the LLJ layer, where wind speeds are lower and spatial variability is higher, this growing horizontal distance results in increasing discrepancies in both wind speed and wind direction between the two measurement systems.

The larger discrepancies observed in LLJ conditions can therefore be attributed to the fundamentally different measurement principles. Although radiosondes provide good nominal vertical resolution, their wind data are commonly low-pass-filtered by manufacturers to suppress artefacts caused by irregular sonde motion. As a result, standard radiosonde wind observations are reliable only on vertical scales larger than approximately 300 m [48]. Fine-scale vertical structures, such as the sharp wind speed gradients associated with LLJs, are therefore difficult to capture accurately with radiosondes, whereas UAV-based measurements are able to resolve them directly.

These results indicate that direct comparison of UAV-based wind profiles with radiosonde observations does not always provide an optimal verification framework. This limitation arises not only because the UAV maintains a fixed GPS position while the radiosonde drifts away from its release point, but also because standard radiosonde wind data lack the vertical resolution required to resolve fine-scale wind structures.

Future work will therefore focus on calibration experiments conducted in the vicinity of meteorological towers and, at higher altitudes, on the use of remote-sensing systems such as LIDAR [11]. In addition, the regression-based wind estimation scheme will be further refined by incorporating additional flight and system parameters (e.g., throttle settings), potentially supported by machine learning techniques, in order to enhance the accuracy and robustness of the method.

5. Summary

This study demonstrates the feasibility and added value of deriving vertical wind profiles from UAV orientation data as an alternative to conventional onboard wind sensors. Based on twelve parallel UAV–radiosonde profile pairs acquired during the WMO’s global demonstration campaign for evaluating the use of UAS in operational meteorology and associated training activities, we provide a quantitative assessment of the accuracy and limitations of this approach in the lower and middle planetary boundary layer.

A key prerequisite of this capability is the dedicated meteorological UAV platform developed for the campaign. The applied UAV, designed, manufactured and developed by the authors, follows a quadrotor configuration with a maximum take-off mass of 2.45 kg and is capable of acquiring vertical atmospheric profiles up to 3000 m AGL under a wide range of weather conditions. The system integrates flight control, power management and environmental sensing in a compact, weather-resistant airframe designed for repeated vertical profiling under operational conditions. Stable ascent and descent rates, reproducible trajectories and continuous recording of orientation and control parameters form the physical basis for wind estimation without dedicated anemometric hardware. This hardware–software co-design distinguishes the presented system from generic UAV applications and establishes a reproducible framework for meteorological deployment.

The comparative analysis shows that, across most wind speed regimes, the central 50% of UAV–radiosonde wind speed differences remain within the WMO OSCAR threshold requirements. This indicates that UAV-based wind profiling already meets operational performance levels for a wide range of applications, including the initialization and constraint of numerical weather prediction models. Systematic deviations are physically interpretable and emerge primarily in strongly sheared boundary-layer flows.

The low-level jet case was deliberately selected as a stress test, representing an extreme wind regime with sharp vertical gradients. Its analysis illustrates a key advantage of UAV-based profiling: the ability to resolve fine-scale vertical wind structures that are often smoothed in standard radiosonde products. In LLJ conditions, RMS deviations peak in the 5–10 m/s wind speed range corresponding to the jet core, while differences remain small at low wind speeds and outside the jet layer. This pattern reflects both intrinsic flow structure and the fundamentally different sampling geometries of a stationary UAV and a drifting radiosonde and highlights the limited effective vertical resolution of standard radiosonde wind data in strongly sheared layers. The fact that the UAV system remains safe and the wind estimates remain consistent under such conditions supports the robustness of the approach in less demanding regimes.

Beyond the present validation, the study establishes a development pathway for UAV-based wind profiling. The current regression framework already yields operationally meaningful accuracy, but can be further improved by incorporating additional flight-state parameters (e.g., throttle, attitude dynamics) and machine-learning-based calibration strategies. Planned intercomparisons with meteorological towers and remote sensing systems such as Doppler LIDAR will provide physically co-located reference data, enabling a more rigorous separation of methodological and atmospheric effects.

Together, these results position UAV-based wind profiling as a complementary observing technique rather than a surrogate for radiosondes. By combining a purpose-built meteorological UAV platform with high-resolution in situ profiling and adaptive calibration strategies, UAV-based systems have the potential to become an integral component of future multi-platform atmospheric observing networks, particularly in the critical 100–500 m layer where conventional observations remain sparse.