The first scientific flight with the WinDarts took place on 18 September and the last on 26 September. Over the course of 9 d, 11 flights were conducted, with some days featuring more than one flight. An overview of all successful flights is presented in Table . In most flights except for flight 20220925.1335, the WinDart identified as WD1-1 was tethered above the WinDart labelled WD1-2, resulting in WD1-1 taking off first and landing last. The total duration of deployment across all flights amounted to almost 3 d and 10 h, 38 min, 10 s.

The WinDart's naming convention is DS.MPWD.b1.yyyymmdd.hhmm.nc, where:

DS is the institute identifier (Dynamics and Self-Organization).

The ground weather station measurements ranging from the 13 to 28 September are divided into daily data sets, with a new file starting every day at 00:00:00 UTC. Thus, resulting in 16 NetCDF files and 16 CSV files. The naming convention follows that of the WinDarts.

In Sect. , we will provide a complete description of the file structure, detailing how the data is organised within the NetCDF and CSV formats. Meanwhile, in the following section, we present an explanation of the data, using WD1-1 and flight 20220920.0750 as an example, as it includes one of the most complete measurements.

In this context, “Level 1 data” refers to the raw, minimally processed data, i.e. the parsing of data, synchronisation between different sensor timestamps, filling of missing values and generation of validity identifiers. Each data set and instrument underwent quality control and was standardised into a standard format for subsequent release and analysis. This format adheres to the Climate and Forecast (CF) convention for units and nomenclature. Detailed specifications and guidelines for the CF convention can be found in the official documentation at https://www.earthdata.nasa.gov/about/esdis/esco/standards-practices/climate-forecast-metadata-conventions (last access: 2 April 2026).

All measurements were synchronised to GNSS-derived Coordinated Universal Time (UTC). Likewise, all times in this manuscript are presented in UTC. During September, Finland observes Eastern European Summer Time (EEST), which is UTC+3, meaning local time was 3 h ahead of UTC during the campaign.

Each first-generation WinDart is equipped with a BeagleBone Blue (BB) single-board computer and an Arduino Mega 2560. The BB lacks a real-time clock; thus, the data-transmission timestamps (“log_time” in data files) are initially generated using Python’s perf_counter() function. The synchronised timestamps (“time” in data files) for each data record are calculated using linear interpolation within a data chunk. This estimates when a specific measurement occurred between the start and end times of its data transmission. Positional time series recorded by the GPS refer to GNSS-derived UTC directly and do not undergo any further synchronisation.

The ground weather station employed the same acquisition codebase as the MPCK⁺ (for more details, we refer the reader to ). The system clock was used to record events, and during post-processing, the corresponding timestamps were calibrated against GNSS-derived UTC timestamps using the onboard “GPS” sensor (see Table ). This was achieved by referencing all available GPS timestamps and applying a linear regression to correct for clock drift and offset. As all sensors within the system referenced the same internal clock, this method provided consistent synchronisation across all data streams. The resulting time alignment achieved sub-second accuracy, typically accurate to within a few milliseconds.

This synchronisation strategy results in consistent timestamping across all sensors within each WinDart and facilitates coordinated measurements with the ground weather station.

Instruments were calibrated before the campaign. Defective data were removed from further analysis and from this manuscript according to a set of a priori, uniformly applied quality control criteria. Data were marked for exclusion when any of the following conditions held: (i) parsing failures (checksum/format/length), (ii) timestamp issues, (iii) sensor dropouts or communication faults, and (iv) out-of-range values for the measured quantity. For WinDarts, explicit out-of-range limits for temperature, relative humidity, and pressure are stated in the NetCDF attributes for each quantity, together with per-sensor quality-flag variables denoted as, variable_name_qc_flag, where 0 indicates valid data, 1 denotes a parsing, timestamp, or communication fault, and 2 indicates values outside the specified valid range. The variable attributes also include the percentage of data excluded due to out-of-range values, i.e. quality-flag 2, which for most cases is 0 %. For all variables, samples affected by parsing failures, timestamp issues, or sensor dropouts/communication faults are stored as NaN. The attributes of the variables include the number of valid samples and the total sample count. For WinDart data, excluding GPS-related time series, the percentage of valid data is 100 % for all quantities. For ground-weather-station data, again excluding GPS/GNSS-related time series, the percentage of valid data is > 99 % for all time series. The GPS/GNSS-related time series are parsed from NMEA 0183 sentences, in which null fields may occur and indicate invalid data or that no data are available at a given epoch. During conversion to NetCDF, we have replaced null fields with fill values (i.e. NaN) and are counted as “invalid” data points, which explains the large number of invalid data points in GPS/GNSS-related time series. We did not observe a systematic dependence on environmental or operational conditions for parsing failures, timestamp issues, or sensor dropouts/communication faults. Possible systematic errors can be identified owing to the WinDarts' redundant measurement design, as illustrated and discussed in the examples below.

In cases where a device malfunctioned at some point after the beginning of the flight, we documented the failure in the metadata of the corresponding NetCDF file and indicated it in Table  by labelling it as “(failed)” or excluding it from data availability.

Figure  illustrates two flights with failures. The upper plot depicts flight 20220921.0716, where both WinDarts experienced battery failures mid-flight, resulting in the simultaneous failure of all instruments on each WinDart. For WD1-1 the failure occurred at 10:18 UTC and for WD1-2 at 11:18 UTC. The lower plot shows flight 20220926.0530, where the GPS on WD1-1 failed during the flight.

The label BMP indicates that altitude was computed from the barometric pressure measured by the BMP. The GPS on WD1-2 failed at the start of both flights.

Figure  presents the time series of pressure, temperature, relative humidity, and altitude recorded by all sensors onboard WD1-1 during flight 20220920.0750, along with the particle size concentration measured by the OPC. For pressure, temperature, and relative humidity, a reference sensor was selected based on the highest accuracy and resolution specified by the manufacturer (see Table ): BMP for pressure, TMP for temperature, and SHT (heater off) for relative humidity. These reference measurements are shown in black in the figure. To evaluate inter-sensor consistency, we calculated the offsets of each sensor relative to the corresponding reference, expressed as ΔT and ΔRH for temperature and relative humidity, respectively. The temporal evolution of these offsets is displayed in panels zoomed around zero, and the mean offset over the entire flight is indicated in the panel labels. Temperature and relative humidity sensors exhibited the most significant offsets. For temperature, where TMP was used as the reference, SHT and BME showed mean offsets of 0.2 and 0.4 K, respectively, both within the specified nominal accuracy (Table ). The heated SHT3 sensor displayed a positive mean offset of 3.6 K, as expected. The OPC temperature sensor, housed inside the WinDart body and OPC body itself, shows an elevated mean offset of 11.7 K. This offset was anticipated given the sensor location and instrument design. By contrast, the BMP sensor exhibited a mean offset of 2.4 K, which exceeds its specified accuracy and cannot currently be explained. For relative humidity, SHT was chosen as the reference (non-heated) sensor. The SHT3 (reference heated sensor) reported systematically lower humidity, with a mean offset of -16.9 %, consistent with expectations for heated sensors. The OPC also showed a strong negative mean offset (-29.2 %). To evaluate the accuracy of the reference choice, we computed a corrected relative humidity for all sensors using the mixing ratio as a conserved quantity. Specifically, the mixing ratio was calculated from the heated sensor's (e.g. SHT3) temperature and humidity measurements, and then used together with the TMP temperature and BMP pressure to recover a corrected relative humidity

To compute the mixing ratio we used metpy.calc.mixing_ratio_from_relative_humidity() and to compute the relative humidity we used metpy.calc.relative_humidity_from_mixing_ratio(), both from the MetPy package (version 1.7.1).

During several flights, the WinDarts entered cloud layers. Initial cloud penetration was noted by visual inspection during flight, although this method is inherently uncertain, and was later confirmed through particle concentration measurements. Figure  compares the corrected relative humidity with OPC particle concentrations for WD1-1 during flight 20220920.0750 between 08:15 and 09:50 UTC. The scatter plot includes only concentrations exceeding 20 counts per cm³, with data points color-coded by particle size bin from the smallest to the largest.

According to the flight log, cloud entry occurred at 09:03 UTC. However, Fig.  indicates an earlier transition, with an increase in particle concentrations across size bins beginning at approximately 08:30 UTC. This shift suggests the presence of cloud droplets within the OPC detection range, demonstrating the effect of cloud passage on the measured particle size distribution. Furthermore, no saturation or supersaturation was detected by either reference relative-humidity sensors, even though the OPC and flight log indicate entry into cloud. This discrepancy can be explained by the sensors' slow time response to short transients, degraded accuracy near saturation, and/or suboptimal ventilation in WD1. Future designs should mitigate these issues by adding active aspiration (e.g., a fan) to improve exposure and reduce response time.

Figure  illustrates the time series of the air velocity vector in the platform (WinDart) frame, denoted as U=(U,V,W), as measured by the SVM (five-pressure-channel Pitot tube, Vectoflow GmbH) without corrections for platform motion. The Pitot tube measures five differential pressures, which are then processed using Vectoflow Post-Processing libraries to obtain the wind velocity vector U. The Vectoflow function used the air temperature and air pressure time series recorded by the TMP and BMP sensors, respectively. At this stage, no additional post-processing techniques are applied to the velocity measurements. The components (U,V,W) represent the wind-direction, transversal-to-wind-direction, and vertical-direction components of U. Due to the fins installed in their tails (illustrated in Fig. ), the WinDarts align with the wind direction.

The derivation of the true wind vector from the measured flow requires correction for the platform's motion and orientation. The necessary steps are as follows:

Coordinate system transformation: The velocity vector measured by the SVM in the platform's body-fixed coordinate system must be rotated into an Earth-fixed reference frame (e.g., North-East-Down, NED). This transformation is performed using a rotation matrix (R) derived from the BNO1's orientation data (roll ϕ, pitch θ, heading ψ).1Umeasured, NED=R(ϕ,θ,ψ)⋅U

In this Level 1 data set, all information required to perform platform motion correction is provided. We explicitly note that the velocity data shown in Fig.  have not been corrected for motion, as such correction involves several choices regarding interpolation and filtering that depend on the specific use case, which are beyond the scope of a Level 1 dataset. However, more importantly, we believe that measurements from the BNO sensors do not provide a sufficiently reliable basis for the automatic correction of velocity data, as they occasionally exhibit drift (as shown in more detail below). In some time windows the velocity corrections based on BNO sensors may still be useful, provided the data are carefully checked. Nevertheless, Level 1 velocity measurements can be directly used only for analysing flow structures at scales unaffected by platform motion, as discussed below, unless users apply the provided correction procedure.

Figure  illustrates the time series of Euler angles measured by the BNO1 sensor for flight 20220920.0750. Only BNO1 data are shown here because this sensor was mounted inside the instrument box and usually more reliable source than the BNO2. We suspect that the errors in BNO2 arose from the long communication cable connecting the sensor to the Arduino in the electronics case. The time series analysed in Fig.  spans 08:20–13:00 UTC, which is also the interval used to compute the mean and standard deviation reported in Fig. . It can be seen that the yaw angle measured by the BNO sensors occasionally drifts, most likely due to high noise. We are also unsure why rare, spuriously large yaw values appear, although data corruption during logging cannot be ruled out. Overall, the performance of the BNO sensors used in the first-generation WinDarts is unsatisfactory and should be improved in future deployments.

Figure  shows the power spectral density (PSD) of the Euler angles and velocity components for the same period. Distinct spectral peaks emerge at frequencies above ∼0.1 Hz in both the angle and velocity signals, in particular at 0.2 Hz and just below 2 Hz. For statistically stationary, homogeneous, and isotropic turbulence, the inertial subrange of the energy spectrum scales as k-5/3 ; under Taylor’s hypothesis this implies an f-5/3 scaling in frequency space. Assuming this holds for the WinDart measurements during PaCE, the spectrum of longitudinal velocity U exhibits a slope close to -5/3 for frequencies below 0.1 Hz (spatial scales of order 100 m). Discrepancies at frequencies above 10 Hz can also be contributed to the finite response of the Pitot system, for example the pressure-tube length and associated damping. Following these arguments, one can restrict analyses to frequencies < 0.1 Hz without correcting the velocities for platform motion. Prior work indicates that platform motion associated with such distinct oscillations does not bias the inertial-range scaling or dissipation-rate estimates at other (larger or smaller) scales .

Figure  displays the longitude and latitude coordinates of WD1-1 throughout flight 20220920.0750, highlighted in grey in the figure. This spatial representation shows the flight path taken during the data collection.

Figure  shows the time series of air temperature, wind speed, relative humidity, and atmospheric vertical electric field recorded by the ground weather station during 20 September, when the WinDarts performed flight 20220920.0750. For each quantity, the primary sensor was chosen as the one offering the highest accuracy and resolution according to the specifications provided in the manufacturers' manuals: the Lufft WS500UMB for temperature and relative humidity, the Metek uSonic3 Class A-MP for wind speed, and the Campbell Scientific CS110 field mill for the electric field. Traces of a diurnal cycle are captured in temperature and relative humidity measurements.