The DH2 embodies many improvements over the original DataHawk sUAS (Lawrence and Balsley, 2013) based on hundreds of flight hours conducted over a variety of geographical and meteorological regimes. Similar to the SUMO, the original DataHawk design used a commercial molded expanded polyolefin (EPO) foam airframe (the Hobby Zone Stryker), which was attractive due to the very low initial cost. However, field experience revealed shortcomings in ruggedness that led to frequent repairs and a short lifespan, increasing operational and maintenance costs. Although much of this damage could have been avoided by selecting a smooth, forgiving landing area and by using a skilled radio control (RC) pilot, these luxuries were often limited in the field campaigns of interest. The airframe was also found to be difficult to operate in high winds. This was most critical during launch and landing, making operation in windy conditions difficult and thereby restricting the conditions that could be sampled. Beyond flight operations, the one-piece molded airframe occupied a large volume in relation to its wingspan, requiring a correspondingly large container for shipping. This made it expensive to bring more than a few airframes to any field campaign, undercutting the advantages of a low-cost aircraft for redundancy and maintaining availability throughout a lengthy campaign. Finally, another limiting quality of the original airframe was that, as with many off-the-shelf products, the long-term supply was unpredictable.

Based on these experiences, the DH2 airframe (see Fig. 1) was designed with the following characteristics and procedures to reduce overall cost and to improve field operability:

gust-insensitive aerodynamic design, with no wing sweep or dihedral, and a vertically symmetric tail to eliminate the roll moment due to sideslip, resulting in neutral lateral stability and a natural tendency to weathervane into a gust rather than to roll away from it;

The resulting aircraft is very rugged and is rarely damaged from hard landings in rugged terrain. Typically, accumulation of abuse results in a loosening of the exterior tensioning tape, but this is easily replaced with new tape to restore the rigidity of the airframe. If repairs are needed, spars and foam sections can be easily cut out and their replacements glued in place.

There are a variety of avionics on board modern sUASs. Servos for control surfaces and speed controllers for the propeller motor have advanced rapidly and now contain programmable microprocessors to set a variety of operating modes and safety limits. These are relatively independent of other avionics, and there are a wide variety of off-the-shelf options to choose from. Similarly, manual flight control through a RC radio link has several sophisticated commercial options. More complicated is the choice of autopilot avionics and associated signal conditioning and data handling for on-board scientific sensors.

When the original DataHawk was developed, there were no suitably small and low-cost autopilot systems available, so one was developed in-house as part of a Ph.D. thesis (Pisano, 2009). At the time of the DataHawk re-design, many of the original autopilot avionics components had become obsolete, and a re-designed custom autopilot was developed. This process was undertaken with two primary considerations: (1) developing hardware architecture to keep up with the constant innovation (and obsoleting) of key autopilot components (e.g., inertial sensors and GPS receivers), and (2) providing a software foundation to support continuous advances in measurement and operational techniques required by the scientific community.

With these considerations in mind, the DH2 took a modular approach, separating the functions of the microprocessor, power conditioning, flexible connection to peripherals, and multiplexing between autopilot and RC manual control of the control surfaces and propulsion between multiple boards. The processor was upgraded to a 32-bit ARM microcontroller clocked at 180 MHz, with a floating-point co-processor. This enables updates to many different components to be localized to that corresponding board without requiring wholesale changes elsewhere. It also enables components to be located optimally on the airframe, helping to reduce interference of motor currents on the magnetometer and reducing multipath reflections on the GPS antenna. Figure 1 shows where the components of the autopilot and sensors are in the airframe.

Flight software presented a more difficult trade off. While commercially available autopilot software is extensively tested, it can be daunting to modify such a large code base without intimate knowledge of its architecture. Simultaneously, building a custom code base has the advantage of complete version control and comes with intimate knowledge that enables modifications and customization. As a result, the choice was made to develop custom software for the DH2.

The autopilot processor also handles sensor data, storing it at native rates on a microSD card and sending a subset of these data to the ground station at lower rates for real-time sensor monitoring during flight. This process is complicated by the asynchronous nature of many of the sensor data streams and by the many different sensor data rates. The highest data rate is 800 Hz. Other sensors are sampled at 100 and 5 Hz. These data are buffered into 4 K byte blocks for writing to the SD card in three different messages. Each message has a processor clock time tag, and the 5 Hz GPS data contains the GPS time-of-week (TOW) time reference. This allows the data to be time aligned in post-processing to the 5 Hz GPS TOW resolution (0.2 s). Real-time telemetry sends nine different data packets at 5 and 0.5 Hz rates for display on the ground station during the flight. The autopilot hardware has been designed to maximize sensor interface flexibility by providing breakout connectors for many different interfaces, including dedicated connectors for three SPI buses, three I2C buses, six UARTs, and one CAN bus, with a hardware matrix re-assignment for various power voltages and signal assignments to nine uncommitted connectors for analog inputs, timer input and/or output, etc.

The autopilot takes a vector field control approach (Lawrence et al., 2008), causing the aircraft to be attracted laterally to specified circles or lines. Vertically, the aircraft tracks specified rates of climb and/or descent, bounded by specified ceiling and floor parameters. For example, a repeating helical vertical profile is provided by selecting a horizontal circle center location and radius, climb and/or descent rates, and ceiling and floor altitudes. Lower-level control loops track compass heading, airspeed, elevation angle, and bank angle to track the vector field, with active compensation for current wind conditions that modifies the commanded compass heading to produce the desired GPS course heading. Airspeed is sensed with a miniature pitot-static tube. Heading, elevation, and bank angles (aircraft attitudes) are estimated by fusing a 3-axis accelerometer, gyroscope, magnetometer, and “moving-base” differential GPS in a simplified Kalman filter algorithm that runs at 100 Hz. The latter two measurements enable reliable attitude estimation in high-latitude locations where the local magnetic field vector is nearly co-linear with the gravity vector.

Experience operating in the restricted airspace R-2204 at Oliktok Point, Alaska, in proximity to wind profiling radars and an Air Force early warning radar, prompted several avionics modifications to avoid anomalies in flight control. Early campaigns there experienced a range of intermittent anomalies, from sensor glitches to GPS mistracking to catastrophic processor execution halt. In response, a hardware multiplexing scheme was developed to allow manual control of aircraft elevons and propulsion to override the autopilot by direct RC command, independent of the state of autopilot processor execution. This provides a fail-safe backup in case of autopilot failure for any reason. Also, an in-flight processor reset capability was added, where the entire state of the flight control system is continuously stored in non-volatile memory so that the processor can be reset, following which the previous flight state can be restored for a smooth continuation of the flight. These resets can be automatically generated by detections of peripheral sensor anomalies or watch-dog time-out if the processor execution stops. Resets can also be manually generated from the ground station. Software protections include detection of a plugged pitot-static airspeed sensor (e.g., from rain drops or icing) and detection of erroneous GPS tracking to prevent upsetting of the state estimation and control system. Mitigations of such “off-nominal” operation challenges further enhance the ruggedness of the system over off-the-shelf options, and these mitigations are enabled by the custom hardware and software development used in the DH2.

The DH2 is primarily equipped to make detailed measurements of the thermodynamic structure of the atmosphere. To support such measurements, the system has carried a variety of sensors throughout its history. The most recent version of the DH2 has included a Vaisala RSS-421 pressure, temperature, and humidity (PTH) sensor suite embedded in the airframe foam, with the sensors extended into the streamflow that passes over the aircraft. The RSS-421 is similar to Vaisala sensors that are commonly used as radiosondes (RS-41) and are identical to the sensor suite integrated into the Vaisala dropsonde system (RD-41). The RSS-421 is unshielded on the DH2, similar to the RS-41 application of these sensors; the silver solar reflective coating on the temperature sensor helps mitigate solar effects. The platinum resistive temperature sensor on the RSS-421 offers 0.01 ∘C resolution and measurement repeatability of 0.1 ∘C, with a response time of around 0.5 s at typical airspeeds. The capacitive silicon pressure sensor has a resolution of 0.01 hPa, with a repeatability of 0.4 hPa. Finally, the thin-film capacitive relative humidity (RH) sensor includes active sensor temperature monitoring and correction, offering a resolution of 0.1 % RH, a repeatability of 2 % RH, and a temperature-dependent response time that ranges from approximately 0.3 s (at 20 ∘C) to 10 s (at -40 ∘C). Previous versions of the aircraft also employed an iMET-1 radiosonde sensor system developed by interMet Systems, though this sensor is not currently used in the DH2.

In addition to the Vaisala sensor system, the DH2 carries a custom fine-wire array that was developed at the University of Colorado. This consists of 5 µm diameter platinum sensor wires, one operated as a cold-wire thermometer and one as a hot-wire anemometer using a custom electronics board. The array also includes a Sensiron SHT-85 temperature and humidity sensor. This array was modified for use at high latitudes following a test campaign on the Svalbard Archipelago. Initially, the shroud was constructed of aluminum, but this caused multipath issues with the differential GNSS antennas integrated on the wings for high-latitude operations. This issue is exacerbated by the relatively low position on the horizon of the GNSS satellites at high latitudes. The design was modified to have a foam-covered plastic shroud which mitigated these GPS issues while still retaining the insolation-shielding properties of the original design. The new shroud design is also a result of detailed wind tunnel studies characterizing the secondary turbulence generation of fine-wire protections against contact with airborne particles. Despite the small scale of these protective obstructions, it was found that the additional turbulence generated has enough cascading energy at larger scales to affect the parameterization of geophysical turbulence. Thus, many of the protections used previously (small shields upstream of the wires, rear-facing wires, etc.) were removed, and the shroud diameter was increased from 1 to 3 cm. Comparisons between free-stream fine-wire placement with those inside the new shroud in DH2 flight tests showed negligible impact on the portion of the inertial sub-range used for turbulence parameterization. Because of this new design, fine-wire breakage does occasionally occur in-flight if precipitation is present and sometimes upon landing where snow or vegetation fragments can be kicked up, but generally the fine wires are robust enough to withstand operation in rugged terrain. The fine wires themselves are produced in batches using the Wollaston wire technique in the laboratory at CU. At a cost of < USD 5 each, wire breakage is not a cost driver, and wires are easily replaced in the field.

The voltage signals from the fine-wire electronics are converted to fluctuations in relative wind velocity and temperature through post-flight calibration, as detailed in Section 3. Spectral analysis can then be used to fit a Kolmogorov inertial sub-range model to the power spectral density as a function of frequency, and mean velocity is used to convert frequency to wavenumber. These spectral fits can then be converted to infer information on turbulent characteristics of the atmosphere, such as kinetic energy dissipation rate ε (from the hot wire) and temperature structure parameter CT2 (from the cold wire) (e.g., Frehlich et al., 2003).

Contemporary UASs like the MMAV (van den Kroonenberg et al., 2008), MASC (Wildmann et al., 2014), BLUECAT (Witte et al., 2016), SUMO (Båserud et al., 2016), Skywalker X6 (Calmer et al., 2018), and OVLI-TA (Alaoui-Sosse et al., 2019) have shown proof of concept for turbulent wind measurement using high-cadence, multi-hole pressure probes and fine-wire (Witte et al., 2016) sensors typically for measurements in the atmospheric boundary layer. However, the authors report that, due to the elevated noise floor of the multi-hole pressure sensors, the effective bandwidth of the sensors is limited to 40–100 Hz. This inhibits most UASs from measuring small-scale, weak turbulence structures typically found in the free atmosphere. The DH2 is equipped with a custom fine-wire anemometer and thermometer that sense airspeed and temperature at a cadence of 800 Hz. The low white-noise floor of the custom fine-wire turbulence sensors on the DH2 enables the DH2 to measure turbulence in scales as small as ∼0.0375 m (15 [m s-1]/400 [Hz]; assuming 15 m s-1 nominal flight speed).

Finally, the DH2 carries a pair of infrared temperature sensors. These sensors offer information on surface heterogeneity below the aircraft as well as on cloud cover above. Such information can be useful when attempting to associate changes in atmospheric conditions with surface features such as coastal boundaries, leads in sea ice, lakes or ponds, or vegetation coverage. The sensors are based on a custom design that utilizes the 10TP583T thermopile in combination with amplification and compensation from an integral case-temperature thermistor to provide an approximate optical temperature of the area in the sensor's approximately 90∘ conical field of view. One sensor is mounted with a view above the aircraft, and one is mounted with a view below, providing temperature variation information from the sky and from the surface (see Fig. 2). The sensor time constant of 15 ms and the 100 Hz sampling enable fast variations of ground features to be captured at the typical flight speeds of the aircraft.

Data is logged by the autopilot to an integrated microSD card at 800 Hz for the hot-wire and cold-wire signals, at 5 Hz for the GPS signals and the RSS-421 signals, and at 100 Hz for the rest of the measurements. SD card write failures or in-flight autopilot resets can cause these signals to become unsynchronized, so all signals are time-aligned in post-flight processing to GPS time within 200 ms.

A few factors may impact the uncertainty values given in the RD-41 dropsonde datasheet for the RSS-421 and discussed in Sect. 2.2, as the RD-41 is designed as a one-time use sensor. Vaisala includes an option to regenerate the humidity sensor through a heating cycle to avoid the impacts of aging on the sensor accuracy. Under standard operating procedure, this process is conducted at least daily during DH2 field campaigns. However, on MOSAiC, the RSS-421 sensors were frequently failing irrecoverably after undergoing this process; so, given the limited number of sensors on board and the inability to get more, the decision was made to generally forego this step. Second, the DH2 moves at a higher airspeed than the RD-41 descends close to the surface, which produces increased aspiration over its sensors. Additionally, as mounted on the DH2, the RSS-421 does not have any solar shielding (similar to the RS-41 radiosonde), so solar effects could impact its measurements in certain cases, though flight data from MOSAiC does not show a significant dependence of temperature on solar angle (<0.1 ∘C of variation present across all solar angles on the sensor). Lastly, flying in certain weather conditions can result in wetting (e.g., the summer fog of MOSAiC Leg 4) or icing (e.g., the cold winter of MOSAiC Leg 3) of the sensor, which could impact measurement quality.

Finally, the DH2's IR sensors have been calibrated to provide only a relative measurement of infrared temperature. This can help one distinguish ground or sky features as mentioned in the previous section, but the CU IR sensors do not currently provide an accurate determination of optical temperature. Figure 2 provides an example of the perspectives offered by this sensor, leveraging data from a flight near Oliktok Point in Alaska, where the periods of flight over different surface features can be clearly seen on the flight trajectory colorized by the IR sensor temperature data. Additionally, it is important to note that the atmosphere is not entirely transparent to these sensors (3–15 µm spectral range), meaning that, at altitudes significantly different to that of the object whose temperature is being sensed, atmospheric contributions to the measured temperature may impact the readings.

The high-rate fine-wire sensor measurements are calibrated against co-located (but slower) reference sensors in post-flight data analysis. Voltages from the cold-wire temperature sensor are calibrated against the reference temperature from the SHT-85 sensor, located approximately 3 cm downstream inside the protective shroud on the file wire module. As a result, the calibrated cold-wire temperature inherits the uncertainty of the SHT-85. Due to the differing time constants between the cold-wire (about 0.5 ms) and the SHT-85 (about 2 s), these signals can be offset relative to each other when the ambient temperature changes (e.g., in vertical profiling). Here, it is important to include both ascent and descent profiles as part of the calibration so that lag-induced offsets caused by these differing sensor time responses cancel each other out. The calibration improves with a larger range of temperature values (e.g., several ∘C or more) so that average signal excursion dominates the turbulent fluctuations, reducing uncertainty in the calibration curve fit.

Figure 3 shows a comparison between the different temperature sensors carried by the DH2, based on data from an extended-level flight leg in the Arctic boundary layer (Jozef et al., 2021). This example shows the different response times of the individual sensors and the reporting frequencies of each. The cold-wire sensor, recorded at 800 Hz, is able to record very fast fluctuations in temperature, though system noise becomes evident in this particular case around 100 Hz. The SHT-85 has a much slower response time, producing a roll-off in the spectrum above 0.1 Hz, and suffers from a relatively high level of signal quantization that causes spectral flattening above 1 Hz. Finally, the RSS-421 is shown to have a response roll-off starting around 0.5–0.7 Hz due to the inherent time constant of the sensor.

The hot-wire sensor voltage is calibrated against pitot-static airspeed. Both sensors are located on the top of the aircraft at about the same longitudinal and vertical positions, with approximately 10 cm lateral offset. The measured pitot-static differential pressure P is calibrated to first-order airspeed v using the dynamic pressure formula 1P=12ρv2, with an estimate of the local air density ρ derived from altitude and the US standard atmosphere. This value is used for autopilot airspeed control and wind-aware guidance. A second-order correction to this pitot airspeed is conducted in post-flight analysis by comparing mean airspeed to extrema of GPS speeds during circular trajectory segments, adjusting pitot airspeed to lie midway between these GPS extrema. Only the turbulent fluctuations in airspeed are sensed by the hot-wire instrument, because an auto-zero process is active in flight to keep the hot-wire voltage near the midpoint of the measurement range. Auto-zero adjustments are also recorded so that re-calibration against pitot airspeed can be computed whenever the adjustments change (although this happens rarely during flight). Calibration for the hot-wire data is calculated by comparing spectral data from the pitot airspeed and the hot-wire voltage and adjusting the hot-wire scale factor from V to m s-1 to agree. The hot-wire spectrum is relatively free of propeller vibration noise compared to the pitot data, providing a wider portion of the inertial subrange for an estimation of turbulent kinetic energy dissipation rate. Since the pitot-static tube is small (< 15 cm total tubing lengths to the sensor) and since the differential pressure sensor has a 15 kHz bandwidth, spectral roll-offs due to sensor dynamics are not seen up to the 400 Hz Nyquist rate in the pitot spectrum, e.g., during gliding descents with no propeller vibrations. However, these vibrations limit the portion of the inertial sub-range that can be used for hot-wire spectral calibration against pitot spectra to frequencies typically less than 100 Hz.

Calibrated velocity and temperature fluctuations, respectively, are used to parameterize turbulence intensity in terms of kinetic energy dissipation rate ε and temperature structure parameter CT2. These parameters are computed via spectral processing in post-flight analysis. The fast fine-wire response, high sample rate, and low electronics noise floor enable the high-spatial resolution of these turbulence parameters. For example, if 1 s time records of the 800 Hz samples are used, spectral analysis provides up to 2.6 decades of the inertial subrange to fit with the f-5/3 characteristic Kolmogorov cascade (Kolmogorov, 1962; Frehlich et al., 2003), providing turbulence estimates averaged over a spatial interval of 15 m horizontally or 1 m vertically (assuming 15 m s-1 airspeed and 1 m s-1 ascent rate). Figure 4 shows a representative power spectral density of temperature fluctuations (blue line) and the fractional decade frequency bin averages (red dots) along with the Kolmogorov cascade fit (black line) and the standard deviation of this fit (dashed lines). The level of the fit is then converted to turbulence parameterization (CT2 in this case) according to Frehlich et al. (2003). Details of this process – such as the removal of spectral artifacts by choice of which bin averages (green dots) to use in fitting – are currently in preparation for publication, but similar methods can be found in Luce et al. (2019).

Wind retrieval from a moving platform is a complex topic. Briefly, the DH2 has used both the “standard” approach, using attitude estimates to rotate body-frame-relative wind measurements into Earth-frame coordinates to combine with GPS velocity measurements in the wind triangle to derive wind estimates, and a “hybrid” approach that relies primarily on airspeed magnitude and GPS velocity, with only secondary use of attitude estimates. Both methods are susceptible to errors when the vehicle makes rapid maneuvers, e.g., during the downwind leg of tracking a circle in high winds, requiring judicious data excision of some intervals before applying the wind estimation algorithms.

The standard approach leverages the equations documented in the literature for wind estimation from aircraft. From the perspective of UASs, this technique is laid out clearly in van den Kroonenberg et al. (2008), where the zonal, meridional, and vertical wind components are defined as 2u=vAg-UAD-1[(cos⁡θsin⁡ψ)+tan⁡β(sin⁡ϕsin⁡Θsin⁡ψ-cos⁡ϕcos⁡ψ)+tan⁡α(cos⁡ϕsin⁡Θsin⁡ψ+sin⁡ϕcos⁡ψ)],3v=uAg-UAD-1[(cos⁡θcos⁡ψ)+tan⁡β(sin⁡ϕsin⁡Θcos⁡ψ-cos⁡ϕsin⁡ψ)+tan⁡α(cos⁡ϕsin⁡Θcos⁡ψ+sin⁡ϕsin⁡ψ)],4w=-wAg-UAD-1[(-1sin⁡θ)+tan⁡β(sin⁡ϕcos⁡θ)+tan⁡α(cos⁡ϕcos⁡θ)], where UA is the true airspeed measured by the aircraft's air data system, D is a function of the aircraft's angle of attack (α) and sideslip (β) angles: 5D=1+tan⁡2α+tan⁡2β, and vAguAg and wAg are the eastward, northward, and downward velocities of the aircraft relative to the ground, as measured by GPS, and θψ and ϕ are the aircraft pitch, yaw, and roll angles, respectively, as measured by the inertial measurement unit. Unfortunately, the DH2 does not carry a sensor to measure angle of attack or sideslip; so, for the purpose of estimating winds, those relative wind angles are assumed to be always constant and set to 0 (although the angle of attack could be set to any constant value, if desired, to account for an estimated average in-flight attack angle). This assumption makes the wind estimates insensitive to high-frequency lateral turbulent motions in the atmosphere, as the aircraft does not instantaneously weathervane into the relative wind (estimated time constant is about 1 s), although the longitudinal turbulent components of the wind are not attenuated up to the 400 Hz Nyquist rate of the pitot-static sensor.

These calculations are also sensitive to misalignment between the axis of the aircraft's airspeed sensor and the UAS inertial measurement unit as well as to differing time delays between the GPS and IMU-derived variables. Additionally, they are very sensitive to biases in airspeed. To account for these issues, the measurements from the aircraft are put through an optimization routine that varies UA, θ, and ψ, with the latter two undergoing a full rotation to account for the impact of an adjustment to an individual axis on the values for the other two axes. To accomplish this, winds are calculated for each combination of variables, and the variance in the wind estimate is minimized, as the impact of angular offsets or incorrect UA is to create steps in the calculated winds as a function of heading, which increase variability in the derived winds.

Another approach to retrieving wind estimates has also been used on the DH2. This derives from the “airspeed-only” approach (Lawrence and Balsley, 2013) that uses the geometry of the wind triangle along with measured GPS velocity and pitot-static airspeed to constrain the horizontal wind vector to a circle at each time step. Wind estimates from the previous time step are projected onto this constraint circle along the direction of the current airframe compass heading, reducing the one-parameter family of solutions for the wind to a single solution at the current time step. This method reduces the sensitivity to variable delay in sensor data but can be biased by poor previous wind estimates. Methods to counteract this error involve forward- and backward-in-time wind estimate updates to cancel the directional bias. Both these wind retrieval methods are currently in development and validation by comparison with nearby radiosonde winds. However, the raw data from many of the previous campaigns is available for others to use in pursuing wind estimation approaches as well.

To assess the field performance of the DH2 sensors, measurements from the aircraft (Jozef et al., 2021) are compared to radiosonde-based observations obtained during the MOSAiC campaign (Maturilli et al., 2021). The data from each DH2 sensor and the radiosonde parameters were averaged over 10 m altitude bins, starting with an altitude of 30 m and extending to the top altitude of the DH2 flight. A paired t test was chosen to investigate if there is a mean difference between the radiosonde data and the data from various sensors on the DH2. Except for the derived standard wind speed estimate (detailed in Sect. 3.3), the true mean difference between the radiosonde and DH2 observations was found to be not zero (i.e., the null hypothesis of zero mean difference was rejected) at the 95 % significance level. There is minimal usefulness in knowing that the two sensors are not absolutely the same; this is already assumed. However, knowing a range for the actual difference between the radiosonde and DH2 is of interest. Therefore, a confidence interval was computed to determine this actual difference between the sensors, given the same 95 % significance level. For each measurement, the standard deviation of the difference between the DH2 and radiosonde data bins are given in Table 2 along with the minimum and maximum values showing the confidence interval. The data used in this comparison are limited to radiosonde data taken within an hour of DH2 data-point time, span the Arctic melt season (5 April–26 July 2020), and are from flights conducted in a variety of atmospheric conditions.

The data in Table 2 show that the two temperature sensors present on the DH2 show similar errors relative to the radiosonde data, which exceed the repeatability value for the RSS-421 (0.1 ∘C) and SHT-85 accuracy of ±0.1 ∘C. This difference could be because of the effects mentioned previously (increased sensor aspiration, solar affects, wetting or icing of the sensor), but it is also plausible that, given the difference in time (up to one hour) and lateral position (up to 2.3 km) of the measurements, a ∼0.3 ∘C true difference in temperature is present. In the first panel of Fig. 5, a small difference (approximately the 0.3 ∘C shown in Table 2) can be seen over most of the profile, though this difference is slightly larger or smaller at certain points during the flight. Pressure shows good agreement between the RSS-421 and radiosonde, slightly exceeding the repeatability value (0.4 hPa) given in the RSS-421 datasheet. This similarity can be seen in the second panel of Fig. 5; little deviation between the two pressure sources can be seen. Greater differences both between the RSS-421 and SHT-85 and between each sensor and the radiosonde data are present in the relative humidity data, as seen in Table 2 and the third panel of Fig. 5. The SHT-85 differences exceed the stated accuracy (±1.5 % RH) by a small amount, which seems reasonable given the difference in the sensors and the position and time of the measurement. However, the RSS-421 has significantly more deviation from the radiosonde, well exceeding its repeatability value of 2 % RH. Outside of factors common with the other measurements (differences in time and position of measurement), the larger difference in RH may be due to the lack of reconditioning, as mentioned in Sect. 2.2. The significant deviation of the RSS-421 from the other sensors is apparent in the example profile shown in the third panel of Fig. 5. Visualizations of the comparison data presented in Table 2 can be found in de Boer et al. (2022b).

Both wind estimation techniques can be seen compared to the radiosonde wind estimates in Fig. 5, panels four (wind speed) and five (wind direction). For this example flight, the two wind estimation techniques are similar to one another and the radiosonde estimates, though they do deviate somewhat from the radiosonde wind velocity and direction estimates at certain points in the profile. Table 2 shows the agreement for the winds computed using radiosonde data points taken within one hour of DH2 data points. From the confidence intervals calculated in the MOSAiC radiosonde comparison (detailed earlier in this section), the DH2 standard approach shows very good agreement (<0.12 m s-1 difference) with the wind speed estimates from the radiosondes. The hybrid approach is within 1 m s-1 of the radiosonde estimate but differs more than the standard approach. For wind direction, the standard approach has a wider confidence interval for the true difference from the radiosonde than the hybrid approach, but less difference between the two techniques is discernable here; both range from approximately 1 to 3∘ offset from the radiosonde. Given the difference in time and physical position between the DH2 and radiosondes, both wind estimation approaches seem reasonable for both wind speed and direction.