The front shell – comprising the scoop and payload compartment – was shaped and integrated with the CSWX's outer shell to produce a smooth and continuous-surface aerodynamic enclosure that minimizes step changes and flow separation. The front shell is fully detachable to facilitate the installation of different sensor package configurations, as well as calibration and maintenance tasks. The CSWX's front shell is equipped with multiple thermohygrometers for measurement redundancy: three iMet-XF bead thermistors and three IST HYT-271 capacitance humidity sensors. Following the sensor placement methodology described by , the sensors were arranged in an inverted “V” shape within a cylindrical chamber.

Figure  illustrates a sectional view of the CSWX front shell showing the airflow across the scoop and internal sensor chamber. The scoop's geometry and fan placement were designed to ensure airflow aspiration while blocking direct line‐of‐sight to the sensing elements. A small brushless ducted fan, mounted at the bottom of the sensor chamber (see Fig. ), draws ambient air through the scoop at a calibrated constant aspiration speed of approximately 12 m s⁻¹ . This fan-driven flow is supplemented by the forced-air effect of the CSWX flying into the wind via the WVFM. The inlet curvature, chamber diameter, and sensor depth combine to protect the thermohygrometers from solar radiation regardless of solar incidence angle.

To quantify the influence of solar heating on onboard thermohygrometer measurements, we conducted computational flow‐simulation studies in SolidWorks^® followed by a comparison against field data from both the CS3D and CSWX platforms. In the flow simulations, each run replicated similar conditions (solar incident angle, ambient temperature, wind speed, drone pitch, cloud cover, etc.) recorded during corresponding flight days. Both CopterSonde scoops (CS3D and CSWX) were CAD-modeled with all three temperature and three humidity sensors, including the fan for aspiration. The simulations used steady-state conditions with the fan aspiration speed set to 12 m s⁻¹, ambient wind speeds matching the mean conditions observed on each corresponding flight day (ranging from 2 to 12 m s⁻¹), and solar irradiance values between 0 and 800 W m⁻² depending on sky conditions. While the simulated temperature spreads (green boxes in Fig. ) show mean differences of 0.06–0.15 °C relative to the measured CSWX data, we acknowledge that the steady-state approach does not capture transient effects such as turbulent gusts and rapid sun-angle changes, which likely account for the remaining discrepancies in Fig. . These experiments helped establish a baseline distribution of airflow and temperature inside each scoop under purely aerodynamic and solar effects, allowing us to isolate and measure the contributions of solar radiation and aerodynamics to the measurement bias at varying wind speeds and sun angles.

Flow simulations offered our first glimpse into how sunlight and ambient wind interact with the CSWX sensor housing. As shown in Fig. , sunlit areas on the front shell heat up noticeably, with a clear hot spot forming just behind and below the intake, but not reaching the weather sensors located deeper in the scoop. This shows that shielding the instruments from direct and reflected solar radiation is critical to prevent measurement bias, but it also shows that any cover may introduce other sources of error, since heated walls can disrupt local airflow and introduce convective heating, particularly when the Richardson number exceeds order unity (i.e., when buoyancy-driven effects dominate over forced convection). At the aspiration speeds maintained by the ducted fan (∼ 12 m s⁻¹), forced convection is expected to dominate within the sensor chamber, limiting but not entirely eliminating this effect. Re-radiation from heated interior walls may also contribute a small bias, though this is mitigated by the white, high-reflectivity surfaces used inside the scoop. In practice, the final sensor enclosure must therefore have a balance between sufficient sun protection and adequate ventilation, all while maintaining a smooth surface profile that blends with the aircraft's body and does not compromise the UAS's flight performance.

Figure  illustrates the absolute temperature spread among the three onboard thermistors for both CopterSonde configurations (CSWX in red, CS3D in blue) alongside the CSWX simulation baseline (green) across different flight days (D1–D8). On days with negligible solar irradiance and moderate winds (D4 and D7), the CSWX's measured spread collapses to near or even below the simulated baseline (green), demonstrating excellent thermal uniformity; by contrast, the CS3D continues to exhibit some variability (blue), suggesting inherent design‐induced biases unrelated to solar radiation. More broadly, for the CSWX, variations in wind speed – even up to ∼ 12 m s⁻¹ – produce only marginal changes in temperature spread, indicating that convective heat transfer (enclosure to air) diminishes, homogenizing sensor readings once airflow exceeds a minimal threshold (also seen in the thermal comparison from the low-to-high-wind flights in Fig. ). In contrast, full exposures to solar radiation (clear‐sky panels) correlate with pronounced increases in both CSWX and CS3D spreads, with the CS3D showing the largest absolute deviations. The larger CS3D spreads are attributed to differences in scoop geometry: the CSWX features a deeper sensor chamber with a more curved inlet path that increases the separation between the heated outer walls and the sensing elements, as well as hiding the sensors from direct sun radiation exporsure regardless of the incidence angle, thereby reducing both direct and re-radiated solar heating of the sensors compared to the CS3D's shallower duct design. These observations suggest that solar loading is the dominant driver of sensor‐to‐sensor temperature discrepancies on both platforms, whereas wind speed plays a secondary role. The CSWX mitigates both effects more effectively than the CS3D.

Over the past 5 years, our team has developed and refined several algorithms to infer the wind vector using only the CopterSonde's flight data, since no dedicated anemometer is present onboard. These methods exploit the quadcopter's dynamics and kinematics – measured by the autopilot's inertial measurement unit (IMU) – to estimate wind vectors. The most commonly used techniques for the CS3D were based on the linear and quadratic regression curves as a function of the wind-induced tilt of the platform, assuming horizontally-steady vertical flights. These algorithms were initially implemented and evaluated on the CS3D and then ported to the CSWX with the parameters retuned in open-field tests against meteorological towers and Doppler wind lidars .

However, the linear and quadratic methods were limited to horizontal wind estimation. To resolve the full wind vector – including the vertical component – we must leverage rotor thrust information. By accurately measuring each rotor's angular velocity and mapping it to thrust (through propeller characterization), the CopterSonde can infer the drag force acting on its airframe. Combining the known weight vector, estimated thrust, and measured accelerations into the dynamic model allows for the computation of the complete wind vector. The formulation of these relationships in a simplified linear time-invariant (LTI) dynamic model is as follows: 1P˙=VV˙=g+1mRBIT+1mD⇒x˙=Ax+Buy=Cx, where P is position, V velocity, T total thrust, D aerodynamic drag, RBI is the body-to-inertial rotation matrix, mg weight (mass times gravity) of the quadcopter UAS. In the LTI form, x is the state vector, u the input vector, and y the measured outputs, with system matrices A, B, and C being the state, input, and output weights, respectively. shows a more detailed deduction and formulation of the presented equations.

A critical requirement for this method is the use of an accurate propeller model, for which we adopted the characterization provided by . However, their thrust formulation depends on the inflow velocity to the propellers (perpendicular to the rotor disk), which is equal to the airspeed of the CSWX. The airspeed of the CSWX is the difference between the CSWX ground velocity vector and the wind vector; the latter is unknown, and we aim to measure it in the first place. Therefore, we opted to use a Linear Extended State Observer (LESO) for the estimation of non-measurable parameters . In Eq. (), the unknown parameter is the aerodynamic drag D, which is a function of the wind-induced airspeed and acts as the coupling between the wind vector and the platform's translational dynamics through the Rayleigh drag equation. Subsequently, the LESO extends the state-space variables x to include an estimation process for D as an integrator. After applying these control-theory operations to Eq. (), the resulting LTI+LESO system is: 2dx^dt=(A-LC)x^+Bauh+Lyy=Cx^, where x^ = [P;V;D]T is the state estimate, u is the rotated thrust vector combined with the weight vector – where thrust is computed from measured rotor RPM using the propeller model of , iteratively refined with the current airspeed estimate – and h is the rate of change of the wind. Assuming that the mean wind changes slowly over time, then h = 0. This quasi-steady assumption is appropriate for the profiling ascent rates used (3.5 m s⁻¹) and WVFM function enabled. As a result, the platform traverses slowly enough that the mean wind speed profile evolves on longer timescales than the LESO's 10 Hz update rate. Additionally, it is assumed that the WVFM's time response is fast enough to catch up with the mean angular wind shear and keep the CSWX pointing into the wind, making h stay close to zero relative to the CSWX's reference frame. The LESO then initializes with a first guess of wind velocity estimate (usually set to zero) and iteratively refines it using the wind and thrust estimate computed in the previous step. With sufficiently high sampling rates (∼ 10 Hz), this approach quickly corrects and converges to stable drag estimates. The matrices in Eq. () are defined as: 3A=010001m000,Ba=001001,4C=100010,andL=βp1βv1βp2βv2βp3βv3.

The observer gain L determines the convergence speed and noise sensitivity; this is a user-defined parameter chosen so that the eigenvalues of (A-L,C) lie in the left half of the complex plane, ensuring system stability and estimation errors decay to zero . The CSWX is capable of sampling the input parameters for the LESO estimation at a rate of 10 Hz

Figure  demonstrates the LESO-based CopterSonde's wind‐sensing performance compared to data collected with a Doppler wind Lidar (DWL) system. The cumulative scatter plots aggregate over 31 000 data points from 16 profiles, revealing strong correlations (R = 0.79 vertical, R = 0.90 horizontal) and low RMSEs (0.49 m s⁻¹ vertical, 1.03 m s⁻¹ horizontal) between LESO estimates and DWL measurements. Although both systems were tested over a relatively narrow range of horizontal wind speeds, the LESO begins to underestimate at higher wind speeds (see the deviation from the red one-to-one line at higher wind speeds in Fig. b). This is likely because the simplified dynamic model underperforms in fully capturing the platform's nonlinear aerodynamic behavior at high speeds.

To achieve higher resilience to high wind conditions, a powerful propulsion system alone is not enough. The airframe must also be designed to minimize wind-induced drag at high wind speeds without sacrificing sensitivity for wind estimation under calmer conditions. Tilting or angling the UAS body relative to its rotor disk can significantly enhance high-speed flight performance by reducing aerodynamic drag. Traditional quadcopters typically tilt their entire body to produce a desired inclination of the thrust vector to compensate for wind effects or shift the UAS position in 3D space. This inclination increases the exposed frontal area, drag, and reduces aerodynamic efficiency. In contrast, a built-in tilted-body quadcopter can increase maximum achievable speeds as much as 12.5 %, as demonstrated by . However, recommends limiting the airframe angle to below 20° to preserve flight stability and maneuverability. Based on this guidance, the CSWX airframe was designed with a fixed airframe angle of 15°, leaving a safety margin with the upper limit.

To quantify the CSWX's wind sensitivity over the CS3D, we performed 10 side-by-side vertical profiles (up to 1000 m) colocated with NSSL's Collaborative Lower Atmosphere Profiling System (CLAMPS) DWL reference system , sensing wind speeds up to 22 m s⁻¹. Subsequently, using this dataset, we tuned our three wind estimation algorithms – linear, quadratic, and LESO – using a Differential Evolution optimizer (DEO) to minimize root-mean-square error (RMSE) relative to the DWL observations. For each CopterSonde version, RMSE values from all the methods were averaged over 2 m s⁻¹ wind-speed bins to produce the comparison plot in Fig. . Averaging across all methods ensures a fair comparison of the aircraft's aerodynamic performance independent of the performance of any single wind-estimation algorithm. The results show that the CSWX outperforms the CS3D at low wind speeds (demonstrating enhanced sensitivity), matches performance at mid-range winds, and again delivers lower RMSE at higher speeds, evidencing the benefits of its refined aerodynamic design.

Figure  presents two examples of colocated vertical soundings from CopterSonde (CS3D and CSWX), radiosonde, and DWL. In the high‐wind profile (left), all instruments align closely, with the CSWX maintaining the best agreement with the reference instruments at the upper end of wind speeds. In the low‐wind profile (right), the CSWX's LESO‐based estimates closely track the references, benefiting from the LESO algorithm's ability to isolate wind‐induced aerodynamic effects and the tilted airframe's enhanced sensitivity to gentle breezes.

Over the last five years and through numerous field experiments as referenced in Sect. , the CS3D has been thoroughly characterized, enabling us to consider its maximum safe wind tolerance of 20 m s⁻¹ as accurately established. Figure  presents a case where the CSWX and CS3D, both colocated and simultaneously, measured wind profiles of a nocturnal low-level jet (LLJ). In panel (a), wind estimates from each UAS show that the CS3D reached its wind limit of 20 m s⁻¹ at approximately 275 m, automatically triggering return-to-launch (RTL) mode, whereas the CSWX sustained winds approaching 24 m s⁻¹ up to about 520 m.

Given that the CSWX was a less-well-characterized platform, its ascent was terminated by operator intervention for caution rather than by reaching an unproven estimated wind threshold. Figure b shows the instantaneous throttle values ranging from 0 to 1 for each platform during ascent. Maximum available throttle (Th_M) is internally calculated by the autopilot based on the current UAS attitude and available power to achieve a stable flight. The throttle out (Th_O) is the amount of throttle requested by the autopilot to the rotors, while the throttle hover (Th_H) is the estimated throttle required to maintain a stationary hover.

To quantify thrust performance Th_P of each CopterSonde, we computed the difference between Th_M and Th_O and then normalized it by the difference between Th_M and their corresponding Th_H, as expressed by: 7ThP=ThM-ThOThM-ThH. Here, the numerator Th_M - Th_O represents the excess thrust available above the thrust required for the current operating condition, while the denominator Th_M - Th_H quantifies the total thrust margin above the hover thrust (the minimum thrust needed to maintain flight in calm conditions). Therefore, Th_P becomes a dimensionless metric of how much of the available thrust reserve is still usable to counter external disturbances, allowing for consistent comparison across flight regimes and platforms.

Figure c, shows the resultant Th_P for both CSWX and CS3D in red and blue, respectively, for this particular example. Once the CS3D's Th_P falls to approximately 0.4 (the predefined safe margin), it automatically initiates RTL. We therefore adopted this thrust-based safety cutoff as a high-wind performance metric to evaluate the CSWX, estimate its maximum wind tolerance, and predict the altitude at which the CSWX would have triggered RTL, assuming that the wind profile approximates a quadratic curve.

The quadratic curve fitting was used to determine Th_P = f(h) and U = f(h) from Fig. , where h is the altitude above ground in meters. The resulting polynomials are: 8ThP(h)=-6.831×10-7h2-1.252×10-4h+0.96,9U(h)=-1.338×10-5h2+39.8×10-3h+5.924.

By setting ThP(h) = 0.4 and solving for h we obtained the estimated altitude at which the CSWX would have reached its maximum wind tolerance. We then applied the estimated max altitude to the wind-speed fit U(h) to obtain the corresponding safe mean-wind limit. For the profile shown in Fig. , this procedure yielded hsafe ≈ 818 m and Usafe ≈ 29.5 m s⁻¹. The absolute mean-wind limit (Th_P = 0) corresponds to Umax⁡ ≈ 33.5 m s⁻¹.

Next, we evaluated whether the CSWX's stored energy is sufficient to reach the safe altitude hsafe under sustained high-wind conditions. The CSWX's total energy consumption ET was obtained by summing the integrals Eqs. () and () over the ascent and descent legs. 10ET=∫0TaPaU(h(t))dt+∫0TdPdU(h(t))dt, where Ta and Td are the total ascent and descent times, respectively, which can be calculated using the UAS's fixed ascent and descent speeds. For the presented theoretical flight case up to hsafe = 818 m and assumming quadratic increase of wind speed to 29.5 m s⁻¹ from Eq. (), it yields ET ≈ 82 Wh. This is below the LiPo battery's total energy capacity of 130 Wh. Thus, we anticipate that the CSWX can safely reach 818 m and return to its launch point with roughly 35 % of the total energy still available. Under the same wind conditions, this represents almost 200 % increase in maximum altitude compared to the CS3D.

Unfortunately, we could not empirically validate these estimates because no sufficiently strong LLJ occurred during this study to test the CSWX's wind limits. However, achieving a top speed of 35.6 m s⁻¹ during stress-flight tests and using the well-characterized CS3D as a proxy gave us confidence in the feasibility of these performance predictions.

At the KAEFS site, in addition to the Oklahoma Mesonet tower, we deployed an OTT Pluvio² L precipitation weighing gauge (pluviometer) . Because the CSWX must withstand both high winds and precipitation during severe weather, it was designed to be nearly sealed, with only a few strategically placed vents for airflow and heat dissipation. To evaluate its performance in wet conditions, we conducted flights on a rainy day at KAEFS. Overcast skies with a low cloud ceiling of approximately 700 m limited our maximum flight altitude to around 300 m due to poor visibility.

Figure  shows the precipitation rate recorded by the pluviometer (red line) during CSWX operations, with vertical dashed lines marking each flight. Although the gauge indicates almost no rainfall between 19:00 and 21:00 UTC, ambient humidity was elevated. After 21:00 UTC, the CSWX encountered moderate rain during its flights, which it handled successfully. Figure  presents the temperature and relative humidity profiles measured by the CSWX for the last four flights between approximately 21:00 and 23:00 UTC. These profiles show that the CSWX operated in highly humid and rainy conditions and produced smooth, physically plausible temperature and relative humidity curves without exhibiting obvious rain-induced artifacts (e.g., sudden jumps or sensor saturation). However, no independent reference instrument (e.g., a colocated radiosonde) was available during these flights to quantitatively validate the absolute accuracy of the measurements. Therefore, these results demonstrate the platform's operational resilience in wet conditions but should not be interpreted as a full accuracy assessment under precipitation.