KASPR, a state-of-the-art cloud scanning radar, is capable of collecting Doppler spectra and radar moments through alternate transmission of horizontally (H) and vertically (V) polarized waves and simultaneous reception of co-polar and cross-polar components of the backscattered wave with the beamwidth of 0.32∘. Therefore, a full set of polarimetric radar observables are available including the radar reflectivity (ZHH), differential reflectivity (ZDR), differential phase (φDP), co-polar correlation coefficient (ρhv), linear depolarization ratio (LDR), and cross-polar correlation coefficient (ρhx), along with the Doppler velocity and spectral width. The specific differential phase (KDP) is estimated using an iterative algorithm proposed by Hubbert and Bringi (1995). The data post-processing details are described in Oue et al. (2018). KASPR was calibrated using a corner reflector technique, providing reliable reflectivity data to calibrate the data of the other two radars. The detailed configurations are also available in Kumjian et al. (2020) and Kollias et al. (2020b).

During the radar measurements on 20 February 2019, KASPR executed a scanning strategy that consisted of surveillance (plan position indicator, PPI) scans at a 15∘ elevation angle, a zenith-pointing PPI, hemispheric range–height indicator (HSRHI) scans at four azimuth angles, and a 5 min vertically pointing (VPT) mode during which Doppler spectrum data were collected. This pattern was repeated and took approximately 15 min to complete. During a 15 min cycle, two 15∘ PPI scans were included, so we had the 15∘ PPI scans every ∼7 min which were used to produce quasi-vertical profile products. The PPI and HSRHI scans were performed in a full polarimetry mode with scan speeds of 6 and 2 s-1, respectively, to collect data with a 30 m range-gate spacing, 0.6∘ PPI azimuthal spacing, and 0.3∘ HSRHI elevation spacing. The VPT mode was executed with only horizontally polarized waves transmitted and both horizontally and vertically polarized waves received. During the VPT mode, the Doppler spectra were collected every second with a 15 m range-gate spacing and 0.04 m s-1 velocity bin spacing. The zenith PPI scans were used to estimate a systematic bias of ZDR. The ZDR values presented in this study were corrected for the systematic biases.

The system was initially developed as an airborne radar and was integrated on the Center for Interdisciplinary Remotely-Piloted Aircraft Studies Twin Otter aircraft (Mead et al., 2003). In 2017, the system was refurbished by installing 61 cm parabolic dish antennas and upgrading all the C-FMCW electronics, including a new metal frame to hold the antennas, the server computer and the power supplies, to make it suitable for ground-based observations and easy shipping.

This radar system is capable of collecting Doppler spectra with spatiotemporal resolutions similar to those of KASPR (Table 1) and located next to KASPR, which allows good beam matching and reliable DWR measurements. The data during the cases were collected every 4 s at 30 m vertical spacing with a beamwidth of 0.3∘ (Table 1).

The MRRPro is the latest version of the MRR developed by Metek GmbH that has evolved to be a powerful standalone profiler for investigations of precipitation and cloud dynamics with very low installation and logistics effort. The MRRPro features a high-performance processing unit which significantly improves the options in the operating parameters (Table 1). During the observation in this study, the MRRPro collected Doppler spectra at a 60 m range-gate spacing every 4 s up to the maximum observation range of 7 km. The Nyquist velocity was 12.08 m s-1 during the observations producing the velocity bin spacing of 0.192 m s-1.

A Parsivel optical disdrometer measures terminal velocity and the horizontal size of individual precipitation particles passing through a sheet of light (30 mm wide, 1 mm high, and 180 mm long) with a 650 nm laser diode with a power of 3 mW (Löffler-Mang and Blahak, 2001). The total measuring surface has an area of 54 cm2. The measured size and velocity are classified into 1 of 32 size bins ranging from 0.062 to 24.5 mm and 32 velocity bins ranging from 0.04 to 20.5 m s-1 every minute.

The multi-angle snowflake camera (MASC) is located adjacent to the Parsivel. The MASC consists of three cameras that are separated by an angle of 36∘, each pointed toward the focal point about 10 cm away (Garrett et al., 2012; Garrett and Yuter, 2014). On top of each camera rests a 2700 lm light-emitting diode. The focal point lies within a ring that has two near-infrared emitter–detector pairs arranged in arrays that are separated vertically by 32 mm. The arrangement of the emitter–detector pairs allows for a trigger depth of field of 3100 mm2 but because of the camera field of view and depth of focus, only about 11 % of the images taken are in focus. Falling hydrometeors larger than 0.1 mm are recorded, and their fall speed is calculated as the time difference between triggering each emitter–detector pair.

KASPR reflectivity measurements were well calibrated using a corner reflector technique (Lamer et al., 2021). Therefore, systematic offsets for the MRRPro and ROGER total reflectivities have been corrected by comparing them with the KASPR reflectivity at cloud bases from a different non-precipitating cloud case. The MRRPro and ROGER reflectivity and mean Doppler velocity data were interpolated into the KASPR VPT data resolution (15 m range and 1 s time spacings).

Gaseous attenuation needs to be considered and corrected when using short-wavelength radars (Lamer et al., 2021). The MRRPro's K-band (24 GHz) frequency is the lowest in the present study; however, the 24 GHz frequency is very close to a peak in the water vapor absorption spectrum (e.g., Liebe et al., 1993; Rosenkranz, 1998). Therefore, the water vapor attenuation for MRRPro could also be significant. We corrected the MRRPro, KASPR, and ROGER reflectivities for water vapor attenuation based on the Rosenkranz (1998) results, using sounding profiles launched twice daily (00:00 and 12:00 Z) at Upton, 21 km east of the observatory. The estimated column-integrated two-way attenuations at K, Ka, and W bands for our case study were up to 0.7, 0.2, and 1.2 dB, respectively.

Another source of the gaseous attenuation we should consider is oxygen (e.g., Liebe et al., 1993). Although the attenuation in oxygen may not be as large as that in water vapor, it may be non-negligible. We also estimated the attenuation by oxygen (i.e., dry air) for each of the three frequencies using the sounding profiles and corrected the MRRPro, KASPR, and ROGER reflectivities. The estimated column-integrated two-way attenuations for dry air at K, Ka, and W bands were generally 0.1, 0.2, and 0.3 dB, respectively.

Liquid water, which was expected to be present in precipitating clouds as supercooled droplets producing riming, can also be a cause of significant attenuation. Riming commonly occurs in snowstorms observed along the US northeast coast, indicating the presence of significant amounts of supercooled cloud water in the snowstorms (e.g., Colle et al., 2014). However, it was difficult to identify liquid cloud layers and liquid water content and estimate specific attenuation at each range bin in the ice clouds. Moreover, attenuation by ice particles might be significant if the large amount of ice were produced in the clouds and the radar beam passed through the ice layers. Tridon et al. (2020) proposed a relative path-integrated attenuation (PIA) technique to retrieve liquid water content using DWR profiles. A key idea of this technique is that the DWR from dual-frequency radars near cloud tops, where it is expected that small ice crystals are in the Rayleigh scattering regime for both radar wavelengths, is mainly due to the PIA associated with liquid cloud droplets and ice particles. The DWR attributed to the total attenuation should then be equal to the DWR plateau near the cloud top. We applied the technique of Tridon et al. (2020) to the DWR from KASPR and ROGER measurements to find the DWR plateau near the cloud top as follows:

The measured DWRKaW values are averaged over 450 m (30 gates) and 20 s (20 rays). The range window was adapted for this study.

Another error source of the DWR analysis is radar beam mismatching. The three radars were located at the same observation site; the distances between those radars were less than 5 m; therefore, we expect that the beam mismatch due to location is small. On the other hand, a difference in beamwidths (Table 1) is another possible cause of beam mismatching. The KASPR and ROGER beamwidths are well matched, while MRRPro's beamwidth is 5 times larger than those of KASPR and ROGER. The beamwidth differences between MRRPro and KASPR and between MRRPro and ROGER might result in larger variabilities in DWRs.

Similarly to the reflectivity measurements, the MRRPro and ROGER mean Doppler velocity data were interpolated into the KASPR VPT data resolution. The observed mean Doppler velocities from the three radars were corrected for air density changes based on the sounding profiles and adjusted to the surface.

The polarimetric radar observables such as ZDR and KDP are more prominent when they are collected at lower-elevation scans, whereas the DWR data were collected by vertically pointing measurements. To compare those two data sets from the different types of scans, we employed a quasi-vertical profile (QVP) technique proposed by Ryzhkov et al. (2016). The QVP technique azimuthally averages polarimetric radar variables for each conical PPI scan at non-zero elevations to produce these variables in a height-versus-time format. The QVPs have high vertical resolutions allowing for capture of important polarimetric radar signatures and their evolution (e.g., Griffin et al., 2018, 2020; Kumjian and Lombardo, 2017; Troemel et al., 2019). We use the PPI scans at an elevation angle of 15∘ every 7–8 min with a scan rate of 6∘ s-1. Since the slant range resolution of the 15∘ PPI data is 30 m, the corresponding QVP data have vertical spacing of approximately 10 m and a maximum height of 7.8 km. Note that the actual vertical resolution of QVP is determined by the vertical size of the radar resolution volume, which increases with distance from the radar (Ryzhkov et al., 2016). The use of conical PPI at a higher elevation angle (15∘) for QVP reconstruction ensures relatively high horizontal resolutions at lower altitudes (11 km at the height of 2 km) that facilitates direct comparison with the DWR profiles from the three radar vertically pointing measurements. The KASPR QVP data were interpolated into the KASPR VPT data resolution, similarly to in Oue et al. (2018). Because a single PPI scan was performed every 7 min while the KASPR 5 min VPT dwell collecting profile data every second was performed at a 15 min interval, a single KASPR QVP corresponds to about 150 DWR profiles.

Based on the DWRKaW height–time plots, we selected four regions as shown in Fig. 4d, each of which had similar characteristics in terms of DWRs, MDV, and polarimetry to identify ice particle types and their growth processes. We first present traditional triple-frequency DWR–DWR diagrams (DWRKW versus DWRKKa in Fig. 7) for each selected region. The DWRs from Region A and Region B tend to be distributed toward the model low-rime-degree lines (smaller DWRKW at a given DWRKKa), while those from Region C and Region D were distributed toward the higher-rime-degree regions (larger DWRKW at a given DWRKKa). These are consistent with MASC ice particle observations. Although the distribution of the DWRs for each region seems to be significantly separated, most of the data overlap, making it hard to distinguish the growth processes and types. This is, in part, because K-band (24 GHz) and Ka-band (35 GHz) measurements are not sufficiently separated in the frequency domain.

Besides the insufficient frequency separation, there are data points that deviate from the model lines in the DWRKW-versus-DWRKKa field in each region. There are several causes of such deviations (e.g., Lamer et al., 2021). The most likely cause is unaccounted for attenuations particularly at Ka and W bands due to supercooled cloud water or ice or both. The ceilometer backscatter measurements shown in Fig. 5f, in addition to the MWR LWP measurements, suggest that thin supercooled liquid cloud layers were indeed present at least around the large Doppler spectrum width layer. Unfortunately, the ceilometer backscatter information is insufficient to provide a complete mapping of such layers because of complete signal extinction caused either by the ice clouds or by underlying liquid layers themselves. Ice particles could also cause signal attenuation (Battaglia et al., 2020; Tridon et al., 2020) particularly for the shorter-wavelength radars. Although the DWR plateau-based PIA technique has corrected the ROGER reflectivity for those attenuations related to the KASPR reflectivity (Sect. 3.1), the attenuation in the KASPR reflectivity itself cannot be accounted for in this study. This factor also causes underestimation of the PIA-corrected ROGER reflectivity.

Secondly, the beam mismatch could be significant when the radar beams penetrate fine narrow fallstreaks, even though the radars were collocated within 5 m of each other. As mentioned previously, the KASPR and ROGER beamwidths are well matched (0.3∘), while the MRRPro's beamwidth is 5 times larger (1.5∘). The radar sampling volumes, which are larger at higher altitudes, cannot resolve the small-timescale and small-spatial-scale phenomena, and the difference in beamwidth is a source of uncertainty. Moreover, the ice particle models may not represent the whole gamut of ice particles possibly present in the clouds.

Observed MDV is mainly attributed to the vertical air motion and the particle fall speeds, which are sensitive to particle size, rime degree, size distribution, and type and can provide additional information to distinguish ice types and processes. Kneifel and Moisseev (2020) demonstrated that MDV is a function of rime fraction. We further illustrate that MDV coupled with DWR shows a good indicator of degree of riming. Figure 8 shows the observed DWRKaW as a function of KASPR MDV together with the model plot with different rime degrees (lines). Most of DWRKaW values from Region A are less than 7 dB and are located between the middle-rime-degree (yellow and cyan) and low-rime-degree (red and magenta) lines, suggesting light riming of small aggregates (Fig. 8a). These particles could grow keeping a similar degree of riming by aggregation, as the data points from Region B are shifted toward larger DWRKaW values between the middle-rime-degree and low-rime-degree lines (Fig. 8b). It is possible that the turbulence layer at around a 3 km altitude (Fig. 5b and c) contributed to light riming. The turbulence also contributed to the wide distribution of MDV.

DWRKaW from Region C generally follows the low-rime-degree particle lines; DWRKaW increased from near zero to 10 dB while MDV changed from near zero to -0.8 m s-1. Some data points are shifted toward the middle-rime-degree particle lines (i.e., faster downward motion at a given DWRKaW). These data clusters suggest that the aggregation dominated, but some particles started riming. Region D, which is located below Region C, also has generally two data clusters. A smaller data cluster closely follows the low-rime-degree lines, as the DWRKaW increased 2 to 9 dB while the MDV changed from -0.6 to -1.3 m s-1. The other population, which has higher occurrence, is generally along the middle-rime-degree lines; the DWRKaW increased from 3 to 12 dB while the MDV changed from -1.8 to -2.5 m s-1 in the middle of the population. The left edge of the second data population is closer to the higher-riming-degree (blue and green) lines. Those downward MDVs belonging to the two populations are consistent with fall velocities of aggregates and heavily rimed particles, respectively, as reported by Locatelli and Hobbs (1974). These characteristics suggest that aggregates produced near the cloud top at 6 km rimed during the falling as particle fall speeds quickly increased. These distinct separations of the particle populations associated with the particle growth processes are not clearly found in the triple-frequency DWR field in Fig. 7, whereas these are shown not only in the fields of DWRKaW versus KASPR MDV but also in the DWRKW-versus-MDV (Fig. 8e) and DWRKKa-versus-MDV (Fig. 8f) diagrams.

Adding polarimetric information supports this interpretation and gives further insights into particle microphysics in terms of their shapes. Figure 9a–d and e–h are the same as Fig. 8a–d, but the color shades represent KASPR QVP ZDR (Fig. 9a–d) and KDP (Fig. 9e–h), respectively. For the data following the low-rime-degree lines for Region C and Region D, ZDR values decreased as DWR increased (Fig. 9c, d), while KDP values slightly decreased by approximately 0.2∘ km-1 (Fig. 9g, h). It can be interpreted that small ice particles producing near-zero DWRs were horizontally oriented oblate particles in the dendritic crystal growth zone (temperature of -15 to -10 ∘C), which produced large ZDR values and then aggregated into large crystals as DWR increased. On the other hand, the individual frequency pair DWR for vertically pointing measurements also strongly depends on particle aspect ratios (Matrosov et al., 2019). The impacts of particle aspect ratio on DWRKaW values could be as high as ∼3 and ∼5 dB for particle distributions with a median volume size of 1 and 2 mm, respectively (Matrosov, 2021). The increase in DWRKaW in the diagrams can include both the particle size and the particle shape effects.

The DWR–MDV diagrams suggest that as DWRKaW increased, the MDV corresponding to the low-rime-degree particle populations in both Region C and Region D reached ∼-1 m s-1, consistent with the fall speeds of low-rime-degree aggregates. This effect is more likely due to the increase in size rather than in aspect ratio. During the aggregation process, the size distribution of snowflakes evolves in such a way that the concentration of smaller, higher-density particles decreases whereas the number of larger, lower-density snowflakes increases. This is a primary reason for the reduction in both ZDR and KDP due to aggregation, although the increase in the average aspect ratio and possibly more chaotic orientation additionally contribute to such a reduction. The KDP values could also be accounted for by changes in the number concentration of the horizontally oriented oblate particles (with aspect ratio <1); its increase contributes to increasing KDP.

In the DWR–MDV data clusters following the high-rime-degree particle lines in Region C, ZDR and KDP quickly decreased as the DWRKaW and the magnitude of MDV increased; ZDR values decreased from 2 to 0.5 dB, and the KDP values decreased from 0.4∘ km-1 to near zero (Fig. 9c, g). Although the increase in the DWRKaW includes the effects of both size and aspect ratio as discussed above, the increase in the magnitude of MDV can represent the increase in size. The ZDR and KDP values are lower than those from the cluster following the low-rime-degree model's lines at a given DWRKaW. Lower KDP and ZDR values suggest particle growth by heavier riming, which tends to produce more spherical particles.

These ZDR and KDP characteristics shown in both low-rime-degree and high-rime-degree particle data groups in Region C are very similar to those in Region D, but the KDP and ZDR values in Region D are generally lower at a given DWRKaW, with a mean MDV of -3.5 m s-1 (Fig. 9d, h). The lower KDP and ZDR in Region D represent further particle growth which is accompanied by the decrease in their density, increase in their aspect ratios, and more random orientation.

It is interesting that for DWRKaW values of less than 5 dB in Region C, the observed ZDR values with faster fall speeds (corresponding to the high-rime-degree particle lines) are larger than those with slower fall speeds (corresponding to the low-rime-degree particle lines) at a given DWRKaW (Fig. 9c). This suggests that riming first worked to fill the gaps of branches of dendrite crystals, resulting in increasing the mass of individual crystals without significant change in their aspect ratio, and thus ZDR increased. This type of riming would not significantly contribute to the increase in KDP (Fig. 9g), likely due to low concentration of such particles. This characteristic is consistent with the early stage of riming reported by previous studies (e.g., Oue et al., 2016; Li et al., 2018).

Compared to Region C and Region D, the polarimetric observables in Region A and Region B (Fig. 9a, b, e, and f) do not show clear trends with changes in rime degree, and the dynamical oscillation shown in Fig. 5a–c results in an uncertainty in the particle identification for Region A, particularly when DWRKaW values are smaller than 5 dB and MDV varies between -3 and 0 m s-1. Adding polarimetric variables together with temperature information facilitates the interpretation of the microphysics. ZDR values in Region A are positive but smaller than 1 dB and smaller than those from the later fallstreaks (Region C and Region D, for a given DWR), suggesting an aggregation process which was accompanied by a decrease in particle density, an increase in their aspect ratios, and more random particle orientations compared to Region C and Region D. In contrast, KDP is larger than that from the later fallstreaks. The large KDP and smaller ZDR values in Region A suggest aggregation intensified by a higher number concentration of ice crystals. The increase in ice number concentration can be explained by two processes. One possible cause is that near the dendrite growth regime (around -15 ∘C), dendritic ice crystals were nucleated. The dendritic branches could work to facilitate interlocking (Pruppacher and Klett, 2010). This is a well-known characteristic in winter storms that has been reported by many previous studies using polarimetric radar measurements (e.g., Kennedy and Rutledge, 2011). Another process is seeding from above (e.g., Griffin et al., 2018; Oue et al., 2018), which is more likely to contribute to an increase in ice concentration for this case. The cloud top height during observations in Region A and Region B reached 10 km, approximately 4 km higher than in Region C and Region D (Figs. 5 and 6). This fact suggests that a higher concentration of ice particles aloft seeded in Region A. Moreover, a possible light riming in the turbulence region could increase the mass of individual particles, hence KDP, as the cluster extended to the middle rime degree included large KDP values.

The particles were further growing at lower altitudes as DWRKaW increased with decreasing ZDR in Region B. However, a sublimation process near the ground could also be plausible. The nearest soundings at Upton (12:00 Z, black lines in Fig. 4e) showed a dry air condition at the lower altitudes. This sounding time was ∼5–6 h before the radar observation, but the dry air condition could still have been present near the ground, thus favoring sublimation in the lower altitudes of Region B. Due to sublimation some branches and/or edges of aggregate particles could have disappeared, resulting in decreasing the mean volume diameter. The classical aggregation process could have stopped with KDP remaining relatively large because it usually decreases proportionally to the mean volume diameter. Decrease in IWC attributed to the sublimation might have been minor with any noticeable impact on KDP. These processes related to the sublimation are represented by a cluster with high KDP in Region B, where the DWRKaW values slightly increased while KDP values remained high compared to in Region A. The sublimation also contributed to decreasing particle fall speed, as shown by a minor decrease in the magnitude of MDV in the data group, but the MDV probably resulted from some balance between the fall speed increase due to aggregation and its decrease due to sublimation. The classical diabatic sublimation cools and moistens the ambient air. Therefore, the sublimation subsided as the cloud base descended with time (Fig. 5f) and snow particles in the fallstreaks eventually reached the ground.

The MDV measurements also have frequency dependencies because of the complex interplay between non-Rayleigh effects and the PSDs. Figure 10 shows dependencies of the ice particle types on the diagrams of DWRKaW versus differential MDV (dMDV = KASPR MDV - ROGER MDV) for Region C and Region D, together with the scattering calculations using the particle models. Similarly to the DWR–MDV diagram in Fig. 9, Region C includes a cluster with a higher number of occurrences along the particle lines of low-to-middle rime degrees, and a lower-frequency cluster extends toward the high-rime-degree particle lines. Region D has more data points for the high-rime-degree particle population.

Region D also includes large dMDV values greater than 0.6 m s-1 for DWRKaW values between 3 and 10 dB (Fig. 10b). It is possible that the larger values of dMDV correspond to an increase in the particle sizes and not to changes in the degree of riming. The ZDR values corresponding to these large dMDV values (Fig. 10c) are approximately 0.7 dB, suggesting that the particles were non-spherical, possibly contributing to the decrease in DWRKaW compared to the spherical particles.

As the scattering calculations show, distinguishing among different degrees of riming requires accurate measurements of MDV with an error of a few hundredths of 1 m s-1 and exact range-time-bin gate matching for lower DWRKaW (<5 dB). Although the vertical air motion contributions in MDV from each radar are canceled out in dMDV, subgrid-scale turbulence, the wide range of particle fall speeds, and different sampling times for the observations (1 s for KASPR in VPT mode and 4 s for ROGER) can all be sources of uncertainties. This limitation may affect the scatterplot distributions, e.g., with some points clustering outside the envelope of the model's lines. This limitation also affects Region A (not shown).

The particle properties retrieved from the ground-based measurements including fall speed, size, aspect ratio, and area ratio are the result of ice growth processes in the clouds aloft. The Parsivel and MASC observations allowed us to evaluate the radar-based particle characteristics described above. The Parsivel and MASC collocated with the radars collected precipitation particles after 18:13 and 18:16 UTC, respectively. The snow images from the MASC were quantified by measurements of aspect ratio and area ratio, and their time series were presented in Fig. 11. Figure 11a and b present frequencies (color shades) together with median values (black line) observed for a 20 min time range every 1 min.

We also estimated the mass-weighted mean diameter for Parsivel-measured PSD. The ice particle mass was estimated using a methodology proposed by von Lerber et al. (2017). The methodology is based on a theory that individual particle mass can be expressed based on a hydrodynamic theory derived by Böhm (1989) using Reynold's number and the Best number (e.g., Mitchell, 1996; Mitchell and Heymsfield, 2005; Heymsfield and Westbrook, 2010). The equation of mass (Eq. 5 of von Lerber et al., 2017) indicates that the mass can be a function of fall velocity, area ratio, and size. In the present study, the area ratio is derived from the MASC images, and the fall velocity and size are estimated from the Parsivel measurements. The Parsivel-observed particle diameter and fall speed are fitted to a form of V=aDb, where a and b are constants using the 20 min integrated data. Previous studies have pointed out that Parsivel's velocity and even size measurements for snow include large uncertainties owing to the sampling limitation (Battaglia et al., 2010). Before estimating the relationships, we removed the following apparently unrealistic velocity values: (1) those exceeding 1.5 m s-1 associated with particles having a diameter of less than 1 mm in agreement with Locatelli and Hobbs (1974) and (2) data outside upper and lower boundaries of the V–D relationships. The upper boundary was determined based on Locatelli and Hobbs (1974) V–D relationships for rimed aggregates, and the lower boundary was determined based on Szyrmer and Zawadzki (2010) V–D relationships for unrimed aggregates. The Parsivel-measured size was adjusted to the maximum dimension using a technique proposed by von Lerber et al. (2017). Figure 11c presents the time series of the estimated water-equivalent mass-weighted mean diameter from the Parsivel-measured PSD.

These time series are consistent with the fallstreaks reaching the ground. Aspect ratio represents oblateness of particles relating to ZDR and partly contributing to KDP. Korolev and Isaac (2003) reported that mean aspect ratios of ice hydrometeors observed in situ from aircraft sampling are often around 0.6, while those of heavily rimed particles such as graupel increase toward 1. Depolarization-based radar retrievals of snowflake aspect ratios near the ground indicated mean intrinsic aspect ratios of about 0.4–0.5 (e.g., Matrosov et al., 2020). Area ratio in the current study is defined as the ratio of the area of the snowflake, which is found by counting all white pixels in a black-and-white image, to the area of the circumscribing circle defined by the maximum diameter from the MASC. The area ratio increases with riming (von Lerber et al., 2017).

The aspect ratio was relatively low before 18:30 UTC, when the median value was less than 0.6. At the same time, the area ratio was also relatively low, with the median area ratio smaller than 0.5. This period corresponds to a time when fallstreaks included in Region A and Region B reached near the ground, consistent with the radar MDV–DWR characteristics. The mass-weighted mean size was approximately 0.4 mm, consistent with the scattering model calculations shown in Fig. 2a. It should also be noted that aspect ratio estimates from in situ data (e.g., from Parsivel and/or MASC measurements or aircraft-based particle probes) are inferred from 2D particle projections, so these estimates usually overestimate actual (i.e., intrinsic) aspect ratios, which are defined as true minor-to-major particle dimension ratios (Jiang et al., 2017; Matrosov et al., 2017).

The median in situ aspect ratio exceeded 0.6 between 18:30 and 19:00 UTC while area ratio also increased. The water-equivalent mass-weighted mean diameter increased after 18:33 UTC, as it exceeded 1.3 mm between 18:38 and 19:48 UTC except at 18:58, 19:22, and 19:28 UTC. Those large-diameter periods correspond to times where fallstreaks included in Region D reached the ground. The ground-based characteristics suggest that the snowflakes were heavily rimed, larger aggregates, consistent with the observed characteristics of the radar MDV, DWR, and polarimetric observables.