On the quality of RS41 radiosonde descent data

Abstract. Radiosonde descent profiles have been available from tens of stations for several years now – mainly from Vaisala RS41 radiosondes. They have been compared with the ascent profiles, with ECMWF short-range forecasts and with co-located radio-occultation retrievals. Over this time our understanding of the data has grown, and the comparison also shed some light on radiosonde ascent data. It has become clear that the fall rate is very variable and that it is an important factor, with high fall rates being associated with temperature biases, especially at higher altitudes. Ascent winds are affected by pendulum motion, on average descent winds are less affected by pendulum motion and are smoother. It is plausible that the true wind variability in the vertical lies between that shown by ascent and descent profiles. The discrepancy indicates the need for reference wind measurements.


As radiosondes are designed to measure during the ascent, it is useful to consider how they differ from dropsondes which 31 always measure on descent. Dropsondes are launched from aircraft and are mainly used for sampling around tropical cyclones 32 and for field experiments. Radiosondes typically have the temperature and humidity sensors mounted diagonally above the 33 radiosonde body whereas dropsondes (e.g. Hock and Franklin, 1999) have the sensors underneath -in each case to sample air 34 undisturbed by the radiosonde body. The AVAPS (Advanced Vertical Atmospheric Profiling System) processing system used 35 by many dropsondes includes an 'inertial' correction for the delayed response to horizontal wind shear (Appendix of Hock 36 and Franklin, 1999). Modern radiosondes are usually on a line 30-55 m below the balloon whereas dropsondes are only 1 m 37 or less below a parachute. As noted by Wang et al (2008) 'The dropsonde fall rate is much smoother than the radiosonde 38 ascent rate because of the radiosonde's pendulum effect and self-induced balloon motion'. Typically dropsondes fall at about 39 10 m s -1 , just after launch it can be about 20 m s -1 before the parachute opens fully. As discussed below, radiosonde descent 40 can be much faster (to 100 m s -1 or more if no parachute is used) shortly after balloon burst. There has been some use of 41 controlled descent, by partial deflation of the balloon, for measurement of stratospheric humidity (Hurst et al, 2011). Zhang 42 et al (2019) tested the use of a low density 'hard ball' to give more consistent drag than a parachute when deriving the vertical 43 velocity of the air using a radiosonde descending from a height of about 10 km. 44 45 Figure 1 shows BUFR (Binary Universal Form for Representation of meteorological data) descent reports over Europe for 46 September-November 2019 (descent data were also available from New Zealand, not shown). BUFR allows the reporting of 47 high vertical resolution radiosonde data (Ingleby et al., 2016). Geller at al., (2021) found that in mid-2020 44% of operational 48 radiosonde stations were providing high vertical resolution ascent data. Since 2019 descent data has become available from 49 more European stations and a few in the Americas. After launch the balloon is advected horizontally by the wind, especially 50 at upper levels, and typically travels 50 to 300 km before burst with the larger distances being more common in winter (Seidel 51 et al., 2011).

54
There were 14 stations from Germany, 6 each from UK and Norway and 2 each from Finland and Portugal. 55 Figure 2 gives an indication of the number and vertical extent of descent profiles. Larger balloon size and fill volume is used 56 to achieve higher altitudes. On average, radiosondes that achieve higher altitudes drift further horizontally, resulting in the 57 radio signal to the launch station being lost at higher altitudes on descent due to obstruction by terrain or signal attenuation. 58 This can be seen clearly in the UK results which have been split into automatic and manual launches: the manual launches use 59 larger balloons and the number of descent reports starts to decline earlier, below 9 km. Automatic launchers are documented 60 by Madonna et al. (2020). Some of the other countries use a mixture of manual and automatic launchers, but with little or no 61 difference in balloon size. A small proportion of ascents do not have a corresponding descent report, often due to a fault 62 developing with the radiosonde before or upon burst, leading to an automatic termination. 63     On ascent the sensor boom projects above the radiosonde, so that it samples air that has not flowed over the body of the 108 radiosonde. On descent, with a working parachute, it should be in a similar position -so it may sample air that has flowed 109 over the radiosonde body, which has the potential to introduce biases or contamination. It is not known how a radiosonde 110 descending without a parachute is orientated, or if it may be tumbling. 111

Types of parachute and string length 112
For some manual launches a parachute (if used) is attached to the line not far below the balloon. Figure 4 shows that sometimes 113 the two can become entangled after balloon burst. For automated launches a parachute (if used) is stored within the balloon 114 and this seems to cause fewer entanglement problems. This can be used for manual launches too, and has been used at 115 Lindenberg for some years, but there is a small additional expense. In general, most of the parachutes are quite basic and do 116 not include a hole. Air can build up inside the parachute and suddenly spill out. It is clear from some of our results that 117 parachutes do not always open as intended. 118 In earlier decades the string connecting the balloon and the radiosonde may have been 10 m or less, but in the stratosphere the 119 balloon gets larger and there can be intermittent influences of the balloon wake upon the instruments (WMO, 1994; Luers and 120 radiosondes ascending to 30 km or higher. In practice an 'unwinder' is used to increase the line length shortly after the 122 radiosonde launch (WMO, 2018b). The Vaisala unwinder for the RS41 gives 55 m when extended (Vaisala, 2017). We note 123 that while longer lines benefit stratospheric temperature measurements they cause larger amplitude pendulum motion in the 124 winds. 125

Preparation of profile reports 126
Data values are transmitted to the ground receiver every second and are processed by Vaisala MW41 software. Raw data 127 values, both ascent and descent, indexed by time are stored locally (the GRUAN archive makes the one second data available 128 for GRUAN sites). The MW41 software looks for a sustained decrease in altitude to determine the time of burst. In the past, 129 all later data would usually have been discarded but there is now an option to continue processing and to produce a separate 130 BUFR descent message using sequence 3 09 056 (WMO, 2019) designed for descent data. In 2019, as an interim measure, the 131 dropsonde sequence 3 09 053 was being used. As the timeliness of ascent data is critical for data users, it is preferable to 132 transmit the ascent profile as soon as possible after burst, followed by the descent data sent once the profile is completed. 133 BUFR from the European stations involved in this study is generally provided every two seconds (about 10 m separation in 134 the vertical during ascent). 135 The MW41 horizontal winds are derived primarily from Doppler processing of the Global Positioning System (GPS) signals 136 but the GPS locations are also used (GRUAN processing only uses the GPS positions). There is very good vertical resolution 137 but it also means that the winds sample the pendulum motion of the radiosonde -this is probably a mixture of planar and 138 conical pendulum motion. The period of the pendulum motion is a function of the length of line between the balloon or 139 parachute and the radiosonde. The processing attempts to filter out the pendulum motion (discussed briefly in Dirksen et al 140 2014), but the filtering is imperfect as discussed below. 141

Descent fall rates 142
The balloon and gas are chosen so that the ascent rate is about 5 m/s on average -however there is usually notable high 143 frequency variability probably due to pendulum motion. Especially in the stratosphere, there can be lower frequency signals 144 due to gravity waves and both of these features can be seen in Figure 5 (grey line, ascent). After the balloon bursts the 145 radiosonde falls very fast (over 70 m/s in this case) and then slows abruptly -presumably when the parachute opens fully. 146 After this there is a little high frequency variability (but much less than on the ascent) and a gradual decrease in descent rate 147 as the air density increases. Looking at a sample of Lindenberg descents over several weeks (supplementary material), some 148 exhibit an abrupt deceleration and others do not. Figure 6 shows descent rates from the station at Sola in Norway, without 149 parachutes. These do not show the abrupt deceleration, but do show a large range of descent rates. The slower descents tend 150 to have larger high frequency variability. We tentatively suggest that in these cases, the remains of the balloon are acting to 151 slow the descent and there is some pendulum motion. The variability in the descent rate may be due to variability in the mass

Motion of radiosonde during descent 175
A radiosonde as it ascends through the atmosphere can be thought of as a pendulum with a moving pivot (Marlton et al, 2015). 176 As the radiosonde encounters small scale turbulence which is ubiquitous in our atmosphere it causes the radiosonde beneath 177 to swing. The periodicity is a function of string length l given by 178 where g is the acceleration due to gravity. Different radiosonde manufacturers supply different string lengths to their 180 radiosondes, with the aim of removing the radiosondes sensors from the wake effects (Luers & Eskridge 1998  due to transmission drop out. The issue becomes more noticeable in descent data as the radiosonde is now even further from 196 its ground station. This means a traditional Fourier transform method is not appropriate. Thus, a Lomb periodogram is chosen 197 (Lomb 1979), which can generate periodograms which have irregularly sampled data. To ensure that we focus on the motion 198 of the radiosonde we use the processed horizontal wind components to remove the wind field from our raw GPS readings 199 leaving the motion of the radiosonde beneath that balloon. 200  variability. An oscillation is still present indicating that some of the balloon remnants are acting as a parachute. For the 217 parachuted RS41s the spectral width in oscillation widens significantly indicating that there is variation in the motion behavior 218 of the radiosonde. As discussed earlier this may be due to how and when the parachute deployed and if any of the parachute 219 remains entangled with the parachute rigging. The latter is hard to determine without retrieving the radiosonde which is seldom 220 done. We can get a better understanding of the variation in oscillation by looking at individual ascents. levels. One surprise was that the descent profiles (in red) fit the background more closely than the ascent profiles (in black), 283 particularly at upper levels. Comparing individual ascent/descent profiles the descent winds generally appear smoother and 284 this appears to be the cause of the better fit to background. Figure 12 shows the raw 1-second data for a single profile (faint 285 line) and the data after smoothing to remove the pendulum motion (bold line). In this case the smoothing was performed using 286 the GRUAN algorithm (Dirksen et al, 2014), whereas the BUFR reports have smoothing applied by Vaisala MW41 software 287 which is similar but not identical. In both cases the smoothing is a time filter applied to ascent and descent data in the same 288 way. The period of the pendulum motion depends on the length of the pendulum which will be approximately the same for a 289 parachute as for the balloon. (Of course, if there is no parachute there is a different scenario for the descent.) Because the 290 radiosonde is falling faster than it ascended, a filter based on a fixed time interval corresponds to a larger height interval on 291 the descent. Note also that the MW41 processing does not include an inertial correction as used in the AVAPS dropsonde 292 processing (Sect. 1), this counteracts time-lag effects which will be largest when falling fastest. As shown in Figure 12   The green symbols show average values for bins of 10 m s-1. 308

Temperature comparison 309
Firstly, we note that at about 50 hPa, in the extratropics, the ECMWF background is too cool by about 0.5°C (this can be seen 310 against the RS41 ascent data in Figure 14). This is recognised as a model error, due mainly to excessive humidity and hence 311 extra long-wave cooling as shown by Shepherd et al (2018). More recent work on the analysis system has approximately 312 halved the short-range forecast bias (Laloyaux et al., 2020). To provide reassurance ascent/descent pairs were compared to 313 radio occultation (RO) retrievals ( The clearest difference between ascent and descent is that at upper levels the descent temperatures are higher than the ascent 321 values ( Figure 14). This has been noted previously, for different radiosonde types, see section 6. At 10 hPa the descent-ascent 322 difference is over 1.5°C for the Norwegian stations and about half that for the UK and German stations. For Finland the 323 highest standard level reached is generally 20 hPa and the difference there is about 1°C. One hypothesis advanced was that 324 this could be a time-lag effect. However, descending from 30 to 100 hPa the mean temperatures were approximately constant 325 or increasing slightly (Figure 15) suggesting that another explanation is needed. A convincing link to the radiosonde fall speed 326 was found (Figure 16). There is no clear link to the time of day (and solar radiation) as shown by the different coloured 327 symbols in Figure 16. The fall rate correction put forward in the next section (derived originally for a single radiosonde station) 328 does a good job of removing most of the bias (Figure 17 Returning to Figure 14 the large top-level ascent-descent difference in the Norwegian data has disappeared by 300 hPa, but 331 the smaller top-level Finnish difference becomes an offset of 0.2 or 0.3°C throughout the troposphere. The important 332 difference seems to be that the Norwegian radiosondes have a pressure sensor, but the Finnish radiosondes do not. Without a pressure sensor the pressures must be computed and biases in the temperature will feed into later biases in the pressures -334 discussed in more detail in the next section. A smaller version of the same effect can be seen between the German data (with 335 pressure sensors) and the UK data -without a pressure sensor these have an offset of about 0.1°C in the troposphere: smaller 336 than the Finnish data because the UK radiosondes have parachutes.     As the comparison of descent data with NWP model suggests, there is a positive temperature bias for the data measured by 362 descending Vaisala RS41 radiosondes. This bias is bigger in the stratosphere than in the troposphere and is more significant 363 for the data taken from radiosondes without parachutes. 364 As the descent rate often exceeds 50 m s -1 , occasionally even 100 m s -1 , frictional heating seems to be a reasonable explanation 365 of the observed bias. A related issue is recognized for sensors on aircraft, which also measure temperature while moving with 366 high speed relative to the free air (WMO 2018b, section 3.3). For aircraft the kinetic energy is transferred to internal heat 367 mostly by adiabatic compression. In the case of radiosondes we expect that most of the conversion is done by direct collisions 368 of air and sensor molecules (friction), but it is also possible that the effect is done by adiabatic compression in the boundary 369 layer of the sensor. We use a quadratic relationship on descent rate (DR) -this arises from a simple energy balance, independent 370 of the energy conversion mechanism: 371 = .

(2) 372
A is a coefficient, determined below. This is similar to the equation for the heating of aircraft temperature sensors -see 373 Appendix. It is also linked to the 'viscous dissipation' or 'compressional heating' mentioned by Wagner (1964) for 374 rocketsondes (launched high in the atmosphere by a rocket they measure on the descent, slowed by a parachute). This 375 relationship was examined by comparing the descent temperatures with ascent temperatures from the same radiosonde. Most 376 of the data were from the Praha-Libus (Prague) upper air station: 554 descents with average length of descent 23 km. From 377 these data there was a sample of about 528 000 comparable levels. The data covers the period from July 2019 to January 2020. 378 This station used radiosondes RS41-SG, without pressure sensor or parachute -so similar to the soundings from Finland in 379 section 3. 380 The time and space difference between ascent and descent measurement of a particular level starts at 0 km and seconds at the 381 moment of balloon burst and can rise up to 2 hours and 150 km for lower troposphere levels. We expect that this difference 382 will result in deviations of atmospheric measurements, but according to the long-term data there wasn't expected any bias 383 caused by this difference in the stratosphere. For the lowermost layers (below 4 km) was expected warm bias for 06 and 12 384 terms, and cold bias for 00 term due to the diurnal variation. 385 In the next step the sample was binned by DRintervals used were 0-5 m/s, 5-10 m/s etc. There is clearly a quadratic 397 dependence of ΔT on DR in Figure 21 (average ΔT for these bins). The standard deviation of ΔT shown with grey lines is 398 almost independent of DR. The black line is the best estimate with A= 4.05•10 -4 . 399 For DR greater than 110 m/s the fit is slightly less good but the sample size is small with data available from less than 3 % of 400 examined soundings. When equation [2] with A= 4.05•10 -4 is applied as a temperature correction, the root mean square ΔT 401 is lowered from 1.22 °C to 1.06 °C, indicating that the correction explains 24.4 % of the variance seen. The coefficients were also calculated separately for the data from each time of the launch 411 -00, 06 and 12 UTC soundings -the estimates of coefficient A range from 3.9•10 -4 to 4.3•10 -4 412 (   Figure 22 shows the mean and SD of ΔT as a function of height before and after applying the correction. We can see that bias 419 was almost completely removed, except for the lowest layers, where the bias is expected due to diurnal effects. Another notable 420 result was that ΔT SD for heights above 20 km was significantly lowered.   It can be seen from the results in Table 4 that the exact value of the correction coefficient is slightly uncertain. As there is a lot 431 of noise in the data due to other reasons of ΔT than just friction, we would need a larger sample of data to investigate further. 432 433

Indirect effect of heating 434
Some radiosondes measure atmospheric pressure using a sensor and the geopotential height is calculated using hydrostatic 435 equilibrium: dP = -ρ(H) g dH, where density of air ρ depends on pressure, temperature and humidity of the air. This method is 436 used for processing of the data from RS41-SGP radiosondes. The RS41-SG type of radiosonde measures height using GPS, 437 and the pressure is calculated with the very same equation. 438 As discussed in the last subsection, during the descent the radiosonde overestimates air temperature, mostly in the stratosphere, 439 where the descent speeds are high enough to have an impact. This overestimation of temperature leads to underestimation of 440 air density. For the RS41-SGP it means that (negative) height increments are smaller than they should be and thus for the 441 certain pressure level, higher altitudes are reported than they should be. As the height errors accumulate during the descent, 442 the shift of height still remains in the troposphere levels, where direct heating impact is negligible. 443 For the RS41-SG radiosondes, the effect is very similar, resulting in an underestimation of pressure increments, causing a 444 lower pressure for a given height. And vice versa, lower height for the given pressure. Illustration of this effect can be seen in 445 Figure 23. 446 The shift of the profile is visible only if we use as a vertical coordinate the variable which is calculated, not directly measured. 447 As most applications (including many NWP systems) use pressure as vertical coordinate, the effect can be seen for RS41-SG 448 radiosondes. It should lead to an increase in SD when comparing variables to the NWP model, but also to increase of 449 tropospheric temperature bias due to the temperature gradient in the troposphere (as can be seen in Finnish data compared to 450 ECMWF in Figure 14). 451 The effect is clearly visible in Figure 24. For the Praha-Libus data sample ascent and descent levels were matched both using 452 height (blue for bias, red for SD) and using pressure (grey for bias, yellow for SD). In the stratosphere the direct heating of the 453 sensor has a major effect on T differences and the lines are almost the same. In the troposphere, the friction is much lower due 454 to the slower DR and for pressure-matched levels, the shift of the profile caused by accumulated pressure errors is responsible 455 for the majority of bias (difference between gray and blue line). Up to 11 km there is also visible worsening of SD for pressure-456 matched profiles due to this effect.  The positive effect of the pressure correction was checked using Praha-Libus temperature data. Figure 26 shows the bias (solid) 474 and SD (dashed) relative to the ascent for the two versions of temperature descent data (all the data used Tcor according to Eq. 475 (2)). Green lines are for the data matched by reported pressure and red for the data matched by pressure recalculated using 476 corrected temperature. The negative effect of accumulated pressure error due to friction was removed by the pressure 477 correction. The red lines on figure 26 are almost identical with blue and red lines on figure 22 using data matched by height. 478 Overall, it appears that the pressure errors arising from stratospheric heating of the temperature sensor can largely be removed 479 by using corrected temperatures in the hydrostatic calculations.  assimilating European RS41 descent data for 20 January to 28 April 2020. The large-scale impact was very small as expected, 486 but over Europe there were modest improvements in the fit of the 12 h forecast to radiosonde ascent data ( Figure 27). There 487 were improvements over Germany (not shown) and the impact was mixed over Scandinavia. The decision was taken to 488 assimilate only the German descent data for the time being -this is the best subset, because they have parachutes and pressure upper-level descent winds may be reviewed in the future. Note that at the upper-levels the ascent and descent are close together 493 in space and time and so one may not want to assimilate both ascent and descent profiles. As discussed in section 4, some of 494 the bias problems would be reduced if height was used as the vertical coordinate rather than pressure -however this would 495 involve significant work and testing, so there are no plans to do so in the near future. 496 497 Figure 27. Effect of assimilation of all descent data (wind, temperature, humidity; blue line) or just descent winds (red line) 498 2020-01-20 to 2020-04-20. Results are shown for temperature and vector wind fit of 12 hour forecasts to European radiosonde 499 ascents, normalised by the fit of the control forecasts (so values less than 100% indicate improved forecasts). 500

Discussion and conclusions 501
The most obvious difference between ascent and descent data is that the descent temperatures are higher at upper levels. This 502 had been noted before for different radiosonde types, by Tiefenau and Gebbeken (1989) and Vencat Ratnam et al. (2014). The 503 latter did not discuss the cause but Tiefenau and Gebbeken (1989)  that the ascent temperatures were too low due to sampling the balloon wake and adiabatic cooling of the gas within the balloon. 505 Whilst wake effects cannot be completely discounted, our results suggest that the descent temperatures are too high and that 506 this is closely linked to the descent rate ( Figure 16). In Figure 21 the dependence on descent rate appears quadratic. Vaisala 507 are working on updated processing to address the temperature bias and other issues. There has been considerable discussion 508 on the source of the ascent/descent temperature differences. Whilst we cannot definitively explain the heating mechanism a 509 plausible hypothesis is a conversion of kinetic energy via frictional heating. Clearly the falling radiosonde (plus balloon 510 remains and parachute if fitted) are slowed by friction otherwise it would accelerate to much higher speeds during the descent. 511 Future work might include testing radiosonde sensors in a wind tunnel with a flow of 20 m/s or more to see if the heating is 512 replicated (care would be needed with the reference temperature). We have not been able to find such tests in the literature. 513 There is a paper by de Podesta et al. (2018) about the effect of sensor diameter on temperature errors but they were looking at 514 lower flow rates. 515 516 Another difference, that doesn't seem to have been reported before, is that on average the descent winds are smoother than the 517 ascent winds. In part this seems to reflect the fact that ascents are generally more affected by pendulum motion, however 518 inertial effects and the filtering applied to 'remove' pendulum motion also play a role. The smoother descent winds have a 519 closer RMS fit to the NWP winds, but we cannot currently say whether the ascent or descent winds are more accurate. Most 520 studies of radiosondes concentrate on the temperatures and humidities and the winds are somewhat neglected; the use of long 521 strings improves stratospheric temperatures at the expense of increasing the pendulum motions. For aircraft the winds have 522 more than twice the impact of the temperatures on the quality of short-range forecasts (Ingleby et al, 2021) and forecast 523 sensitivity diagnostics suggest that the same is true of radiosondes (Pauley and Ingleby, 2021), partly because satellite 524 instruments primarily provide temperature and humidity data. Experience shows that GPS winds are generally good quality 525 and biases do not seem to be a problem. GPS can provide high vertical resolution winds -but this makes pendulum motion 526 more obvious and avoidance or removal of pendulum motion deserves more attention. Sako and Walterscheid (2016) discuss 527 empirical filtering of wind profiles from radiosondes and Jimsphere balloons used specifically for wind measurement. It seems 528 likely that dropsonde wind profiles are closer to the true winds (Wang et al, 2008) than radiosonde ascent or descent winds. 529 The two balloon ascents of Kräuchi et al. (2016) largely eliminate pendulum motion but need more evaluation. 530 531 Some aspects of the descent data can be improved by estimating and removing heating effects due to high fall rates (on the 532 temperature, and also on the pressure for radiosondes without a pressure sensor). The descent characteristics are more variable 533 than ascent rates in that for balloons with parachutes, the manner in which the parachute deploys can affect the amplitude of 534 the pendulum motion and the descent speed. It is also likely that there can be improvements in the filtering of pendulum 535 motion. Vaisala are working on these aspects but are not yet ready to give a timescale for changes. In principle users could 536 apply bias corrections, but improving the winds is difficult if they have already been filtered. On the whole, it is simplest to 537 stick to the current practice of manufacturers providing best estimate profiles, but more details of the processing would be 538 https://doi.org/10.5194/amt-2021-183 Preprint. Discussion started: 30 June 2021 c Author(s) 2021. CC BY 4.0 License.
welcome and this area should be kept under review. On a similar note there is a question of whether there should be a GRUAN 539 descent product for the RS41 -more work on the uncertainties would be needed for this. There is the wider question of how 540 much the lessons learnt from the RS41 descent are applicable to other radiosondes such as the Meteomodem M10. There is 541 some evidence that pressure sensor accuracy is worse whilst falling fast, but more work on this is needed. However the fall 542 speed should have very little effect on the accuracy of GPS derived positions, because the GPS satellites are moving much 543 faster anyway. 544 545 There is evidence that use of parachutes and/or pressure sensors gives some improvement to the descent data (this will reduce 546 with improved processing/bias correction). There is also the possibility of installing extra receivers so as to obtain more descent 547 data from the lower troposphere (this has been demonstrated in Corsica, Peyrat, pers. comm. 2020). Whether the extra costs 548 are worthwhile would need to be assessed. We note that the impact of extra radiosonde profiles over well-observed Europe 549 will be less than the impact of extra profiles near remote islands or ships. In May 2021 descent data was received from several 550 European ships in the North Atlantic and also a station in Antarctica. ECMWF and DWD have started operational assimilation 551 of a subset of descent profiles -excluding the stratospheric segment with higher average fall rates (arguably it would be better 552 to exclude data based on the actual descent rate, but this is not reported, it is also desirable to exclude values where there are 553 particularly large accelerations, e.g. Fig. 5). The US Navy global model is assimilating all available descent profiles (Pauley,554 pers. comm. 2021), we are not aware of other NWP systems using them yet. NWP systems generally use pressure as the vertical 555 coordinate for radiosonde data, arguably there would be advantages in using height instead. There has been much more use of 556 NWP model fields in this investigation than is traditional for development/validation of in situ observations (but routine now 557 for new satellites -Newman et al. 2020). This means that a much larger sample can be examined. Note that traditional 558 radiosonde intercomparisons (e.g., Nash et al, 2011) can't be used to assess descent data because the multi-radiosonde rig used 559 has various implications for the descent including possible entanglement.