A quality control method for high-resolution radiosonde temperature and wind data

Fang Yuan, Zijiang Zhou, Jie Liao, Qinglei Li

National Meteorological Information Center, China Meteorological Administration, Beijing, China

Keywords:High-resolution Radiosonde data Quality control TIPEX-III

ABSTRACT This study proposes a method to derive the climatological limit thresholds that can be used in an operational/historical quality control procedure for Chinese high vertical resolution (5–10 m) radiosonde temperature and wind speed data.The whole atmosphere is divided into 64 vertical bins, and the profiles are constructed by the percentiles of the values in each vertical bin.Based on the percentile profiles (PPs), some objective criteria are developed to obtain the thresholds.Tibetan Plateau field data are used to validate the effectiveness of the method in the application of experimental data.The results show that the derived thresholds for 120 operational stations and 3 experimental stations are effective in detecting the gross errors, and those PPs can clearly and instantly illustrate the characteristics of a radiosonde variable and reveal the distribution of errors.

1.Introduction

A network of radiosonde observations can provide a threedimensional structure of the atmosphere, and such radiosonde data have proven useful in a variety of fields, including operational weather forecasting ( Ingleby and Edwards, 2015 ), atmospheric structure research( Seidel et al., 2012 ), global/regional numerical forecasting ( Bai et al.,2017 ), cloud structure studies Wang and Rossow (1995),and verification of satellite measurements ( Kuo et al., 2005 ).Compared with the traditional TEMP reports WMO (2015) that only comprise observations of standard and significant levels, radiosonde data with a high vertical resolution (50 m or finer) can reveal profiles of atmospheric variables in detail.Therefore, they can be used for research on turbulence, wind shear, and gravity waves.Ingleby et al.(2016) showed that the scientific value does not increase linearly with the number of vertical levels,but very high-resolution data allow different users to process the data according to their own needs.

To fulfill the benefits of radiosonde data, quality control (QC) procedures are required before data application.Collins (2001) classified errors in meteorological data into three categories: random error, gross error, and systematic error.The task of QC is to detect gross errors and large random measurement errors.By analyzing the spatiotemporal distribution of those errors, the systematic errors that may result from defects of measurement devices or faulty operations of data processing might be revealed.

In order to develop an automatic, timely, and effective QC algorithm that can be used operationally, an indicator that is robust to outliers is required.In addition to the rather broad but still useful “gross limit ”(e.g., Durre et al., 2008 ; Loehrer et al., 1996 ), the climatological limit check is a preferable choice.For a normal distribution, the sigma test is usually used to flag the data if they are more than six standard deviations(STDs) from the respective mean ( Durre et al., 2010 ).Although useful,as pointed out by Durre et al.(2008),this approach has some limitations.Besides the need for sufficient amounts of data, the large number of gross outliers would contaminate the mean value and STD, and it is unsuitable for variables whose distributions are zero-bounded and highly skewed,such as wind speed.

Graybeal et al.(2004) obtained the climatological limits by ranking all the values and then subjectively analyzing the 10 lowest and highest values (approximately the 0.002 and 99.998 percentiles, respectively, in their sample) to screen for outliers.Houchi et al.(2015) proposed a statistical quality control (SQC) method based on the percentiles and mean values of the probability density function at each vertical level (1 km vertical bin size), to isolate and eliminate outliers in high-resolution wind data.In our tests, this SQC method proved effective in outlier detection.However, it is unsuitable for real-time operational use because of the data quantity requisition, and more importantly, it cannot satisfy the need for timeliness.To this end, in this study, we provide an approach to derive the thresholds from the vertical percentiles of temperature and wind speed, and the thresholds can be used for the QC of operational real-time data and field experiment data.

Table 1 Definitions of the 64 vertical bins.

2.Data and method

The high-resolution dataset used in this study is from the Chinese operational upper-air observing system, which is made up of 123 sites all over the country.Observations are usually made twice a day, at 0000 UTC and 1200 UTC, and at some stations there is a third additional observation at 0600 UTC during the flood season.Occasionally,there might be a fourth additional observation at 1800 UTC.Radiosonde TEMP reports are also used here for comparison with the high-resolution data.To verify the effectiveness of this method in field experiments,the temperature radiosonde data from the third Tibetan Plateau (TP)atmospheric scientific experiment (TIPEX-III, http://data.cma.cn/tipex)( Zhao et al., 2018b ) are employed.

For the application, firstly the atmosphere is divided into tens of vertical bins (i.e., layers).Houchi et al.(2015) set the vertical bin size(i.e., thickness of a layer) as 1 km.Here, in this study, we enhance the vertical stratification to obtain more detailed information, especially in the lower level, where the meteorological elements vary dramatically with height.Furthermore, we use pressure instead of altitude as the coordinates because in the high-resolution radiosonde data file there are only pressure and geopotential height data but without altitude (i.e.,height above sea level or ground).As shown in Table 1, for the lower level (below 700 hPa), the bin size is set to 20 hPa, roughly equivalent to 0.25 km.So, the lowest 17 bin ranges are [1040, 1020), [1020, 1000),…, [720, 700).For the levels from 700 hPa to 10 hPa, the bin size is set roughly equivalent to 0.5 km, and above 10 hPa observation data are rare and the bin size is 1 hPa.According to Table 1, the 18th to 27th bins are [700, 660), [660, 620),…, [340, 300), the 28th to 37th bins are[300, 280), [280, 260), …, [120, 100), the 38th to 44th bins are [100,90), [90, 80), …, [40, 30), the 45th to 54th bins are [30, 28), [28, 26),…, [12,10), and the highest 10 bins are [10, 9), [9, 8), …, [1, 0) (units:hPa).

The second step is to collect data for each vertical bin.Theoretically,the rate of ascent of the air balloon is set to a constant value (approximately 6.6 m s), and the radiosonde samples every second, and therefore one launch will obtain thousands of observation data.For example,the launch from Beijing station at 0000 UTC on 24 May 2013 (hereafter BJ-2013-05-24-00 launch) reached a height of 3.2 hPa, and gathered data for 7326 levels.A level here means a group of data, including time departure from the launch, pressure, temperature, relative humidity, wind speed and direction, geopotential height, and balloon location information.Each vertical bin will collect any temperature or wind data whose corresponding pressure falls into the corresponding range.Take the 23rd bin as an example.The bin range is [500 hPa, 460 hPa), and for the BJ-2013-05-24-00 launch, there are 85 temperature data with pressure in the range.For the whole year of Beijing station, there were 821 launches and a total of 65,757 temperature data fell into the 23rd bin.

The most important step is to calculate the percentiles, mean values,and STDs in each vertical bin.Following Houchi et al.(2015),the

n

values in each vertical bin (

x

,

i

= 1, …,

n

) are sorted in ascending order,and then for any given percentile

p

,the percentile value

x

p

is defined as a linear interpolation between the closest ranks as follows:

Table 2 Percentiles and corresponding profile numbers.

Fig.1.Percentile profiles (PPs) of temperature for the year 2013 (solid lines,different colors denote different percentiles) and thresholds based on PPs (magenta dashed lines) and based on the sigma test (green dashed lines).

where

i

= 1,…,

n

.The percentile values of all vertical bins then form a profile.The percentiles and corresponding profile numbers used in this paper are the same as in Houchi et al.(2015),as shown in Table 2 ,and the minimum and maximum values correspond to the 0 and 100%percentiles (i.e., the 1st and 17th percentiles).We calculated the annual percentile profiles (PPs) of temperature and wind speed for every operational station over the period of 2010 to 2018.Fig.1 shows an example.Solid lines are the PPs of temperature in 2013, and different colors denote different percentiles.As mentioned in the introduction, we derived the fixed thresholds for each station that can be used in the operational QC procedure.The criteria are interpreted as follows.Take wind speed as an example.Consider two variables for each vertical bin: Δ

x

is the difference between two adjacent profiles, and

r

i

is the ratio of the difference to the inner profile:

where

i

= 16, 15, …, and 16 means it starts from the 16th profile (i.e.,the 99.9% profile).The maximum value in each vertical bin is not considered since it usually contains many gross errors if the sample size is big.For wind speed, if Δ

x

is less than 20 m s,and meanwhile

r

is less than 30%, select

x

i

as the preliminary threshold; otherwise, check the inner profile (i.e., the 15th profile).In most cases, the 14th (99.7%)profile would be enough, but for some extreme conditions, manual inspections are needed.Note that the vertical bins are independent of each other, and the chosen threshold in a vertical bin is irrelevant to that in other bins.

In operational use, it needs an expanding tolerance to avoid excessive QC.Here, add double STDs as the tolerance to the chosen percentile values while the STD is less than 15 m s1 ; otherwise, the tolerance is 30 m s.It is important to note that the mean and STD values are calculated from 98% (i.e., excluding the largest 2% wind speed) of the total data since these data are sampling solely the (presumably reasonable) core portion of the distribution.Manual inspection is necessary to avoid the anomalous STD caused by too many errors.Finally,use a five-point least-squares fitting algorithm to smooth the threshold profiles.

The thresholds for temperature follow the same algorithm, except for the different parameters.The absolute value of Δ

x

i

should be no more than 5°C, and the tolerance is two STDs or 20°C if the STD is larger than 10°C.The STD values are calculated from data that are larger than 1% and smaller than 99% of profiles (i.e., 98% of the total data).Notice that the temperatures have both upper and lower thresholds.We use the 23rd bin that ranges from 500 hPa to 460 hPa in Fig.1 to explain how to operate the algorithm.According to Eqs.(1) and (2),the 2nd and 3rd percentile values are − 41.65°C and − 40.73°C, then | Δ

x

i

| is 0.92°C, less than 5°C, and

r

is 2.26%, which means we can use the 2nd percentile value as the preliminary lower threshold (i.e., − 41.65°C).On the other hand, the 15th and 16th percentile values are − 1.82°C and − 1.24°C, so the | Δ

x

| here is 0.58°C, and

r

is 31.8%, over the 30% limit.Then, we have to check the 14th percentile value, which is − 2.10°C.The associated | Δ

x

| between the 14th and 15th percentile values is 0.28 and

r

is 13.3%, which means the preliminary upper threshold is − 1.82°C (i.e.,the 15th percentile value).For the tolerance, the STD of this vertical bin is 9.66°C, less than 10°C, so the tolerance of this layer is 2 STDs, i.e.,19.32°C.Finally, the thresholds for the 23rd vertical bin are − 61.97°C and 17.50°C.The parameters for Δ

x

and are

r

are mainly based on trial and experience, and taking temperature as an example, we tried a range for Δ

x

from 3°C to 6°C and for

r

from 10% to 50%, and visually checked hundreds of observation profiles to examine the effectiveness of the parameters.We aim to make the thresholds equivalent to the magnitude of the sigma test (i.e., departure of six STDs from the respective mean;Durre et al., 2010 ) but better-fitting in detail.Here, we still use data from Beijing station as an example ( Fig.1 ).Note that the 0 and 100%percentiles are the observed minimum and maximum values in each vertical bin, and all the observations fall into this range.It can be seen that the thresholds based on PPs (magenta dashed lines in Fig.1 ) are more consistent with the 0 and 100% percentile profiles than the thresholds based on the sigma test (green dashed lines).

The final thresholds that will be used in the QC procedure are the maximum threshold values of the nine years (2010–2018) in each corresponding vertical bin.

Fig.2.(a) Percentile profiles of wind speed for the year 2013 (solid lines, different colors denote different percentiles) and climatological limit check thresholds(black dotted line) of Beijing station (WMO No.54511) and (b) climatological limit thresholds of temperature (black dotted line) and an observation profile on 29 March 2016 of Naqu station (WMO No.55299).

3.Results

3.1. Performance of the quality control procedure

Under normal circumstances, from the distribution of PPs, it is easy to figure out the regular vertical variations of a variable.Fig.2 (a) shows the PPs of wind speed calculated from a total of 821 cases from Beijing station in 2013, and, as can be seen, the wind speed increases with height in the middle and low levels, peaks at 200 hPa, and then decreases with height.The rightmost (i.e., 17th) profile is the maximum value of the wind speed in each vertical bin, and the sharp angles indicate some gross errors that may be caused by the poor performance of the radar system.The black dotted line is the climatological limit threshold derived from the algorithms described in Section 2,and it shows that most of the gross errors can be excluded by the climatological limit.Fig.2 (b) shows an example for temperature.It is clear to see that the observation profile on 29 March 2016 from Naqu station (red line) exceeds the climatological limits of this station (black dotted line), and therefore will be rejected by the QC system.Such a situation is usually caused by the failure of temperature sensors, and should be excluded from the dataset before application.To summarize, from Fig.2 we can see that this method can not only exclude the gross errors but also give an intuitive view on if and how much an observation deviates from the normal range.As pointed out by Ciesielski et al.(2012),manual inspection is also necessary because subtle errors in sonde data are often difficult to identify with objective procedures, and useful parameters or software could facilitate this remarkably.The climatological limits here could not only serve as a threshold in the QC system but also illustrate the normal range of a station, which would be highly conducive to a quick and subjective judgment.

3.2. Systematic error of wind speed

When analyzing wind speed, we found that at some stations the PPs are quite chaotic, especially the outermost four profiles ( Fig.3 (a) is an example).Further analyses show that there are a few anomalous observations, like the red line in Fig.3 (b), in which a large portion of the values exceed 120 m s1.Meanwhile, we recalculated the wind speed from the elevation angle, azimuth angle, and distance ( Li et al., 2018 ),and found that the recalculated wind speeds distributed in a much more reasonable range (except for some gross errors, which were caused by the erroneous radiosonde positions).More importantly, the recalculated wind speeds are consistent with the TEMP report (green dots in Fig.3 (b)).The consistency reveals that the fallacious high-resolution wind speed observations are likely caused by software bugs that lead to the wrong wind speed calculation from the radiosonde positions.The PPs in Fig.3 (c) are derived from the recalculated wind speeds.Compared with Fig.3 (a), except for the maximum (17th) profile, there is substantial improvement between 700 hPa and 100 hPa, and the outermost six profiles (from 90% to 99.9%) are consistent with each other.Nevertheless, the distribution of PPs in Fig.3 (c) is unsatisfactory.There are some obvious gross errors revealed by the maximum profile, and the large wind speeds near the ground indicate poor performance of the wind-finding system, which may be caused by the malfunction of the L-Band radar at low elevations ( Chen et al., 2017 ).This means the QC procedure must be applied to the radiosonde positions before calculating the wind speed.

Fig.3.(a) Percentile profiles of wind speed for the year 2018, (b) profiles at 1200 UTC 8 March 2018: observed high resolution wind speed (red), recalculated wind speed (blue) from observed position of the balloon, the wind speed from the TEMP report (green dots), and (c) percentile profiles of recalculated wind speed for the year 2018 at Wudu (WMO No.56096) station.

3.3. Application in field experiment data

Owing to the high elevation and naturally harsh environmental conditions of the TP, and the less-developed logistics, observational data for the TP are scarce.So, field experiment data are extremely valuable.By using Vaisala portable radiosonde systems, TIPEX-III conducted intensive routine radiosonde observations at Shiquanhe (WMO No.55228),Gaize (WMO No.55248), and Shenzha (WMO No.55472) stations in the western TP at 0000, 0600, and 1200 UTC every day from June to August in 2014.Simultaneously, the Institute of Plateau Meteorology of the China Meteorological Administration carried out radiosonde observations as joint research at Jinchuan (WMO No.56168) and Jiulong stations (WMO No.56462) ( Zhao et al., 2018a,2018b ), and 709 soundings were collected.Fig.4 shows the distribution of the TIPEX-III stations.Three stations are located in the western TP along the east–west direction and two stations are located in the central region of Sichuan Province (red stars).Blue dots in Fig.4 are operational radiosonde stations, so we can see that the TIPEX-III stations are a good supplement to the observation network.

In this study, we analyzed the quality of temperature data from TIPEX-III with the methods described in Section 2,but there with a slight modification.Because there are no new data updated to the dataset,when calculating Δ

x

i

,

i

starts from the 17th profile (i.e., the maximum profile) and the tolerance is not needed.Fig.4 (a–e) exhibits the PPs of five stations, and the black dashed lines denote the thresholds used in the limit check of the QC applied to the dataset (the algorithm is beyond the scope of this paper).As can be seen, in most circumstances the percentile distribution is relatively “narrow ”, indicating small STDs in the vertical bins and few gross errors.Fig.5 (e) is one of the few exceptions.The six outermost profiles (from the 99%, i.e., 12th, percentile)vary from 20°C to 60°C from the 200 hPa to 50 hPa level, with a maximum value of 56.6°C at about 200 hPa.The variations of PPs reveal that there are either systematic errors or incorrect profiles.After further analyzing the observational profiles, we find the incorrect ones as shown in Fig.5 (f).The upper part of the profile seems to moves to the right for 110°C on the graph, and is rejected by the thresholds of the limit check(black dashed lines in Fig.5 (e)).

Fig.4.Distribution of radiosonde sites: operational sites (blue dots) and TIPEX III sites (red stars with station IDs).

4.Conclusions and discussion

In this paper, to carry out a QC procedure and evaluate the Chinese high vertical resolution (5–10 m) radiosonde data, we divided the whole atmospheric layer into 64 vertical bins, constructed the PPs, and then derived the fixed thresholds as the climatological limits for each station that can be used in real-time operational QC procedure.Values falling out of the thresholds would be flagged.The thresholds also illustrate the normal range of an element at a station.By comparing a specific profile with the thresholds, one can tell whether and how much an observation deviates from its normal range, and it is especially helpful when manual inspection is needed.The application of this method to TP field data validates the effectiveness of the algorithm in experimental data.Moreover, the thresholds can be beneficial in the QC procedure of other upper-air data such as aircraft reports and wind profiler reports, which are too scarce to create their own QC parameters.

Fig.5.Percentile profiles of temperature from TIPEX-III for the stations (a) Shiquanhe (WMO No.55228), (b) Shenzha (WMO No.55272), (c) Jinchuan (WMO No.56168), (d) Jiulong (WMO No.56462), and (e) Gaize (WMO No.55248), and the associated limit check thresholds (black dashed line) and (f) an observation profile on 12 July 2014 of Gaize station.

The distribution of PPs itself contains a lot of useful information.In this study, the PPs of wind speed exhibited untidy variations, and through further inspection, the systematic errors probably caused by a data processing software bug were confirmed.Fortunately, this problem can be solved by recalculating the wind data according to the elevation angle, azimuth angle, and distance.The PPs could clearly and instantly reveal the gross errors and systematic errors (if any), which makes it a good tool for final checks before releasing a dataset since human mistakes can happen during any procedure.

Funding

This work was supported by the National Innovation Project for Meteorological Science and Technology [grant number CMAGGTD003-5] and the National Key R&D Program of China [grant number 2017YFC1501801].