the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 08 Oct 2026
Bias Assessment and Random Forest-Based Correction of Temperature Observations from the Level-Drift Phase of Ascent-Drift-Descent Radiosonde System
Abstract. To facilitate the quantitative application of Ascent-Drift-Descent Radiosonde System (ADDRS) temperature data from the level-drift phase in numerical weather prediction (NWP) models, this study performs a comprehensive quality assessment of the temperature observations using the fifth generation European Centre for Medium-Range Weather Forecasts reanalysis (ERA5) temperature field data as a reference. The results indicate that, due to solar radiative heating, the temperature biases between the level-drift observations and the ERA5 are positive during daytime and tend to increase with the solar elevation angle (SEA). During nighttime, temperature biases are slightly negative. Statistical analysis of temperature biases reveals that the biweight mean values are 3.48 K (daytime) and −0.91 K (nighttime), with corresponding biweight standard deviations of 5.3 K and 2.3 K. Thus, the daytime biweight standard deviation meets the World Meteorological Organization (WMO) "threshold" target, while its nighttime counterpart meets the WMO breakthrough target. Furthermore, a random forest-based temperature bias correction model is developed. After correction, the biweight mean values of temperature biases decrease to 0.52 K (daytime) and −0.23 K (nighttime), and the biweight standard deviations decrease to 2.03 K (daytime) and 1.54 K (nighttime). Both evaluation metrics meets the WMO breakthrough target, with the nighttime metrics notably exceeding it. The probability distribution function of the corrected biases aligns more closely with a normal distribution, demonstrating the effectiveness of the random forest-based model in correcting the biases of level-drift temperature observations. This research establishes a critical foundation for the future application of ADDRS data assimilation in NWP models.
- Preprint
(8983 KB) - Metadata XML
- BibTeX
- EndNote
Status: open (until 12 Nov 2026)
- RC1: 'Comment on egusphere-2026-4913', Anonymous Referee #2, 23 Oct 2026 reply
-
RC2: 'Comment on egusphere-2026-4913', Anonymous Referee #1, 26 Oct 2026
reply
This manuscript evaluates temperature observations from the level-drift phase of the Ascent-Descent-Descent Radiosonde System (ADDRS) at four stations in Guangdong Province and develops a random forest model to reduce departures from ERA5. The topic is highly relevant to Atmospheric Measurement Techniques because prolonged stratospheric balloon observations can provide valuable information in data-sparse regions. The manuscript is well-structured and written in clear language. However, several methodological flaws must be properly addressed prior to publication.
Major Revisions
1、In Section 4, the authors manually applied scaling factors of 0.8 and 1.4 to the model predictions after inspecting the test-set performance. This constitutes a severe case of test-set leakage, and the selection of these factors lacks a sound physical or statistical justification. Standard data partitioning must follow a strict independent "training-validation-test" mechanism, where all hyperparameters or post-processing correction coefficients are determined based strictly on the validation set, leaving the untouched test set exclusively for final, one-time independent evaluation. Furthermore, the use of such piecewise linear adjustment coefficients appears physically arbitrary. It is recommended that the authors incorporate a day/night indicator directly as an input feature during model training; if a single model struggles to capture both regimes, separate models for daytime and nighttime should be considered.
2、The training target of the random forest model is the observation-minus-background (OMB) departure (T_ADDRS - T_ERA5), which fundamentally represents a collocation departure rather than the true ADDRS measurement bias. It lumps together sensor errors, ERA5 errors, and spatio-temporal representativeness/interpolation errors. Training a model on this target and subsequently claiming improved agreement with ERA5 carries elements of circular reasoning. More importantly, because the observed ADDRS temperature serves simultaneously as an input feature and an algebraic term within the target (OMB = T_ADDRS - T_ERA5), the model is highly susceptible to "shortcut learning." That is, the model may simply learn to reconstruct an ERA5-like temperature via its own algebraic mapping rather than genuinely correcting sensor biases. The authors must incorporate independent validation using third-party observational products (such as GNSS-RO temperature profiles, which are typically unbiased in the relevant altitude layers) to prove the objective validity of the corrected results.
3、Line 285 notes the emergence of secondary peaks in the model-predicted departures shown in Figure 10f—an unexpected and insightful phenomenon worth deeper exploration. However, based on the preceding bias analysis, the internal linkage between the model's three input features (temperature, geopotential height, and solar elevation angle) and these secondary peaks remains unclear. The authors should provide a detailed provenance analysis to clarify how the random forest model captures these secondary modes.
4、Section 3 currently focuses primarily on the relationship between the SEA and OMB. To robustly justify the subsequent selection of "sounding temperature, geopotential height, and SEA" as joint input features, Section 3 should be supplemented with bivariate scatter plots and statistical analyses illustrating the relationships between sounding temperature/geopotential height and OMB. This would not only provide a solid intuitive basis for feature selection but also help explain the secondary peak phenomenon mentioned above.
5、The PDF evaluation in Figure 12 lumps daytime and nighttime data together, which directly obscures the clear daytime/nighttime bimodal and secondary peak structures observed in Figure 7. Consequently, the current results are insufficient to demonstrate that the model successfully corrects the complex bimodal bias distribution. It is strongly recommended that the residual PDF evaluations in Figure 12 be presented strictly separated for daytime and nighttime conditions.
6、The SHAP analysis results in Section 5 indicate that the contribution and sensitivity of the "level-drift temperature" to the model output far exceed those of the solar elevation angle. This phenomenon requires a sound physical explanation. I believe this may be explained by the analysis of the relationship between the "level-drift temperature" and OMB addressed in the fourth point above.
Minor Revisions
1、Abstract, line 25: Change "Both evaluation metrics meets" to "Both evaluation metrics meet."
2、Line 40: Correct the citation formatting error: change "Zhou et al,. 2023" to "Zhou et al., 2023."
3、Line 44: Standardize the citation format for "Zhang 2022" (e.g., "Zhang, 2022").
4、Figure 8 caption: Change "Mean, SD and RMS indicates" to "Mean, SD and RMS indicate."
5、Line 265: The claim that random forests "do not suffer from overfitting" is overly absolute; it is recommended to soften this to "are less prone to overfitting."
6、Line 278: The unit for mean squared error (MSE) should be K^2 rather than K.
7、References: The publication year for the citation of Zou and Zeng is erroneously given as 2026; it should be corrected to the actual publication year (2006).
8、Numerical statements regarding WMO breakthrough and threshold requirements must be supported by appropriate official or literature citations.
Citation: https://doi.org/10.5194/egusphere-2026-4913-RC2
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 247 | 67 | 34 | 348 | 30 | 39 |
- HTML: 247
- PDF: 67
- XML: 34
- Total: 348
- BibTeX: 30
- EndNote: 39
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
General
Following the presentation of the ADDRS balloon sounding technique in Cao (2026) (amt-19-3231-2026) and other publications, it is only natural that the quality and applicability of the data obtained using this measurement method is being discussed in further detail. The manuscript is well structured and written in clear, easy-to-understand language.
Specific comments
It should set out more clearly what the WMO criteria are against which the data is compared, and what exactly they mean. No reference is provided; general information is only given in Section 3.3 (l251-2). A consistent explanation is needed of which quantity or (statistical) parameter forms the basis for the comparison with the criteria – and thus for the assessment of data quality – and why, i.e., what is being compared to what (mean bias relative to ERA5, or the dispersion (standard deviation) of the bias, in each case compared with the WMO criteria, or a combination of both?). An example with a clear wording is the report of WMO's 2022 Upper-Air Instrument Intercomparison Campaign (IOM 143).
As far as I can see, the radiosonde models used for this study are not introduced.
Section 3.1
It is stated that, based on a previous study, ERA5 reanalysis data exhibit high accuracy and can serve as a reference. Since the assessment of the data quality resulting from the presented study depends, in principle, on the quality of the ERA5 data, it should be underpinned what is meant with "high accuracy" and to provide at least a few rough figures quantifying the uncertainty of the ERA5 data.
It is stated that ERA5 data on an hourly basis are used for comparison, after spatial and wighted temporal interpolation. What is the spatial resolution? The limited spatial and temporal resolution of the ERA5 data will involve a 'base' scatter. Can the extent and (possible) systematics of that portion of the scatter in the O-B analysis data be estimated?
It is noticable that the TA-TE scatter even at nighttime (roughly within +/-5 K, Fig. 6 in Section 3.2) is considerably larger than the T measurement reproducibility of a modern radiosonde which is in the order of +/-0.5 K. This actually raises the fundamental question of how statistically valid the T-correction method can be when it is based on comparisons between measurements that are very local in terms of time and space, and average model data, which 'creates' noise in the data to be corrected.
The prominent bimodal, i.e. systematic, pattern in the data (Figs. 5-7) are unexpected. They are in the order of 8 K. This is considerable, and my first thought was that this does not reflect the real or 'natural' behaviour state of the atmosphere. The need for further research is stated only in the conclusions. However, I feel it is needed to identify the source and to accordingly handle the data sets before use for further analysis. Otherwise, there remains a reduced level of confidence in the quality and suitability of the input data for this study.
At least, one could start with identifying which of the possible effects are not the cause. Can the pattern be resolved by plotting data subsets categorized e.g. by station or season, or by the two main wind directions, which seem to prevail during the course of the year of collected data suggested by Fig. 3/4 (b) and (c)? Is it in the ADDRS or in the ERA5 data?
What is the potential impact on the following statistical analyses? I am no expert on statistical learning theorybedenken
or on working with such models. But I assume that with the training the random forest model is in particular able to ‘simulate away’ also this effect. What would be the impact if the correction were applied to measurement data that have a more typical or expected, non-bimodal distribution? In other words, how reliable is the correction? If such concerns are unfounded, the authors should address this convincingly.
Section 3.2
Under stratospheric conditions at altitudes of the drift phase there might be some variation of the effective (perpendicular) actinic flux on the sensitive structures of the radiosonde, mainly at times immediately after sunrise or before sunset. However, the actual solar sensor heating will be much stronger affected by the variability of the ventilation in surrounding air. To illustrate this: We would get double solar heating by doubling the irradiance (linear effect), but this would not happen to that extend during a real daytime stratospheric balloon sounding at constant altitude. However, ventilation varies between several meters per second and virtually zero, driving considerable variability of the effective solar heating of the T-sensor due to the highly non-linear efficiency of convective cooling, depending on the actual ventilation. The scatter in the TA-TE data reaching 20 K or more for some data points is probably due to this effect. Furthermore, ventilation decreases with altitude (note that pressure and thus ventilation decrease by a more than a factor of two in the altitude band of 25-30 km where the bulk of the data during the drift phase is measured). Note that the heated parts of a radiosonde, in particular the temperature sensor and sensor boom, with more or less intense spinning around the vertical axis, cannot be considered as a plane surface, so there is much less of a 'cosine effect' with solar elevation with regard to the effective irradiance.
The effective ventilation of the radiosonde floating with the horizontal wind is mainly due the pendulum motion. The pendulum motion is excited by small direction changes of the balloon trajectory caused by wind shear and turbulence. Wouldn't it be conceivable that, for this reason, the ventilation which limits the solar heating of the temperature sensor is, on average, influenced by seasonally varying wind conditions at the sounding altitudes? That might for example explain the larger TA-TE in Winter daytime (Fig. 6(l) and (m)) which significantly contribute to the first (and larger) maximum at ~40°SEA in Fig. 6(a). Is there a correlation between floating altitude and season?
Sections 4,5
Given my limited expertise, I would prefer to refrain from commenting on sections 4 and 5. In my view, the explanation could be more detailed to aid understanding.
Further detailed comments
l22 The target values should be specified here, including the altitude range and application area for which the criteria are defined
l25 "meets" -> "meet"
l54 The same as l22; a small table summarising the WMO target criteria information, to which the manuscript refers repeatedly, seems appropriate somewhere in the introduction. Is this information from WMO's Observing Systems Capability Analysis and Review Tool?
l67 The paragraph suggests that, after correction, radiosonde humidity data measured during the level-drift phase in the stratosphere would be of sufficient quality so that it can be assimilated. Is this correct?
l67-70 Something seems to have got a bit mixed up here: solar radiative heating at nighttime?
Here, biases are compared to targets (not standard deviations of biases as f.e. in the abstract)
l69 There is some inconsistency (for example with the abstract) with regard to which of the two magnitudes, bias and standard deviation, are compared to the WMO criteria.
Fig. 1 Panel (b) might be removed after adding the numbers and station names given in panel (b) next to the station marks in panel (a).
Fig. 2 Panels (e), (f): Use "Lon /°" and "Lat /°" as axis titles, respectively.
l126-7 Given that the "horizontal linear patterns" apparently indicate a systematic non-natural effect, a slightly more detailed discussion of the possible causes helps to legitimate the decision to remove these data from further analysis and to bolster confidence in the quality of the remaining data.
l149 "analyze" -> '"estimate" or "determine" (?)
l156 "sessions" -> "soundings" (?)
Fig. 6 (See also above general comment to Section 3.1) I find it striking that the bimodal pattern (see also Fig. 5 and comment to l162-3) at nighttime are overall weaker or even not present during the summer months April to August.
l229 Can mean T deviations of 1 K to 2 K in atmospheric radiosonde measurements be classified as "weak"?
It seems obvious that positive biases are caused by solar irradiation. But there is no discussion on the possible mechanisms behind the negative nighttime biases.
l230-2 Looking at the "Mean" curve in Fig. 8(a) comprising eight data points with significant fluctuations I would be cautious about identifying three positive peaks and attributing them to a specific cause. Peaks of comparable size are present also in the nighttime curves (see -60° SEA), but are unidentified. Isn't that stochastic?
l232-3 The comment regarding the difference between O_SD and SD (which is to be expected) stands on its own, without further comment.
l369-71 What exactly means 'uncertainties' here?