A Deep Convolutional Neural Network for Retrieving Tropospheric Temperature and Moisture Profiles from Refractivity over Tropical Oceans: Framework Development and Characterization

Muralidharan, Santosh; John, Harris; Mishra, Deepak

doi:10.5194/egusphere-2026-2027

Preprints

https://doi.org/10.5194/egusphere-2026-2027

Preprints

29 May 2026

| 29 May 2026

A Deep Convolutional Neural Network for Retrieving Tropospheric Temperature and Moisture Profiles from Refractivity over Tropical Oceans: Framework Development and Characterization

Santosh Muralidharan, Harris John, and Deepak Mishra

Abstract. Retrieving profiles of temperature and water vapor from atmospheric refractivity over tropical oceans constitutes an inherently underdetermined problem. Conventional one-dimensional variational methods resolve this through numerical weather prediction (NWP) priors, potentially propagating model biases into retrieved profiles and limiting the independence desirable for climate monitoring applications. We present a deep learning retrieval that substitutes learned statistical constraints for model-dependent priors. A convolutional neural network, trained on approximately 20,800 high resolution radiosonde profiles —combining reference-grade GCOS Reference Upper-Air Network (GRUAN) measurements with quality-controlled operational GCOS Upper Air Network (GUAN) and field campaign data— predicts dry refractivity and partial pressure of dry air as intermediate targets; temperature, water vapor pressure, and relative humidity are derived analytically. The model achieves water vapor pressure root-mean-squared errors of ~0.5 hPa near the surface (decreasing with height) and relative humidity errors below 6 % from 100 m to 10 km, performance comparable in magnitude to the mean measurement uncertainty of state-of-the-art radiosondes. Temperature errors are ~1.5 K near the surface, improving to ~0.5 K in the 10–15 km range where dry refractivity dominates. Evaluation against more than 27,400 geographically independent radiosonde ascents across the tropical Pacific, Indian Ocean, and Atlantic—including observations from the 1992–93 TOGA-COARE field campaign, two decades predating the training period—demonstrates robust generalization within the tropical marine atmosphere. The framework accepts only refractivity vertical structure as input, with no dependence on geographic coordinates or NWP background states. This paper establishes the retrieval framework and characterizes performance under ideal input conditions; Part 2 addresses application to satellite observations.

Received: 09 Apr 2026 – Discussion started: 29 May 2026

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.

Download & links

Preprint (PDF, 2425 KB)

Supplement (2804 KB)

Download & links

Santosh Muralidharan, Harris John, and Deepak Mishra

Status: final response (author comments only)

RC1:
'Comment on egusphere-2026-2027', Anonymous Referee #1, 24 Jun 2026
General Comments
The manuscript by Muralidharan et al. presents a neural-network-based retrieval framework for deriving thermodynamic profiles from GNSS-RO refractivity observations over tropical oceanic environments. The topic is relevant for data assimilation and for demonstrating the potential impact of GNSS-RO measurements within the global climate observing system, as well as for assessing the value of upper-air reference observations as a collateral outcome of their investigation. The proposed deep-learning methodology itself is not fundamentally novel, as similar machine-learning approaches have already been explored for a variety of atmospheric retrieval and inversion problems. Nevertheless, its application to GNSS-RO retrievals and upper-air observations remains potentially valuable and the study hasmay have the potential to provide insights that extend beyond the machine-learning methodology itself. However, the current experimental design does not yet fully exploit this opportunity.

A key limitation is the absence of a rigorous comparison with established 1D-Var GNSS-RO retrieval frameworks, which remain the standard physically constrained approach for combining observations with prior information. Without such a benchmark, it is not possible to assess whether the proposed method provides a genuine improvement over existing methodologies in terms of retrieval accuracy, vertical structure representation, or uncertainty characteristics.

Another key limitation is the restricted geographical representativeness of the validation framework. The analysis relies primarily on island-based tropical stations, which may not adequately sample the variability of open-ocean atmospheric conditions. In particular, boundary-layer processes and moist convection regimes in the tropical Atlantic and broader oceanic regions are not explicitly tested. This raises concerns about the claimed generalization to “tropical ocean environments”.

The study does not also convincingly demonstrate true spatial or instrumental generalization beyond a relatively constrained set of island-based and campaign-oriented radiosonde observations. Although the use of multiple datasets spanning different periods is presented as a strength, in practice the evaluation framework mixes heterogeneous observing systems without a systematic stratification of measurement quality, correction level, or representativeness errors.

In particular, the role of GRUAN observations as a reference standard is not fully exploited or critically assessed. While GRUAN is appropriately treated as a high-quality dataset, the analysis does not isolate the specific benefit of using GRUAN versus non-GRUAN radiosondes, nor does it quantify how mixed-quality observations may influence both training and evaluation outcomes.

Moreover, the inclusion of historical datasets (e.g., TOGA-COARE) together with modern operational radiosondes introduces potential inhomogeneities due to known long-term changes in radiosonde instrumentation and correction practices (e.g., solar radiation bias, humidity sensor dry bias). These issues are not explicitly addressed through homogenization or sensitivity analysis.

Therefore, the manuscript contains a promising methodological contribution and is potentially suitable for publication in AMT. However, substantial additional analysis is required before the conclusions can be considered robust.

Specific comments are provided below.

Specific Comments

1. Although 1D-Var is discussed conceptually, no quantitative comparison is provided under consistent observational conditions. Given that 1D-Var is a standard physically based retrieval method for GNSS-RO, such a benchmark is necessary to evaluate:

retrieval accuracy,

sensitivity to prior information,

vertical structure representation,

and robustness across regimes.

In this first manuscript (where the authors note that a second paper will include real GNSS-RO measurements), it is essential to provide statistical comparisons with GNSS-RO data produced by the authors for the exact same region analyzed, using a suitable collocation criteria. Without this, the claimed advantages of the neural network approach cannot be verified.

2. The dataset includes radiosonde data from multiple instruments and processing pipelines. However, systematic errors (such as solar radiation bias and humidity sensor dry bias, which are well-known and characterized in the literature for older radiosonde models like the RS80 and RS92) are not explicitly addressed. The statistical analysis seems to work independently of potentially inaccurate input data. Furthermore, it is unclear whether the algorithm can perform bias adjustments by learning from relationships with modern radiosondes. While this could potentially solve several homogenization problems, such a major claim must be demonstrated. To prove this, the authors need to conduct additional experiments where data from recent radiosonde models are separated from older ones. Indeed, including data from historical campaigns (e.g., TOGA-COARE) introduces temporal inconsistencies due to changes in instrumentation and calibration practices. Currently, no homogenization strategy or sensitivity analysis is provided to evaluate the impact of these differences on retrieval performance and generalization.

3. In analogy to the previous comment, GRUAN observations are appropriately recognized as high-quality reference measurements; however, their role is not isolated in the analysis. The manuscript does not quantify the benefit of GRUAN versus non-GRUAN radiosondes, nor does it assess whether mixing different observational standards introduces systematic effects in both training and evaluation. A stratified analysis separating GRUAN-only, corrected operational, and uncorrected radiosonde datasets is required to properly assess robustness and avoid conflating observational quality with model skill. For example, statistics based on GRUAN station using the GRUAN data products and using, for the same stations, data from other archives such as IGRA, maintained by NOAA-NCEI.

4. A more detailed justification of the dataset selection strategy would help readers understand the intended scope and applicability of the retrieval framework. The validation framework relies primarily on island-based radiosonde stations located within broadly similar tropical environments. While the use of geographically distinct stations is valuable, validating the model on island sites that share comparable climatological regimes with the training dataset does not, by itself, demonstrate true spatial generalization. To support the claimed transferability of the approach, the authors should provide a detailed statistical characterization of the radiosonde profiles used for training and validation, including the distribution of the key thermodynamic variables. Without such an analysis, it remains unclear whether the validation dataset is sufficiently independent from the training dataset in terms of atmospheric state space. The possibility that the reported performance largely reflects climatological representativeness rather than genuine model generalization cannot be excluded. Indeed, the high representativeness of tropical observations over climatological timescales has been documented extensively in the literature (e.g., Weatherhead et al. (2017)). The manuscript would also benefit from validation over additional tropical and subtropical oceanic regions. The tropical Atlantic is notably underrepresented despite offering an opportunity to test the robustness of the retrieval framework under different thermodynamic and dynamical conditions. The use of additional island stations (e.g., St. Helena), coastal African sites, or independent ship-based radiosonde observations, largely available through NOAA-NCEI IGRA would considerably strengthen the evidence for geographical generalization. Likewise, the rationale for excluding the large number of available ship-based radiosonde observations should be clearly explained.

5. The impact of radiosonde drift is not considered in the validation framework, despite its potential importance when using island-based observations as proxies for open-ocean conditions. Recent studies have shown that representativeness errors can depend substantially on balloon drift characteristics, particularly when sondes launched from islands are advected over the surrounding ocean. While island observations may provide a reasonable characterization of the free troposphere, this assumption is less robust within the planetary boundary layer, where horizontal gradients and local surface influences can be significant. Given the emphasis placed on the representativeness of island-based radiosonde observations for tropical oceans, the manuscript would benefit from a quantitative assessment of balloon drift statistics, including typical displacement distances and their potential impact on retrieval validation.

6. The authors appropriately acknowledge the importance of observational uncertainties and discuss uncertainty estimates associated with some of the datasets used in the study, particularly GRUAN observations. This is appreciated and demonstrates awareness of the issue. However, the treatment remains largely qualitative and does not fully exploit one of the key strengths of GRUAN, namely the availability of traceable and well-characterized uncertainty estimates for each profile and measurement level. In particular, the manuscript frequently compares retrieval RMS values with observational uncertainties, implicitly suggesting a level of equivalence between these quantities. From a metrological perspective, however, retrieval error, measurement uncertainty, and representativeness error are fundamentally different concepts and should not be interpreted interchangeably. GRUAN was specifically developed to provide reference observations with traceable uncertainty estimates and documented uncertainty budgets, which distinguishes it from most operational radiosonde archives. The issue is further complicated by the fact that the study combines GRUAN observations with several non-GRUAN datasets for which uncertainty information is either unavailable. Consequently, it remains difficult to assess how observational uncertainty contributes to the reported retrieval performance and what specific benefit is obtained from including reference-quality observations within the training and validation framework. While a full uncertainty propagation through the neural-network retrieval may be beyond the scope of the present work, a more rigorous discussion separating observational uncertainty, representativeness error, and retrieval error would substantially strengthen the interpretation of the results and the scientific value of the GRUAN component of the study.

7. The manuscript would benefit from substantial restructuring and condensation. Several sections currently contain material that appears misplaced with respect to their intended purpose. For example, parts of the Introduction read more like a discussion of results and limitations, whereas some methodological descriptions are embedded within broader contextual sections. As specific examples, the material currently presented in lines 169-191 appears more appropriate for the Discussion section, while parts of lines 196-210 would fit better within the section describing the methodologies. More generally, the manuscript is considerably longer than necessary and contains a substantial amount of repetition. The Introduction could be streamlined to focus more clearly on the scientific motivation and knowledge gap, while several interpretative statements could be moved to the Discussion. Likewise, the Conclusions and Limitations sections could be significantly shortened and made more focused on the key findings and demonstrated evidence. A more concise structure would improve readability and make it easier for readers to distinguish between established facts, methodological choices, results, and interpretation.

8. Figure may be improved in their quality; in particular Figure 6, would benefit from improved color scaling and normalization choices. The current visualization may compress or mask important differences between datasets or model configurations, thereby limiting interpretability of spatial or vertical error structures. A more carefully tuned colormap range (or the use of anomaly-based scaling / normalized difference plots) is recommended in order to avoid visually minimizing meaningful differences between experiments and to better highlight regional or regime-dependent variability. Finally, Table 1 is not needed. Instead, some references associated with the GRUAN observations appear somewhat incomplete.
Citation: https://doi.org/10.5194/egusphere-2026-2027-RC1

RC2: 'Comment on egusphere-2026-2027', Anonymous Referee #2, 02 Jul 2026

Recommended decision: Major revision

General comments

This is an interesting paper. The manuscript addresses an important and timely problem in atmospheric measurement science: how to retrieve temperature and water-vapor/moisture profiles below 20 km from refractivity N over tropical oceans. The authors develop a deep convolutional neural network framework that uses radiosonde-derived refractivity as input and radiosonde thermodynamic variables as targets. The paper uses high-resolution radiosonde records, GUAN/GRUAN-related observations, and field-campaign soundings from DYNAMO and TOGA-COARE to evaluate whether the model can retrieve T and W without using an NWP background during inference. This topic is suitable for AMT because it concerns retrieval methodology, characterization, and potential observational applications.

2) The key and unique contribution is the attempt to train a model that takes N as input and retrieves T and W below 20 km. In my view, the novelty lies not only in the use of a CNN. The more important contribution is whether the authors have found a defensible way to learn the information contained in refractivity and to separate the temperature, pressure, and moisture contributions in a physically meaningful way. Therefore, two questions (as mentioned in the paper) are central for this paper. i) what is the retrieval accuracy? ii) how much does the AI approach learn beyond climatological covariance, radiosonde sampling priors, and simple regression relationships? I expect authors to provide solid evidence to demonstrate i) and ii), so that this approach will be useful for the future inversion for deriving T and W using real RO observations.

My evaluations for i) and ii) presented in the current paper are:

i) Evidence for accuracy. The manuscript provides useful evidence: RMSE, bias, standard deviation, profile examples, confidence intervals, station/campaign evaluations, and comparisons among training, GUAN, DYNAMO, and TOGA-COARE datasets. These results show that the framework can reproduce radiosonde thermodynamic profiles from radiosonde-derived N under ideal input conditions. However, this is still a closed-loop evaluation. Because the input N and target T/W are derived from the same radiosonde ascents, the evidence for accuracy does not yet demonstrate retrieval performance on actual GNSS-RO refractivity profiles.

Please explain how the current setting can be adapted to actual GNSS-RO refractivity profiles?

I understand that for real RO data, vertical smoothing, horizontal averaging, super-refraction, critical refraction, lower-tropospheric N biases, noise, and missing low-level information may substantially change retrieval errors.

However, at least, you may add vertical N uncertainty to your test set then do the 1D var exercise.

ii) Evidence for how much the AI model learns. The paper provides evidence that the model learns nontrivial empirical structure, particularly through cross-station/campaign tests, profile examples, and the perturbation experiment. These tests are useful, but they do not quantify the incremental value of the AI approach. To support the claim that the AI model has learned useful retrieval information, the paper needs baseline comparisons and ablation tests. At a minimum, the authors should compare against altitude-only climatology, basin/season climatology, multilinear regression, simpler MLP/random-forest retrievals, direct prediction of T/W/RH, and the same CNN without the intermediate targets or wavelet-transform input. Without these tests, it is difficult to know how much skill comes from the proposed AI architecture and how much comes from the limited tropical state space and strong T-W-N covariability.

Here I summarize my claim-evidence assessment

Claim or issue	Evidence provided	Assessment	Needed revision
Retrieval accuracy	RMSE, bias, standard deviation, confidence intervals, profile examples, and independent station/campaign evaluations.	Useful for ideal-input characterization, but not sufficient for real RO retrieval accuracy.	Add RO-like smoothing/noise/bias tests and real GNSS-RO validation.
AI learning	Cross-station evaluation, WCTN/intermediate-target design, perturbation experiment, and profile examples.	Suggestive, but not quantitative. The paper does not show the gain over simple baselines.	Add climatology, regression, MLP/random forest, direct-output CNN, and ablation tests.
NWP independence	The retrieval does not use an NWP background during inference.	Partly supported. The model still depends on the empirical radiosonde prior used in training.	Use “NWP-independent during inference” and avoid “prior-free” language.
Satellite RO applicability	The paper motivates GNSS-RO use and states that real RO validation will be Part 2.	Not yet demonstrated. Radiosonde-derived N is not real RO N.	Keep claims restricted to framework characterization under ideal input conditions.
Uncertainty	RMSE, bias, standard deviation, and confidence intervals.	Incomplete. RMSE is not per-profile retrieval uncertainty and CIs likely assume independent profiles.	Add uncertainty propagation, block bootstrap, sensitivity tests, and per-profile uncertainty estimates.

Key problems that should be addressed

1) The writing sometimes presents claims more strongly than the evidence supports. The manuscript should distinguish what is demonstrated from what remains a future application. The current evidence supports the ideal-input radiosonde-to-radiosonde framework characterization. It does not yet support broad claims of general-purpose retrieval, independent observational products, input-source independence, or satellite RO applicability. The authors should revise the abstract, introduction, discussion, and conclusions so that each claim is tied directly to the evidence presented. (Some of the related lines can be seen on 16-40, 23-31, 33-40, 181-190, 605-618, 741-744, 888-895, 950-972, 1019-1025, and 1067-1073.

2) Ideal inputs and retrievals do not prove that the approach works for real RO data. The paper uses radiosonde-derived refractivity as the input. This is internally consistent and useful for testing the framework, but it is not equivalent to GNSS-RO refractivity. Real RO profiles have vertical smoothing, horizontal averaging, tracking errors, super-refraction, critical refraction, lower-tropospheric negative N bias, and missing/low-level data. These effects directly influence the N-to-T/W inversion. Therefore, the present results should not be described as validation for real satellite RO retrievals. (see 16-40, 33-40, 83-113, 315-323, 605-618, 745-781, and 1067-1073).

3) The paper does not explain why temperature retrievals below 20 km are not better. Below 20 km, temperature retrieval from refractivity is underdetermined because N combines dry-air and water-vapor contributions. If the model retrieves temperature with only limited improvement, the paper should explain whether the limitation comes from T-W ambiguity, the dominance of moisture variability in the lower troposphere, radiosonde representativeness, the architecture, the intermediate-target design, or the limited information content of N. The manuscript needs a physical interpretation of where the retrieval contains information and where it is mostly constrained by the learned empirical prior.

4) Several results are overstated. Terms such as “independent observational products,” “prior-free,” “input-source independent,” “platform-agnostic,” and “general-purpose learned inversion operator” should be revised. More defensible language would be “NWP-independent during inference,” “trained with an empirical radiosonde prior,” and “evaluated under ideal radiosonde-derived input conditions.” (some of the lines includes 16-40, 145-152, 169-173, 181-190, 203-210, 605-618, 820-860, 888-895, 950-972, 1019-1025, and 1067-1073)

Detailed comments

Lines 16-40: Abstract and scope. The abstract should clearly state that the evaluation uses radiosonde-derived refractivity as ideal input. The phrase that the framework is independent of NWP background states is acceptable only for inference. The radiosonde training distribution still constrains the model. The abstract should not imply satellite RO validation.

Lines 23-31: Accuracy versus measurement uncertainty. Please do not compare RMSE directly with radiosonde measurement uncertainty as if they are the same quantity. RMSE includes retrieval error, representativeness error, sampling effects, and possible residual QC effects. The coverage factor should be defined consistently. The paper should avoid implying that the retrieval reaches the measurement limit unless this is demonstrated by uncertainty propagation.

Lines 33-40 and 249-263: Ideal-input framework characterization. The scope paragraph is important and scientifically honest. The paper should move this idea into the abstract and conclusions. The present paper establishes a framework for behavior under ideal input conditions; it does not yet demonstrate performance for real RO profiles.

Lines 83-113: Lower-tropospheric GNSS-RO challenges. The GNSS-RO motivation should introduce super-refraction, multipath, horizontal gradients, critical refraction, vertical smoothing, and lower-tropospheric N biases earlier. These are not secondary details; they directly determine whether the proposed N-to-T/W retrieval can work for real RO data.

Lines 114-152 and 203-210: Prior information. The manuscript should not frame the approach as prior-free. The relevant comparison is not between an NWP prior and no prior. It is an NWP model prior versus an empirical radiosonde sampling prior. Both are priors, each with error structures and representativeness limitations.

Lines 169-173: Radiosonde reference statement. The statement that colocated radiosondes are superior to any NWP analysis or reanalysis at the same footprint is too strong. Radiosondes are traceable in situ measurements, but they are point observations with drift, sensor errors, and representativeness limitations. Please revise to a more defensible statement.

Lines 181-190: Claim of largest independent assessment. The claim that this is the largest independent in situ assessment should be substantiated with a short comparison to prior studies or softened. “To our knowledge” is not sufficient unless the relevant literature is adequately covered.

Lines 266-279: WCTN input and 150 m dilation. The wavelet-transform input is central to the method. The rationale for the 150 m dilation should be given in the main text, and the authors should show performance without WCTN and with alternative dilation scales. Otherwise, it is hard to know whether this design is physically necessary or simply a tuned feature representation.
Lines 284-298 and 508-510: Vertical grid definition. The vertical grid appears inconsistent. The text refers to 100 m to 20 km, 10 m spacing, and 2,000 vertical elements. Please define the exact lower boundary, upper boundary, spacing, endpoint convention, and number of levels.

Lines 293-298 and 861-887: Intermediate targets. The choice of dry refractivity and dry-air partial pressure as intermediate targets may be useful, but this has not been proven. Because these variables are not independently observed components of total N, their retrieval still depends on learned prior information. Please add ablation experiments against direct T/W/RH prediction and against a single multi-output network.

Lines 315-323: Validation wording. Because input N and target T/W are derived from the same radiosonde ascent, this is not validation against independent observations of the same atmospheric state. Please use “evaluation under internally consistent ideal-input conditions” or similar wording unless independent real-RO validation is shown.

Lines 324-413 and 493-514: QC reproducibility. The data description is extensive, but the QC implementation is not fully reproducible. Please provide profile counts before and after each QC step, RH clipping counts, altitude-threshold exclusion counts, and objective criteria for visually identified anomalies.

Lines 414-419: Island radiosonde representativeness. A small island should not be described as a large ship permanently anchored in the ocean. Island boundary layers, terrain, surface heating, and local circulations may affect low-level profiles. Please quantify whether errors change when the lowest 0.5-1 km is excluded.
Lines 474-477: Geographic independence. Manus and Nauru are not geographically independent if they appear in both the training and TOGA-COARE evaluation sets. Please report metrics excluding these locations or describe the test as temporally/instrumentally independent rather than geographically independent.

Lines 523-565 and 1080-1120: Reproducibility of architecture and training. The main manuscript should include sufficient information for scientific interpretation: the number of parameters, normalization, loss weights, optimizer, learning-rate schedule, train/validation/test splits, early stopping, random seeds, and hyperparameter ranges. Code, trained weights, preprocessing scripts, and evaluation scripts should be archived in a persistent repository with a DOI.

Lines 563-565 and 672-687: Confidence intervals and dependence. The confidence intervals appear to assume independent profiles. This is likely unrealistic for repeated soundings from the same stations, campaigns, seasons, and synoptic regimes. Please recompute uncertainty using station/campaign/day block bootstrap or a comparable dependence-aware method.

Lines 577-585: Need for baseline models. A similar RMSE between the training and evaluation data is encouraging, but it does not prove that the model learned a general inversion. Please compare against altitude-only climatology, season/basin climatology, multilinear regression, MLP/random forest, and direct-output CNN baselines.

Lines 605-618: Independent observational products. The sentence that all model retrievals constitute independent observational products is too strong. They are independent of NWP input during inference, but not independent of the radiosonde training prior. They are also not yet validated against real satellite-observed refractivity.

Lines 624-658 and 688-710: Temperature and moisture interpretation. The interpretation of retrieval skill should be strengthened. Please explain why temperature retrievals below 20 km are not better and identify altitude ranges where N contains useful independent information versus ranges where the output is dominated by the learned empirical prior.

Lines 728-740: Regime-dependent statistics. The statement that differences are less than 5% under deep-convective and dry conditions should be quantified rather than inferred primarily from visual inspection. Please provide statistics for moist, dry, convective, suppressed, day/night, boundary-layer, and free-tropospheric regimes.

Lines 741-744 and 995-1004: DYNAMO temporal overlap. The temporal overlap issue should be handled with a sensitivity test. Please exclude all DYNAMO evaluation profiles from headline statistics or report results with and without them.

Lines 745-781: Perturbation experiment. The perturbation experiment is useful, but it should not be interpreted as evidence of 10 m resolving power. A two-grid-level RH spike is an idealized diagnostic. Please repeat the test after RO-like smoothing and realistic N perturbations.

Lines 782-819: What the model learns. This is one of the stronger parts of the manuscript because it recognizes that the model learns training-distribution statistics rather than explicit physical laws. Please connect this discussion more directly to limitations in retrieving sharp inversions, boundary-layer caps, and low-level humidity structures.

Lines 820-860: Robustness claims. Cross-radiosonde and cross-campaign generalization are useful, but they are not equivalent to transferability to satellite RO observations. Please moderate claims of spatial, temporal, and instrument robustness.

Lines 888-895 and 1019-1025: NWP independence. The paper should use “NWP-independent during inference” rather than “genuine NWP independence.” The model inherits radiosonde sampling, processing, QC, station distribution, and climate-regime priors.

Lines 950-972 and 1067-1073: Platform-agnostic statement. The architecture may accept input profiles from different sources, but performance is source-dependent. GNSS-RO profiles have different vertical resolution, horizontal averaging, bias, noise, and missing-data characteristics. Please revise the platform-agnostic wording.

Specific comments

Lines 16-40: Add one sentence in the abstract stating that real GNSS-RO validation is not included in this Part 1 study.
Lines 18-19: Use consistent hyphenation: “one-dimensional” and “model-dependent.”
Lines 23-24: Use “high-resolution” as a compound adjective.
Lines 23-31: Define whether the reported radiosonde uncertainty is standard uncertainty or expanded uncertainty, and do not imply that RMSE equals measurement uncertainty.
Lines 33-40: Add a sentence such as: “Part 2 will determine how this performance changes for actual GNSS-RO inputs.”
Line 83: Capitalize as “Global Navigation Satellite System.”
Lines 88-90: Consider changing “of the order of a few hundred meters” to “on the order of a few hundred meters.”
Line 106: Check the spelling and capitalization of “PlanetIQ” and the cited author name.
Lines 114-144: Revise the 1D-Var discussion so that the contrast is not NWP prior versus no prior, but NWP-model prior versus empirical radiosonde prior.
Lines 145-152 and 203-210: Replace “prior-free” language with “NWP-independent during inference” or “constrained by an empirical radiosonde prior.”
Lines 169-173: Revise the radiosonde-superiority statement and acknowledge representativeness limitations.
Lines 181-190: Substantiate or soften the “largest independent in situ assessment” claim.
Line 199: Insert a space in “domain.The.”
Lines 207-210: Revise the sentence beginning “The approach seeks...” for grammar and clarity; it may be better as two sentences.
Line 247: Change “The Part 2” to “Part 2.”
Lines 249-263: Move the main limitation of ideal-input testing into the abstract and conclusions.
Line 255: Avoid “etc.” in a formal methods section; list the factors or write “among others.”
Lines 266-279: Give the rationale for 150 m WCTN dilation and show sensitivity to the dilation scale.
Lines 284-298: Define the exact vertical grid and resolve the apparent inconsistency in the number of levels.
Lines 286-287: Use consistent section-reference formatting: “Sect. 2.3, Sect. 2.4, and Sect. S1.”
Lines 293-298: Do not imply that Nd and Pd are straightforward targets simply because they are linearly related to N; their decomposition still depends on a learned prior.
Lines 307-313: Please check Eq. (8). The denominator inside the Haar wavelet argument appears to use h rather than the dilation a.
Line 311: Correct “and and Zb” to “and Zb.”
Lines 315-323: Avoid using “validation” without qualification because the input and output come from the same radiosonde ascent.
Line 316: Insert a space in “RHprovided.”
Lines 324-413: Provide station-by-station counts before and after QC and document objective criteria for visual anomaly removal.
Lines 361-365: The text mixes standard and expanded uncertainty. Please define k = 1 and k = 2 consistently.
Lines 414-419: Remove or revise the “large ship” analogy; quantify island effects in the lowest 0.5-1 km.
Lines 474-477: Do not describe Manus and Nauru as geographically independent if they are included in both training and evaluation locations.
Lines 493-514: Quantify sensitivity to RH clipping, gap interpolation, truncation below 100 m, and exclusion of profiles not reaching 20 km.
Lines 511-514: Change “Unrealistically high positive (negative)” to “Unrealistically high positive (negative) RH values”; change “was small” to “were small.”
Lines 523-565: Add number of trainable parameters, normalization strategy, train/validation split, early stopping criteria, optimizer, loss weights, random seeds, and hyperparameter ranges.
Lines 532-543: Resolve the inconsistency between 0-20 km and 100 m-20 km output ranges.
Lines 563-565 and 672-687: Recompute confidence intervals using a station/campaign/day block bootstrap or another method that accounts for dependence.
Lines 577-585: Add baseline models to quantify the actual AI contribution.
Lines 605-618: Revise “independent observational products”; the retrievals are not independent of the prior radiosonde training.
Lines 624-658: Add a compact main-text table showing RMSE and bias for 0.1-2, 2-5, 5-10, 10-15, and 15-20 km for each evaluation dataset.
Lines 688-710: Bring a concise station-by-station summary into the main text, including the best and worst sites and possible dependence on basin, season, sonde type, and local time.
Line 704: Clarify the meaning of “~15% RMSE 10-12 km.”
Lines 728-740: Quantify the less-than-5% statement with regime-conditioned statistics instead of relying on visual inspection.
Lines 741-742: Change to “It should be noted, however, that a temporal overlap exists...”
Lines 741-744 and 995-1004: Support the conclusion about temporal overlap with a sensitivity test excluding DYNAMO evaluation profiles.
Lines 745-781: Repeat the perturbation experiment after RO-like vertical smoothing and realistic refractivity noise.
Lines 782-819: State more clearly that sharp real thermal and moisture structures may be smoothed by the learned prior.
Lines 820-860: Tone down claims of robustness and distinguish radiosonde generalization from satellite RO transferability.
Lines 861-887: Add an ablation test against direct prediction of T/W/RH and against a single multi-output network.
Lines 888-895 and 1019-1025: Use “NWP-independent during inference” consistently and avoid “genuine NWP independence.”
Line 929: The phrase “behavioral property of the model” is acceptable but should be connected to retrieval information content.
Lines 950-972: The limitations section should include missing baselines, no per-profile uncertainty, no real-RO validation, correlated evaluation data, subjective QC, and incomplete code/model availability.
Lines 1000-1004: The statement about temporal overlap not conferring a measurable advantage appears repeated; remove one occurrence.
Lines 1021-1022: Use “NWP-derived” consistently.
Lines 1067-1073: Revise the platform-agnostic statement because performance is source-dependent.
Line 1069: Use “general-purpose” as a compound adjective.
Lines 1080-1120: Strengthen the code and data availability statement. Trained weights, preprocessing scripts, architecture, normalization constants, and evaluation scripts should be placed in a DOI repository.
Line 1131: Change “reference radio soundings” to “reference radiosonde soundings.”

Final statement

Overall, I find this manuscript interesting, relevant to AMT, and potentially publishable after major revision. The key value of the paper lies in its attempt to learn an N-to-T/W retrieval relationship at distances below 20 km without using an NWP background during inference. However, the present evidence supports only an ideal-input framework characterization using radiosonde-derived refractivity. It does not yet demonstrate that the method works for real GNSS-RO refractivity profiles. The authors should revise the claims, add baseline and ablation experiments, improve the treatment of uncertainty, explain the information content and limitations of temperature retrieval below 20 km, and strengthen reproducibility. If these issues are addressed, the paper could make a meaningful contribution to atmospheric retrieval methodology and to the development of future GNSS-RO profile retrieval.

Citation: https://doi.org/10.5194/egusphere-2026-2027-RC2

Santosh Muralidharan, Harris John, and Deepak Mishra

Supplement

https://doi.org/10.5194/egusphere-2026-2027-supplement

Santosh Muralidharan, Harris John, and Deepak Mishra

Viewed

Total article views: 173 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	Supplement	BibTeX	EndNote
110	55	8	173	32	7	9

HTML: 110
PDF: 55
XML: 8
Total: 173
Supplement: 32
BibTeX: 7
EndNote: 9

Views and downloads (calculated since 29 May 2026)

Month	HTML	PDF	XML	Total
Jun 2026	46	22	5	73
Jul 2026	42	22	3	67
Aug 2026	22	11	0	33

Cumulative views and downloads (calculated since 29 May 2026)

Month	HTML	PDF	XML	Total
Jun 2026	46	22	5	73
Jul 2026	42	22	3	67
Aug 2026	22	11	0	33

Viewed (geographical distribution)

Total article views: 140 (including HTML, PDF, and XML) Thereof 140 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 18 Aug 2026

Short summary

We trained a deep learning model on high resolution radiosonde data from tropical islands to extract temperature and moisture profiles from atmospheric refractivity. Tested on 27,400+ geographically independent radiosonde measurements across three ocean basins and decades, it achieves accuracy close to measurement uncertainty for moisture. With only refractivity as input, model is applicable in principle to GNSS RO satellite observations over tropical oceans.


Total:	0
HTML:	0
PDF:	0
XML:	0