the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 23 Mar 2026
Global-ABLWind: a global atmospheric boundary layer wind speed profile dataset derived from Aeolus and surface ancillary information
Abstract. Accurate wind speed profiles within the atmospheric boundary layer (ABL) are essential for understanding atmospheric processes, climate change, and wind energy assessment. However, existing global ABL wind products lack either sufficient vertical resolution or accuracy, limiting their ability to resolve wind structures throughout the ABL. Here, we propose a physics-constrained machine learning framework designed to reconstruct continuous ABL wind speed profiles by integrating physically interpretable bias-correction mechanisms with dynamical constraints from Aeolus L2C observations. The proposed method enables the reconstruction of high-accuracy wind speed profiles at a vertical resolution of 100 m across the full ABL (0–2 km), overcoming the trade-off between accuracy and vertical resolution that characterizes existing products. Independent validation against RS observations demonstrates that the proposed method achieves high accuracy across all ABL heights. It has an overall correlation coefficient (R) of 0.92 and a root mean square error (RMSE) of 1.94 m s-1, outperforming the original Aeolus L2C product (R = 0.90, RMSE = 2.23 m s-1). Further comparisons at 100 m vertical resolution with the fifth generation ECMWF reanalysis (ERA5) and the power law method confirm the superior accuracy of XGB-Wind, especially in the near-surface layer (0–500 m). Applying the proposed framework to the full Aeolus mission period (from July 2020 to April 2023), we generate a global high-resolution ABL wind speed profile dataset, termed Global-ABLWind. This dataset provides 100 m vertical resolution wind profiles with enhanced accuracy, continuous ABL coverage, and reduced data gaps on a global scale. The dataset is freely available at https://doi.org/10.5281/zenodo.18286457 (Tong et al., 2026) and represents a valuable remote sensing resource for boundary-layer wind studies and wind-related environmental applications.
- Preprint
(8512 KB) - Metadata XML
-
Supplement
(7685 KB) - BibTeX
- EndNote
Status: open (until 21 Aug 2026)
- RC1: 'Comment on essd-2026-73', Anonymous Referee #1, 29 May 2026 reply
-
RC2: 'Comment on essd-2026-73', Anonymous Referee #1, 29 May 2026
reply
This manuscript presents Global-ABLWind, a global atmospheric boundary layer wind speed profile dataset derived from Aeolus L2C observations, radiosonde measurements, and ERA5 ancillary variables for July 2020 to April 2023. The stated aim is to reconstruct wind speed profiles at 100 m vertical spacing within the atmospheric boundary layer using an XGBoost-based framework with near-surface bias-correction terms and upper-level Aeolus dynamical constraints. The dataset is potentially valuable because global high-vertical-resolution ABL wind profiles are difficult to obtain from existing observing systems, and the combination of Aeolus wind lidar information with radiosonde and reanalysis constraints addresses a relevant community need. This manuscript can be published after considering the following comments concerning clarifications or extensions:
- 3.2.2 states that 10 m RS wind speed is used as the bottom constraint during model construction because ERA5 10 m wind correlates weakly with 100 m RS wind speed. Sect. 4.3 then states that the global product uses ERA5 10 m wind speed as the alternative input where RS is unavailable. The reported model-validation metrics are derived from a configuration using RS 10 m wind, whereas the released global product depends on ERA5 10 m wind. The rationality of this substitution needs further clarification.
- 4.3 states that Global-ABLWind maintains the same spatial and temporal resolution as Aeolus L2C. The product is therefore along Aeolus observation profiles rather than a globally gridded continuous field. Several figures use global map shading that visually resembles gridded continuous coverage. The author should provide a clearer explanation in the manuscript.
- The bias-correction method relies on quadratic relationships between binned Ray/Mie wind biases and selected variables. The statistical basis of these relationships is unclear because the manuscript does not specify how bins were defined, how sample sizes vary by bin, whether the R² values are based on raw samples or binned means, or whether these fitted corrections were estimated on independent data from the validation folds.
- The term “missing rate” appears throughout the manuscript, but it is not always clear whether it refers to missing height levels within an observed profile, missing Aeolus profiles, missing valid retrievals, missing model outputs, or missing map bins after aggregation.
- The relationship between the Ray-based and Mie-based products is not sufficiently explicit. Whether they use the same processing procedure is not clearly stated. The main text is based on Ray wind, while the supplement presents Mie-derived results. However, the manuscript does not make sufficiently clear whether the released dataset should be regarded as one primary Ray-based product with a secondary Mie-based companion, two parallel products, or two alternative realizations of the same product. This distinction is important for users deciding which files to use under different atmospheric or cloud/aerosol conditions.
- The inclusion of R and RMSE in a hyperparameter table is ambiguous because they are performance metrics rather than hyperparameters. The RMSE units are not shown in the table.
- Terminology is inconsistent in several places. The most conspicuous example is Fig. 12 and Fig. S10, where the legend reads “Global-PBLWind” while the manuscript and data record describe “Global-ABLWind”. The title of section 3.2 is “XBG-Wind”, while elsewhere in the manuscript it is “XGB-Wind”. Fig. 8 also has panel-lettering inconsistencies between the visual layout and caption. Tables 1 and S1 contain the misspelling “learing_rate.” The title of the subgraph in Figure 2 does not match the content. Please carefully review the entire manuscript.
Citation: https://doi.org/10.5194/essd-2026-73-RC2 -
RC3: 'Comment on essd-2026-73', Anonymous Referee #2, 13 Aug 2026
reply
Review of ‚Global-ABLWind: a global atmospheric boundary layer winds speed profile dataset derived from Aeolus and surface ancillary information’ by Tong et al., submitted to ESSD for potential publication
Summary statement:
This study proposes a valuable addition to globally available airflow data for many potential applications. It produces and documents a dataset of horizontal wind speed at 100m vertical increments from the surface to 2000m asl with global coverage from the Aeolus mission’s active wind-lidar observational system for the period July 2020 to April 2023. Machine learning techniques, more specifically extreme gradient boosting (XGB,) is used to improve currently available products in the lowest layers near the surface by learning from the space and time matched global radiosonde dataset (IGRA) regarding data availability and precision. In addition to describing their methodology and reporting the statistics of their ‘Global-ABLWind’ product, an attempt is made to describing the time-averaged circulations/ phenomena found in the lower troposphere for the observational period.
The methodology is well described and thought-out, I enjoyed reading most sections of the manuscript and believe that the product has a substantial enough merit to warrant publication in ESSD. I only have a few scientific, and some editorial comments and suggestions. The question of how and if to address the latter is at the discretion of the authors, but I believe it would benefit the publication.Overall, I recommend acceptance of the manuscript following minor revisions addressing the comments listed below.
General scientific comments:
- Uncertainty in radiosonde reference data: Judging from Figs. 3c and 8b/ 8d and the accompanying text I am convinced that the XGB method significantly improves the match between the L2C-Rayleigh winds and the radiosondes (RS). However, the authors neither investigate nor mention the uncertainty of the RS profiles in the lowest hundreds of meters above ground. Often, RS profiles even do not contain data near the surface and/or observations are discarded in post-processing because of potential artifacts due to observer effects (when the observer releases the balloon), local obstruction/ deflection of the airflow by buildings and topography (such as hills, mountain chains etc), and delays in receiving the radiocommunication data from the ascending balloon. In other words: the RS data itself, which the authors here consider as the reference or ‘truth’, suffer from the largest observational error in the exactly the lowest heights approx. < 200 to 300m. I realize that the authors carefully screened the RS profiles and discarded potentially problematic data, but even if RS wind speeds are reported and used, the uncertainty near the surface is substantial. I would like to see some discussion of this i) measurements uncertainty by e.g. a comparison with tower locations (if existent in the vicinity) or ground-based wind lidars. I realize that these comparison can only be done for a limited time and spatial extent of the Global-ABLwinds, and ii) how this uncertainty compares to the improvement from the LC2-Ray to XGB winds.
- Wind speed vs wind direction: While the focus of the study is on wind speed, the direction of the airflow and its vertical directional shear is also important for many potential applications including those in the wind energy sector. Observing height-dependent wind direction in the lowest troposphere is much more difficult than solely the magnitude of the wind vector, but I would appreciate if the authors discussed/ showcased the usefulness of their method to also compute improved wind directions having in mind applications for which directional shear is important (cyclogenesis, orographic effects, ocean-air interaction, wind energy, aviation meteorology etc).
- Statistics of Global-ABLWind product: While I enjoyed reading almost all sections, I found Fig. 12 and section 4.4 the weakest parts of the manuscript and wonder if they can be replaced by more meaningful analyses. For Fig. 12 and accompanying text in Section 4.3, it is entirely unclear where these stations are located, why they were selected, and how representative these annual (?, averaging interval never described) are at a global scale. Winds change diurnally, synoptically (3 to 5 days in mid latitudes), and seasonally, so presenting a crude comparison of averaged profiles is not helpful in estimating the merit of the Global-ABLWind product. Only when RS observations deviate substantially from those of Ray L2C and other products for the lowest heights near the surface, the improvements by XGR should be largest. So, instead of a bulk analysis, I suggest replacing the bulk statistics with conditionally sampled ensemble averages across methods when i) speed and directional shear in the lowest heights are most important to the atmospheric science community (examples see comment B), or ii) the physical representation and interpretation of the found circulations/ phenomena (e.g. monsoons) have changed before and post application of the XGR method. In other words, I suggest the authors showcase the improvements in their data product to the reader in a more convincing manner using concrete case studies (with limited space and time domains), instead of global maps and profiles.
General editorial comments:
- Redundant text: While the language is appropriate for an international audience and terminology is correctly used, the current drafts suffers from substantial repetitions, which are a bit annoying. To give concrete examples: i) Section 4.3 starts with a summary of the key methodological steps (Ln 398 to 403), ii) Ln 355-358 repeats information mentioned before; iii) Ln 295 to 302: mere repetition. These are just a few examples. I suggest making the writing more concise, and shorten the manuscript. I encourage the authors to go through the manuscript and reduce the redundant text passages to the absolute minimum.
- Writing style: The authors spend a lot of text on restating the statistics mentioned in the figures/ tables, and describe the figures (e.g. Ln 335-358,….) in unnecessary detail instead of summarizing and interpreting the findings. I suggest moving all this information into the captions (if not already present) to improve the flow of information for the reader, and to make the text more concise.
Minor comments:
- Please use Rayleigh winds instead of ‘Ray winds’. While the world is full of abbreviations and acronyms, using the full word improves readability.
- Round statistics to significant digits: Given the uncertainty in both the spaceborn wind lidar and the RS observations and the inherent variability of the turbulent airflows, I am convinced that wind speeds and all their statistics (MAE, STD, etc) are not significant to within 0.01 m/s, which is 1 cm/s. Please assess both the accuracy and precision of the boundary-layer winds are report all statistics (including figures, text, tables) only for significant digits.
- Use of terminology of ‘boundary-layer winds’: While one expects the lowest part of the troposphere to be/ contain the ABL, the authors never actually evaluate the ABL height or demonstrate typical ABL features such as nocturnal jets, or speed increase above, or vertical shear, or turbulence intensity/ TKE. Particularly above the oceans/ polar regions the ABL heights should substantially lower than 2000m asl, so I think a demonstration/ analysis should be added to support their claim of producing ABL winds.
Citation: https://doi.org/10.5194/essd-2026-73-RC3
Data sets
Global atmospheric boundary layer wind speed profile dataset derived from Aeolus observations from July 2020 to April 2023 Zhe Tong et al. https://doi.org/10.5281/zenodo.18286457
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 316 | 251 | 27 | 594 | 376 | 27 | 37 |
- HTML: 316
- PDF: 251
- XML: 27
- Total: 594
- Supplement: 376
- BibTeX: 27
- EndNote: 37
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
This manuscript presents Global-ABLWind, a global atmospheric boundary layer wind speed profile dataset derived from Aeolus L2C observations, radiosonde measurements, and ERA5 ancillary variables for July 2020 to April 2023. The stated aim is to reconstruct wind speed profiles at 100 m vertical spacing within the atmospheric boundary layer using an XGBoost-based framework with near-surface bias-correction terms and upper-level Aeolus dynamical constraints. The dataset is potentially valuable because global high-vertical-resolution ABL wind profiles are difficult to obtain from existing observing systems, and the combination of Aeolus wind lidar information with radiosonde and reanalysis constraints addresses a relevant community need. This manuscript can be published after considering the following comments concerning clarifications or extensions: