the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 11 Sep 2026
Physics-Guided Diagnosis of CYGNSS Wind Mismatch under Tropical-Cyclone Wind–Wave Disequilibrium
Abstract. The Cyclone Global Navigation Satellite System (CYGNSS) provides extensive ocean-surface wind observations within tropical cyclones, yet its reference-relative mismatch can vary systematically with storm structure and the evolving sea state. This study develops a physics-guided diagnostic framework using 212,697 unique CYGNSS–tropical-cyclone matchups from 147 storms during 2022–2024 across the western North Pacific, Bay of Bengal, and North Atlantic. CYGNSS observations are combined with ERA5 atmospheric and wave fields and IBTrACS best-track information to construct six descriptors of derived wave-height departure, young-wave state, bulk wave steepness, transient wind forcing, transient sea-state response, and storm-relative proximity. A supervised Wind–Wave Disequilibrium Index (WWDI∗) is derived using training-only robust normalization and leave-one-component-out evidence weighting, with all parameters frozen before temporal and geographical evaluation.
Across four storm-independent and observation-independent cross-basin transfers, WWDI∗ retains positive associations with absolute CYGNSS–IBTrACS mismatch. Observation-level Spearman correlations range from 0.219 to 0.381, whereas 10-bin regime-scale R2 values range from 0.639 to 0.753. Wave steepness receives the largest training-derived weight in both development basins, followed by young-wave state and transient significant-wave-height response. Permutation testing, storm-wise bootstrap resampling, and sensitivity experiments confirm the persistence of the regime-scale relationships. After controlling for ranked tropical-cyclone intensity, storm-relative radius, their quadratic terms, and their interaction, the residual association reverses sign in all four transfers, with residual Spearman coefficients from −0.645 to −0.543. This shows that the positive unconditional relationship is embedded within the joint storm-intensity–radius structure rather than representing an independent monotonic effect.
Independent comparison with 5,771 CMEMS significant-wave-height matchups gives an ERA5 RMSE of 0.631 m and Pearson correlation of 0.942. Diagnostic machine-learning experiments further show that local atmospheric wind provides substantial transferable information beyond tropical-cyclone intensity and radial position, with additional benefit from bulk wave-state variables in three of the four transfers. A training-derived WWDI∗-conditioned adjustment reduces raw external RMSE by 29.8–40.4 %; relative to a simpler global training-bias correction, the additional point-estimate improvement is 1.7–13.0 %, with storm-wise bootstrap support for the reciprocal WP–BOB transfers.
These results identify transferable tropical-cyclone wind–wave regimes associated with systematic changes in reference- relative CYGNSS mismatch. WWDI∗ is therefore interpreted as a supervised physical diagnostic of mismatch-prone storm environments rather than as a causal index or an operational wind-retrieval correction.
- Preprint
(25648 KB) - Metadata XML
- BibTeX
- EndNote
Status: open (until 06 Nov 2026)
-
RC1: 'Comment on egusphere-2026-5300', Anonymous Referee #1, 12 Oct 2026
reply
-
AC1: 'Reply on RC1', injamam masum, 21 Oct 2026
reply
Dear Referee 1,
Thank you very much for your careful and constructive comments. We have addressed all of the points raised and performed additional analyses that substantially clarified the interpretation of the study. Please find our detailed point-by-point response in the attached supplement. We greatly appreciate your helpful suggestions.On behalf of all co-authors,
S. M. Injamamul Haque Masum
-
AC1: 'Reply on RC1', injamam masum, 21 Oct 2026
reply
-
RC2: 'Comment on egusphere-2026-5300', Anonymous Referee #2, 12 Oct 2026
reply
Dear authors and editors,
This manuscript presents a study of errors in CYGNSS retrieved wind speeds under tropical cyclones. The authors aim specifically at diagnosing the mismatches between CYGNSS and other existing wind sources (ERA5 and IBTrACs), with a focus on sea-state conditions and wind-wave disequilibrium that stems from the legitimate and interesting idea that CYGNSS, as a GPS sensor, is very sensitive to wave and swell conditions under high winds.
The study is extensive, with a wide range of interesting and consistent methodological methods to prove the solidity of the results. The study of basin to basin transferability, specifically, that allows to support the fact that measured differences in retrieved wind speeds are not notably dependent on the geographical specificities of the measurement area, are interesting and insightful. The diversity of validation and statistical correlation/dependence methods is impressive and insightful.
However, in my humble opinion, the study is undermined by questionable and confusing choices, that make any clear conclusion on the contribution of sea state, and more generally on the conditions that cause errors in CYGNSS winds, difficult to draw. For these reasons, detailed hereafter, I recommend a major revision of the results.
The first and main issue is the choice of measuring local wind errors as a difference with IBTrACs maximum wind. As detailed in the comments below, this introduces an immediate local bias in all observations, as errors will grow with distance to the maximum wind location. This general bias induces, in addition, system-scale and TC-case-wise biases, as the structure of the TC (notably profile steepness and wind field asymmetry) will strongly influence the mean errors over one case (a steeper/more intense TC will show a stronger average mismatch). This problem is actually visible in several of the presented results, notably the case study where the dependence of errors on the radial distance is very clear (Fig. 12h). The authors discuss this shortly but dismiss the problem as a "local issue" that does not prevent their index from being relevant as a "system-scale / bulk state reference mismatch". However, in my opinion, this judgment is invalid, for the reasons listed above: individual errors propagate on the system scale and thus on general results, due to TC structure. This choice of using the difference with IBTrACs is not really justified by the authors and appears as confusing to me: existing studies all rely on local wind measurements (Powell and Ruf 2025 use ERA5 winds, Ricciardulli et al. compare with other satellite wind fields, Ruf et al. with SFMR, etc.), and ERA5 winds are used in this study, but never as a basis to compute wind speed mismatch, which I find a bit puzzling.
The second problem, connected to the first, is the choice to ignore or remove the local wind intensity from the construction of the sea state index. While the link between CYGNSS mismatch and intensity is documented (Ricciardulli et al., 2021, notably) and is discussed in some of the presented results in this manuscript (cf. confounder analysis and machine learning benchmark that both show that the WWDI index does not notably improve the correlation compared to intensity and wind structure parameters alone), the wind speed is not explicitly included as part of the designed index. Consequently, the analysis relies only on wave parameters and draws a link between CYGNSS mismatch and wave steepness. One could argue that this link is entirely due to the primary dependence on intensity, which link with wave steepness is strong. As such, the observed link between WWDI variables and CYGNSS mismatch consequently does not allow to describe the dependence of mismatch on sea state variables. This comment is backed up by the analysis of Figure 6, where the error evolution compared with sea state appears as reversed when the wind contribution is removed. In my opinion, the study should seek to explain this "intensity-free" tendency, rather than only considering WWDI as a "bulk state" index that, as such, loses a lot of meaning and interpretability.
Some other general remarks:
- figures are not linked clearly enough with the corresponding text. Some figures are not even discussed in the text, or barely referenced, and others would benefit from a more explicit and thorough referencing
- In several results, authors point at a variability in the error/mismatch, without actually interpreting, or trying to explain it. For instance, the fact that some basin-to-basin transfers behave differently from others, possibly pointing at basin-specific conditions, are only mentioned without any interpretation. Same with the comparison with maximum intensity location, which shows larger errors for some basins occurring after the main intensity peak. While the main issue of IBTrACs referencing makes any interpretation complicated, the fact that none is proposed to explain results that appear as interesting left me with questions
Detailed comments are listed hereafter (see attached pdf for detailed locations):
35 : a short definition could be useful here
48–49 : what is the advantage of CYGNSS for TC observation ? Coverage, resolution, frequency (saturation/cloud penetration) ? What are the defaults ? This paragraph states that CYGNSS is useful but does not back up the statement.
64 : vague... surface waves, swell ?
74 : Which "conditions" exactly ? The previous sentence lists a set of parameters that describe the sea state. We understand that the disequilibrium is a departure from a steady state of balance between wind and waves, but this should be stated more clearly (maybe explicitly defined above, cf previous comment)
94 : this might not be clear without acronym explanation. Maybe use an expression like "basin to basin transferability" instead ?
111–112 : this is very vague at this stage. Maybe reference section 2.4 ?
118–122 : the high sensitivity of GPS signal to sea surface conditions might gain to be clarified, here or in the introduction. Most satellite surface wind measurements (microwave scatterometers, SAR) are not as affected by surface wave conditions, so it might be useful to explain clearly that this is a specificity of CYGNSS due to its frequency band, which makes it more sensitive than other usual measurement satellites.
134 : what does this flag consist of ?
155 : I am not familiar with ERA5 data, but in what conditions would observations "lack finite values" ? Is it near land ? What does it refer to ? Please specify and maybe explain shortly in the text.
212–213 : How much of the dataset is affected and filtered as physically non plausible ? Specifying at each step of selection how many samples are removed from the dataset could be useful to realize the impact/utility of each filtering step
214 : what is a duplicate observation ? Does it mean the exact same observation and if so, why does CYGNSS record duplicates ? How many data does this represent over the entire dataset ?
288–292 : this list is very unclear at this stage of the article. enumerating the methods that are described in following sections might be confusing. Authors should. at least, reference the section describing these methods in this sentence.
Fig. 2 (caption) : the first TC-WWDI index does not appear on the figure. I am guessing it should be on the right in the second box. Maybe add it for clarity ?
303–305 : It is unclear to me why ETC is used as the principal response variable when it is a reference-relative mismatch, while EERA5 is a point-wise error. What is the meaning of ETC, given that a mismatch at a given point between CYGNSS and IBTrACs doesn't necessarily mean that there is an error in either of the two sources, but maybe that the observation is located away from the maximum wind location ?
350 : This figure is not described or interpreted anywhere. The link between the text and the different panels is unclear: "dimensionless scaling" is shown in panel (a), but what exactly is the point of comparing Hw* and Aw ? The plot is not interpreted at all. "weak wind behaviour" is shown, it seems, in panel C, but again what is the point of the figure ? What is the conclusion ? Panels (b) and (d) regarding transfers are not even mentioned in the text. Why does the WP-BOB transfer show a lower correlation, for instance ? This figure should be referenced in the corresponding paragraphs, along with the corresponding interpretation showing how it backs up/validates your methodology.
A few other remarks on these plots:- The y-axis labels on panels b and c are not very clear: R2 regarding to what ? correlation between what and what ? This could be more explicit.
- The form chosen for panels (b) and (c) is not optimal: The bar plots for panel (b) are all between 0.8 and 1, making the differences hard to see. Maybe 95% intervals could also be added to show the spreading, instead of just the mean, replacing the bar plot with a box plot including mean/median, quantiles and min/max
same remark for panel (c): the line plot for a comparison like this is not the best choice, a box plot would give more information and with more clarity.
- Panel (d) should probably be separated from this figure, as it shows the WWDI* index which is introduced much later in the text.408 : this variable is not defined
409 : the criteria for invalidation are unclear. How can Qj,98 be smaller than Qj,2, since they are defined to be otherwise ? What is Nj,finite ? Please give more details.
444–445 : this is equal to TC-WWDI, right ? If so, why not relating WWDI* to TC-WWDI more explicitly ?
468–473 : what about the ATL basin ? If it is only for validation and not a training dataset, it should be specified.
498 : what motivates this choice ? What is the advantage of the Spearman correlation compared to the Pearson correlation ?
504 : I am not sure about the meaning of this expression: "the larger of 30 observations". Is it a typo ? Does it mean "more than 30 observations" ? If not, larger than what ?
504–505 : this in unclear to the reviewer. The previous paragraph states that "external observations are divided in 10 equal count bins", meaning each bin represents 10% of the external dataset. Yet, here, you suggest that a bin could contain less than 0.2% of the external dataset. I most likely misunderstood something... Could you clarify ?
506–507 : There are many justifications regarding the interpretation of the meaning of variables (lines 505,460,428,419,140,28,1148,1103,1021,825,788, and more). Is there a specific reason why you insist on that ?
If there isn't, then I suggest avoiding repetitions. In the present example, it is clear based on the previous variable definitions that bin-wise defined variables are "regime-scale" and not point-wise observation level measurements.716–717 : this difference among basins doesn't seem specific to dimensionless wave height. The behaviour of wave age is very similar and the wave steepness WP-BOB transfer seems more dissociated from other transfers than for the two other variables. Could you explain why you insist on dimensionless wave height ?
720–721 : What justifies this assertion on the figure ? This is unclear to me. Is it because of the error bars variability ? Please specify/detail.
721–722 : maybe expliciting the previous sentence will clarify this, but from the figure it is not obvious why TC-WWDI would not be considered as an individual mismatch diagnostic as well as a regime separator. The error bars are pretty separated from one bin to the next so it seems that the individual errors are basically related to the regime, and thus to TC-WWDI. More explicit links between the figure and your assertions would help identify and understand this distinction.
723–724 : this sentence is unclear. is it a justification for the previous assertion ? what is the signification/interpretation of "asymmetric, distribution-free confidence intervals" ?
Fig. 4 (caption) : the last panel is not discussed anywhere... It shows a clear correspondence between WWDI and WWDI*, that could also feed the discussion in the following paragraph as to why the basic sea state variables receive negative weights : the overlapping correlation between temporal diagnostics and instant wave parameters in the WWDI* weighting procedure explains the proximity between TCWWDI and WWDI*
742–744 : While this indeed proves the regime-scale character of WWDI*, the very large spreading also tends to indicate that the designed variables might not be the proper ones to explain the mismatches. A question that arises is the one of the intensity bias, which is the object of the next subsection. A strong intensity bias of CYGNSS with other sources has been noted in previous papers (cf. fig. 11 of Ricciradulli et al., 2021: https://www.mdpi.com/2072-4292/13/24/5110)
While the GPS frequency indeed points at the importance of the sea state itself in the wind estimate, which justifies the present study, the reason why the authors didn't include intensity itself directly in the TC-WWDI* diagnostic is unclear.
Did the authors conduct a sensitivity study to assess the weight in the WWDI* of intensity alone ? If so, could they discuss it and, at least, mention it in the text ?
cf next comment as well regarding intensity bias treatment745 : This section, as well pointed out by the authors, highlights the clear colinearity/correlation between storm intensity and sea state properties. When dissociated through residual analysis, the dependence of sea state on mismatch is reversed. The meaning of this is unclear to me: does it mean there is no dependence on sea state once the structural dependence is removed, or does it mean that, for a given intensity and structure, the bias decreases with TC-WWDI*, i.e., smoother and smaller waves have larger mismatch at equivalent intensity ?
The efforts of the authors to explicit the impact of intensity on the sea state diagnostic are appreciable. However, as they show that this link is indeed very strong and important, one could ask why intensity was not considered as the primary variable for mismatch evaluation, and why sea state was not considered only as complementary to intensity ? This result questions the choice of WWDI* as a mismatch diagnosis, since it does not include intensity itself and choses to approach the mismatch distribution under the angle of sea state variability only. It seems that, seeing these results, a TC-WWDI* including both intensity as a primary variable and intensity-normalized sea state as a secondary driver might help better dissociate the mismatches. Is there a fundamental reason why the authors chose not to include intensity in the analysis ? Did the authors conduct sensitivity analyses including intensity into WWDI* ?752 : the term "controlling" is quite unclear, even with the methodology paragraph in mind. What exactly does it mean ? How are the scatter plots in Fig.6 obtained ? How do you manipulate the rank distributions to obtain the "controlled" data ? Apologies if I missed something.
761–765 : while this interpretation seems right, it rises the question asked in the previous comment: then, why use sea state to summarize the whole TC structure, and not use TC structure as a primary driver and normalized sea state separately, in addition ?
766 : same remark as for previous figures: instead of discussing the results and then referencing the figure separately, the discussion above should be done in conjunction with the figure. Each result discussed section 4.3 is illustrated in Fig. 6, so it would help to reference them within the paragraph. Please modify accordingly
772 : again, even though this time the text thereafter kind of follows the figure, it would be preferable to give explicit links with the visuals: correlation values, location on the plots, etc. to help the reader follow your interpretation of the figure
776–784 : it would be nice to have more details about the actual reasons why these discrepancies are observed. How do you explain that WP transfers in wave height departure have lower correlation with ATL and BOB than ATL and BOB with each other, for instance ? Are there studies of sea state properties that allow to interpret these differences ? Is it due to basin size, synoptic wind regimes, coastal configuration ? Do you have insights in these regards ?
825 : unit ?
851 : since ERA5 data are used point-wise and not as system-averaged, and since the point of your study is to study how the sea state under tropical cyclones affects the CYGNSS mismatch, it would be interesting to have insights on the representation of sea state by ERA5 in terms of TC conditions: does ERA5 feature the asymmetry of wave fields under TCs ? Does it correctly feature the dependence on intensity regime ? I guess there is already existing studies on the subject... Maybe reference and discuss them could be useful. But this is just a suggestion, and the results in this section are already clear and satisfying as they are.
853–855 : this point is repeated 5 or 6 times throughout the manuscript. It could be removed here
883–886 : same remark as before, the figure could be referenced within the paragraph to back up your points visually, rather than independently at the end.
889 : The choice of the IBTrACs-CYGNSS wind mismatch rather than the ERA5-CYGNSS mismatch is confusing to me. Why prefer a bulk difference rather than a local one ? This introduces additional biases in the analysis: for instance, as intense TCs have a more peaked profile, they necessarily have a stronger mismatch out of the eyewall. Weak TCs are more asymmetric, and consequently also display a mean bias in their mismatch simply due to the average gap between maximum wind and local wind in the TC wind field. This introduces additional sources of bias in the mismatch between intensity categories. This choice is not really justified in the manuscript. It would be nice to have more details about why the authors chose IBTrACs over direct ERA5 comparison.
927–929 : Do you have ideas regarding why the mismatch appears to be larger after the peak ? wave steepness and wave age don't show the same tendency.
932 : centred --> centered ?
Fig. 10 (caption) : Why is there no WP-->ATL transfer for panels b-d ?
Fig. 11 (caption) : This figure shows almost identical behaviours of CYGNSS-IBTrACs mismatch and TC intensity. This is consistent with the comment posted above about the choice of computing differences relative to IBTrACs, which introduces a bias in mismatch due to the intensity itself (higher intensity = steep profile = larger mismatch). Most results in this manuscript can thus be interpreted as being, to the least, influenced by your methodological choice. This is a considerable limitation. It also questions the choice of not including the intensity itself in the diagnostic, as it seems to be an important driver, as said in previous comments (but this would require a prior correction of this methodology-induced bias)
Fig. 12 (caption) : This figure and section again highlight the limitations of the methodology. The mismatch and WWDI* radial distributions (g,h) clearly highlight the problem of chosing IBTrACs to assess the mismatch: mismatch depends primarily on the distance to the center, rather than on WWDI and wave conditions that show a very different dependence on radius. This should by itself be a sufficient motivation for revising the mismatch definition: the mismatch is induced by its own definition. This pointwise issue also induces a system-scale bias, which makes the authors justification regarding the interpretation of WWDI* as a system-scale diagnostic questionnable: Since the radial distance strongly determines the bias locally, the radial steepness of the wind profile will strongly influence the system-scale bias. Panel (c) tends to confirm that, showing very large biases when the TC reaches its peak intensity (although the intensity curve is not shown). An additional panel showing the ERA5 mismatch would be really useful to assess if, yes or no, the bias is more related to the wind distribution or to the wave properties. As such, conclusions cannot be made clearly, in my opinion.
Other remark: the colormap limits on panel (e) are clearly not adapted, nothing can be seen except the old waves in the beginning of the life cycle.1000–1002 : if the wind wave disequilibrium and the temporal lag between wind and wave adjustment was the reason for the CYGNSS mismatch, the transient response variable should carry a more important weight and be the major source of mismatch. Instead, and as noted several times through the manuscript, the main driver in your WWDI* and WWDI indices is consistently wave steepness. Moreover, wave steepness is closely related to the wind speed itself, and in many results the direct link between the CYGNSS-IBTrACs mismatch and the intensity is clearly shown. Hence, it seems that in the present state your study shows more evidence regarding the connection between TC intensity and mismatch, than regarding the role of the sea state. To the least, the evidence about the specificity and importance of wave properties in mismatch characterization is not, in my opinion, sufficiently supported by the results. Even though the hypothesis that sea state explains part of the CYGNSS mismatch is most likely true, the choices made by the authors (mainly, not including the local wind contribution in the WWDI* index, measuring the bias as a system-scale difference with IBTrACs, and treating the sea state index as a systemic diagnostic including TC structure and intensity), do not allow to draw this conclusion properly.
1086–1089 : This limitation is indeed the main and most problematic one. The three studies cited here do not use IBTrACs as a reference for bias correction: Ricciardulli et al 2021 use radiometer/scatterometer data, and Powell and Ruf 2025 use ERA5 winds. The choice of authors to use the IBTrACs error as the main mismatch diagnostic is thus puzzling to me. As discussed in previous comments, the main problem with that choice is that it introduces a mean bias in the analysis, due to the fact that bias will naturally be larger as observations are further from the maximum wind; and that this mean bias induces a systemic bias which is that the wind field asymmetry and the shape of the wind profile will then be primary drivers of the mean mismatch at the scale of a TC acquisition. This invalidates in itself the the idea of the authors of diagnosing a "reference-relative mismatch", since the reference itself induces a system-scale mismatch. What led the authors to chose to use IBTrACs for bias measurement, even though ERA5 local winds were available and have already been used in Powell and Ruf 2025 is one of my main interrogations
1114–1117 : The fact that wave steepness is the main driver, while it is also the most related to local wind speed, indicates in itself the contribution of wind speed to the mismatch.
In my opinion, this is how I am led to interpret your study based on your results:
comparison between CYGNSS and Vmax means that points far from the center will have larger bias, and that bias will increase with intensity, since the difference with Vmax is overall larger as Vmax increases. Since the wind is not included in the analysis, the difference in wind speed will be relfected mostly by the wave steepness: steeper waves ~ higher winds ~ larger mean difference with Vmax1138–1140 : If WWDI* does not provide an improvement on the system-scale compared to bulk wind data, then its use as a specific diagnostic of sea state is not justified, even though it is designed as systemic and not local. The point of this work being to prove the importance of including sea state diagnosis in the characterization of CYGNSS wind mismatchs is not supported by this result.
1148–1150 : what "environmental regimes" were identified here as systematically inducing larger mismatches ? Apart from the very general relation between wave steepness and mean bias, which was discussed above as being a potential reflection of the wind bias induced by the authors definition of mismatch, the authors do not discuss in their figures the existence of specific regimes/combinations of variables leading to higher mismatches. The local variability is ignored in the study (which is a fair choice), but the system-scale changes are not analyzed in depth, as the analysis only provides general error diagnostics and does not draw conjectures on the reason for basin-scale or bin-wise differences in general.
The authors take a lot of precautions to describe their "interpretation" of the result at the end of all sections, but as a result, the designed WWDI* appears simply as a "bulk diagnostic" of sea state that contains inherent biases and does not actually bring information on specific conditions leading to mismatches. The rare occurences showing specific behaviours are not interpreted, such as the reason why the bias is larger after the peak intensity than before, or why some basins some transfers appear to be different from others (cf. comment in section 4.4)-
AC2: 'Reply on RC2', injamam masum, 21 Oct 2026
reply
Dear Referee 2,
Thank you very much for your detailed, thoughtful, and constructive comments. We have addressed all of the points raised and performed additional analyses that substantially clarified the interpretation of the study, particularly regarding local CYGNSS–ERA5 disagreement, storm-structure/reference-scale mismatch, and the role of wind–wave descriptors.
Please find our detailed point-by-point response in the attached supplement.
We greatly appreciate your careful review and helpful suggestions.
On behalf of all co-authors,
S. M. Injamamul Haque Masum
-
AC2: 'Reply on RC2', injamam masum, 21 Oct 2026
reply
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 238 | 130 | 64 | 432 | 54 | 49 |
- HTML: 238
- PDF: 130
- XML: 64
- Total: 432
- BibTeX: 54
- EndNote: 49
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
The manuscript presents a diagnostic framework for CYGNSS wind mismatch in tropical cyclones. My main concern is that the response combines cyclone wind-field structure with differences between wind products. The present evidence does not sufficiently establish how much WWDI* contributes beyond this structural contrast and sampling geometry. I therefore recommend major revision.