In their manuscript “The impact of Aeolus winds on surface wind forecast over tropical ocean and high latitude regions”, the authors analyze the impact of Aeolus data assimilation in the ECMWF model on surface wind in the tropics and high latitudes. This is done by comparing surface buoy and weather station data with two model runs. One with and one without assimilating Aeolus data. The comparison is done by means of normalized difference between the root mean square errors, triple collocation analysis and correlation analysis.

The paper in general is well structured and well written. Unfortunately, the performed analysis and interpretation has some deficiencies which need to be addressed before the publication can be recommended.

The major point concerns the triple collocation. This method is usually used to assess the quality of three independent datasets. In the present study, the authors use it to compare two model runs from the same atmospheric model (which are not independent from each other) with an independent set of in-situ measurements. There is no explanation at all why this method should also be applicable to non-independent datasets and to forecasts.

A second general point to explain would be, why the authors have used the relatively “coarse” ECMWF model for their analysis. They point out, that surface winds are especially important for ocean shipping and wind farming, which is why they focus on surface winds in tropical and polar regions. However, for these use cases often short-term forecasts and higher-resolved regional models are used. So why not perform the analysis on such model runs?

As a third deficiency, I see the interpretation of the results from the tropical region. All comparison methods show a strongly varying impact over time. Nevertheless, the authors claim an improvement of the prediction, which from my point of view cannot be derived from the provided plots.

A detailed list specific comments is given below.

L64: Why are you are limiting your study to tropical and polar regions?

L70: I guess the datasets you are using are also included in the model reanalysis? So, what is the additional benefit of your study? The data assimilation in reanalysis at least somehow deals with the representativeness error between “large” scale model and point wise in situ measurement. You are barely touching this issue in your analysis.

L72: Here you state it yourself: triple collocation analysis is good for the evaluation of three independent data sets. Whereas in your study, I would not say that the two model runs are independent. So why can you use triple collocation?

L75: Here you are talking about representative errors, whereas later on you state “the error covariances are free of representation error”.

Section 3 – Method: The information provided in this section is very short. Could you please expand a bit more on the derivation of the different equations and especially the reasoning why you use two different methods and how they compare with each other. Especially regarding the triple collocation method, I would hope for a bit more explanation, why it can be used under the current circumstances. As far as I know, the method usually uses three independent datasets. However, the two ECMWF model runs are by no means independent. Why do you still think the method is applicable and what does this mean for the equations. Do equations (5) – (7) till hold for the strongly correlated datasets 2 & 3?

L167: Are you now talking about triple collocation or NDRMSE? Which Figure are you describing? Please reference to make it easier for the reader to follow you.

L167 – 175: The results are strongly scattering and changing from positive to negative from time-step to time-step. So, I would be careful with saying the forecast is improved. Your analysis is not providing any reliable result here. Same on Figure 3, btw.

L196/197: As you write in L361f, this is expected. So please explain this to the reader already here.

L200: Again, I think it is problematic to talk about error reductions. The lines cross each other multiple times and there is no clear trend for one or the other.

L220: “The correlations do not reveal much improvement in forecast skill between the two forecasts.” I think this is the correct phrasing also for the two chapters before. No clear evidence for an improvement.

L232ff & L258ff: Any idea why? I guess somehow related to the storm season?

Figures 10 - 12: Can you elaborate a bit more on the temporal evolution of the CTV? Is this expected? Why?

L272f: This is a surprising result that the meridional wind with and without Aeolus assimilation are barely correlated anymore. Any idea why? At these high latitudes, the Aeolus measurement geometry starts to become favourable for measuring the meridional wind. Thus, I would expect a stronger impact on the meridional wind in this region compared to the tropics. However, when looking at the in-situ data, there is only a marginal improvement. So, I do not really understand what is going on here.

L334ff: This might hold for this one time step, but as mentioned before this is strongly varying depending on forecast time and I would be careful with such a general statement.

L342ff: Again, the question: What is the benefit of your study compared to the ECMWF study? It seems to me you have strong issues with the representativeness and the spatial coverage (and number if samples), which are not present in the ECMWF study.

L353ff: The solar background issue might be one reason. Another, physical reason might be related to the seasonality of the atmosphere. Winter is storm season in the high latitudes and correct wind information might have stronger impact on the long term in strong wind conditions. Thus, I would suggest to also look at different wind speed regimes before concluding this to be an instrument problem. Different wind speed regimes / synoptic situations might also be interesting especially for the use cases in shipping and energy production. Strong winds are the most problematic for these use cases. Thus, a more precise prediction of strong wind cases would be very beneficial.

L388f: Since future spaceborne DWL programmes have not been mentioned throughout the whole paper, I wonder if a statement on this topic is well funded here. If the authors want to comment on this topic, they should also include a statement on what would be necessary to further improve the surface wind forecasts they analyse (e.g. wind vector instead of HLOS, higher temporal and/or spatial resolution). I would recommend to completely remove this sentence.

General comments:

  This study examined the impact of Aeolus winds on surface wind forecast from the OSEs using ECMWF model. In general, the impact is quite small and not statistically significant at least in the tropical regions. The impact in the NH is negligible at forecast lead times < 192h.  The triple colocation analysis results look very noisy and hard to interpret.  It is not clear how the correlation between the OSEs will help to justify the performance of the OSEs.  In summary, I do not see any significant impact of Aeolus winds on the forecast of surface winds from the OSEs.

Specific comments:

Abstract, line 13: It is not clear how do you get this conclusion: “The results show that with Aeolus data assimilation, the tropical sea surface wind forecast could be slightly improved at some forecast time steps.”   This is the opposite to the statement from line 175: “Unfortunately, the NDRMSEs are not statistically significant at a 95% confidence interval for all three tropical ocean regions.”

Section 2.1:  The resolution of the ECMWF model version is different from that of Rennie et al (2022). Have you re-tuned the specified observational error for Rayleigh and Mie winds for these specific OSEs in this study? It will be helpful to add some information about how many Aeolus Rayleigh and Mie winds are assimilated into the OSEs in the lower troposphere.

Line 120: Please explain more why those stations with weak correlations (R < 0.5) with the analysis should be removed.  Add numbers of stations were removed.

Line 199, fig. 4:  The triple collocation analyses of the two OSEs (no and with Aeolus) show no evident difference to me. Are the differences are statistically significant? This also applies to Figs. 10, 11.12, 17,18,19.

Line 200: The errors from the two OSEs are <1.0 m. Can you explain more why the errors are so small, considering a typical error of ~ 1.4 m/s of radiosonde near the surface.

Line 220: It is not clear how “The correlations can reveal improvement in forecast skill between the two forecasts?” Please explain.

Line 229:  Are the positive impact statistically significant? Also applies to Fig. 14.

Lines 232:  The seasonal variations of error reductions may not necessarily solely due to the quality of Aeolus winds. Other factors may also contribute to this.

Line 252: why the initial error from OSE with Aeolus is so small, only ~ 0.2 m/s?

Lines 335, 387: This is the opposite to the statement from line 175: “Unfortunately, the NDRMSEs are not statistically significant at a 95% confidence interval for all three tropical ocean regions.” How can you get the statement: “the research findings of this study demonstrate the potential of Aeolus observations on surface wind forecasts with the ECMWF model over the tropical ocean”?