Interesting and well-written report, however further insight could have been good, for example using the  Desroziers method to quantify the relative impact of the different observations used in the experiments.

row 72 is vague  as it proposes that ZTD is the only source of moisture data used operational and should be clarified!

This is an excelent manuscript, almost ready to publish.

It addresses a topic that is currently at the forefront of GNSS meteorology: The use of ZTD delay gradients estimated from GNSS data in numerical weather prediction models via data assimilation, on top of assimilation of ZTD.

By consideration of both a dense and a sparse GNSS network dataset, it is demonstrated that assimilation of ZTD gradients (which are cheap to obtain) improves the resulting analyses in a way that would otherwise require a denser network (which is in most cases expensive to obtain) if only ZTD is assimilated.

A few things should be improved:

Specify in more detail the type and amount of "conventional data" that was assimilated.

In line 73. Consider changing to: "..only source of GNSS moisture data..", unless the conventional data did not include any humidity data.  

Figure 2 is confusing.  As I understand the text DA is done every 6 h. I take it is done simultaneously for the different types of observations (depending on the experiment), using a 6 h old forecast from the previous run with that experiment as first guess. 

But why are then the conventional observations represented by tilting arrows, and the GNSS data by vertical. Somehow horisontal is time in the figure, but the arrows are separate in time then.

I would expect there to be a "DA 2" in the box named 20130506 06 UTC, a "DA 3" in the box named 20130506 12 UTC, and so forth.

Review comments to the paper ”Assimilation of GNSS Zenith Delays and Tropospheric Gradients: A Sensitivity Study utilizing sparse and dense station networks”, manuscript ID egusphere-2025-19.

General comments:

The topic of the paper is really interesting and could potentially be very valuable for areas with sparse networks of GNSS receivers. I write “potentially” since from what I read from the paper it is not possible to determine very much from the results. More specific comments will follow but as the paper is written it is not suitable for publication. In fact it is rather poorly written with very little explanations of figures and results which makes it impossible to understand what is plotted and how it should be interpreted. The recommendation is therefore major revisions with a much better description of what is verified and why. I would as well like to see results for forecasts longer than 5 hours.

Specific comments:

1. A minor comment from the introduction: “ZTDs are the only source of moisture data used operationally” – This is not true. Moisture is obtained from radiosondes as well as from several satellite observations. Nowadays, many operational NWP models also include relative humidity at 2 meter height from synop stations.

2. Line 123: Does entire Figure 1 show the model domain? If so (1), there is a lot of area that is not covered by GNSS observations. If so (2), it does not match the domain shown in Figure 4.

3. Line153: The CV5 option does not mean anything unless you are familiar with WRF. Explain further or remove.

4. Line 154: A 6 hour assimilation cycle is not a rapid update cycle. In fact it is the opposite. Most NWP models nowadays run with 3 hour cycling or a rapid update cycle of 1 hour.

5. Lines 155-165: This is a bit hard to follow since the authors have named the experiments after the included observations. The observations are described in the following section. I would suggest to describe the observations first, i.e. switch sections so that 3.2 becomes 3.1 and vice verse.

6. Line 165, Figure 2. A spinup of 12 hours is mentioned in the figure caption. Why is this spinup run and is 12 hours enough? It should be explained in the text and not in the figure caption.

7. Line 159: Why only run 5 hour forecasts? In an impact experiment you would normally want to run at the very least 12 hour but up to 24 is recommended in order to see how persistent the impact from the observations is. If there is a big impact on the analysis and very short forecasts that quickly disappears it can be an indication that your observation error is wrong and you give too much weight to the observations.

8. Section 3.2: Why was the resolutions 0.5 degrees for the dense network and 1.0 degrees for the sparse network selected?

9. Line 178: “a station specific bias correction” – Is this a variational bias correction or a fix one? In case of variational, what are the predictors and if fixed, how was it derived?

10. Lines 179-180: The observation errors, how were they selected or derived?

11. Line 180: “we set up a thorough network of surface reports” – What does this mean? That you install your own observations?

12. Entire section 4: The authors compare the RMSE and the improvement of RMSE. The RMSE of what? Is is the comparison of the analysis and observations or forecast and observations? If the latter, what forecast lengths?

13. Lines 193-204 and Figure 3: The authors state that they “clearly demonstrate” a positive impact. I don’t really see how. First, Figure 3 shows RMSE change (again of what) in percent while the text describes the same change in mm. Please be consistent. Secondly, the figure caption need to describe the figure better, e.g. what is NG and EG (one could guess but still)?

14. Line 206: “we extend the analysis to include 18 independent GNSS stations” – Does this mean the these are in addition to the other stations or instead of the other stations? And why only 18? There are 430 available stations, you select 250 which means you should have an additional 180 stations to use for verification.

15. Line 214: “assimilation of ZTD and TGs significantly enhances that accuracy” – Have you tested the significance or is it just a feeling? If you write this it needs to be validated.

16. Line 234: Again, significant, what does it mean?

17. Figure 4: The unit should be at the large color bar. The panels are labeled a-j but these are never referred to.

18. Line 246: Perhaps it could be of interest to separate the very lowest model levels and the slightly higher ones, i.e. separate the boundary layer and the free atmosphere?

19. Lines 251-252: SSIM index parameter – Please explain shortly what this is and how it is calculated. If the reader is not familiar with SSIM the number 0.98 does not mean much without reference.

20. Figure 5: Again, what is shown? RMSE of what?

General Remarks  

The manuscript provides a relevant overview of the state of the art regarding the assimilation of ZTD and tropospheric gradients, although the operational use of ZTD in major NWP centers could have been detailed further. The evaluation of the complementarity of gradients is an interesting topic, and the idea of two observation densities is insightful. The additional impact of gradients seems very positive and advocates for operational use, but the protocol implemented is not sufficiently detailed, and the results are sometimes lacking in clarity. The following review raises several questions, the answers to which could be usefully added to the manuscript.

Questions about the Dataset Used  

- What conventional data were used?  

- Was it possible to use satellite data in the assimilation cycle? If so, why did you choose not to use them?  

- How was the geographical thinning performed to reduce the resolution of observation sets? Was it based solely on interdistance or also on quality criteria?  

- Did you conduct prior monitoring of ZTD and gradient data? If so, what were the rejection thresholds in RMS and bias? Otherwise, what type of quality control is performed in the assimilation cycle?  

- Why did you choose such a limited sample of excluded stations?  

Questions and Comments on the Impact of the Data  

- Why did you limit the forecast range to that of the assimilation cycle? This seems insufficient for drawing solid conclusions, knowing that a positive short-range impact calculated in the space of added observations can be misleading and unsustainable. At a minimum, forecasts with a 24-hour range should be performed to draw definitive conclusions.  

- The reference for RMSE calculations is unclear. It should be specified in every occurrence within the text and figures what is actually being calculated.  

- Unless I missed it, why did you not use radiosonde data to evaluate the forecast impact in the various configurations?  

- Even though it is not the central topic of the study, it would be interesting to have details on the impact of gradients on the humidity field according to vertical levels. Where are they most informative? Perhaps adding the curve for the ZTD_1.0° experiment in Figure 5 would help?  

- Do you think ZTD and gradients would have had the same impact if satellite observations had been assimilated? The same question applies to radar data.