Retrieving Precipitable Water Vapor from Real-Time Precise Point Positioning Using VMF1/VMF3 Forecasting Products

Round 1

Reviewer 1 Report

Dear Lily Zhao, dear Dr. Dimitrios Katsanos,

 

The manuscript entitled “Retrieving precipitable water vapor from real-time precise point positioning using VMF1/VMF3 forecasting products” by Peng Sun and 4 co-authors deals with an interesting topic; the assessment of forecast zenith hydrostatic delays and mapping function coefficients provided by Vienna (in essence, three numbers per station per epoch), that are necessary for the atmospheric delay modelling and parameter estimation during GNSS data analysis. The authors did two things. First, they implemented in their software all three atmospheric models and run it for 41 stations for 5-6 months. Second, they compared zenith hydrostatic delays from radiosondes to zenith delays from these forecast products, over 12 months, for 403 stations. The article is well-written, structured, and easy to read. The authors worked with GPS data (5-6 months), VMFx_FC (12 months) and radiosonde data (12 months). No new concept, model, or evaluation methodology developments were carried out within this work. The authors do observations based on the analysis they have carried out, but they do not explain why. While the article has some scientific value and lots of work has been devoted to its preparation, in its current status there are many open questions, all of which are overshadowed by the lack of novelty. I have no good answer to the question I asked myself after reading the article “what have I learned?”, since the conclusion is fairly expected. I strongly recommend a major revision; the final decision rests with the editors.

 

My major concerns follow.

To me, it is unclear what problem the authors try to solve by the work described in this article.

The data that the authors have used to evaluate ZHD, mf_h and mf_w is not enough. The dominant signal in all three has a frequency around 1cpy, for the vast majority of stations. Therefore, the authors should use at least 2.5 years of data so that they will be able to assess the seasonal signals as well.

In assessing the bias and RMS, the authors do not mention whether these values are statistically significant or not. It is very hard for the readership to understand whether these differences are large or not.

The statistical measures they use include averaging over many stations and many epochs and are not appropriate to reveal the subtle differences that are expected between these forecast models.

The authors pit different forecast models against each other, but do not do the most obvious test, that is, to compare these models against an analysis with the “final” setup using their software. Why?

While there is no absolute way to assess the quality of ZTDs, there are some ways to assess the station coordinates estimated from GNSS data analysis (height/etc. repeatability). Since the atmospheric delays are correlated with the height estimates, it would be illuminating to show whether there is a significant difference in this regard. Moreover, those interested in ZTDs are but a small fraction of those interested in positioning; therefore, such an assessment would broaden the readership of the article.

 

More detailed comments follow

23: The authors mention “GNSS”, however, to my understanding they used “GPS-only”. This might confuse the readership. Please, make the necessary changes to reflect the actual work carried out.

48: In equation 1, the authors show that the delay induced by the electrically neutral atmosphere is only dependent upon the elevation angle. However, it is common practice to estimate atmospheric delay gradients in microwave space geodetic techniques such as VLBI and GNSS, since more than two decades. This renders the instantaneous STD a function of elevation and azimuth. Often, such a parameterization is not sufficient; therefore, one should include the residuals in the equation as well.

52: VMF1 and VMF3 are not empirical models. Perhaps the authors intended to refer to GPT3 and its predecessors.

55: State-of-the-art high-resolution weather models such as ERA5 and MERRA2, are very accurate in terms of pressure prediction over areas that do not feature steep orographic gradients; hence, parsing such ZHDs in lieu of in situ pressure logs rarely yields a statistically significant improvement. I would not say that “having in situ met data is a condition to accurate ZHD”.

59: The ZWD WRMS between GNSS, VLBI, and meso-beta scale weather models is typically at the 1 cm level, or better. Also, the water-vapor-weighted mean temperature is usually obtained from weather data. It would be useful for the readership if the authors stated what is the ZHD/ZWD accuracy they aim at.

102: There are several global weather models that provide accurate real-time and forecasts for the fields necessary to estimate zenith delays and mapping function coefficients. Ray-tracing has nothing to do with GNSS data analysis. The reader might get confused by the phrasing used here.

106: The authors could mention other similar services provided by UNB and GFZ.

166: Section 2 feels like a filler that can be pushed to an appendix, to the supplementary material, or skipped altogether. A reference to a review article would be sufficient in my opinion. Which software did the authors use? Perhaps a reference to that would be better instead.

192: The authors could explain what do the coefficients presented in equation 10 represent. Also, they use the same symbol for elevation angle and water vapor pressure.

212: Why?

213: Is it the entire 2020, or the first six months?

217: Equation 11 could be skipped since it is a standard since more than 35 years.

247: Is the uncertainty of the estimated ZWD utilized in the derivation of the statistics? If not why?

250: The authors use MATLAB’s “jet” colormap, which is a poor choice to illustrate bias and RMS. I recommend a polar colormap for the bias and sth like “hot” for the RMS.

265: So, the authors simply guess? The underlying weather fields through which ray-tracing is performed for the derivation of p2 and p3 are identical. Perhaps, the authors could assess whether there is a relation between some statistical values (bias, rms, etc.) and the heigh difference between the actual height of the stations and the projection thereof to the model orography.

267: The reader might wonder what is the difference between the ZHD/ZWD predicted from ECMWF and the zenith delays from the GNSS data analysis. Perhaps, the authors could include such a figure.

277: How can the readership know whether a difference of 0.4mm is significant in this context?

279: Is the title of this section identical to the one of the previous on purpose?

282: Why do the authors use the empirical model GGNTm to obtain the Tm? Is it to refer to their previous work? Since GGNTm is derived from ERA5, isn’t it a bit inconsistent with VMFx products that are derived from ECMWF’s operational model? The readership might get a bit confused here.

302: Many (if not most), jump to the conclusions after reading the title of an article. Perhaps the authors could add a reference or two here, so that a reader would not get any false impressions.

324: Is there any real information in this sentence?

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Review of the paper by Sun et al: Retrieving precipitable water vapor from real-time precise point positioning using VMF1/VMF3 forecasting products.

The authors assess the accuracy of zenith hydrostatic delays (ZHD) from the forecast VMF1 and VMF3 products and the accuracy of precipitable water (PW) from real-time PPP determination with forecast VMF1/VMF3. For the ZHD, they find from comparisons against radiosonde data an agreement at the level of 5mm. In the second part, they compare zenith total delays against IGS values finding an agreement of better than 9 mm, corresponding to an PW error better than 2.4 mm. Their conclusion is that the forecast VMF products are accurate enough for real-time determination of PW.

The research has been carefully carried out, but the readers might have questions: If the forecast VMF products would not be available, then one would have to use empirical models, such as GPT2 for the ZHD and the mapping functions. How much worse would the PW values be?

The authors are comparing against IGS zenith total delays. Do we know how the IGS solution is modelling the a priori ZHD? I think it would be very informative if the authors could calculate their own "final" version and compare it against their real-time PPP solution. Then it would be clearer which contribution comes from the forecast VMF1 and VMF3 products.

 

Minor comments:

line 139: sentence not complete

Table 1: random walk for ZWD: do also write 0.02m/sqrt(s)

Equation 10: what is n?

line 202: it might make sense to provide more details about the vertical reduction model although Kouba is cited.

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

I am fine with the revised version and the rebuttal letter. I believe the manuscript has been sufficiently improved to warrant publication.