This paper presents a seismic velocity analysis technique based on a novel non-rigid matching scheme for matching seismic traces in a CIG gather after PSDM. The goal is to derive depth corrections on seismic reflections that can be used for velocity model updating. The paper presents the technique and demonstrates it by applying it to four different settings across the Sunda subduction zone and the Hikurangi margin, each with different challenges for velocity analysis and seismic imaging. The introduction and presentation of the analysis is reasonably well written and easy to follow. However, as discussed below, the writing in later sections, especially the Discussion, needs major revision. Overall, the technique seems very reasonable and effective and is potentially a valuable tool for velocity analysis and prestack depth migration in settings that are not conducive to layer-based models. At its core, the paper is acceptable and should be published. However, I have a number of concerns about the presentation and the basis and validity for many of the claims made here. I suggest major revisions are needed to fix these issues. Below are my biggest issues. There are also many additional comments written on “stickly notes” on the manuscript pages.

Main issues:

• The description of the technique is reasonably well presented and illustrated, but there are two things that are not very clear. First, what is the need for deriving the NRM displacement field and then calculate an NRM residual depth error? It seems like this is simply flattening the arrivals in the gathers. Isn’t this effectively the same as picking the arrivals and flattening them to a common depth? Is the NRM technique mostly a convenient way to derive the residual depth errors, or is there an analysis here that makes it superior? More explanation is needed for what the NRM does beyond simply picking the events to derive a depth offset.

Secondly, I did not find an explanation for how the residual depth errors are used to update the model. Table 1 describes the second to last step as “Update the Tomography Model Properties that will Minimize the CIP-gather RMO”. I did not find any description of this. How exactly is this done, and is this the only way in which the residuals are used? I thought that this is usually done with some constraint such as a linear or hyperbolic relationship with offset. How is it done here?

• The comparison to the initial model is a problematic basis for making any significant conclusions. I certainly see that there is value in using the initial model to demonstrate the changes over the iterations; however, it is not clear that it represent anything that would make a useful basis for conclusions beyond showing improvement. Were the velocities considered to be the best possible that could be derived from a particular technique, and can that be demonstrated? Are the wide-angle data comparable? The wide-angle data are likely to be affected more by anisotropy than near-zero offset data because their raypaths have a significant horizontal component. It is fine to use an initial model to show how the technique is applied and how it changes from start to finish, but it can’t be used to claim that the change in velocity has any geological significance as it is for some of the conclusions. It would be best if it were stated directly that the comparisons between initial and final models are only to demonstrate how the technique improves the results over the iterations and not to show that it is superior to any other technique. You can make it clear that the initial model does not represent a best result from another technique and that you are not claiming this is the degree of improvement that can be expected beyond other techniques.

• There should be some discussion of how anisotropy or effects of streamer feathering may contribute to residual errors or can be accommodated with this technique. There is a brief mention of anisotropy, but no mention of streamer feathering. These are obvious factors that can impact arrival positions within the gathers and skew the velocity models.

• My biggest issue with this paper is with the writing, mostly in sections 4 and 5. There are many vague and unsubstantiated claims made throughout these sections, and some in section 3. For example, in lines 368-372, what determines “large-scale length” what are “significant” velocity corrections? It is not clear why the scale length is appropriate for the first iteration and why there is a corresponding reduction in scale length with subsequent iterations. What makes observed velocity updates more “pronounced”? What defines a “fluctuation of velocity changes” and what determines when they are or are not “related” to “subsurface structures” and what constitutes structures that relate to these changes? This is just three sentences. There are more such issues throughout. There are far too many claims like this that are hard to determine what is meant and what the basis for them is. I have noted many of them on the manuscript, but this section needs to be rewritten.

There are few topic sentences. Without them it is hard to tell what the goal of the paragraph is and where the discussion is leading. The discussion wanders through a number of topics that are not very closely related to each other or seemingly to the rest of the paper. It is not clear why these topics are being discussed and whether they are to help establish the validity of the technique or the results from its application along the Sunda trench or Hikurangi. It is not clear why a new setting (Hikurangi) is added in the discussion rather than as part of the results with the Sunda results.

Overall, this is appears to be a valid technique and the paper does a reasonable job of presenting it, illustrating how it is applied, and showing the expected results. What I don’t understand are the goals of the discussion and how it adds to the paper. The discussion and conclusions need to be significantly revised before this is published.

The use of PSDM images is unfortunately not yet widespread in seismic processing even though the method has well defined and robust procedures for more than 25 years. However, Geomar has a long tradition in seismic processing and imaging and the geophysical group there has used PSDM in many locations and settings to provide novel information and significant scientific results.

The Geomar group in this manuscript present a new approach to produce the macro velocity field necessary for PSDM which is in fact the most critical and arguably complicated step towards producing high-quality PSDM images. The scope of the manuscript is focused on providing the evidence of the superior performance of the NRM as an automatized method to iteratively produce a final model and the resulting images are little discussed in terms of the geological significance of the structures. 

I think the work is a valuable contribution that should be published, but it requires a number of revisions and clarifications before it can be accepted.

The accompanying review by N. Bangs provides very valuable commentaries that I fully subscribe and I have a few more to add:

Comment 1)

There seems to be a misunderstanding in terms of the resolution of the technologies available to build macro-velocity models:

It is stated in Line 61 "The starting velocity model is normally retrieved from seismic data semblance velocity analysis (Neidell and Taner, 1971) of either non-migrated CMP gathers, pre-stack time migrated CIP analyses, wide-angle travel-time tomography, or full-waveform inversions if no additional in situ information is available (Gras et al., 2019; Górszczyk et al., 2019)"

In reality, the uncertainty intrinsic to define the reflector position and velocity above makes reflection-based travel time tomography a method with undefined accuracy that may be improved by using body waves. Further, full-waveform inversions (FWI) is not used to build initial models, but to increase the resolution of models created by some form of travel time tomography, potentially including those obtained from inversion of body waves or the method described in this manuscript.  

Comment  2)

In line with my comment 1, the statement in Line 73 “The precision may be improved by setting a smaller vertical and lateral picking interval to maximize the reliability of the tomography. On the other hand, accuracy is strongly limited by signal interference, background noise, side echoes, and accurate depth error information “ and the statement in Line 77 “we improve the accuracy and precision of the depth error estimation without any hyperbolic assumption or predefined depth horizons of the subsurface structure” fall in the same category. This is that apparently the overall message collectively conveyed in the introduction of the manuscript is that reflection tomography provides the highest possible resolution for velocity determination, which is incorrect. My comment is not in detriment of the work presented, but an introduction should necessarily place the scope of the work in the right technological perspective.

Comment  3)

Section 2 . "Non-rigid and warping matching techniques” describes several different NRM technologies but it is unclear the advantages of some of the them over others, and which one they finally choose and why. In particular the last 2 paragraphs (lines 116-130) could be recast to clarify their choice.

Comment  4)

In Fig 2. “(d) Residual move-out picks calculated by recursive summation of the relative depth errors (b) at predefined depths to get the cumulative depth error. “

I may have gotten it wrong, but should not the vertical axis have a scale that is not the same km of a) and c)

Comment  5)

Line 153-154. “An application of the NRM field to flatten the synthetic gather requires a recursive depth variant correction.“

Since the manuscript deals with the method it would be great that the application is somewhat further detailed.

Comment  6)

Line 171. “Of importance for tomography is not the waveform, but only a correct depth error estimation along reflected events. “

This is again related to my comment 1 and 2, clarify that this is travel time inversion based on near vertical reflections-only.

In fact the authors may want to explore the potential of Vp uncertainty estimation. Since they have developed a method that largely automatizes velocity picking they could design for instance a Monte Carlo based strategy to evaluate the quality of the inversion solution.

In the current manuscript the initial model is only briefly  described and there is no information how much influence the choice of an initial model may have in the performance of NRM and thus on the final velocity model.

Comment  7)

The following two sentences in the manuscript link to my previous comment 5 on initial models.

Line 207-210. “. The initial velocity model for the reflection depth tomography was merged from a velocity tomographic inversion of a collocated 2-D refraction seismic line covered by 46 ocean bottom seismometers (OBS) with a spacing of 6 km (Planert et al., 2010) and a manually estimatedvelocity model for the near seafloor structure.”

What is a manually estimated velocity?

Line 213-214 “. As a consequence, the approximated velocity at shallow depth was additionally smoothed before merging with the wide-angle velocity model and was used as the initial velocity for the NRM-tomography. “

What is the approximated velocity?

I think it would be good to introduce the initial model earlier in the manuscript and to explore its significance. In their example offshore Indonesia the seismic line had been previously intensely studied and a good quality model obtained from ocean bottom seismometers existed, which is not typically the case for most streamer lines. What would be the approach then? To obtain an initial model by another approach? or can NRM work with a simple 1D initial model? would this method work for land data where the initial model may be much more complex? 

Comment  8)

Fig 4. Water Vp of 1590 m/s appears unusually large. In apparent accordance, the CIP gathers in Fig 5 (a) and (b) show un-flattened seafloor reflections. Is that Vp realistic? it is relevant because affects raytracing from across the water layer.

Comment  9)

If they need to condense material to include new or expanded sections I would recommend removing one example offshore Java, or possibly the line from Hikurangi which I imagine that was collected with 6 or 8 km streamer, but the effect of a different acquisition is not discussed and the geology is also rather briefly described.

Also sections 3.2 and 3.3 descriptions might be condensed without harming the manuscript scope.

I summary I recommend that the manuscript may be accepted after some clarifications are made, so that the scope and the description of the methodology, which is the core of the contribution, are further explained, and that  the role of the initial model is integrated in the discussion.

C. R. Ranero