Dear Authors
Thank-you again for your revised manuscript. Both referees have now reviewed this revised version. As a result I am asking for minor revisions. Below are the referee comments on the revised version, and some editorial comments of my own (as sent to you previously). Please address all of these.
Yours sincerely
John Huthnance (editor)


Referee 1 comments
The authors have addressed my comments satisfactorily. The presentation is also improved with the revised figures. I think the paper can be published in OS.

A few minor technical issues:
- There are several blank pages.

- li340. I think the description of the data slope spectrum is still confusing for the reader. Perhaps the narrative could be: obtain the horizontal wavenumber spectrum of the seismic reflection amplitude; multiply this by (2\pi k_x)^2 to produce a so-called "slope spectrum".

- li 421, which generate data

- Conclusions: perhaps rename as "Summary".

- li 437-438: the attenuation of the random noise is repeated in two sentences.

- references: The authors do not give DOIs for the references.

- conference papers: In physical oceanography papers, citation of conference papers, expanded abstracts are not very meaningful. We even avoid citing unpublished papers, theses etc.

- I was hoping the paper was going to be shorter but it came back with 20 figures.
Font size in Figs 18, 19 and 20 can be larger. Sub-panel labels (a, b, c, d) can be moved to side or into the panels.

Thank you.


Referee 2 comments
General comments– The paper has an improved balance of content, though I still think much of the
detail of the filtering method could be moved to supplementary material as it will be of low interest
to the principal readership of this journal.

Editor to advise if American spelling is acceptable and some phasing is still awkward in places but it does not detract from the understanding. <>

Comments and corrections.
line 8 punctuation

line 23: “dropping equipment” ? Rephrase as this is ambiguous

line 25: first published use of technique was by Gonella, J., Michon, D., 1988. Deep internal waves measured by seismic-reflection within the eastern Atlantic water mass. Comptes Rendus de l Academie Des Sciences Serie II 306, 781–787 (in French with English abstract). However it was Holbrook who first used the term seismic oceanography.

Line 27/28: “The differences in temperature and salinity between water column generate ...” poor english

line 29: need reference here suggest Ruddick, B., Song, H., Dong, C., Pinheiro, L., 2009. Water
column seismic images as maps of temperature gradient. Oceanography 22 (1), 192–205.

line 31: need reference for “conventional oceanographic methods” and link to the equipment that was dropped in line 23.

line 34: redundant word “careful”

line 39: “...higher exploration expenses...” ambiguous as higher than what

line 45: the vertical resolution of 1.5 m is based on the theoretical Rayleigh limit however in a 3D environment it is unlikely that you can achieve this due to interference from out-of-plane scattering.

Line 49/50: the maximum impedance contrasts are smaller

line 53 & 64: most noise sources are not random so the authors need to define which noise sources that consider as problematical, especially for SO data where a dominant noise is caused by pressure variations at the receiver due to waves on the surface. This peaks at frequencies below ~5 Hz so with sparker SO you should do better then airguns as you can set your low-cut filter at a higher frequency without damaging your primary signal bandwidth. I think what the authors are looking at here is noise generated from the ship as it has a distinct time characteristic and I would not be surprised if this was during periods of tidal flow against the direction of travel. So perhaps the ships propellers were run at a faster rate to maintain speed over the ground which has resulted in more cavitation. However, that period does not match with the statement that the ship speed was 5.5 knots. Suggest authors include auto-correlation functions both in space and time of the red boxes (Fig 3) indicated to justify that the synthetic random noise used in training datasets is appropriate.

Figure 2 – does this show the whole line or just the portions after the shelf sections were omitted?

Line 163 no comma in velocity value and possibly need to correct style to m s-1 and other units as necessary. Also this is ambiguous as ‘water velocity’ in the context of oceanography describes the mass movement of a body of water. An acceptable term is ‘sound-speed’. As the temperature and salinity of the water is known (XCTD) then authors could do better than just use a generic sound speed by using the equations of state.

Line 173; define what you mean by very small as it looks like the calculated reflection coefficients are at the level of the minimum resolution of the XBT and given that you have two XCTD why did you use a constant density for those?

Line 209: justify why you used 1000 kg m^-3 for the density as true density is provided as part of the model – why does this change make the synthetic more SO-like if you don’t apply a scaling an shift to the compressional wave velocities too. It sounds like a fix to overcome inefficiencies in your modelling process eg calling depth time and followed by a convolution. Given the level of adjustment and modification (see line 224), I see no purpose for using these sophisticated models as a set of time-shifted spikes (eg an enlarged version of Fig 16) would have probably done equally as well.

Line 210: were the simulations acoustic or elastic?

Line 247-260: you don’t actually know the ‘ground truth’ as the section without added noise is, as you correctly point out, still contaminated with noise and will contain inter-layer peg-legs and out-of-plane events which you algorithm will also treat as primary signal.

Line 275: what is x and y in equation 5

line 339 suggest a sub heading here to indicate you have changed to dealing with the SO data

line 342 “and” → “to reveal”

lines 342-344 suggest: Holbrook et al. (2013) suggested analysing the slope of the spectrum for the complete data before calculating the slope of the spectrum from the water reflections because the noise that needs to be suppressed is more evident in the spectrum of the complete data.

line 369 “clearly imaged” → “improved”

Line 375-378, 385-387 & 389-393: I think you are making too much of this as the differences
between the approaches is marginal as shown by your fig 20. Essentially both methods worked on the data to a sufficient degree to enable meaningful estimates of the slope of the internal wave to be discerned. If you feel you want to emphasise the benefit of D2 vs D1 then show a difference plot of the two.

Line 397: what sound speed model did you use for conversion?

Line 406: would be useful to quote horizontal limit of resolution due to the Fresnel zone which I estimate, by eye-balling your data, to be ~0.02 cpm so above this value you would expect the amplitude to the perturbations to be strongly filtered so the fact the signal becomes noise limited is to be expected and further noise reduction will not provide any useful information. I note that your data appear to have a lower frequency content than I would expect – how do you explain that? What were the acquisition parameters? As these directly effect the data signal bandwidth.


Editor’s comments.
Line 67. “McCormack” (spelling)

Lines 95-98. The text and figure 1 do not match well here despite the figure being cited: no “residual learning” in figure 1; no ReLU in text; no “shortcut” in figure 1.

Line 103. “by adding noise-free seismic data (𝑦) and noise (n)” -> “as a sum of “true” seismic signal (y) and noise (n)”

Lines 108-116. This section is difficult for oceanographers owing to the use of several unfamiliar and unexplained terms: residual learning, residual network, network depth (and why should it increase?), residual unit, trainable nonlinear reaction diffusion, batch, batch normalization (there is a gap to its explanation), mini-batch.

Lines 133-140. More difficulty for oceanographers owing to unfamiliar and unexplained terms: rectified linear units (previously mentioned but not explained), activation function, why add nonlinearity? Is “3x3xc” and reference to colour (c=3) necessary since (line 140) c = 1? Why 64 feature maps?

Line 159. “multiplying by the square root of time”

Line 166. “Internal waves in the research area propagate”

Line 172. “(assuming a constant density 1 g/cm3)” is only necessary for the XBT data. If you applied it to the XCTD data, did you ignore conductivity?

Line 189. “that” -> “those”.

Lines 225, 227-228. What do you mean by “epoch”, “iteration”? “is a process” does not explain them.

Line 233. “or” -> “of”

Line 234. What is the purpose of using the “Adam optimizer” ?

Line 332. “point” -> “patch”.

Line 340. “data slope” and “horizontal slope” make no sense. You have not given the definition asked for by Referee 1.

Lines 395-396. “. . scaled the seismic sections again, multiplying the signal by the square-root of return time at each time step to make the amplitude correction.”

Pages 40-42, 45 are blank

Dear Authors
Thank-you for your re-revised manuscript with responses to the referees. I am now asking for minor revisions, as below. This means that I (but not reviewers) should see it again: the main reason is to check that it is intelligible to oceanographers. You argue reasonably that the machine learning aspects should be presented, but there is little value in that if readers cannot understand the words or “jargon” used. Please make every attempt to use standard English and explain technical terms.

Detailed comments.
Line 30. Better “. . (2003). Spatial differences in temperature and salinity generate . .”
Lines 42-43. Better “. . streamers several kilometres long, . .”
Line 47-48. Given your response to Referee 2 Q9, maybe better “. . data with vertical resolution theoretically as fine as 1.5 m were acquired . .”

Lines 100-102. This sentence is not clear. What does “it” refer to? What is a “residual block”? Either “residual blocks” or “a residual block”. Surely machine learning is wanted and not a “problem” “to overcome”?

Lines 236-238. These two sentences might better be at line 234 before first use of “epoch”. If I understand the sentence at lines 236-237, then better “Each training cycle using one mini-batch is an “iteration” and each training cycle using all the training data is an “epoch”.”

Lines 256-258. “Eighty-six 300 × 300 size test data were available, and we divided the test data into 50 × 50 size patches, which is the same size as the training data patch. Then, we discarded the remaining data divided by 128 (mini-batch size) for computational efficiency; thus, the number of test data points was 3,072.” I do not follow this arithmetic. There can be 86 x 36 = 3096 patches in your test data. What exactly was discarded and what was divided by 128? Do you mean “points” or “patches”?
Line 293 and figures 9, 13. Does “before” mean before the first epoch? If so, please make this clear in the text, the figures and especially the figure captions. At present “before” might mean after the 39th epoch, indeed that is what the figure captions definitely imply.

Line 311. “points” –> “patches”?

Line 409. “multiplying the time” –> “multiplying the spectrum by time”? It would help if you made clear that “time” here is a proxy for reflector distance.

Line 437. “which” –> “and”. [What you have said here is that an air gun source can provide information with higher vertical resolution. Only use “which” immediately after the item it refers to.]

Data Availability. You should put the data in a long-term accessible data bank that does not depend on the authors.

References
Need to add Gonella and Michon (1988).

Dear Authors
Thank-you for your re-revised manuscript. I am asking for “technical corrections” according to the detailed comments below, after which you should enter the paper to the editorial system for publication (I do not need to see it again). There will be copy editing and you should check that this keeps your intended meaning. Thank-you for publishing in Ocean Science.
Yours sincerely
John Huthnance (editor)


Detailed comments.
Line 37. “the internal wave” -> “internal waves” (more than one to have a spectrum!)
Line 41. Better “lower” –> “coarser”
Line 136 “rectified linear units”, line 139 “activation function” seem to refer to the same thing according to lines 140-141 “ReLU . . is used for the activation function” and Figure 1 caption “ReLU is rectified linear unit activation function”. Please use consistent names and define what you mean when you first use a name.
Line 222. Delete “with”?
Line 257. Delete first “of”?
Line 300. “. . amount of noise . .”
Line 437. I prefer “higher” –> “finer”

Figure 10 caption line 1. I think you mean “. . before applying the D1 model and after the 40th epoch. The”
Figure 14 caption line 1. I think you mean “. . before applying the D2 model and after the 40th epoch. The” Line 3: “D1” –> “D2”