OBS noise reduction from horizontal and vertical components using harmonic-percussive separation algorithms

Zali, Zahra; Rein, Theresa; Krüger, Frank; Ohrnberger, Matthias; Scherbaum, Frank

doi:https://doi.org/10.5194/egusphere-2022-823

Preprints

https://doi.org/10.5194/egusphere-2022-823

Preprints

04 Oct 2022

| 04 Oct 2022

OBS noise reduction from horizontal and vertical components using harmonic-percussive separation algorithms

Zahra Zali, Theresa Rein, Frank Krüger, Matthias Ohrnberger, and Frank Scherbaum

Abstract. Records from ocean bottom seismometers (OBS) are highly contaminated by noise, which is much higher compared to data from most land stations, especially on the horizontal components. The high energy of the oceanic noise at frequencies below 1 Hz complicates the analysis of the teleseismic earthquake signals recorded by OBSs.

Previous studies suggested different approaches to remove low frequency noises from the data, but mainly focused on the vertical component. The records of horizontal components, crucial for the application of many methods in passive seismological analysis of body and surface waves could not be much improved in the teleseismic frequency band. Here we introduce a noise reduction method, which is derived from the harmonic-percussive separation algorithms used in Zali et al., (2021) in order to separate long-lasting narrowband signals from broadband transients in the OBS signal. This leads to significant noise reduction of OBS records on both the vertical and horizontal components and increases the earthquake signal to noise ratio without distortion of the broadband earthquake waveforms. This is proved through synthetic tests by measuring SNR and cross-correlation coefficient where both showed significant improvement for different realistic noise realizations. The application of denoised signals in surface wave analysis and receiver function is discussed through synthetic and real tests.

Received: 23 Aug 2022 – Discussion started: 04 Oct 2022

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this preprint. The responsibility to include appropriate place names lies with the authors.

Download & links

Preprint (PDF, 19748 KB)

Notice on discussion status
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
Preprint (19748 KB)

Supplement (4378 KB)

Download & links

The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.

Journal article(s) based on this preprint

01 Mar 2023

Ocean bottom seismometer (OBS) noise reduction from horizontal and vertical components using harmonic–percussive separation algorithms

Zahra Zali, Theresa Rein, Frank Krüger, Matthias Ohrnberger, and Frank Scherbaum

Solid Earth, 14, 181–195, https://doi.org/10.5194/se-14-181-2023,https://doi.org/10.5194/se-14-181-2023, 2023

Short summary

Zahra Zali, Theresa Rein, Frank Krüger, Matthias Ohrnberger, and Frank Scherbaum

Interactive discussion

Status: closed

RC1:
'Comment on egusphere-2022-823', Anonymous Referee #1, 14 Nov 2022
The manuscript covers a noise reduction method for OBS (ocean bottom seismometer) data. OBS are inherently noisier than their land counterparts due to the noise sources in the oceans. It is crucial to clean the contaminated time series to be able to improve analysis methods, e.g. tomographic imaging, receiver function analysis, etc.

The overall method of HPS (Harmonic-percussive separation) has two steps: SIM and MED. First, SIM (similarity matrix) separates the monochromatic and the harmonic noise. In the second step, MED (median filtering) is applied along the time axis of the spectrogram to suppress percussive events and enhance harmonic ones. The method has been previously used on harmonic volcanic tremors and was extended in this manuscript to OBS data.

Further, the manuscript shows how the method on the example of real (DOCTAR experiment) and synthetic data is improving the signal-to-noise ratio.

Generally, I would have liked to see this method applied to a larger dataset. DOCTAR only covered a small area. I would expect that the signals and noise throughout the network are similar to each other. Noise in other parts of the Atlantic Ocean, the Pacific Ocean or the Indian Ocean might show other challenges. It would have been nice to see the method applied to other OBS experiments. Other than that I have only some minor specific comments:

L.148 - 154: You are mentioning the head buoy as a source of harmonic noise, but what about the flag? Can it also play a role or is the strumming of the head buoy overshadowing the flag signal?

L. 351 - 353: You are using qseis for generating synthetic seismograms. Were Source Time Functions (STF) used for synthetics? It is a minor point (and no action is required). There is probably no considerable difference between the tests, but a more realistic STF might even improve some of the cross-correlations.

L. 356-358: Your noise sources were picked at the beginning (N1), during (N2) and after (N3) tidal currents. How do you ensure the signals are not "contaminated" with other noise sources, such as storms or ships? Would that even matter for the analysis?

L.542-551: It would be nice to mention how efficient the algorithm is.

Figure 1: Why don't you show the hydrophone channel? I think it would be nice to see it as a comparison.

Figure 2: This figure is a bit confusing because not all the acronyms are in the caption (e.g. X’, H, R…). They are defined in the text of the manuscript but it would be nice for completeness to have everything in the caption.

Figure 3: Here, it looks like the N3 noise source is close to the earthquake (or the arrow is even pointing at the earthquake). Did you ensure that the noise, which was added to the synthetics, was earthquake free?
Citation: https://doi.org/10.5194/egusphere-2022-823-RC1
- AC1: 'Reply on RC1', Zahra Zali, 17 Nov 2022
  
  We appreciate your comment on applying the HPS denoising algorithm to other OBS data with different noise sources. We have already used the presented algorithm to denoise data from the “KNIPAS” project (Schlindwein et al., 2018) and have found significant improvement when calculating receiver functions (Rein et al., EGU22, manuscript in preparation). We are therefore quite confident that the algorithm is able to significantly reduce noise from different sources as far as they are long-lasting narrowband signals, which is the signature of many important OBS noise signals.
  Here we show one denoising example from the KNIPAS data set is one part of another paper under preparation. We have restrained from including more examples for different data sets in the submitted manuscript to avoid the need for presenting the different experiments and noise conditions that would inflate the size of the paper significantly.
  Figure1. Comparison of the spectrogram and waveforms (white traces) of SO and HPS signals from KNIPAS data (Rein et al., EGU22, manuscript in preparation). The comparison illustrates the noise reduction on the horizontal (H1 and H2) and vertical components.
  L.148 - 154:
  According to Essing et al., (2021), the noise source of the flagpole is most likely depending on the orientation of the OBS instrument, since it is fixed directly on the frame of the OBS. Essing et al. (2021) have analyzed the tremor noise sources in detail and did not observe any dependency of the tremor signal on OBS orientation. Therefore the head buoy is most likely the predominant noise source for the tremor events.
  L. 351 - 353:
  For generating the synthetics with qseis, we used a normalized square half sinus as STF with a duration of 2 s. A more realistic STF would change the source spectrum but not its wideband (transient) characteristic, which is the basis of the separation of the earthquake signal from long-duration narrowband noises using our HPS denoising algorithm.
  L.356-358:
  We ensured that these noise scenarios are not contaminated with other noise sources acting at frequencies below 1 Hz, however, even if other noise sources would exist, the analysis is independent of the type of noises. As stated above, the denoising algorithm will remove mostly the long-lasting narrowband noise type, which is typical for OBS recordings.
  L.542-551:
  Thanks for the remark. We added the below sentence to mention the efficiency of the algorithm.
  An example of one day OBS signal with a sampling frequency of 100 Hz is presented on the GitHub page. The average computation time for this example is about 7 minutes on a PC with an Intel core i7 (six-core) processor of 2.2 GHz and 16 GB of RAM.
  Figure 1: We haven’t shown the hydrophone channel since we feel it does not provide much additional information here. The main purpose of this study is to reduce noise from horizontal components and make use of consistent patterns determined from the spectrograms. Other algorithms (Crawford and Webb, 2000; Bell, et al., 2015) designed for denoising only the vertical component of OBS recordings certainly make use of the hydrophone channel. However, in this study, the hydrophone is not needed for the denoising of the vertical and horizontal components. Therefore we have decided to show only those components, which are used.
  Figure 2: Thanks for the comment. The caption is modified and the information is added.
  Figure 3: The earthquake shown in Figure 3 is the synthetic earthquake, which was added to the real noise data at the position of noise type N3. In this Figure we show the improvement of the HPS noise reduction algorithm on the R and T components, using the illustrated synthetic earthquake at N3 as an example. However, for N1-N3, we have ensured to only add earthquake-free noise data to the synthetic earthquake. We have changed the illustration of arrows in figure 3 to clearly point to the noise and not the earthquake signal.
  References:
  Bell, S. W., Forsyth, D. W., & Ruan, Y.: Removing noise from the vertical component records of ocean-bottom seismometers: Results from year one of the Cascadia Initiative, B. Seismol. Soc. Am., 105(1), 300-313, https://doi.org/10.1785/0120140054, 2015.
  Crawford, W. C., & Webb, S. C.: Identifying and removing tilt noise from low-frequency (< 0.1 Hz) seafloor vertical seismic data. B. Seismol. Soc. Am., 90(4), 952-963, https://doi.org/10.1785/0119990121, 2000.
  Essing, D., Schlindwein, V., Schmidt-Aursch, M. C., Hadziioannou, C., & Stähler, S. C.: Characteristics of Current-Induced Harmonic Tremor Signals in Ocean-Bottom Seismometer Records, Seismological Society of America, 92(5), 3100-3112. , https://doi.org/10.1785/0220200397, 2021.
  Rein, T., Zali, Z., Krüger, F., & Schlindwein, V. (2022). Receiver Function analysis of noise reduced OBS data recorded at the ultra-slow spreading Knipovich Ridge (No. EGU22-3755). Copernicus Meetings.
  Schlindwein, V., Krüger, F., Schmidt-Aursch, M.: Project KNIPAS: DEPAS ocean-bottom seismometer operations in the Greenland Sea in 2016-2017, Alfred Wegener Institute, Helmholtz Centre for Polar and Marine Research, Bremerhaven, PANGAEA, https://doi.org/10.1594/PANGAEA.896635, 2018.
  
  Citation: https://doi.org/10.5194/egusphere-2022-823-AC1
RC2:
'Comment on egusphere-2022-823', Anonymous Referee #2, 22 Nov 2022

This paper describes an interesting method for separating harmonic and "percussive" signals on ocean-bottom seismometer (OBS) data. OBS data have strong harmonic noise and this method could be of great use in identifying and analyzing percussive signals such as earthquakes.

The authors present the parameters used in their algorithm in a haphazard manner and often before they have explained why these parameters are needed. For example, on line 214 they indicate that they divide the frequency content of the signal into two ranges and they give the values of the ranges, but they don't explain why this division is needed and how they choose the ranges until lines 271-272 and 301-303.

The authors should reorganize the text so that the need for parameters and the criteria used to select these parameters is explained from the beginning. This will allow others to more easily understand and profit from their algorithm. I also recommend that the authors make a table of these parameters.

The figure captions often contain expository text that should be in the main text and lack specific information about the figures themselves (see below).

There are some language issues that do not prevent understanding but slow down reading, including extraneous or missing "the"s and overuse of "in order to". Below is an incomplete list of more complicated examples, with suggested corrections:

- L313: "part of that is still remained" -> "part of it remains"

- L314: "The signals ... that don't ... in the spectrogram are difficult to be captured by our HPS algorithm so part..."

-> ""Signals ... , which don't ... in the spectrogram, are difficult to capture using our HPS algorithm, so part...""

# SPECIFIC CORRECTIONS/COMMENTs

## MAIN TEXT

L53-59: The details of microseism noise aren't relevant to the performed

analysis/results.

L82: The IG wave signal used by Crawford and Webb was not recorded by a hydrophone,

but by a differential pressure gauge. Differential pressure gauges, nano-precision

bottom pressure recorders or broadband hydrophones can be used to measure the

IG wave signal, though I'm not sure if broadband hydrphones are sensitive

enough below their corner frequency.

L136-141: These details of the LOBSTER OBSs development aren't relevant to the

method or the data presented.



L175-176: Repeats previous lines.

L180-192: In this description of the MED algorithm, it is not clear which bits

are information about median filters and which are intrinsic to the

specified algorithm.

L210-212: The statement that most OBS noises are narrow-band is false for tilt,

microseism and compliance noise.

L301-303: This should be explained in the algorithm section

L298-300: This should be written more clearly

L 214-215: Why 0.1 to 1 Hz? Why two ranges? This is only explained ~60 lines later.

L 235: Why 2%? Is this a parameter you set? Or an observation of some separation

in S-values?

L306: "will be separated in the": is this another step? or the output of this step?

L320: Does the phase component have a name? (since the amplitude is name "V")

Eq 7: Use the same emphasis in the equation as in the text (N and N' are bold in

the text, but italicized in the equation)



L337: the window length and overlap should be in the parameter table and the chosen

values explained.



L339: is the frequency resolution relevant?

L341: Should your choice of a kernel size of 80 (parameter table!) be used by

other users, or should they run their own tests?



L354: 11.5 km deep oceanic crust: Are there 11.5 km of sediments over this crust?

5 km water + 6.5 km sediments? or do you mean that the bottom of the oceanic

crust is 11.5 km beneath sea level? Or something else?

L361: "Here only those events were used": Is this a subset of the 46 events you name

above, or was this part of the selection criteria that led to 46 events being

chosen. If the former, how many events were used finally?



L380: "improvements of the method" -> "improvements obtained using the method", I think.

L388: Does a high correlation coefficient really demonstrate that there is no waveform

distortion? What is the threshold correlation coefficent for which this is true?



394-396: Unclear.

403-4: Repeats what you already wrote.

L406-7: Isn't 4D just an overlapping plot of the lines in Figure 4C? If so, you

don't need to describe it in such detail.

L438: The sentence starting with "Group velocity curves..." seems out of sequence

L440: "noise situations N1-N3": Be consistent in your naming, you refer to N1-N3

as "situations" here and in the figure captions, "scenarios" on line 357

and "type" on line 519



L445: "in the range of the signal frequencies" repeats, remove it.

"0.05 to 0.2 Hz": you give a frequency range here but the figure only shows periods.

L446: You state that longer signal periods can not be recovered, but it appears

in the figure that they can for N1 and N3, as you state for N3 on lines 447-8

L488-9: "became a broader peak..." Compared to SO? or to P_{410}S?

L508-511: These sentences are not specific enough, they read more like a summary than

a conclusion.



L514-515: Unless, I'm mistaken, this is the first time you mention extracting

microseism signal. If so, this should be mentioned in the discussion,

not the conclusion.

L519-537: Much more specific and detailed than in the discussion, should be put

in the discussion and simply referred to here.



L539-540: "and has especially application in noise reduction of OBS signals": seems

to just repeat the first half of the sentence.



## REFERENCES:

Has some non-standard journal references:

- "Journal of largescale research facilities" -> "J. ... large-scale ..."

- "J. Geophys. Res - Sol Ea" -> "... Solid Earth" (?)

- Essing et al. and Negi et al.: "Seismologial Society of America" -> "Seism. Res. Lett."



Missing DOIs for some articles, including:

- Beyreuther et al, 2010

- Duennebier & Sutton, 1995

- Friedrich et al., 1998

- Langston, 1979

- Romanowicz et al., 1998

- Silver and Chan, 1991

## FIGURES

Figure 2b: put units on axes of spectrogram plots

Figure 3:

- Remove ("SNR is defined as..."), already stated in article text

- Remove (or place in article text) sentences starting by "The spectrograms

clearly show..." and "The whole amplitude and the phase information..."

Figure 4:

- Remove (or place in article text) the three sentences starting with:

- "We see significant improvements..."

- "The HPS signal has significantly lower..."

- "A high noise reduction is seen..."



Citation: https://doi.org/10.5194/egusphere-2022-823-RC2
- AC2: 'Reply on RC2', Zahra Zali, 28 Nov 2022
  
  We appreciate your comment on explaining the parameters of the algorithm in a more structured manner so that the reader can easily understand them. We reorganized the manuscript and add more information at the beginning. However, we keep the detailed explanation about the reason for the two frequency ranges in the method section after we explain SIM. The reason is that the reader needs to know how SIM works to clearly understand it. Now we mention this at the beginning of the method section as well so the reader knows that he/she will have a better understanding of this until the end of the section. We created a table, which contains all the parameters, which we used in our study. The explanation of the parameters is presented in the text.
  - Figure captions: We modified figure captions; removed the unnecessary information which exists in the text, and added more information about the figure itself.
  - Some language issues: We applied the suggested corrections as well as some further language modifications.
  L53-59:
  We shortened the text and removed details of microseism noise.
  L82:
  Thanks for this remark. We replaced “hydrophone data” with “differential pressure gauges”.
  L136-141:
  We removed this part from the manuscript.
  L175-176:
  The sentence is removed.
  L180-192:
  This is true. However, in this subsection, we only described the MED itself as the subsection title shows as well. In the following subsection 3.4 we described how we used MED and SIM in our algorithm.
  L210-212:
  We modified the sentence. OBS noise signals are more narrow-band compared to earthquake signals. The different frequency characteristic of earthquakes and the OBS noise is an important feature that makes HPS suitable for separating them.
  L301-303 & L298-300 & L 214-215:
  We reorganized and modified the text so some information has been moved to the beginning part of the algorithm description. However, for the reader, it is necessary to know how the SIM works to clearly understand the reason for dividing the frequency range into two parts. At the beginning of the Sect. 3.4 we mention the reason for this division briefly. Later we explain it in more detail after the reader knows how the algorithm works. The ranges are now shown in the table parameters as well.
  L 235:
  We use a threshold for picking the highest similarity. We choose the upper 2% of the time frames with the highest S values as the similar frames. We modified the sentence and added the term “the upper” to make it clear.
  L306:
  This is the output. We modified the sentence to make it clear.
  L320:
  As it is explained in the text (line 224), equation 1 separates X into its amplitude (V) and phase components (by looking at the equation it is clear that the phase component is: exp(1j* phi) ) where phi is the phase of X (written in the text line 227). Naming the amplitude component as V helps to better understand figure 2, but it is not needed to write a specific name for the phase component.
  Eq 7:
  Within the whole manuscript, we used bold for the variables in the text and used italic for the equations.
  L337:
  In section 3.6 it is explained that for extracting narrow-band signals a high-frequency resolution is needed in the spectral domain. So it is clear that a long time window should be used for the STFT. We mentioned our recommendation, however, one can use other sizes for the FFT window as far as it is large enough to create sufficiently high resolution in the frequency domain to be able to capture the narrow band nature of noise signals. We have mentioned our choice both in the text and table for allowing the reader to reproduce the results.
  L339:
  Yes, this is relevant since we want to emphasize that the algorithm doesn’t destroy the low-frequency content and that the corresponding waveforms are well-preserved/reconstructed after denoising.
  L341:
  We added more information about the kernel size and how to choose it. 80 is our recommended size but users may want to capture more noise at the cost of probable minimal waveform distortion, so they can choose a larger size. Using the exact recommended size is not critical for the algorithm, but the user should be able to understand the effect of this parameter and tune it based on the application. So we provide the information here and explain how to choose the kernel size. We also present the chosen value in the table.
  L354:
  We agree to describe the structure more precisely. We defined the Moho at a depth of 11.5 km, meaning that the Moho is at 11.5 km below sea level (water depth is 4.9 km, and oceanic crust thickness is 6.6 km). We adapted it accordingly in the main text.
  L361:
  This information is given in Table S1, which is referenced in line 370 in the main text. From all 46 events, some were used for SW analysis, some for RF analysis, and some for both. 9 out of the 46 events were used neither for the SW, nor the RF analysis.
  L380:
  It is now modified to “To quantify the improvements obtained when using our method”.
  L388:
  We agree that a high correlation coefficient solely doesn’t demonstrate that there is no waveform distortion. Also, there isn’t any specific threshold for which the coefficient shows good preservation since there is always some noise remaining after denoising and the amount of the remaining noise depends on the type of noise. However, a high correlation coefficient is an indication of signal preservation. Along with all the other tests, it helps to demonstrate the wanted earthquake waveform preservation. We adapted the text accordingly.
  394-396:
  We modified the unclear part and added more explanations to make it clear.
  403-404:
  We keep this since it is important to mention the peak on the arrival time of seismic phases and emphasis that the energy of seismic phases is preserved.
  L406-7:
  Figure 4d is not an overlapping plot of the lines in Figure 4C, but it is a comparison of the synthetic signal with the trace showing the “difference of SO and synthetic” and the “difference of HPS and synthetic”.
  L438:
  This sentence is removed.
  L440:
  Thanks for this remark. Now we used “situation” in all cases.
  L445:
  The dispersion maps show that noise energy in the range of the signal frequencies is removed successfully for periods between 5 and 20 s. Longer signal periods that are weakly visible in the noise-free image (Fig. 5d) can only partially be recovered.
  L446:
  We modified the sentence and mentioned that they can be only partially recovered.
  L488-9:
  Compared to SO. We changed the sentence to clarify the comparison between SO and HPS.
  L508-511:
  We modified these sentences and now they fit better in the conclusion section.
  L514-515:
  We mention this in the conclusion since this was not the purpose of the study, however, this could be an application of this algorithm. We don’t mention it in the discussion because we didn’t specifically extract the microseism signal but one can do so by applying a bandpass filter to the extracted noise signal.
  L519-537:
  We agree with the comment. We moved some details to the discussion and shortened the conclusion section.
  L539-540:
  The algorithm can extract and separate different signals in the OBS recordings. As mentioned in the previous comment, one can extract the microseism signal for further study on it. This is one application of this algorithm where the extracted signal is the wanted signal. Another application, which we focus on in this study, is noise reduction of OBS recordings where the extracted signals are considered as noise for the study of teleseismic earthquakes.
  References and figures:
  Thanks for the suggestions and corrections. We applied all. We also added other missing DOIs.
  
  Citation: https://doi.org/10.5194/egusphere-2022-823-AC2

Interactive discussion

Status: closed

RC1:
'Comment on egusphere-2022-823', Anonymous Referee #1, 14 Nov 2022
The manuscript covers a noise reduction method for OBS (ocean bottom seismometer) data. OBS are inherently noisier than their land counterparts due to the noise sources in the oceans. It is crucial to clean the contaminated time series to be able to improve analysis methods, e.g. tomographic imaging, receiver function analysis, etc.

The overall method of HPS (Harmonic-percussive separation) has two steps: SIM and MED. First, SIM (similarity matrix) separates the monochromatic and the harmonic noise. In the second step, MED (median filtering) is applied along the time axis of the spectrogram to suppress percussive events and enhance harmonic ones. The method has been previously used on harmonic volcanic tremors and was extended in this manuscript to OBS data.

Further, the manuscript shows how the method on the example of real (DOCTAR experiment) and synthetic data is improving the signal-to-noise ratio.

Generally, I would have liked to see this method applied to a larger dataset. DOCTAR only covered a small area. I would expect that the signals and noise throughout the network are similar to each other. Noise in other parts of the Atlantic Ocean, the Pacific Ocean or the Indian Ocean might show other challenges. It would have been nice to see the method applied to other OBS experiments. Other than that I have only some minor specific comments:

L.148 - 154: You are mentioning the head buoy as a source of harmonic noise, but what about the flag? Can it also play a role or is the strumming of the head buoy overshadowing the flag signal?

L. 351 - 353: You are using qseis for generating synthetic seismograms. Were Source Time Functions (STF) used for synthetics? It is a minor point (and no action is required). There is probably no considerable difference between the tests, but a more realistic STF might even improve some of the cross-correlations.

L. 356-358: Your noise sources were picked at the beginning (N1), during (N2) and after (N3) tidal currents. How do you ensure the signals are not "contaminated" with other noise sources, such as storms or ships? Would that even matter for the analysis?

L.542-551: It would be nice to mention how efficient the algorithm is.

Figure 1: Why don't you show the hydrophone channel? I think it would be nice to see it as a comparison.

Figure 2: This figure is a bit confusing because not all the acronyms are in the caption (e.g. X’, H, R…). They are defined in the text of the manuscript but it would be nice for completeness to have everything in the caption.

Figure 3: Here, it looks like the N3 noise source is close to the earthquake (or the arrow is even pointing at the earthquake). Did you ensure that the noise, which was added to the synthetics, was earthquake free?
Citation: https://doi.org/10.5194/egusphere-2022-823-RC1
- AC1: 'Reply on RC1', Zahra Zali, 17 Nov 2022
  
  We appreciate your comment on applying the HPS denoising algorithm to other OBS data with different noise sources. We have already used the presented algorithm to denoise data from the “KNIPAS” project (Schlindwein et al., 2018) and have found significant improvement when calculating receiver functions (Rein et al., EGU22, manuscript in preparation). We are therefore quite confident that the algorithm is able to significantly reduce noise from different sources as far as they are long-lasting narrowband signals, which is the signature of many important OBS noise signals.
  Here we show one denoising example from the KNIPAS data set is one part of another paper under preparation. We have restrained from including more examples for different data sets in the submitted manuscript to avoid the need for presenting the different experiments and noise conditions that would inflate the size of the paper significantly.
  Figure1. Comparison of the spectrogram and waveforms (white traces) of SO and HPS signals from KNIPAS data (Rein et al., EGU22, manuscript in preparation). The comparison illustrates the noise reduction on the horizontal (H1 and H2) and vertical components.
  L.148 - 154:
  According to Essing et al., (2021), the noise source of the flagpole is most likely depending on the orientation of the OBS instrument, since it is fixed directly on the frame of the OBS. Essing et al. (2021) have analyzed the tremor noise sources in detail and did not observe any dependency of the tremor signal on OBS orientation. Therefore the head buoy is most likely the predominant noise source for the tremor events.
  L. 351 - 353:
  For generating the synthetics with qseis, we used a normalized square half sinus as STF with a duration of 2 s. A more realistic STF would change the source spectrum but not its wideband (transient) characteristic, which is the basis of the separation of the earthquake signal from long-duration narrowband noises using our HPS denoising algorithm.
  L.356-358:
  We ensured that these noise scenarios are not contaminated with other noise sources acting at frequencies below 1 Hz, however, even if other noise sources would exist, the analysis is independent of the type of noises. As stated above, the denoising algorithm will remove mostly the long-lasting narrowband noise type, which is typical for OBS recordings.
  L.542-551:
  Thanks for the remark. We added the below sentence to mention the efficiency of the algorithm.
  An example of one day OBS signal with a sampling frequency of 100 Hz is presented on the GitHub page. The average computation time for this example is about 7 minutes on a PC with an Intel core i7 (six-core) processor of 2.2 GHz and 16 GB of RAM.
  Figure 1: We haven’t shown the hydrophone channel since we feel it does not provide much additional information here. The main purpose of this study is to reduce noise from horizontal components and make use of consistent patterns determined from the spectrograms. Other algorithms (Crawford and Webb, 2000; Bell, et al., 2015) designed for denoising only the vertical component of OBS recordings certainly make use of the hydrophone channel. However, in this study, the hydrophone is not needed for the denoising of the vertical and horizontal components. Therefore we have decided to show only those components, which are used.
  Figure 2: Thanks for the comment. The caption is modified and the information is added.
  Figure 3: The earthquake shown in Figure 3 is the synthetic earthquake, which was added to the real noise data at the position of noise type N3. In this Figure we show the improvement of the HPS noise reduction algorithm on the R and T components, using the illustrated synthetic earthquake at N3 as an example. However, for N1-N3, we have ensured to only add earthquake-free noise data to the synthetic earthquake. We have changed the illustration of arrows in figure 3 to clearly point to the noise and not the earthquake signal.
  References:
  Bell, S. W., Forsyth, D. W., & Ruan, Y.: Removing noise from the vertical component records of ocean-bottom seismometers: Results from year one of the Cascadia Initiative, B. Seismol. Soc. Am., 105(1), 300-313, https://doi.org/10.1785/0120140054, 2015.
  Crawford, W. C., & Webb, S. C.: Identifying and removing tilt noise from low-frequency (< 0.1 Hz) seafloor vertical seismic data. B. Seismol. Soc. Am., 90(4), 952-963, https://doi.org/10.1785/0119990121, 2000.
  Essing, D., Schlindwein, V., Schmidt-Aursch, M. C., Hadziioannou, C., & Stähler, S. C.: Characteristics of Current-Induced Harmonic Tremor Signals in Ocean-Bottom Seismometer Records, Seismological Society of America, 92(5), 3100-3112. , https://doi.org/10.1785/0220200397, 2021.
  Rein, T., Zali, Z., Krüger, F., & Schlindwein, V. (2022). Receiver Function analysis of noise reduced OBS data recorded at the ultra-slow spreading Knipovich Ridge (No. EGU22-3755). Copernicus Meetings.
  Schlindwein, V., Krüger, F., Schmidt-Aursch, M.: Project KNIPAS: DEPAS ocean-bottom seismometer operations in the Greenland Sea in 2016-2017, Alfred Wegener Institute, Helmholtz Centre for Polar and Marine Research, Bremerhaven, PANGAEA, https://doi.org/10.1594/PANGAEA.896635, 2018.
  
  Citation: https://doi.org/10.5194/egusphere-2022-823-AC1
RC2:
'Comment on egusphere-2022-823', Anonymous Referee #2, 22 Nov 2022

This paper describes an interesting method for separating harmonic and "percussive" signals on ocean-bottom seismometer (OBS) data. OBS data have strong harmonic noise and this method could be of great use in identifying and analyzing percussive signals such as earthquakes.

The authors present the parameters used in their algorithm in a haphazard manner and often before they have explained why these parameters are needed. For example, on line 214 they indicate that they divide the frequency content of the signal into two ranges and they give the values of the ranges, but they don't explain why this division is needed and how they choose the ranges until lines 271-272 and 301-303.

The authors should reorganize the text so that the need for parameters and the criteria used to select these parameters is explained from the beginning. This will allow others to more easily understand and profit from their algorithm. I also recommend that the authors make a table of these parameters.

The figure captions often contain expository text that should be in the main text and lack specific information about the figures themselves (see below).

There are some language issues that do not prevent understanding but slow down reading, including extraneous or missing "the"s and overuse of "in order to". Below is an incomplete list of more complicated examples, with suggested corrections:

- L313: "part of that is still remained" -> "part of it remains"

- L314: "The signals ... that don't ... in the spectrogram are difficult to be captured by our HPS algorithm so part..."

-> ""Signals ... , which don't ... in the spectrogram, are difficult to capture using our HPS algorithm, so part...""

# SPECIFIC CORRECTIONS/COMMENTs

## MAIN TEXT

L53-59: The details of microseism noise aren't relevant to the performed

analysis/results.

L82: The IG wave signal used by Crawford and Webb was not recorded by a hydrophone,

but by a differential pressure gauge. Differential pressure gauges, nano-precision

bottom pressure recorders or broadband hydrophones can be used to measure the

IG wave signal, though I'm not sure if broadband hydrphones are sensitive

enough below their corner frequency.

L136-141: These details of the LOBSTER OBSs development aren't relevant to the

method or the data presented.



L175-176: Repeats previous lines.

L180-192: In this description of the MED algorithm, it is not clear which bits

are information about median filters and which are intrinsic to the

specified algorithm.

L210-212: The statement that most OBS noises are narrow-band is false for tilt,

microseism and compliance noise.

L301-303: This should be explained in the algorithm section

L298-300: This should be written more clearly

L 214-215: Why 0.1 to 1 Hz? Why two ranges? This is only explained ~60 lines later.

L 235: Why 2%? Is this a parameter you set? Or an observation of some separation

in S-values?

L306: "will be separated in the": is this another step? or the output of this step?

L320: Does the phase component have a name? (since the amplitude is name "V")

Eq 7: Use the same emphasis in the equation as in the text (N and N' are bold in

the text, but italicized in the equation)



L337: the window length and overlap should be in the parameter table and the chosen

values explained.



L339: is the frequency resolution relevant?

L341: Should your choice of a kernel size of 80 (parameter table!) be used by

other users, or should they run their own tests?



L354: 11.5 km deep oceanic crust: Are there 11.5 km of sediments over this crust?

5 km water + 6.5 km sediments? or do you mean that the bottom of the oceanic

crust is 11.5 km beneath sea level? Or something else?

L361: "Here only those events were used": Is this a subset of the 46 events you name

above, or was this part of the selection criteria that led to 46 events being

chosen. If the former, how many events were used finally?



L380: "improvements of the method" -> "improvements obtained using the method", I think.

L388: Does a high correlation coefficient really demonstrate that there is no waveform

distortion? What is the threshold correlation coefficent for which this is true?



394-396: Unclear.

403-4: Repeats what you already wrote.

L406-7: Isn't 4D just an overlapping plot of the lines in Figure 4C? If so, you

don't need to describe it in such detail.

L438: The sentence starting with "Group velocity curves..." seems out of sequence

L440: "noise situations N1-N3": Be consistent in your naming, you refer to N1-N3

as "situations" here and in the figure captions, "scenarios" on line 357

and "type" on line 519



L445: "in the range of the signal frequencies" repeats, remove it.

"0.05 to 0.2 Hz": you give a frequency range here but the figure only shows periods.

L446: You state that longer signal periods can not be recovered, but it appears

in the figure that they can for N1 and N3, as you state for N3 on lines 447-8

L488-9: "became a broader peak..." Compared to SO? or to P_{410}S?

L508-511: These sentences are not specific enough, they read more like a summary than

a conclusion.



L514-515: Unless, I'm mistaken, this is the first time you mention extracting

microseism signal. If so, this should be mentioned in the discussion,

not the conclusion.

L519-537: Much more specific and detailed than in the discussion, should be put

in the discussion and simply referred to here.



L539-540: "and has especially application in noise reduction of OBS signals": seems

to just repeat the first half of the sentence.



## REFERENCES:

Has some non-standard journal references:

- "Journal of largescale research facilities" -> "J. ... large-scale ..."

- "J. Geophys. Res - Sol Ea" -> "... Solid Earth" (?)

- Essing et al. and Negi et al.: "Seismologial Society of America" -> "Seism. Res. Lett."



Missing DOIs for some articles, including:

- Beyreuther et al, 2010

- Duennebier & Sutton, 1995

- Friedrich et al., 1998

- Langston, 1979

- Romanowicz et al., 1998

- Silver and Chan, 1991

## FIGURES

Figure 2b: put units on axes of spectrogram plots

Figure 3:

- Remove ("SNR is defined as..."), already stated in article text

- Remove (or place in article text) sentences starting by "The spectrograms

clearly show..." and "The whole amplitude and the phase information..."

Figure 4:

- Remove (or place in article text) the three sentences starting with:

- "We see significant improvements..."

- "The HPS signal has significantly lower..."

- "A high noise reduction is seen..."



Citation: https://doi.org/10.5194/egusphere-2022-823-RC2
- AC2: 'Reply on RC2', Zahra Zali, 28 Nov 2022
  
  We appreciate your comment on explaining the parameters of the algorithm in a more structured manner so that the reader can easily understand them. We reorganized the manuscript and add more information at the beginning. However, we keep the detailed explanation about the reason for the two frequency ranges in the method section after we explain SIM. The reason is that the reader needs to know how SIM works to clearly understand it. Now we mention this at the beginning of the method section as well so the reader knows that he/she will have a better understanding of this until the end of the section. We created a table, which contains all the parameters, which we used in our study. The explanation of the parameters is presented in the text.
  - Figure captions: We modified figure captions; removed the unnecessary information which exists in the text, and added more information about the figure itself.
  - Some language issues: We applied the suggested corrections as well as some further language modifications.
  L53-59:
  We shortened the text and removed details of microseism noise.
  L82:
  Thanks for this remark. We replaced “hydrophone data” with “differential pressure gauges”.
  L136-141:
  We removed this part from the manuscript.
  L175-176:
  The sentence is removed.
  L180-192:
  This is true. However, in this subsection, we only described the MED itself as the subsection title shows as well. In the following subsection 3.4 we described how we used MED and SIM in our algorithm.
  L210-212:
  We modified the sentence. OBS noise signals are more narrow-band compared to earthquake signals. The different frequency characteristic of earthquakes and the OBS noise is an important feature that makes HPS suitable for separating them.
  L301-303 & L298-300 & L 214-215:
  We reorganized and modified the text so some information has been moved to the beginning part of the algorithm description. However, for the reader, it is necessary to know how the SIM works to clearly understand the reason for dividing the frequency range into two parts. At the beginning of the Sect. 3.4 we mention the reason for this division briefly. Later we explain it in more detail after the reader knows how the algorithm works. The ranges are now shown in the table parameters as well.
  L 235:
  We use a threshold for picking the highest similarity. We choose the upper 2% of the time frames with the highest S values as the similar frames. We modified the sentence and added the term “the upper” to make it clear.
  L306:
  This is the output. We modified the sentence to make it clear.
  L320:
  As it is explained in the text (line 224), equation 1 separates X into its amplitude (V) and phase components (by looking at the equation it is clear that the phase component is: exp(1j* phi) ) where phi is the phase of X (written in the text line 227). Naming the amplitude component as V helps to better understand figure 2, but it is not needed to write a specific name for the phase component.
  Eq 7:
  Within the whole manuscript, we used bold for the variables in the text and used italic for the equations.
  L337:
  In section 3.6 it is explained that for extracting narrow-band signals a high-frequency resolution is needed in the spectral domain. So it is clear that a long time window should be used for the STFT. We mentioned our recommendation, however, one can use other sizes for the FFT window as far as it is large enough to create sufficiently high resolution in the frequency domain to be able to capture the narrow band nature of noise signals. We have mentioned our choice both in the text and table for allowing the reader to reproduce the results.
  L339:
  Yes, this is relevant since we want to emphasize that the algorithm doesn’t destroy the low-frequency content and that the corresponding waveforms are well-preserved/reconstructed after denoising.
  L341:
  We added more information about the kernel size and how to choose it. 80 is our recommended size but users may want to capture more noise at the cost of probable minimal waveform distortion, so they can choose a larger size. Using the exact recommended size is not critical for the algorithm, but the user should be able to understand the effect of this parameter and tune it based on the application. So we provide the information here and explain how to choose the kernel size. We also present the chosen value in the table.
  L354:
  We agree to describe the structure more precisely. We defined the Moho at a depth of 11.5 km, meaning that the Moho is at 11.5 km below sea level (water depth is 4.9 km, and oceanic crust thickness is 6.6 km). We adapted it accordingly in the main text.
  L361:
  This information is given in Table S1, which is referenced in line 370 in the main text. From all 46 events, some were used for SW analysis, some for RF analysis, and some for both. 9 out of the 46 events were used neither for the SW, nor the RF analysis.
  L380:
  It is now modified to “To quantify the improvements obtained when using our method”.
  L388:
  We agree that a high correlation coefficient solely doesn’t demonstrate that there is no waveform distortion. Also, there isn’t any specific threshold for which the coefficient shows good preservation since there is always some noise remaining after denoising and the amount of the remaining noise depends on the type of noise. However, a high correlation coefficient is an indication of signal preservation. Along with all the other tests, it helps to demonstrate the wanted earthquake waveform preservation. We adapted the text accordingly.
  394-396:
  We modified the unclear part and added more explanations to make it clear.
  403-404:
  We keep this since it is important to mention the peak on the arrival time of seismic phases and emphasis that the energy of seismic phases is preserved.
  L406-7:
  Figure 4d is not an overlapping plot of the lines in Figure 4C, but it is a comparison of the synthetic signal with the trace showing the “difference of SO and synthetic” and the “difference of HPS and synthetic”.
  L438:
  This sentence is removed.
  L440:
  Thanks for this remark. Now we used “situation” in all cases.
  L445:
  The dispersion maps show that noise energy in the range of the signal frequencies is removed successfully for periods between 5 and 20 s. Longer signal periods that are weakly visible in the noise-free image (Fig. 5d) can only partially be recovered.
  L446:
  We modified the sentence and mentioned that they can be only partially recovered.
  L488-9:
  Compared to SO. We changed the sentence to clarify the comparison between SO and HPS.
  L508-511:
  We modified these sentences and now they fit better in the conclusion section.
  L514-515:
  We mention this in the conclusion since this was not the purpose of the study, however, this could be an application of this algorithm. We don’t mention it in the discussion because we didn’t specifically extract the microseism signal but one can do so by applying a bandpass filter to the extracted noise signal.
  L519-537:
  We agree with the comment. We moved some details to the discussion and shortened the conclusion section.
  L539-540:
  The algorithm can extract and separate different signals in the OBS recordings. As mentioned in the previous comment, one can extract the microseism signal for further study on it. This is one application of this algorithm where the extracted signal is the wanted signal. Another application, which we focus on in this study, is noise reduction of OBS recordings where the extracted signals are considered as noise for the study of teleseismic earthquakes.
  References and figures:
  Thanks for the suggestions and corrections. We applied all. We also added other missing DOIs.
  
  Citation: https://doi.org/10.5194/egusphere-2022-823-AC2

Peer review completion

AR: Author's response | RR: Referee report | ED: Editor decision | EF: Editorial file upload

AR by Zahra Zali on behalf of the Authors (28 Nov 2022) Author's response Author's tracked changes Manuscript

ED: Referee Nomination & Report Request started (01 Dec 2022) by Simone Pilia

RR by Anonymous Referee #2 (03 Jan 2023)

ED: Publish subject to minor revisions (review by editor) (10 Jan 2023) by Simone Pilia

AR by Zahra Zali on behalf of the Authors (13 Jan 2023) Author's response Author's tracked changes Manuscript

EF by Polina Shvedko (13 Jan 2023) Supplement

ED: Publish as is (18 Jan 2023) by Simone Pilia

ED: Publish as is (05 Feb 2023) by CharLotte Krawczyk (Executive editor)

AR by Zahra Zali on behalf of the Authors (09 Feb 2023)

Journal article(s) based on this preprint

01 Mar 2023

Ocean bottom seismometer (OBS) noise reduction from horizontal and vertical components using harmonic–percussive separation algorithms

Zahra Zali, Theresa Rein, Frank Krüger, Matthias Ohrnberger, and Frank Scherbaum

Solid Earth, 14, 181–195, https://doi.org/10.5194/se-14-181-2023,https://doi.org/10.5194/se-14-181-2023, 2023

Short summary

Zahra Zali, Theresa Rein, Frank Krüger, Matthias Ohrnberger, and Frank Scherbaum

Supplement

https://doi.org/10.5194/egusphere-2022-823-supplement

Zahra Zali, Theresa Rein, Frank Krüger, Matthias Ohrnberger, and Frank Scherbaum

Viewed

Total article views: 476 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	Supplement	BibTeX	EndNote
346	113	17	476	38	1	2

HTML: 346
PDF: 113
XML: 17
Total: 476
Supplement: 38
BibTeX: 1
EndNote: 2

Views and downloads (calculated since 04 Oct 2022)

Month	HTML	PDF	XML	Total
Oct 2022	117	34	5	156
Nov 2022	129	25	11	165
Dec 2022	38	14	0	52
Jan 2023	25	17	0	42
Feb 2023	37	23	1	61
Mar 2023	0
Apr 2023	0
May 2023	0
Jun 2023	0
Jul 2023	0
Aug 2023	0
Sep 2023	0
Oct 2023	0
Nov 2023	0
Dec 2023	0
Jan 2024	0
Feb 2024	0
Mar 2024	0
Apr 2024	0
May 2024	0
Jun 2024	0
Jul 2024	0
Aug 2024	0
Sep 2024	0

Cumulative views and downloads (calculated since 04 Oct 2022)

Month	HTML	PDF	XML	Total
Oct 2022	117	34	5	156
Nov 2022	129	25	11	165
Dec 2022	38	14	0	52
Jan 2023	25	17	0	42
Feb 2023	37	23	1	61
Mar 2023	0
Apr 2023	0
May 2023	0
Jun 2023	0
Jul 2023	0
Aug 2023	0
Sep 2023	0
Oct 2023	0
Nov 2023	0
Dec 2023	0
Jan 2024	0
Feb 2024	0
Mar 2024	0
Apr 2024	0
May 2024	0
Jun 2024	0
Jul 2024	0
Aug 2024	0
Sep 2024	0

Viewed (geographical distribution)

Total article views: 472 (including HTML, PDF, and XML) Thereof 472 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 17 Sep 2024

Download

The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.

Preprint (19748 KB)
Metadata XML

Short summary

The Investigation of the global Earth's structure benefits from the analysis of ocean bottom seismometer (OBS) data that allow an improved seismic illumination of dark spots of crustal and mantel structures in the oceanic regions of the Earth. However, recordings from the ocean bottom are often highly contaminated by noise. We developed an OBS noise reduction algorithm, which removes much of the oceanic noise while preserving the earthquake signal and doesn’t introduce waveform distortion.


Total:	0
HTML:	0
PDF:	0
XML:	0