Multi-station automatic classification of seismic signatures from the Lascar volcano database

Salazar, Pablo; Yupanqui, Franz; Meneses, Claudio; Layana, Susana; Yáñez, Gonzalo

doi:https://doi.org/10.5194/nhess-23-991-2023

Articles | Volume 23, issue 2

https://doi.org/10.5194/nhess-23-991-2023

© Author(s) 2023. This work is distributed under
the Creative Commons Attribution 4.0 License.

https://doi.org/10.5194/nhess-23-991-2023

© Author(s) 2023. This work is distributed under
the Creative Commons Attribution 4.0 License.

Articles | Volume 23, issue 2

Research article

|

03 Mar 2023

Research article |

| 03 Mar 2023

Multi-station automatic classification of seismic signatures from the Lascar volcano database

Pablo Salazar, Franz Yupanqui, Claudio Meneses, Susana Layana, and Gonzalo Yáñez

Download

Final revised paper (published on 03 Mar 2023)
Preprint (discussion started on 14 Jun 2022)

Interactive discussion

Status: closed

RC1:
'Comment on nhess-2022-164', Anonymous Referee #1, 10 Oct 2022

The Authors applied a pre-trained convolutional network (AlexNet) to build a multi-station automatic classification system for volcanic seismic signals. They used data that were recorded by a temporal network installed in the Lascar volcano. The Authors designed and performed four experiments using different datasets with real data, synthetic data, and two different combinations of these (combined 1 and combined 2).

The authors obtain fairly good correct classification performances especially in some experiments (those in which they used a data augmentation technique).

I found the article interesting and worth of publishing in the Natural Hazards and Earth System Sciences journal. I have, however, some general and specific comments, which will require some revision of the article.

General comments:

My feeling is that with some changes in the pre-processing phase the performance can further improve. Furthermore, I find the use of spectrograms in image form rather original, however I do not think it is incorrect.

Specific comments:

The abstract must be improved because the reader does not understand which signals the Authors want to classify and what results they obtain in terms of method performance.

Figures

The labels of the figures must be larger (for example, Fig. 1, 2, 6, 7, 8, 9, 10, 11)

In Fig. 4 the letters a, b, c,

On lines 3, 48, 49 etc. "artificial data" should be replaced with "synthetic data".

Lines 69- 71 This paragraph does not seem clear to me. Please rephrase.

Lines 107 -111 In principle, in step 1 you should first remove the average, then filter (a 10th order filter seems a bit strong to me as a filter for a fairly high frequency band such as 1-10 Hz) and finally resample the series, however I think the ObsPy resampling routines apply an appropriate filter before resampling. Furthermore, looking at figure 2 you can see that tectonic events (TC) and tremor have a significant frequency content in the 10-20 Hz band (this is unusual for volcanic tremor which generally has lower frequencies, between 1 and 6 Hz), therefore, in the pre-processing step it would be more appropriate to filter in the 1 - 20 Hz frequency band.

Lines 117 -123 In step 3 to avoid the imbalance produced on the labeled data, the Authors could create a fifth dataset not applying the data augmentation technique, but selecting a subset of the data so that the amount of data that is selected for each class depends on the number of events in the less populated class; in particular the hybrid events HY class (213 events) that can be used as reference (the number of events for the other classes could perhaps be even slightly higher, for example 250 events for the other classes and 213 for hybrid events).

147 see lines 117 -123

Lines 152 -157 In this paragraph the authors should insert values of performance also in the text and not refer only to table 4.

Line 202 “This will improve the understanding and evaluation of the hazards and risks associated with the activity of volcanoes.” I believe this last sentence is not necessary.

Citation: https://doi.org/10.5194/nhess-2022-164-RC1
- AC1: 'Reply on RC1', Pablo Salazar, 21 Oct 2022
  
  Referee 1:
  Dear Referee 1, thanks a lot for your comments, we are prepared a new documents with all modifications:
  “The abstract must be improved because the reader does not understand which signals the
  Authors want to classify and what results they obtain in terms of method performance.”
  
  Answer: We add some comments in the lines 1-2; line 5, and lines 10-11.
  
  “The labels of the figures must be larger (for example, Fig. 1, 2, 6, 7, 8, 9, 10, 11)”
  
  Answer: We re-scaled the figures and labeled with an adequate font size.
  
  “On lines 3, 48, 49 etc. "artificial data" should be replaced with "synthetic data".”
  
  Answer: artificial data was replaced with synthetic data.
  
  “Lines 107 -111 In principle, in step 1 you should first remove the average, then filter (a
  10th order filter seems a bit strong to me as a filter for a fairly high frequency band such
  as 1-10 Hz) and finally resample the series, however I think the ObsPy resampling
  routines apply an appropriate filter before resampling. Furthermore, looking at figure 2
  you can see that tectonic events (TC) and tremor have a significant frequency content in
  the 10-20 Hz band (this is unusual for volcanic tremor which generally has lower
  frequencies, between 1 and 6 Hz), therefore, in the pre-processing step it would be more
  appropriate to filter in the 1 - 20 Hz frequency band.”
  
  Answer: We added to the routine process your advice. Lines 165-168. We also added a Table 6 that summarized all the results and presents some comments abouts these in the lines 275-278.
  
  “Lines 117 -123 In step 3 to avoid the imbalance produced on the labeled data, the
  Authors could create a fifth dataset not applying the data augmentation technique, but
  selecting a subset of the data so that the amount of data that is selected for each class
  depends on the number of events in the less populated class; in particular the hybrid
  events HY class (213 events) that can be used as reference (the number of events for the
  other classes could perhaps be even slightly higher, for example 250 events for the other
  classes and 213 for hybrid events).”
  
  Answer: This new experiment was created and discuss in the text, lines 243-248; lines 252-254; lines 261-262; lines 269-270; lines 295-298.
  
  “147 see lines 117 -123”
  
  Answer: The new experiment was added to the Results.
  
  “Lines 152 -157 In this paragraph the authors should insert values of performance also in
  the text and not refer only to table 4.”
  
  Answer: The values were added to the paragraph line 250; lines 252-254.
  
  “Line 202 “This will improve the understanding and evaluation of the hazards and risks
  associated with the activity of volcanoes.” I believe this last sentence is not necessary.”
  Answer: This sentence was removed from the text, lines 334-335.
  
  Citation: https://doi.org/10.5194/nhess-2022-164-AC1
RC2:
'Comment on nhess-2022-164', Anonymous Referee #2, 11 Oct 2022

Review of the manuscript nhess-2022-164 by P. Salazar et al.

Multi-station automatic classification of seismic signatures from theLascar volcano database

The topic of the paper interesting, but the manuscript is not fully informative and not well elaborated. Most sections are too compact and are not very informative. The current manuscript requires a quite substantial revision, but it could be in future a good contribution.

My major concern is that the approach and the results are not well presented, elaborated and discussed. These limitations concern the methodological text, the result as well as corresponding figures.

First, I would more extensively describe and discuss AlexNet discussing (a) its implementation, (b) target, and (c) recognized advantages and limitations. These should then be discussed in the light of the analysis of spectrograms, which is here proposed.

Next, authors should better introduce the classification and elaborate on the different classes of signals. This classification is, I believe, inherited from previous manual catalog, where volcano-seismologists introduced the different classes on the base of visual observation of a large amount of data. How is the classification done in the original catalog? What are the typical features? Can one show a compact example of different waveform types and corresponding spectrograms?

Adding some more information on the typical spatial and temporal evolution of these type of events, and the different amount of signals in the original catalog, would also make the paper more interesting and easy to read.

From the technical point of testing the performance of the approach, it would be good to make explicit e.g. the advantage to add merged data-synthetic datasets. Here, I think the first part of the result section could be extended, accompanying the description of each processing with the aim of such test(s).

Authors use only Z component seismograms, and mention 3-component data may be used in future (in a future work). This is acceptable, but the authors should somehow justify this choice.

I am still confused about how the classification performed at different stations is joined into a single catalog. What, e.g. if differently detected and/or classified at different stations?

Results and discussion sections are too compact and not well separated: some of the text in the result section seems to be more adequate for the discussion.

A key issue for the proposed approach, and the interest it may have for a broader volcano-seismology audience, seems to be portability of this method, e.g. to other volcanoes, other time spans, or other monitoring setup. This can only be partially discussed here, as only one volcano monitoring is considered. However, one can still play with different observations. Can we train with one station and process another one?

This type of considerations should be included to extend the discussion, which at the present state is really too short and poor.

Finally, as mentioned above, figures are not fully informative and their layout not optimal. They are often too small, with small captions and need to be substantially improved:

Figure 1 could be more informative, plotting the regions where different type of seismicity was found in the original catalog and to inform the reader where does seismicity take place. Station symbols are too small, as well as labels. The figure size is too small

I would also add a figure showing the time evolution of detections of different classes over time for the study period; such figure could help to summarize these results.

Figure 2 is important and useful, but somewhere hard to read. Labels should be larger, to appreciate the (common) duration of the signal.

Figs. 3-5 can be possibly be merged into a single figure with different panels. Here, again, labels are too small.

Same for Figs. 6 and 7, join in a single figure

And same for confusion matrices (Figs. 8-11), they can be subpanels of a single figure, with no loss of information.

Citation: https://doi.org/10.5194/nhess-2022-164-RC2
- AC2: 'Reply on RC2', Pablo Salazar, 21 Oct 2022
  
  Referee 2:
  Dear Referee 2, thanks a lot for your comments, we are prepared a revised manuscript with all your observations.
  “First, I would more extensively describe and discuss AlexNet discussing (a) its
  implementation, (b) target, and (c) recognized advantages and limitations. These should
  then be discussed in the light of the analysis of spectrograms, which is here proposed.”
  
  Answer: We added some paragraph in order to answer these comments: (a) implementation, (b) target and (c) recognized advantages and limitations. All these aspects are included in the lines 83-113. In the case of discussion, this aspect was included in the lines 304-311.
  
  “Next, authors should better introduce the classification and elaborate on the different
  classes of signals. This classification is, I believe, inherited from previous manual catalog,
  where volcano-seismologists introduced the different classes on the base of visual
  observation of a large amount of data. How is the classification done in the original
  catalog? What are the typical features? Can one show a compact example of different
  waveform types and corresponding spectrograms?”
  
  Answer: This information was added in the lines 131-152. Also the Figures 2 and 3 were added in order to explain the steps.
  
  “Adding some more information on the typical spatial and temporal evolution of these type
  of events, and the different amount of signals in the original catalog, would also make the
  paper more interesting and easy to read.”
  
  Answer: We included a new figure, Figure 5 to show the temporal evolution and the different amount of signals in the catalog. We do not included spatial distribution, mainly because is not the focus of the article and we also have some location for VT events.
  
  “From the technical point of testing the performance of the approach, it would be good to
  make explicit e.g. the advantage to add merged data-synthetic datasets. Here, I think the
  first part of the result section could be extended, accompanying the description of each
  processing with the aim of such test(s).”
  
  Answer: We included some paragraphs in order to give more detail about this in the lines 224-248, lines 181-194.
  
  “Authors use only Z component seismograms, and mention 3-component data may be used
  in future (in a future work). This is acceptable, but the authors should somehow justify
  this choice.”
  
  Answer: We included the lines 123-124 to explain this point.
  
  “I am still confused about how the classification performed at different stations is joined
  into a single catalog. What, e.g. if differently detected and/or classified at different
  stations?”
  
  Answer: This aspect was included in the lines 143-145. Also a new Figure was added Figure 3 in order to explain this.
  
  “Results and discussion sections are too compact and not well separated: some of the text
  in the result section seems to be more adequate for the discussion.”
  
  Answer: Both, Results and Discussion was re-write in order to give more information and cover in the discussion more aspects.
  
  “A key issue for the proposed approach, and the interest it may have for a broader volcano-
  seismology audience, seems to be portability of this method, e.g. to other volcanoes,
  other time spans, or other monitoring setup. This can only be partially discussed here, as
  only one volcano monitoring is considered. However, one can still play with different
  observations. Can we train with one station and process another one?”
  
  Answer: This aspect was discuss in the lines 312-315.
  
  “Figure 1 could be more informative, plotting the regions where different type of seismicity
  was found in the original catalog and to inform the reader where does seismicity take
  place. Station symbols are too small, as well as labels. The figure size is too small”
  
  Answer: The Figure 1 was modified, the sesimicity was not included because is not the focus of our article.
  
  “I would also add a figure showing the time evolution of detections of different classes over
  time for the study period; such figure could help to summarize these results.”
  
  Answer: The Figure 5 was included.
  
  “Figure 2 is important and useful, but somewhere hard to read. Labels should be larger, to
  appreciate the (common) duration of the signal.”
  
  Answer: The former Figure 2, now Figure 4, was modified in the labels associated to the signal.
  
  “Figs. 3-5 can be possibly be merged into a single figure with different panels. Here, again,
  labels are too small.”
  
  Answer: The Figures were modified.
  
  “Same for Figs. 6 and 7, join in a single figure”
  
  Answer: The Figures were modified.
  
  “And same for confusion matrices (Figs. 8-11), they can be subpanels of a single figure,
  with no loss of information.”
  
  Answer: The Figures were modified.
  
  Citation: https://doi.org/10.5194/nhess-2022-164-AC2

Peer review completion

AR: Author's response | RR: Referee report | ED: Editor decision

ED: Reconsider after major revisions (further review by editor and referees) (22 Nov 2022) by Giovanni Macedonio

AR by Pablo Salazar on behalf of the Authors (25 Nov 2022) Author's response Author's tracked changes Manuscript

ED: Referee Nomination & Report Request started (11 Dec 2022) by Giovanni Macedonio

RR by Anonymous Referee #2 (02 Jan 2023)

ED: Publish subject to minor revisions (review by editor) (01 Feb 2023) by Giovanni Macedonio

AR by Pablo Salazar on behalf of the Authors (02 Feb 2023) Author's response Author's tracked changes Manuscript

ED: Publish as is (13 Feb 2023) by Giovanni Macedonio

Short summary

The acquisition of more generalizable models, using machine learning techniques, creates a good opportunity to develop a multi-volcano probabilistic model for volcanoes worldwide. This will improve the understanding and evaluation of the hazards and risks associated with the activity of volcanoes.