Systematic identification of external influences in multi-year microseismic recordings using convolutional neural networks

Meyer, Matthias; Weber, Samuel; Beutel, Jan; Thiele, Lothar

doi:https://doi.org/10.5194/esurf-7-171-2019

Articles | Volume 7, issue 1

https://doi.org/10.5194/esurf-7-171-2019

© Author(s) 2019. This work is distributed under
the Creative Commons Attribution 4.0 License.

Special issue:

From process to signal – advancing environmental...

https://doi.org/10.5194/esurf-7-171-2019

© Author(s) 2019. This work is distributed under
the Creative Commons Attribution 4.0 License.

Articles | Volume 7, issue 1

Research article

|

04 Feb 2019

Research article |

| 04 Feb 2019

Systematic identification of external influences in multi-year microseismic recordings using convolutional neural networks

Matthias Meyer, Samuel Weber, Jan Beutel, and Lothar Thiele

Download

Final revised paper (published on 04 Feb 2019)
Preprint (discussion started on 06 Aug 2018)

Interactive discussion

Status: closed

AC: Author comment | RC: Referee comment | SC: Short comment | EC: Editor comment

- Printer-friendly version

- Supplement

SC1: 'Interactive Figures', Matthias Meyer, 13 Aug 2018
RC1: 'Review of Meyer et al: Systematic identification of external influences in multi-year micro-seismic recordings using Convolutional Neural Networks.', Anonymous Referee #1, 24 Aug 2018
- AC1: 'Response to RC1', Matthias Meyer, 16 Nov 2018
RC2: 'Review comments for esurf-2018-60', Anonymous Referee #2, 13 Sep 2018
- AC2: 'Response to RC2', Matthias Meyer, 16 Nov 2018
EC1: 'Guiding remarks for revision', Jens Turowski, 28 Sep 2018
- AC3: 'Response to EC1', Matthias Meyer, 16 Nov 2018

Peer-review completion

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

AR by Matthias Meyer on behalf of the Authors (16 Nov 2018) Author's response Manuscript

ED: Referee Nomination & Report Request started (28 Nov 2018) by Jens Turowski

RR by Anonymous Referee #2 (05 Dec 2018)

RR by Marine Denolle (17 Dec 2018)

Suggestions for revision or reasons for rejection

Summary:
The team focuses on event classification to discriminate between the natural and the anthropogenic factors in the hope to provide a more accurate assessment of natural events in future classification. The team does so by focusing on a specific source of negative results, the microseismicity recorded by human motions (mountaineers). The team constructs an elaborative plan to classify the events using multiple data sets (images, seismic events, wind, temperature data), proceeds with a rigorous training of the multiple networks, cross-checking of labeling across the data types. The team finds that human motions induce events that would be classified as rockfall without the additional validation with other data types. They also confirm that radiation is correlated with rockfalls and that summer-time events may be unreliable.
The paper is innovative in terms of combination of multiple data types and multiple network architectures. It is well written and clear, and the data and codes are available.

Several small comments:
I read that the seismic data is multi component. But how was it included in the MicroseismicCNNs, in a 3-channel input like in Perol et al, 2018?

P1L4: “wide range of applications”, can you list them instead of just citing?
P2L30: citation for STA/LTA is Allen 1978, as you note later. It’s worth citing it now.
P3L5-18: In general, in the seismology community, some of these discussions have been introduced as the difference between unsupervised learning techniques (auto-correlations, FAST Yoon et al, 2018; PageRank Aguiar and Beroza 2014, sta/lta Allen 1978), and supervised learning techniques (template matching using cross correlation as a distance metric for instance, and other ML classifier).
P3L8-9-10: I am not sure why the sentence on ML technique is separated from neural network techniques. As you are pointing out, the ground truth data, namely the quality of the labeled data in supervised learning, is tricky. It’s okay with large earthquake catalogs (like in the Ross et et al 2018; Kong et al, 2016), but it remains indeed difficult for non-labeled environmental or anthropogenic factor.

Figure 1: I don’t think it’s necessary. Figure 2 contains some of this content.

P5L15: “used” is utilized twice in the same sentence
P14L7: “the the” (removed one of them)
P15L9: “Chollet and others” -> et al.

P17L15-16: why dropping the other classes and why not using a binary classifier?
P20L16: 1.5 x IQR, use a different font for “times” so that it does not look like the “x”
P20L17: “of of”, remove one “of”

Some citations:
Kong Q, Allen RM, Schreier L, Kwon YW. MyShake: A smartphone seismic network for earthquake early warning and beyond. Science advances. 2016 Feb 1;2(2):e1501055.
Aguiar AC, Beroza GC. PageRank for earthquakes. Seismological Research Letters. 2014 Mar 1;85(2):344-50.
Yoon CE, O’Reilly O, Bergen KJ, Beroza GC. Earthquake detection through computationally efficient similarity search. Science advances. 2015 Dec 1;1(11):e1501057.

Marine Denolle

Hide

ED: Publish subject to technical corrections (17 Dec 2018) by Jens Turowski

ED: Publish subject to technical corrections (17 Dec 2018) by Tom Coulthard (Editor)

AR by Matthias Meyer on behalf of the Authors (11 Jan 2019) Author's response Manuscript

Short summary

Monitoring rock slopes for a long time helps to understand the impact of climate change on the alpine environment. Measurements of seismic signals are often affected by external influences, e.g., unwanted anthropogenic noise. In the presented work, these influences are automatically identified and removed to enable proper geoscientific analysis. The methods presented are based on machine learning and intentionally kept generic so that they can be equally applied in other (more generic) settings.