Ground motion prediction maps using seismic-microzonation data and machine learning

Mori, Federico; Mendicelli, Amerigo; Falcone, Gaetano; Acunzo, Gianluca; Spacagna, Rose Line; Naso, Giuseppe; Moscatelli, Massimiliano

doi:https://doi.org/10.5194/nhess-22-947-2022

Articles | Volume 22, issue 3

https://doi.org/10.5194/nhess-22-947-2022

© Author(s) 2022. This work is distributed under
the Creative Commons Attribution 4.0 License.

Special issue:

Earthquake-induced hazards: ground motion amplification and...

https://doi.org/10.5194/nhess-22-947-2022

© Author(s) 2022. This work is distributed under
the Creative Commons Attribution 4.0 License.

Articles | Volume 22, issue 3

Research article

|

22 Mar 2022

Research article |

| 22 Mar 2022

Ground motion prediction maps using seismic-microzonation data and machine learning

Federico Mori, Amerigo Mendicelli, Gaetano Falcone, Gianluca Acunzo, Rose Line Spacagna, Giuseppe Naso, and Massimiliano Moscatelli

Download

Final revised paper (published on 22 Mar 2022)
Supplement to the final revised paper
Preprint (discussion started on 04 Oct 2021)
Supplement to the preprint

Interactive discussion

Status: closed

RC1:
'Comment on nhess-2021-282', Anonymous Referee #1, 16 Oct 2021

The manuscript utilizes machine learning in ground motion prediction and compares ML-based techniques with GMM and ShakeMap. The former has been shown to have better performance than the latter two. I am not surprised by the results, as has been demonstrated by many that ML techniques are advantages over parametric GMMs. ShakeMap is basically also based on GMMs which reply on an input grid-based VS30 map.

General comments:

What is the logic behind the selection of site proxies? For instance, why do you utilize elevation? Is there any physical reasoning? Is it necessary to use elevation, slope (hx and hy) and curvature (hxx and hyy) simultaneously? Could you provide a plot or some discussions on the performance of each site proxy in the best performing model (GPR)? I recommend the authors expend a bit on the performance of these site proxies?

Specific comments:

Line 74-95: these paragraphs have many details. I suggest they better be moved to other sections, rather than in the introduction. The introduction shall serve to intrigue the readers to read the paper, but these paragraphs are too detailed and may be counter-productive.

Citation: https://doi.org/10.5194/nhess-2021-282-RC1
- AC1: 'Reply on RC1', Federico Mori, 15 Dec 2021
  
  Dear Referee,
  we thank you for your thorough assessment of our paper. We have carefully addressed the comments and made corresponding changes to the manuscript. We have modified and added some figures. We have carefully revised the ambiguous statement in the article.
  
  Citation: https://doi.org/10.5194/nhess-2021-282-AC1
RC2:
'Comment on nhess-2021-282', Anonymous Referee #2, 03 Nov 2021

The comment was uploaded in the form of a supplement: https://nhess.copernicus.org/preprints/nhess-2021-282/nhess-2021-282-RC2-supplement.pdf

Citation: https://doi.org/10.5194/nhess-2021-282-RC2
- AC2: 'Reply on RC2', Federico Mori, 15 Dec 2021
  
  Dear Referee,
  we thank you for your thorough assessment of our paper. We have carefully addressed the comments and made corresponding changes to the manuscript. We have modified and added some figures. We have carefully revised the ambiguous statement in the article.
  
  Citation: https://doi.org/10.5194/nhess-2021-282-AC2

Peer review completion

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

ED: Reconsider after major revisions (further review by editor and referees) (19 Dec 2021) by Hans-Balder Havenith

AR by Federico Mori on behalf of the Authors (11 Jan 2022) Author's response Author's tracked changes Manuscript

ED: Referee Nomination & Report Request started (23 Jan 2022) by Hans-Balder Havenith

RR by Anonymous Referee #1 (24 Jan 2022)

RR by Anonymous Referee #3 (01 Feb 2022)

RR by Anonymous Referee #4 (08 Feb 2022)

Suggestions for revision or reasons for rejection

This paper aims to assess the advantages / limitations of machine learning techniques - with respect to more conventional approaches ie. GMM - to produce ground motion prediction maps for large areas considering both topographic and stratigraphic local conditions. To do so, the authors identified key-parameters on which they develop their methodologies. In addition, they provide a comparison of their ground motion prediction maps to the results of advanced numerical simulations based on detailed sub-soil models.

This reviewer acknowledges that the authors took into consideration some of the guidelines of the reviewers of the first step of the review process, which is good. Answering the remarks / questions below might help to further improve the quality of this paper which, in the present state, still requires minor revisions according to this reviewer.

Scientific comments :
Introduction :
- Line 26 : « regarding the spatial correlation … » the meaning of the spatial correlation in this context is not clear to this reviewer
- Line 81 : as already pointed out by reviewers in the first step of the review process, this paragraph should be erased because it is not commented in the text and it does not help to better understand the work.
- Line 85 : this paragraph could be part of the discussion

Data :
- Line 96 : « some data distributions seem to be imbalanced » : the meaning of this sentence is not clear to this reviewer

Method :
- Line 61 : is Vs30 parameter enough to represent sub-soil conditions and take into account possible geological site effects
- Line 194 : the references should be moved to Line 184 according to this reviewer.

Results :
- Line 277 : this statement of a good agreement between geophysical data and ground motion prediction maps is not convincing to this reviewer. In addition, it would be worth showing also the geomorphological maps to allow the authors to draw such conclusions. As previously mentionned by reviewers in the first step of the review process, it is very difficult to visually determine if both maps are coherent (see Accumoli or Amatrice for instance and compare their values to Petrana). Maybe a different definition of the colorbar scales in both Figures could ease the comparison between plots.
- Line 295 : according to this reviewer, the comparison with numerical modeling results, although worthy of investigation, is not described into sufficient details to allow the readers to compare ML results to more advanced numerical simulations results. Some criticisms were already made by reviewers during the first step of the review process. Because this section of the comparison could be an original / novel contribution of the paper, to this reviewer it deserves further explanations : methodology used in the numerical modeling, materials behaviour and properties, mesh resolution, seismic inputs parameters, etc.
Besides, the correspondance between numerical results and this study should be quantified to this reviewer : looking at the shape of the curves, the fit is not so obvious. Although 0.3s might be of interest for this area, what about the response of both methods at other spectral periods ?

Discussion / conclusion
- Line 371 : « it is possible to predict the losses in the area struck by the earthquake in near real time » : the aim of this paper was to produce ground-motion prediction maps over large areas based on ML technique. As mentionned in the text, the link between ground motions and damage is not straightforward. Therefore the content of this sentence calls for some nuance.

Technical corrections :
- Line 11 : what is a « burried morphology » ? Example ?
- Line 56 : « capable of properly representing … »
- Line 69 : « in terms of »
- Line 104 « it seems a hard task to improve the training dataset » ; line 105 « anyway » : to this reviewer, this is more spoken language

Figures / tables :
- Table 1 : hxx refers to the « second order partial derivative of dxx » ? Same for hyy ? Please check.
- Figures 2 and 3 : replace ‘°’ by ‘th’ in the percentiles definitions (25 and 75). Add units of the parameters (25 th, median and 75 th) when relevant.
- Figure 4 : erase « performance performance »

Hide

ED: Publish subject to technical corrections (09 Feb 2022) by Hans-Balder Havenith

AR by Federico Mori on behalf of the Authors (23 Feb 2022) Author's response Manuscript

Download

Article (28143 KB)
Full-text XML

Short summary

This work addresses the problem of the ground motion estimation over large areas as an important tool for seismic-risk reduction policies. In detail, the near-real-time estimation of ground motion is a key issue for emergency system management. Starting from this consideration, the present work proposes the application of a machine learning approach to produce ground motion maps, using nine input proxies. Such proxies consider seismological, geophysical, and morphological parameters.