the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 12 Jul 2022
Machine learning prediction of the mass and the velocity of controlled single-block rockfalls from the seismic waves they generate
Abstract. Understanding the dynamics of slope instabilities is critical to mitigate the associated hazards but their direct observation is often difficult due to their remote locations and their spontaneous nature. Seismology allows to get unique information on these events, including on their dynamics. However, the link between the properties of these events (mass and kinematics) and the seismic signals generated are still poorly understood. we conducted a controlled rockfall experiment in the Riou-Bourdoux torrent (South French Alps) to try to better decipher those links. We deployed a dense seismic network and inferred the dynamics of the block from the reconstruction of the 3D trajectory from terrestrial and airborne high-resolution stereo-photogrammetry. We propose a new approach based on machine learning to predict the mass and the velocity of each block. Our results show that we can predict those quantities with average errors of approximately 10 % for the velocity and 25 % for the mass. These accuracies are as good as or better than those obtained by other approaches, but our approach has the advantage of not requiring to localize the source and an a priori knowledge of the environment, nor of making a strong assumption on the seismic wave attenuation model. Finally, the machine learning approach allows us to explore more widely the correlations between the features of the seismic signal generated by the rockfalls and their physical properties, and might eventually lead to better constrain the physical models in the future.
-
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
(17945 KB)
-
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
- Preprint
(17945 KB) - Metadata XML
- BibTeX
- EndNote
- Final revised paper
Journal article(s) based on this preprint
Interactive discussion
Status: closed
-
RC1: 'Comment on egusphere-2022-522', Anonymous Referee #1, 31 Jan 2023
The manuscript conducted a series of rockfall experiments in the Riou-Bourdoux catchment. The authors utilized the video footage technique to retrieve the rockfall’s velocity and seismic record to determine several parameters. Then, the random forest tree algorithm was applied to predict the rockfall’s mass and velocity. The authors merged published techniques hitting an excellent point in deciphering two critical parameters, mass, and velocity. However, some basic knowledge of rockfall dynamics, further seismic analysis and other machine learning algorithms should be involved to improve the quality of the manuscript. Thus, I make a decision of major revision to the manuscript.
1.Very detailed background information on seismic techniques for landslide and rockfall research in the introduction. After reviewing the introduction, I supposed that the study is about using a new and accurate reconstruction method of rockfalls’ trajectories, then linking rockfalls’ dynamic to describe the recorded seismic parameters. However, based on the manuscript title, several machine learning methods can be chosen. Why a Random Forest? What are the pro and coin of a Random Forest? Authors can remove all extraneous information like unnecessary landslide parts, focus on rockfall background review and involve machine learning.
2.Line 69-70. random forest tree is not an innovative approach.
3.section 2.1. Rockfalls’ physical processes, falling, bouncing, rolling, and sliding associated with the slope angle (Ritchie, 1963). Authors should involve a topographic profile to show the slope angle change along the rockfall trajectory. It is better to put the slope direction map and slope angle map in the supplement material.
4.Figure 1 lacks the color bar for the translation velocity of rocks.
5.section 2.2. Rocks’ geometry is a vital parameter for rockfall trajectory (Caviezel et al., 2021). The authors should put rocks’ photos and dimensions in supplement material.
6.Line 101-102. Where are the four locations with a terrestrial laser scanning device? Please mark them on the map.
7.Can authors offer one video for a rockfall experiment to help the reader understand the whole process?
8.Line 131. Falling is a mechanism term of rockfall linking to the slope angle. Remove free-falling.
9.Line 158-159. What are the exact distances between the source and the receivers? Is that point-to-point distance or the topographic distance? Both distances do not affect the big picture when the source and the receivers are close. However, gradually enlarging the source-to-receiver distance and the seismic wave transmits through high-relief topography. Two different distances may produce the error for the body and surface wave model. The DEM did not cover the GEO1 to GEO8. So, authors can use GEO9 to GEO16 and the early to middle stage of experiments to explain the effect, then put test results in supplement material.
10.section 2.7. The authors cited several references to support the method of Random Forests, but it could be more explicit. What are the criteria, gini or entropy, in this research? Does the author set max depth for each tree? How to generate a result of OOB score in Figure 6? Random Forests is a black box to produce the result whose opaque process means that implementers must fully trust the model result and cannot understand details. Other machining learning techniques like XGBoost, cluster model …etc. needed to be considered to double-check the result of Random Forests. Further, EGU society encourages authors to offer the source code to allow other researchers reproduce the result. I suggested that the authors release the machining learning code when the paper is published.
11.section 3.1. What β values are in the rockfall experiment? When adapting the body wave model, those β values are located in the ideal range or not.
12.Line 238-239. “Low R2 values might be explained by irregular kinematic behaviors such as the block hitting an obstacle (trees, other rocks) ” Also, sliding shows typically in the early stage of a rockfall, with significant impact after energy loss or near the stopping moment. Please offer a video or time-lapse photos to support the result.
13.The dark and light grey lines in Figure 3 are hard to see.
14.section 3.3. what is the distribution of predicted values? Underestimate or overestimate the actual values (velocity/ mass)? An extra figure for predicted values should be included.
15.section 3.4. Lengthy in this section. In Figure 6(a), #13 gets a more substantial OBB score than others and the result. In Figure 6(b), the OBB score of the top 6-10 is close to the top 4-5. Only describing the top 3 parameters of velocity and mass result is enough where their OBB scores are higher than 0.3.
16.Line 285-287. How about the kinetic energy after impact? It is no difference in R2 between 0.39 and 0.34(Figure 4g, 4h). Figure 4 should include a confidence interval around the slope of a regression line.
17.Line 287-289. Although the rock boulder moved in the West-East direction, the seismic wave transmitted from South to North in the initial stage of rockfall and West-East in the middle stage. Different stages may present different features. Putting different stages of data in one figure is unfair. Further, the poorest correlation observed between Viy, and A0 may be caused by the signals near GEO1 and GEO16, which is worth examining the signals from which experiments and what stage of rockfall bring the poorest correlation.
18.Line 308. what is the definition of spectrum width?
19.Line 323-324. “This suggests that the difference of seismic energies recorded at different stations is important information for the prediction of the velocity (not for the mass).” Why not use signals from cluster stations to execute Random Forest and see their difference? For example, GEO1 to GEO3 and GEO11 to GEO 13 can be grouped. Stations in each group inherit similar paths and site effects, which can reveal how distance affects the results.
20.Line 324-325. Huang et al. (2007) already conducted drop experiments of individual rocks, showing that the larger stones generate the extending feature of lower frequency signals.
21.The cause and effect should be clarified between Lines 326-331.
22.Parameters of ES1 to ES5 are crucial to mass and velocity prediction. Let the y-axis of figure2b be a log scale, which highlights the low-frequency band.
23.Perspective: What are the installation criteria, like station geometry, for further research? Is it necessary to have ten stations with linear arrays? Is all station equally important for machining learning? If not, what is the restriction? Can further research do transfer learning from this research? After reviewing the manuscript, I supposed the authors should answer those questions rather than extraneous information in section 5 to enhance the quality of the paper.
Citation: https://doi.org/10.5194/egusphere-2022-522-RC1 - RC2: 'Comment on egusphere-2022-522', Anonymous Referee #2, 18 Feb 2023
-
EC1: 'Comment on egusphere-2022-522', Tom Coulthard, 28 Feb 2023
Dear Authors,
As there are now two reviews submitted for this paper - I suggest you start working on the comments from both reviewers. Both reviews are favourable (suggesting major and moderate revisions) and the reviews both contain a number of specific points about the MS as well as one or two general comments. Regarding these, both reviewers pick up on the need to revisit the introduction/context setting part of the paper. Thank you for your patience waiting for the review process to complete and I look forward to seeing your revisions,
Tom Coulthard
Citation: https://doi.org/10.5194/egusphere-2022-522-EC1 - AC1: 'Comment on egusphere-2022-522', Clement Hibert, 16 Mar 2023
Interactive discussion
Status: closed
-
RC1: 'Comment on egusphere-2022-522', Anonymous Referee #1, 31 Jan 2023
The manuscript conducted a series of rockfall experiments in the Riou-Bourdoux catchment. The authors utilized the video footage technique to retrieve the rockfall’s velocity and seismic record to determine several parameters. Then, the random forest tree algorithm was applied to predict the rockfall’s mass and velocity. The authors merged published techniques hitting an excellent point in deciphering two critical parameters, mass, and velocity. However, some basic knowledge of rockfall dynamics, further seismic analysis and other machine learning algorithms should be involved to improve the quality of the manuscript. Thus, I make a decision of major revision to the manuscript.
1.Very detailed background information on seismic techniques for landslide and rockfall research in the introduction. After reviewing the introduction, I supposed that the study is about using a new and accurate reconstruction method of rockfalls’ trajectories, then linking rockfalls’ dynamic to describe the recorded seismic parameters. However, based on the manuscript title, several machine learning methods can be chosen. Why a Random Forest? What are the pro and coin of a Random Forest? Authors can remove all extraneous information like unnecessary landslide parts, focus on rockfall background review and involve machine learning.
2.Line 69-70. random forest tree is not an innovative approach.
3.section 2.1. Rockfalls’ physical processes, falling, bouncing, rolling, and sliding associated with the slope angle (Ritchie, 1963). Authors should involve a topographic profile to show the slope angle change along the rockfall trajectory. It is better to put the slope direction map and slope angle map in the supplement material.
4.Figure 1 lacks the color bar for the translation velocity of rocks.
5.section 2.2. Rocks’ geometry is a vital parameter for rockfall trajectory (Caviezel et al., 2021). The authors should put rocks’ photos and dimensions in supplement material.
6.Line 101-102. Where are the four locations with a terrestrial laser scanning device? Please mark them on the map.
7.Can authors offer one video for a rockfall experiment to help the reader understand the whole process?
8.Line 131. Falling is a mechanism term of rockfall linking to the slope angle. Remove free-falling.
9.Line 158-159. What are the exact distances between the source and the receivers? Is that point-to-point distance or the topographic distance? Both distances do not affect the big picture when the source and the receivers are close. However, gradually enlarging the source-to-receiver distance and the seismic wave transmits through high-relief topography. Two different distances may produce the error for the body and surface wave model. The DEM did not cover the GEO1 to GEO8. So, authors can use GEO9 to GEO16 and the early to middle stage of experiments to explain the effect, then put test results in supplement material.
10.section 2.7. The authors cited several references to support the method of Random Forests, but it could be more explicit. What are the criteria, gini or entropy, in this research? Does the author set max depth for each tree? How to generate a result of OOB score in Figure 6? Random Forests is a black box to produce the result whose opaque process means that implementers must fully trust the model result and cannot understand details. Other machining learning techniques like XGBoost, cluster model …etc. needed to be considered to double-check the result of Random Forests. Further, EGU society encourages authors to offer the source code to allow other researchers reproduce the result. I suggested that the authors release the machining learning code when the paper is published.
11.section 3.1. What β values are in the rockfall experiment? When adapting the body wave model, those β values are located in the ideal range or not.
12.Line 238-239. “Low R2 values might be explained by irregular kinematic behaviors such as the block hitting an obstacle (trees, other rocks) ” Also, sliding shows typically in the early stage of a rockfall, with significant impact after energy loss or near the stopping moment. Please offer a video or time-lapse photos to support the result.
13.The dark and light grey lines in Figure 3 are hard to see.
14.section 3.3. what is the distribution of predicted values? Underestimate or overestimate the actual values (velocity/ mass)? An extra figure for predicted values should be included.
15.section 3.4. Lengthy in this section. In Figure 6(a), #13 gets a more substantial OBB score than others and the result. In Figure 6(b), the OBB score of the top 6-10 is close to the top 4-5. Only describing the top 3 parameters of velocity and mass result is enough where their OBB scores are higher than 0.3.
16.Line 285-287. How about the kinetic energy after impact? It is no difference in R2 between 0.39 and 0.34(Figure 4g, 4h). Figure 4 should include a confidence interval around the slope of a regression line.
17.Line 287-289. Although the rock boulder moved in the West-East direction, the seismic wave transmitted from South to North in the initial stage of rockfall and West-East in the middle stage. Different stages may present different features. Putting different stages of data in one figure is unfair. Further, the poorest correlation observed between Viy, and A0 may be caused by the signals near GEO1 and GEO16, which is worth examining the signals from which experiments and what stage of rockfall bring the poorest correlation.
18.Line 308. what is the definition of spectrum width?
19.Line 323-324. “This suggests that the difference of seismic energies recorded at different stations is important information for the prediction of the velocity (not for the mass).” Why not use signals from cluster stations to execute Random Forest and see their difference? For example, GEO1 to GEO3 and GEO11 to GEO 13 can be grouped. Stations in each group inherit similar paths and site effects, which can reveal how distance affects the results.
20.Line 324-325. Huang et al. (2007) already conducted drop experiments of individual rocks, showing that the larger stones generate the extending feature of lower frequency signals.
21.The cause and effect should be clarified between Lines 326-331.
22.Parameters of ES1 to ES5 are crucial to mass and velocity prediction. Let the y-axis of figure2b be a log scale, which highlights the low-frequency band.
23.Perspective: What are the installation criteria, like station geometry, for further research? Is it necessary to have ten stations with linear arrays? Is all station equally important for machining learning? If not, what is the restriction? Can further research do transfer learning from this research? After reviewing the manuscript, I supposed the authors should answer those questions rather than extraneous information in section 5 to enhance the quality of the paper.
Citation: https://doi.org/10.5194/egusphere-2022-522-RC1 - RC2: 'Comment on egusphere-2022-522', Anonymous Referee #2, 18 Feb 2023
-
EC1: 'Comment on egusphere-2022-522', Tom Coulthard, 28 Feb 2023
Dear Authors,
As there are now two reviews submitted for this paper - I suggest you start working on the comments from both reviewers. Both reviews are favourable (suggesting major and moderate revisions) and the reviews both contain a number of specific points about the MS as well as one or two general comments. Regarding these, both reviewers pick up on the need to revisit the introduction/context setting part of the paper. Thank you for your patience waiting for the review process to complete and I look forward to seeing your revisions,
Tom Coulthard
Citation: https://doi.org/10.5194/egusphere-2022-522-EC1 - AC1: 'Comment on egusphere-2022-522', Clement Hibert, 16 Mar 2023
Peer review completion
Journal article(s) based on this preprint
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 650 | 380 | 36 | 1,066 | 23 | 24 |
- HTML: 650
- PDF: 380
- XML: 36
- Total: 1,066
- BibTeX: 23
- EndNote: 24
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Cited
4 citations as recorded by crossref.
- Cluster Analysis of Slope Hazard Seismic Recordings Based Upon Unsupervised Deep Embedded Clustering C. Wang et al. 10.1785/0220230011
- Comparing Flow-R, Rockyfor3D and RAMMS to Rockfalls from the Mel de la Niva Mountain: A Benchmarking Exercise F. Noël et al. 10.3390/geosciences13070200
- Rockfall trajectory reconstruction: a flexible method utilizing video footage and high-resolution terrain models F. Noël et al. 10.5194/esurf-10-1141-2022
- Highly energetic rockfalls: back analysis of the 2015 event from the Mel de la Niva, Switzerland F. Noël et al. 10.1007/s10346-023-02054-2
François Noël
David Toe
Miloud Talib
Mathilde Desrues
Emmanuel Wyser
Ombeline Brenguier
Franck Bourrier
Renaud Toussaint
Jean-Philippe Malet
Michel Jaboyedoff
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
- Preprint
(17945 KB) - Metadata XML