Hazard Mapping of the Rainfall–Landslides Disaster Chain Based on GeoDetector and Bayesian Network Models in Shuicheng County, China

Round 1

Reviewer 1 Report

Dear Authors, your manuscript is well written and addresses an interesting topic (although not entirely new). The paper tackles an important issue (Landslides hazard mapping) using known and already tried methods. The method used in this study is not new, in fact, it had been extensively applied in many regions worldwide.

1. The introduction section is poorly written and does not include a thorough review of previous or latest landslide susceptibility methods.
2. There is no information about how landslide inventory was generated and there is redundancy in the text in section 2.2.
3. In Section 2.4, the authors described the landslide conditioning factors, they used 11 landslide condition factors but the overall analysis is done using 90 meters which will affect the final susceptibility results as analysis at this scale will not represent smaller landslides. Results can be improved by using a better spatial resolution of DEM for analysis. Also, authors have used 240 historical landslide points information and did not mention if they have used the centroid of the landslide or any other part of landslide for analysis.
4. In the methodology section, the authors used The GeoDetector model and Bayesian networks which are used in several papers and hence this paper does not represent any novelty in scientific contribution. The author can see recently published paper using the GeoDetector model (Han et al. 2019; Yang et al. 2019). Results show that ROC for Bayesian network is 0.785 and Logistic regression is 0.71 which are quite lower if compared to previous studies in landslides hazard mapping. So, there is no clear novelty and discussion about why the models have lower ROC values.
5. The English language needs to be improved throughout the manuscript; there are various instances of unclear English.

I recommend rejecting this manuscript and encourage resubmission.

Han L, Zhang J, Zhang Y, Lang Q (2019) Applying a series and parallel model and a Bayesian networks model to produce disaster chain susceptibility maps in the Changbai Mountain area, China Water 11:2144

Yang J, Song C, Yang Y, Xu C, Guo F, Xie L (2019) New method for landslide susceptibility mapping supported by spatial logistic regression and GeoDetector: A case study of Duwen Highway Basin, Sichuan Province, China Geomorphology 324:62-71

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The research has been correctly conducted and the authors achieved to a good result through a large complex of interesting and modern analysis techniques.

The overall quality of the presentation could be even improved and in particular I would like to know why did the authors used SRTM DEM and not the more recent and 30 m ASTER one: despite the large area of study, 90 m grid is considered limiting in landslides assessment.

Besides, the authors used throughout the whole paper the term "landslides hazard" while "landslide hazard" would be more appropriate.

Only few points need to be fixed up before the publication and they can be find in the annoyed version of the manuscript.

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

This study examined the potential risk of landslide in Shuicheng County using the Bayesian network model. The paper is well written and all results are well presented. The study is expected to contribute to prediction of landslide under the consideration of coupled rainfall, topples, slides, and debris flow. So the reviewer recommend publishing this paper is journal Water after addressing one minor comment below:

Page 2, Line 82: Please specify what "very steep slope" means for the quantitative description (e.g., angle of slope). 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Author Response

Dear Reviewer,

Thank you for your approval of our manuscript and satisfaction of the revised content we made. 

Meanwhile, thank you again very much for your constructive comments and suggestions which would help us both in English and in depth to improve the quality of the paper.

Sincerely yours,

Jiquan Zhang

E-mail: [email protected]

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

The authors did not revised the paper point by point according to my previous remarks. The major problems, in my opinion, did not addressed: Cross Validation is not set correctly, no justification exist on the use of q in variable selection, etc.

Also, they did not reply/refute my remarks, justifying their changes that are marked with yellow in the new version of the paper.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

The paper is well structured. I only ask for an intervention to improve the literature. I congratulate the authors on their work.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

This paper is timely, appropriate and of great interest to society and the scientific community.

As the authors indicate, the combination of the GeoDetector and BN model is very promising for spatial prediction of geo-hydrological hazards.

The authors validate the methodology used in a specific study area but technically well studied. This methodology could be applied anywhere else in the world.

The article is very well structured, and the figures and tables are adequate and timely.

The results and conclusions are concrete, without superficial paragraphs. The authors "come to the point".

Perhaps the discussion section could be improved, but it is enough.

Finally, the authors must standardize the references and adapt them to the journal's rules.

Comments for author File: Comments.pdf

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 4 Report

Overall, this manuscript provides a thorough analysis of landslide susceptibility and the interaction between different contributing factors to landslides in Shuicheng County, China. The methodology is well described, with the explanations of the machine learning techniques used being clear and thorough. The structure of the manuscript was appropriate, and the results were useful and supported the conclusions drawn well. The quality of English was also generally good, with generally concise phrasing being used throughout. I could only find a few typos, and have listed them where I have seen them. However, the phrasing of the study tends to alternate between present and past tense; please pick the appropriate tense and apply it consistently throughout the manuscript. References were generally provided appropriately, although the bibliography could use a general formatting consistency check.

This manuscript is quite similar to many other studies currently using machine learning techniques to analyse landslide susceptibility in China. However, the specific techniques discussed in this study and the overall strategy of analysing hazard disaster chains with them is novel enough to distinguish this study from others in my opinion. The main aspect of this study that could be improved if the figure quality. Overall, the choice of figures is appropriate, but aspects such as font size, file format, and colour/symbol choices make them harder to interpret than necessary. I think the study would benefit greatly from higher resolution figures, and the use of vector-based images (e.g. .eps files) where appropriate. I understand much of the visual data is raster, but the text and lines could be vectorised so that the images look clean and are readable at any zoom level. Also, I am not sure what the yellow highlighting throughout the manuscript is for.

Below, I provide specific comments for the text and figures. Provided the other reviewers are satisfied with this study, I am happy to recommend it.

Title: Rainfall-Collapse-Landslide-Debris Flow is quite a lot of hyphens. I think the abstract makes it clear how these different keywords factor into the study, so it may be appropriate to shorten the title to “Hazard mapping of landslide disaster chains… etc”.

Line 18: …collapses, landslides, and debris flows; three of the most…

Line 37: Remove full stop immediately before [1,2].

Line 80: …(ROC) curve, and… (please ensure Oxford comma is used consistently throughout the manuscript)

Line 93: In what time period have these 240 collapses occurred?

Line 95: Can you provide a reference for this?

Figure 1: Please ensure the fonts are consistent and appropriately sized across all subfigures. The km bar and co-ordinates are also much smaller than the other text and difficult to read; please fix this. The image contrast is clear for the Shuicheng County. I think you could also remove the word China from the top-right subfigure and just include it in the caption for additional clarity. The size of Figure 6 is also more appropriate for the main data; the other subfigures are relatively unimportant and should be sized such that the main subfigure is as clear as possible.

Figure 2: The subfigure headings are too large and the rest of the text (colorbars, legends, etc) is too small throughout. Replacing (d) with (days) would also clarify the image futher.

Line 144: (Zhu et al. 2018) should be a numbered reference like the others.

Figure 3: Again, please reduce subfigure headings and make the rest of the fonts consistent. The image contrast is good for each factor, and the color bands are well selected and make the data intuitive to interpret from the figures.

Figure 4: This is a good figure for showing the explanatory power of each factor, but you could make the image much smaller while still being clear by using shorter, vertical bars.

Figure 5: I don’t think you need a figure of the general structure of the network, as you have the specific structure of your network in Figure 7 and the structure is dynamically generated. I would just remove this figure and reference Figure 7 personally.

Figure 6: I think you could replace the main subfigure of Figure 1 with this image and refer back to Figure 1 to explain what the training and validation points are. Keep the elevation legend from Figure 1 as well.

Table 4: The formatting of this table could be improved to provide further clarity. I would bold the numbers in the left hand tables where the predicted and actual values matched, e.g. going from top left to bottom right, to make it visually clear what proportion of the predictions were correct. The “Very” column also needs to be “Very Low” for each subtable. I would also change “Level” to “Risk Level” throughout for further clarity. Finally, some extra spacing between the three subtables would further improve readability.

Line 275: Replace girds with grids.

Table 5: I would replace scilicet with i.e. as while the former is technically also correct, the latter is more widely used and known.

Figure 7: The font size is just about okay but could be clearer if slightly larger.

Figure 8: These subfigures are too small to be clear, please increase size. Subheadings are too larger, other text is too small, also (d)-(f) could use a smooth colour gradient such as red-orange-yellow-green-blue for additional readability of the data.

Figure 9: This is the summary image for the overall hazard risk across the area, so it is frustrating that so much of the data is blocked by opaque symbols for disasters we already know the positions of from other figures. I would recommend either using less obstructive symbols such as crosses or dots or using hollow symbols so the underlying data can still be seen. Otherwise please see my comments for other similar figures.

Line 433: Replace Frist with First.

References: Please ensure font size is consistent throughout. Also, journal titles should not be capitalised, journal titles and numbers should be italicised, and article titles should be uncapitalised (with the exception of proper nouns and placenames) throughout. Please ensure that DOIs are provided for all articles for which they exist, and that journal titles are not shortened throughout. Please refer to other published articles from this journal to see what is and isn’t appropriate for this section in terms of formatting.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 5 Report

The paper presents major problems in dealing with geomorphological processes that perhaps the authors are unfamiliar with. Different things are mixed together and treated superficially, already in the title and then in the text, for example talking about "Collapse-Landslide-Debris Flow". What does "collapse" mean? rockfalls, sinkholes, perhaps? These are also landslides .. so why write generically "landslide". The same thing also applies to debris flows.
Mathematical models cannot be defined, which, although they may have their mathematical validity, do not correspond to reality. There is not a geomorphological map showing the landslide processes, nor a photo that can make us understand because landslides then become "disasters", compared to what, to which vulnerable objects?
In the manuscript there is a great terminological confusion, but also in the definition of thematic maps, which suggests a lack of familiarity of the authors with the objects treated, but their affinity with the computer sciences.
Unfortunately the two worlds cannot speak to each other in this way. A more rigorous approach is needed, both in dealing with the general topic and to show the reader and other researchers or scholars that their mathematical method is useful.
In this form, the work is unacceptable because it does not respect journal quality standards.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Dear authors,

I am sorry to reject your article another time. In my opinion the major problems/flaws with your analysis were not addressed, as I explained thoroughly in my first review.

1. There is no external validation https://0-doi-org.brum.beds.ac.uk/10.1073/pnas.102102699, on the use of q in variable selection. See also https://0-doi-org.brum.beds.ac.uk/10.1016/B978-1-55860-335-6.50023-4 about Irrelevant Features and the Subset Selection Problem.
2. Why should one use the threshold of 0.025 and not 0.05 or any other value in variable selection?
3. BN can be used with build-in feature selection (using the Markov Blanket). Why didn’t you use it? 
4. The authors must refer to chapter 7 and especially 7.10.2 of “The Elements of Statistical Learning”, about the correct way to do validation. You cannot use all your data in feature selection and then split them to a training and testing set, you have “data leakage”. The test set should be a random selected set, kept aside during both feature selection and model training.
5. How did the algorithm performed on the training set? The reader should know if BN overfit.
6. There is no description (exploratory analysis, statistics, contingency tables, etc.) of the points with the recorded disasters, as well as the random selected non-disaster points. This is very important in order for the reader to see and understand your data set.
7. From your results in Table 4 you should mention that debris flow has poor results. A naïve model that returns only “No” and “very low” would have an accuracy of 133/142 = 93.7%.
8. The different daily rainfall indices are still highly correlated, as showed in Figure 3. If you reported the correlation of the numeric values we could see these very high values of correlation. Did you try rainfall erosivity as an input variable?
9. In Table 4 please change the “low” to “very low” and report the error metrics that you use in section 3.4
10. NDVI is a continues variable and not a discrete (table 2)
11. Table 6 has a different format of TP, TN etc. than Table 4
12. When I asked about the mean computed percentage risk levels I expected a table such as: low x%, medium y% high z%.
13. If a more “naïve” model as logistic regression offers comparable results (perhaps using cross validation these would be statistical insignifficant) why should we use the proposed methodology that is complex, with arbitrary selected hyper-parameters (as is q in feature selection) and not for example LASSO (least absolute shrinkage and selection operator) that offers variable selection and regularization out of the box?
14. Please keep in mind that you should compare algorithms using exactly the same training – testing sets. CNN results refer to a different data set.

Reviewer 5 Report

The authors have done a considerable and appreciable work of revisiting the initial manuscript, correcting lexical and terminological errors in the classification of landslides and defining the contexts of "disaster".

The work has improved.

- indicate the years of the two google earth images in figure 2;

- in the introduction, between lines 74 and 84, or in the paragraph "2.4. landslides influencing factors", I suggest considering also the following papers in the citations:

Lazzari M., Piccarreta M. 2018 - Landslide disasters triggered by extreme rainfall events: the case of Montescaglioso (Basilicata, southern Italy). Geosciences, 8 (10), 377.  doi: 10.3390/geosciences8100377

Lazzari M., Piccarreta M., Capolongo D., 2013 – Landslide triggering and local rainfall thresholds in Bradanic Foredeep, Basilicata region (southern Italy). Landslide Science and Practice. Volume 2. Early warning, instrumentation and modeling, Margottini et al. (eds) - Springer Series, Proceedings of the Second World Landslide Forum, Rome (ITALY) 3-9 October 2011, pp. 671-678, Vol. 2