Investigation of Landslide Susceptibility Decision Mechanisms in Different Ensemble-Based Machine Learning Models with Various Types of Factor Data

Round 1

Reviewer 1 Report

The study is attempt to use the SHapley Additive exPlanations method to interpret and analyze landslide susceptibility models constructed using ensemble-based machine learning methods combined with different types of factor data. The topic is fit to this journal, but major revisions need to be done before acceptance.

Major concerns

1. What is innovative about this manuscript? Please elaborate on it.

2. The authors collected 214 historical landslide areas using various methods such as Google Earth images, op-214 tical satellite images, and disaster news reports. But detailed information about that is missing. How was it collected? How reasonable are the historical landslide data? Where, when, and how much of the historical landslide was it?

3. As a research article, there are too many references to accept.

4. The figure quality needs to be improved. The legend and text in Fig. 2, Fig.7, and Fig.8 are difficult to be recognized.

Author Response

请参阅附件

Author Response File: Author Response.pdf

Reviewer 2 Report

Overall the manuscript is well written and in detail. The introduction provides a comprehensive overview of landslide susceptibility and the importance of accurate and reliable landslide susceptibility maps (LSMs). It delves into various analysis methods, from qualitative and semi-quantitative to quantitative, emphasizing the role of machine learning (ML) models. The authors present a literature review, detailing the current state of the field, and introduce the objective and organization of the paper.

1) The introduction is quite long and detailed. While thoroughness is appreciated, a more concise introduction might make it more accessible to a broader audience.

2) While the authors do touch upon the novel aspects of the study, this could be made more explicit. Clarifying the unique contributions of the research would help readers immediately understand the innovation

3) While the sampling method is mentioned, some readers may benefit from a more detailed explanation or justification for the specific criteria used in the selection of non-landslide samples.

4) The ML/AI section could be improved by including some latest references: https://www.sciencedirect.com/science/article/pii/S1674775522001093

5) The writing is generally clear but has some repetition and can benefit from careful editing. For instance, lines 907-909 repeat the idea that indicators only explain part of the results.Consider revising for conciseness, removing redundant statements, and improving overall readability.

Moderate editing of English language required

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

Thank you for extending the invitation. I have thoroughly reviewed the manuscript and would like to provide my feedback:

1) As noted by the authors, the SHAP technique represents a burgeoning approach in natural hazard susceptibility studies. They have adeptly demonstrated various SHAP plots in their paper. However, I find it peculiar that the authors have not acknowledged and embraced recent SHAP-based geohazard studies in their introduction, and elucidating how their research output complements, confirms, or contradicts these studies in discussion part. In light of this, I am appending a list of relevant references at the end of my report for their perusal and consideration. Engaging in scholarly discourse on SHAP-based studies is crucial given their novelty, and such a dialogue could enhance the scientific value of this work. Then, please reduce the number of articles cited which embraced very old techniques.

2) I would highly recommend including the mathematical notations of Shapley Additive exPlanations (SHAP) in the manuscript. To that end, I refer the authors to the recommended reference list where they can find suitable sources for incorporating these notations.

3) The illustrations presented in the results section appear somewhat overcrowded. I propose that the authors reconsider their approach and refrain from displaying every type of SHAP plot for each experiment, as evident in Figs. 14-19. Instead, they might consider selecting a representative experiment for both Random Forest (RF) and XGBoost algorithms, focusing on the one that exhibit the highest accuracy, as determined by their expertise.

4) I must mention that the correlation matrix can be rather visually taxing. To address this, I urge the authors to explore alternative visualization methods that are better suited for larger matrices. 

5) In order to bolster the credibility of the negative (non-landslide) samples, it would be beneficial to include them in Fig. 1b. 

6) While I appreciate the use of five distinct conditional probability models, I would like to advocate for subjecting decision tree-based algorithms to feature importance scoring as well. To achieve this, I strongly recommend incorporating permutation feature importance analysis as an additional experiment. Such an approach would provide valuable insights and augment the comprehensiveness of the study.

Recommended references:

10.1016/j.asoc.2023.110324

10.1016/j.soildyn.2023.107994

10.1007/s00477-023-02392-6

10.3390/land12051018

10.1016/j.jenvman.2023.117357

10.1080/19475705.2023.2213807

10.1016/j.catena.2022.106379

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The author has addressed my concerns and the article can be accepted. In addition, please carefully double-check that the references correspond to the text.

Moderate editing of the English language is required.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

Many thanks for extending the invitation to re-evaluate the revised manuscript. I must commend the authors for their diligent efforts in significantly enhancing the paper's comprehensiveness, inclusivity, and informativeness. It's evident that they have taken substantial measures to address the concerns I previously raised. At this juncture, I find myself without further remarks, thereby granting my endorsement for the paper to progress through the peer-review process.

However, I do wish to bring to the authors' attention the importance of adhering to proper proofreading protocols. Specifically, I note that in-text citations should exclusively feature surnames. Additionally, I observed that certain references involve first names rather than surnames. Lastly, please do not forget to remove ® icons on figures. Wishing the authors the best as they finalize this work.

Author Response

Please see the attachment.

Author Response File: Author Response.docx