GIS-Based Landslide Susceptibility Modeling: A Comparison between Best-First Decision Tree and Its Two Ensembles (BagBFT and RFBFT)
Round 1
Reviewer 1 Report
In this manuscript, the best-first decision tree (BFT) and its two ensembles, namely the bagging best-first decision tree (BagBFT) and forest best-first decision tree (RFBFT), were applied to produce landslide susceptibility maps for the city of Zhashui County, China. To this end, the authors initially identified 169 historical landslides and a total of 15 landslide conditioning factors were later taken into consideration. The results revealed that the RFBFT (AUC = 0.895) had the highest predictive performance compared with the others. The application of ensemble learning strategy to increase the accuracy of the resultant maps but the motivation of the manuscript is not convincing. There already exist recent studies applying those techniques, questioning the novelty of the manuscript. I expect to see a novelty in a manuscript dealing with landslide susceptibility because a vast amount of case studies exist on this topic. The result of this manuscript might only be valid for the study area under investigation. It is possible that different results and more effective techniques could be obtained in other locations, depending on site-specific conditions and the quality of the landslide inventory. Additionally, the introduction and results sections contain several potentially arguable statements that may need further explanation and discussion. Given these concerns, I believe that the manuscript in its current form needs major revision to be published in Remote Sensing. My main concerns related to the manuscript are given below:
· The concluding remarks of the abstract are well-written but the abstract should include a statement addressing the novelty of the study.
· The introduction should clearly explain the necessity and originality of the manuscript.
· Please add the “y” letter for “neuro-fuzz(y)” on line 65.
· A literature survey on ensemble learning algorithms in landslide susceptibility mapping practices, and their limitations, as well as the superiority of the proposed model (i.e., BFT), should be included in the manuscript. Please consider the recent literature on more advanced boosting algorithms (gradient boosting, XGBoost, CatBoost etc.) and refer to recent studies.
· The novelty of the work should be explained more clearly in the last paragraph of the introduction. The infrequent use of BFT alone does not always indicate the originality of the study and this paragraph should be restructured accordingly.
· The word “zone” should be replaced with “region” on lines 91-92.
· The authors should provide more explanation on how they identified the historical landslides.
· The NDVI in Figure 2 should be clarified as to whether it is a multi-temporal or single event, and if it is multi-temporal, the pre-processing step to reduce spectral variations caused by different atmospheric and geometric conditions should be provided.
· Referring to the previous question, the year of the NDVI used in the manuscript, and whether the landslide data corresponds to the NDVI data, should be clarified. More information about this issue should be provided.
· The resolution/scale of the thematic inputs should be provided, as well as clear definitions of the sources of these geospatial data.
· The methodology used to produce the LULC map, including the number of samples collected for each class, the accuracy of the produced thematic map, and the test/train ratio, should be explained.,
· The methodology adopted for the selection of non-landslide susceptibility maps should be explained in detail, as this directly affects the quality of the produced maps. Additionally, the geographical location of these maps should be illustrated in Figure 1.,
· The manuscript should include a discussion of hyperparameter tuning optimization to optimize the hyperparameters of machine learning algorithms and the lower and upper limits of the hyperparameters should be included, as they play a critical role in the quality of the susceptibility maps produced. Recent studies on parametrization of hyperparameters for machine learning methods in landslide susceptibility mapping should be cited.
· The geographical location of both landslide and non-landslide bodies with their training and testing samples, as well as the local meteorological stations, should be presented in Figures 1 and 2m respectively. Information about the number of stations used to extract rainfall data and the interpolation method applied should be provided.
· A statistical significance test, such as McNemar’s test or Wilcoxon signed-rank test, should be applied to properly analyze the predictive performance differences between the machine learning algorithms.
· The discretization method used to map landslide susceptibility should be provided.
· The discussion section should include further exploration, method comparison, a summary of results, and a discussion of study limitations.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
The manuscript entitled “GIS-based landslide susceptibility modeling: A comparison between best-first decision tree and its two ensembles (BagBFT and RFBFT)”, by J. Gui, L.R. Alejano, M. Yao, F. Zhao and W. Chen, presents an interesting work.
In general, the manuscript should be acceptable for publication but some serious problems must be repaired prior to publication. It needs some significant improvement. Some suggestions are as follows:
- Please use different terms in the “Title” and the “Keywords”.
- The abstract should state briefly the purpose of the research, the principal results and major conclusions. An abstract is often presented separately from the article, so it must be able to stand alone.
- It would be useful to be described the aim of this paper.
- The English language usage should be checked by a fluent English speaker. It is suggested to the authors to take the assistance of someone with English as mother tongue.
- The buffers in factors “Distance to faults” are so wide. How does a fault cause a landslide at 1 or 2 km away? Please justify. I think that it is wrong.
- The same with and “Distance to roads”.
- The same with and “Distance to rivers”.
- Please justify convincingly why this manuscript (method, thematology etc) connected with RemoteSensing’s content and scope. Perhaps the using of proper from this journal literature would be helpful.
- The authors could make discussion about the relationship between landslide assessment and planning. You could see and use the following publication: “Natural and Technological Hazards in Urban Areas: Assessment, Planning and Solutions” Sustainability 13: 8301.
- When you are using coordinates, please do not use “North Arrow”. This is a mistake in cartography.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
The revisions made by the authors to the manuscript have addressed my previous concerns and I believe the paper is now ready for publication. The quality of the work and the clarity of the writing have improved significantly. I appreciate the authors' attention to detail and dedication to improving the paper.
Reviewer 2 Report
This manuscript presents an improved work.
The manuscript should be acceptable for publication as it is.