Eigenvector Spatial Filtering-Based Logistic Regression for Landslide Susceptibility Assessment

Round 1

Reviewer 1 Report

In this study the Authors implemented eigenvector spatial filtering (ESF) to account for spatial autocorrelation into logistic regression for landslide susceptibility assessment. The study has been carried out on the landslide inventory map and 11 landslide predisposing factors, developing three different models: the eigenvector spatial filtering based logistic regression (ESFLR), the conventional logistic regression (LR) and the autologistic regression (ALR). The three models were evaluated, by means of ROC curve, and compared in terms of prediction capability.

I think this has the potential to be a good contribution to International Journal of Geo-information. However, as is the writing is unclear in many places. First of all I think that some factors have been extracted from a too different scales (i.e. 1:2500000, 1:500000….) in order to be used in the models. A geological and geomorphological setting, in particular a landslides background, is completely absent in terms of typology, which is essential it the assessment of landslide susceptibility. Are the Authors committed that all predisposing factors have to be used? Most of them I think that are not significantly due to their small number (i.e. distance from railway, distance from faults, distance from river) In particular, I should focus the attention on precipitation: what the Authors wants to represent with Precipitation map (Fig. 2i)? Furthermore, which type of curvature they consider, planform or profile?

Finally, all figures have to be improved in order to make them more readable.

That said, I feel that this paper would be accepted after major revision.

 

Some other comments:

Pg. 2 line 80: Which typology of landslides affect the area? Please specify the typology following a classification as Cruden and Varnes 1996 and so on.

Pg. 2 line 81: Which is the classification used to identify the dimension of the landslides (small, large….)?

Figure 1: Please improve the the readability of the figure. Please insert a sketch map of the region.

Figure 2: Please check the caption of the Figure (Figure 1). All figures must be enlarged to improve the readibility. Please insert the unit of measure for each legend (i.e. (m), (°)…..).

Pg. 6 line 116: Please explain the criterion used to do the Rainfall map (Fig. 2i). I highlight that the rainfall could be consider as triggering factor and not as predisposing factor.

Formula 1: Is there a reference for this formula? Whis is the meaning of r value?

Pg. 6 line 150: Please explain the meanings of “nine categories according to the international 150 standard (Fig.2j)”.

Pg. 8 lines 167-168: Please insert the references for TOL and VIF.

Pg. 10 line 213: please check the word “lager”.

Figure 5: usually the ROC plot is represented as Sensitivity vs 1-Specificity

Pg. 16 line 379: which is the criterion used to identify the susceptibility classes? Equal interval, natural breaks, …….

Table 11: I think that it is better that these results have to be showed as histograms in order to stress the validation.

 

English language revision is advised. It would likely help if a native English speaker could read-over once before submitting it again.

Author Response

Response to Reviewer 1 Comments Dear Reviewer, Thank you for your comments about our manuscript. We have carefully examined the comments and questions, and revised the manuscript accordingly. We would like to express our appreciation to your constructive comments and suggestions. Detailed corrections are listed point by point. Point 1: First of all, I think that some factors have been extracted from a too different scales (i.e. 1:2500000, 1:500000….) in order to be used in the models. Response 1: Thank you for you advise. Some data may be a bit rough due to the data accessibility, but for now, it is pity that more appropriate data is not available. We will continue to look for better data sources in further research. We have added this to our conclusion (Pg. 21 line 473-474). Point 2: A geological and geomorphological setting, in particular a landslides background, is completely absent in terms of typology, which is essential it the assessment of landslide susceptibility. Response 2: We fully agree with you on the importance of landslide types. Different types of landslides have different triggers. Unfortunately, however, the type information of landslides in the study area was not recorded due to irregularities in the data collection process. We have added this to the text (Pg. 2 line 80). Therefore, our study focused on exploring the relationship between landslides and a series of trigging factors based on the spatial distribution of landslide events. Point 3: Are the Authors committed that all predisposing factors have to be used? Most of them I think that are not significantly due to their small number (i.e. distance from railway, distance from faults, distance from river). Response 3: We think factors with small number may also make sense in landslide development process. If a factor is not significant enough, the backward stepwise regression method we adopted in the model construction process will help us to delete it and explore the optimal form of regression equation (Pg. 12 line 305). Point 4: In particular, I should focus the attention on precipitation: what the Authors wants to represent with Precipitation map (Fig. 2i)? Response 4: The precipitation map was generated with the daily rainfall data from 2005-2014. Annual precipitation, which reflects regional rainfall intensity and duration in a long term, is a frequently used conditioning factor for landslide research. Areas with more annual rainfall are more prone to landslides, that is why we take it as a factor. We have added this information to the text accordingly (Pg. 6 line 126-130). Point 5: Furthermore, which type of curvature they consider, planform or profile? Response 5: The curvature data used in the paper is general curvature, which can help to analyse the characterization of slope morphology and flow. We have added this information to the text accordingly (Pg. 5 line 111-113). Point 6: Finally, all figures have to be improved in order to make them more readable. Response 6: All figures were prepared at a sufficiently high resolution required by the journal, however, they became less readable due to automatic compression. We will provide a single zip file of all figures. Furthermore, we have enlarged all the figures appropriately and inserted the unit of measure for each legend accordingly. Point 7: Pg. 2 line 80: Which typology of landslides affect the area? Please specify the typology following a classification as Cruden and Varnes 1996 and so on. Response 7: As stated in response 2, the type information of landslides in the study area was not recorded due to irregularities in the data collection process (Pg. 2 line 80). Point 8: Pg. 2 line 81: Which is the classification used to identify the dimension of the landslides (small, large….)? Response 8: We divided the landslides into small-sized (< 50000 m2), medium-sized (50000 ~ 100000 m2) and large-sized (≥100000 m2) according to their covered area. We have added this to the text (Pg. 2 line 84). Point 9: Figure 1: Please improve the readability of the figure. Please insert a sketch map of the region. Response 9: The figures may become less readable due to automatic compression, so we will submit the high-resolution figure separately. And we have inserted the sketch map of the study area together with the original Figure 1(Pg. 3 line 89). Point 10: Figure 2: Please check the caption of the Figure (Figure 1). All figures must be enlarged to improve the readability. Please insert the unit of measure for each legend (i.e. (m), (°)…..). Response 10: We have revised the caption of the figure to “Figure 2”, and all figures have been enlarged appropriately, and the unit of measure for each legend have been added accordingly (Pg. 5 line 99). The figures may become less readable due to automatic compression, so we will submit all figures separately. Point 11: Pg. 6 line 116: Please explain the criterion used to do the Rainfall map (Fig. 2i). I highlight that the rainfall could be consider as triggering factor and not as predisposing factor. Response 11: The Chongqing Institute of Geology and Mineral Resources also provided the daily rainfall data of 1003 rainfall observation stations for the period of 2005-2014. The mean annual precipitation at 78 rainfall observation stations that within and nearby Wulong area was calculated and used to create the precipitation map (Fig.2i) of the whole study area using inverse distance interpolation method. We have added this information to the text accordingly (Pg. 6 line 130-134). We agree with you that rainfall is often the trigger for landslides. Budimir et al. (Budimir, Atkinson, & Lewis, 2015) indicated that rainfall is considered in the literature as both a conditioning factor (long-term indicators, e.g. annual precipitation) and a trigger factor (short-term indicator, e.g. rolling 24-h rainfall). The rainfall used in this study is a long-term data, so we think it is reasonable to be considered as a predisposing factor. Point 12: Formula 1: Is there a reference for this formula? What is the meaning of r value? Response 12: There is no reference for Formula 1. It is a self-defined formula of the landslide affected distance, namely the multiple of the landslide radius (Pg. 6 line 156). Considering different dimensions and value ranges of each factor, all factors was divided into 9 classes, and the frequency ratio (R value) for each class of each factor was calculated and extracted to samples for subsequent model construction. The R value of a certain class is the ratio of the percentage of landslide area covered and the percentage of area covered, defined as the Formula 2. R value takes the prior knowledge of landslide into account, the class with higher R value is more prone to landslide in the case. We have added this information to the text accordingly (Pg. 7 line 175-178). Point 13: Pg. 6 line 150: Please explain the meanings of “nine categories according to the international standard (Fig.2j)”. Response 13: We've explained the nine categories of lithology in detail (Pg. 7 line 166). Point 14: Pg. 8 lines 167-168: Please insert the references for TOL and VIF. Response 14: We have added the reference accordingly (Pg. 9 line 186). Point 15: Pg. 10 line 213: please check the word “lager”. Response 15: We have revised it to “greater” (Pg. 10 line 231). Point 16: Figure 5: usually the ROC plot is represented as Sensitivity vs 1-Specificity. Response 16: Thank you for pointing out the mistake. We have revised the X-axis labels in Figure 5 to “1-Specificity” (Pg. 16 line 372). Point 17: Pg. 16 line 379: which is the criterion used to identify the susceptibility classes? Equal interval, natural breaks, ……. Response 17: This is a custom classification with reference to other scholars. Natural breaks which is a generally used classification method for statistical mapping may be more reliable for identifying the susceptibility classes. So base on natural breaks, we reclassified the landslide susceptibility values into 5 categories: very high, high, moderate, low and very low (Pg. 17 line 396). The landslide susceptibility maps and statistical results were changed and modified in the paper. Point 18: Table 11: I think that it is better that these results have to be showed as histograms in order to stress the validation. Response 18: We have shown the results as histograms, as shown in Figure 9 (Pg. 20 line 441). Point 19: English language revision is advised. It would likely help if a native English speaker could read-over once before submitting it again. Response 19: We have gone through the text and carefully edited English.

Author Response File: Author Response.docx

Reviewer 2 Report

Perhaps if you use colour coding in some tables could be easily read; also differences could be spotted easier (For example, Table 11: show differences Very High, etc. in all LR, ALR, ESFLR by using one specific background colour within a row. 

However in your conclusions, you should say clearly, if ESFLR model is a safe one to use. To my understanding you are still in progress of exploring more factors to be added to your model specification. Are these factors determined or found somewhere in your graphs? You may wish to pre-announce which methods could be more effective or put it in discussion with other scholars and researchers worldwide.

Author Response

Response to Reviewer 2 Comments Dear Reviewer, Thank you for your comments about our manuscript. We have carefully examined the comments and questions, and revised the manuscript accordingly. We would like to express our appreciation to your constructive comments and suggestions. Detailed corrections are listed point by point. Point 1: Perhaps if you use colour coding in some tables could be easily read; also differences could be spotted easier (For example, Table 11: show differences Very High, etc. in all LR, ALR, ESFLR by using one specific background colour within a row. Response 1: Thank you for you advise. As shown in Figure 9, we have shown the results as histograms (Pg. 20 line 441). We think it is also a good way to visually and clearly show the differences between the three models. Point 2: However, in your conclusions, you should say clearly, if ESFLR model is a safe one to use. Response 2: Based on the mathematical foundation, ESFLR is also a safe, reliable and stable method. We have added this in the conclusion (Pg. 21 line 465). Point 3: To my understanding you are still in progress of exploring more factors to be added to your model specification. Are these factors determined or found somewhere in your graphs? Response 3: We will explore factors other than those in our graphs, e.g. land use and rainy seasonal precipitation. Perhaps our original statement was misleading. We have rewritten it (Pg. 21 line 469). Point 4: You may wish to pre-announce which methods could be more effective or put it in discussion with other scholars and researchers worldwide. Response 4: Considering the limitation of the ESFLR method for large landslide data due to the computationally demand, we consulted with other scholars and thought segmented processing approach or fast-ESF method might be helpful, which is what we will try in our next research. We have added this to our conclusion (Pg. 21 line 475-477).

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The paper can be published in the present form

Author Response

Dear reviewer:

Thank you for your recognition of our manuscript.