Review Reports - Comparative Assessment of Quantitative Landslide Susceptibility Mapping Using Feature Selection Techniques

Round 1

Reviewer 1 Report (Previous Reviewer 4)

Comments and Suggestions for Authors

This manuscript aims to systematically compare several feature selection strategies with traditional statistical models for landslide susceptibility mapping. Although the manuscript is generally well organized and the data sources and evaluation metrics are clearly described, the revised version still remains largely at the level of empirical performance comparison. In its current form, the study does not sufficiently articulate a clear scientific problem, nor does it convincingly demonstrate how the reported comparisons advance mechanistic understanding or methodological rigor beyond existing work. As a result, several fundamental issues related to the formulation of the research question, the methodological comparability of the experimental design, and the physical interpretability of the results continue to limit the scientific depth and generalizability of the conclusions, and require substantial further clarification.

Major comments

(1) In the Introduction, the manuscript primarily emphasizes performance improvements of traditional models through feature selection, but it does not clearly articulate a scientific research question or a testable hypothesis. In particular, the authors do not address whether different feature selection techniques operate consistently under common geomorphic and hydrological constraints, or whether they may introduce systematic physical bias. Clarifying this issue is important for moving the study beyond an empirical engineering comparison toward a mechanism-informed analysis.

(2) The results indicate that the correlation-based logistic regression model achieves the highest predictive performance; however, this conclusion is interpreted almost exclusively through statistical metrics such as AUC and recall. The manuscript does not sufficiently examine whether the resulting susceptibility patterns remain consistent with established landslide triggering and control mechanisms. For example, excluding lithology and soil due to weak statistical correlation may reduce the model’s ability to represent structurally weak zones or soft rock formations at a mountainous scale. It also remains unclear whether correlation-based screening implicitly assumes monotonic and linear controls on landslide occurrence.

(3) The study compares Correlation, VIF, and Information Gain feature selection methods in combination with WO, MLR, and LR models, yet the retained factor sets differ considerably in both size and information content. Under these conditions, it is difficult to distinguish whether performance differences among models arise from the selection strategy itself or simply from differences in input dimensionality and data richness. The authors should clarify whether such uneven experimental conditions allow for a strictly methodologically sound comparison, or alternatively provide additional control analyses or discussion.

(4) Correlation analysis is reported as the most effective screening approach, but it inherently prioritizes linear, pairwise relationships. In the context of landslides, which are controlled by strongly coupled geomorphic and hydrological processes, the removal of physically meaningful but weakly correlated variables such as lithology or soil may limit the mechanistic interpretability of the model. Consequently, the designation of a “best” model based solely on predictive performance remains insufficiently supported from a physical-consistency perspective.

(5) VIF and Information Gain address different statistical aspects of feature selection, with VIF controlling linear multicollinearity and IG ranking factors based on sample-level entropy reduction. Neither method explicitly considers spatial autocorrelation or geomorphic continuity, both of which are fundamental characteristics of landslide processes. The manuscript should more clearly discuss the applicability limits of these techniques in spatially structured landslide problems and the potential for misinterpreting statistically important variables that lack clear physical meaning.

(6) The strategy used to construct landslide and non-landslide samples, as well as the spatial constraints applied during sampling, is insufficiently described. Moreover, model evaluation relies on a single 70/30 random data split, without examining the stability of results under alternative partitions. The authors are encouraged to clarify whether the reported model rankings remain robust under spatially structured validation or alternative cross-validation schemes, as this directly affects the generalizability of the conclusions.

(7) Model performance is assessed almost entirely through global metrics such as AUC, recall, and precision, while limited attention is given to whether high-susceptibility zones are physically reasonable in terms of slope gradient, hydrological convergence, or vegetation conditions. Without such checks, statistical optimality does not necessarily translate into suitability for hazard or risk-related applications. Additional discussion on physical plausibility would strengthen the interpretation of the results.

(8) The Introduction focuses mainly on statistical and machine learning approaches and provides limited coverage of recent developments in probabilistic physical modeling and physics-informed or hybrid susceptibility frameworks. Representative studies that integrate uncertainty propagation, physical constraints, or probabilistic stability concepts at regional scales should be incorporated, and the relationship between the present work and these approaches should be more clearly articulated. Without this context, the position of the study within the broader landslide susceptibility research landscape remains ambiguous.

(9) Lastly, this study provides a clear and systematic comparison of feature selection strategies and traditional statistical models for landslide susceptibility mapping, contributing to improved methodological transparency. In recent years, probabilistic physical and physics-informed approaches have attracted increasing attention, as they allow uncertainty in geotechnical, hydrological, and triggering factors to be explicitly considered, offering a useful complement to purely data-driven susceptibility models (e.g., 10.1007/s11440-024-02384-y; 10.1016/j.jrmge.2024.08.005). Consideration of these developments in future work may further enhance the physical interpretability and practical applicability of susceptibility assessment results at regional scales.

Author Response

Dear Reviewer:

The response to reviewer comments/suggestions are as in attached PDF file.

Thank you

Author Response File: Author Response.pdf

Reviewer 2 Report (New Reviewer)

Comments and Suggestions for Authors

This manuscript entitled “Comparative Assessment of Quantitative Landslide Susceptibility Mapping Using Feature Selection Techniques” presents a timely and methodologically sound comparative study on the integration of feature selection techniques with traditional landslide susceptibility models. The research addresses a recognized gap in the systematic evaluation of these preprocessing steps. The experimental design is comprehensive, and the practical recommendations provided are a significant strength. However, the manuscript requires substantial revisions, particularly concerning methodological transparency, depth of analysis, and the quality of graphical presentations, before it can be considered for publication.

Specific Comments：

The procedure for splitting the landslide and non-landslide inventory into the 70% training and 30% testing sets is not described. It is critical to specify whether this was a simple random split, a spatial split, or a stratified sampling approach, as this directly impacts the generalizability of the reported performance metrics.
The methodology for deriving the final factor weights for the Weighted Overlay (WO) model from the three feature selection techniques (Correlation, VIF, IG) is unclear.
The formula for Pearson's correlation coefficient presented on page 9 is incorrect and must be replaced with the standard form.
While the LR model with correlation analysis is identified as the best performer, its AUC-ROC of 75.48% and accuracy of 69.30% are modest. The manuscript would benefit from a discussion on the practical sufficiency of this performance level for real-world planning and the critical trade-off between this interpretable performance and the potentially higher, yet less interpretable, performance of advanced machine learning models mentioned in the introduction.
The validation strategy relies on a random split of the data, which does not account for spatial autocorrelation. The potential for model overfitting and its performance in spatially unknown areas remains unaddressed. A discussion on this spatial transferability, or the use of a spatial cross-validation scheme, is strongly recommended.
In Figure 2, the label "Cases:Model Selection" is repeated and ambiguous. It should be explicitly relabeled to correspond to the three feature selection cases (e.g., "Case 1: Correlation", "Case 2: VIF", "Case 3: IG") for clarity.
The description below Table 8 references "Prediction Range" and "Mean Prediction Value," but these columns are absent from the table itself. The table must be updated to include all discussed metrics.
The criterion for retaining one variable from a highly correlated pair (|r| > 0.7) is not specified. The authors must state whether the decision was based on the variable's higher correlation with the inventory, its presumed physical significance, or another objective rule.
The procedure for the VIF analysis requires elaboration. It should be clarified whether the VIF for each factor was calculated using a model containing all 15 initial factors, and if an iterative elimination process was used, the specific steps (e.g., remove the highest VIF, recalculate, repeat) need to be detailed.
The discretization method (binning) applied to continuous causative factors prior to the Information Gain calculation is a critical step that is not mentioned. The chosen method (e.g., equal interval, quantiles, natural breaks) and its parameters must be described.
The sampling strategy for generating non-landslide points is not detailed. The manuscript must specify how these points were selected (e.g., random sampling, stratified by a stability criterion), what measures were taken to ensure they represent stable conditions, and how spatial balance with landslide points was considered.
The classification scheme used to convert the continuous susceptibility indices (0-1) into the five susceptibility classes is not defined. The method for determining the class breakpoints (e.g., natural breaks, quantiles, equal interval) must be explicitly stated.
Figure 1a does not adequately portray the geological features of the study area as described in the caption. This subfigure should be enhanced to include key geographical elements like rivers, roads, settlements, and a clear visual representation of the different geological units.
Figure 3, which displays the causative factors, is currently a simple array of maps. Its analytical value would be significantly increased by incorporating statistical diagrams (e.g., histograms, boxplots) for each factor, comparing its distribution in landslide versus non-landslide areas.
The Landslide Susceptibility Maps (LSMs) in Figure 4 would be more informative if accompanied by a bar chart quantifying the areal percentage of each susceptibility class for each model. This would facilitate a direct comparison of the models' inherent risk perceptions.
Beyond the recall metric, the spatial predictive capability of the models should be evaluated by analyzing the density of known landslide points within each predicted susceptibility class. A chart or table displaying this density distribution for the final maps is highly recommended.
The provided ROC curves (Figure 5) represent only the test set performance. Including the training set ROC curves (e.g., as dashed lines) would allow for a visual assessment of potential overfitting or underfitting in the models.
The results of the feature selection analyses would be more accessible if presented as dedicated feature importance bar charts for the Correlation, VIF, and IG methods, visually ranking the factors by their selected scores or weights.
A comparative scatterplot of the susceptibility indices from the top-performing models (e.g., LR-Correlation vs. WO-VIF) would be a powerful tool for visualizing the agreement and divergence between different methodological approaches.
The overall quality and professionalism of the figures fall below the standard expected for a scientific publication. A comprehensive revision of all figures is required to improve their clarity, layout, and information density, aligning with the quality seen such as [Reference DOI: 10.3390/rs17122083].

Author Response

Dear Reviewer:

The response to reviewer comments/suggestions are as in attached PDF file.

Thank you

Author Response File: Author Response.pdf

Reviewer 3 Report (New Reviewer)

Comments and Suggestions for Authors

The manuscript presents an interesting topic and has potential for publication; however, several substantial issues must be addressed before the work can be considered for acceptance. Many concerns relate to data accuracy, unclear preprocessing, and incomplete methodological descriptions. At present, the conclusions of the study are not fully supported due to uncertainties in the input data and spatial analyses.

1. Format, style, and mapping standards

(a) Line 84: The abbreviation “AUC” is not fully spelled out at first occurrence. Please check all abbreviation in the manuscript.

(b) Figure 1: The geological map does not follow standard geological mapping rules, including:

Appropriate color schemes
Stratigraphic order from youngest to oldest
Clear and readable legend
Source citation and map scale
Consistency in symbolization

Currently, it is difficult to interpret the stratigraphy and spatial relationships. Please revise using proper geological conventions.

2. Section 2.2 questions about input data quality

(a) Spatial resolution issues:

The datasets have inconsistent spatial resolutions (10 m Sentinel-2A vs. 30 m DEM), yet the manuscript does not describe how these were harmonized.

Did the authors resample?
If so, which method (nearest neighbor, bilinear, cubic)?
Why not use a higher-resolution DEM to match Sentinel-2A when computing TWI and SPI?

This issue may strongly affect model reliability.

(b) Missing dataset dates:

The manuscript does not provide acquisition dates for any of the datasets. Since different environmental factors vary seasonally (vegetation, soil moisture, rainfall), the absence of temporal information raises concerns about dataset compatibility.

The source, compilation method, and date of the landslide inventory are not provided. Besides, the inventory appears to consist of points only, not polygon outlines.

For susceptibility modeling, polygon inventories are standard because they represent actual landslide extents.
If points were used, how was spatial mismatch handled across different resolutions?
Figure 2 indicates that 380 landslide points were used in the analysis; however, Figure 1 does not clearly show or confirm the same number of points.

This omission makes it difficult to assess the validity of the model training and validation.

3. Section 2.3 about landslide susceptibility mapping (LSM)

(a) Figure 2 layout: In the satellite datasets block, two items should be switched to avoid crossing arrows. This will improve readability.

(b) Selection of 15 causative factors:

Line 166: The manuscript does not justify choosing exactly 15 factors.
The original feature distribution for all factors should be included in the Appendix to confirm data quality.
Classifications into five levels require explanation: (i) What method was used? (natural breaks, quantiles, equal interval, expert-based?); (ii) What geological or geomorphological rationale supports these thresholds?

No lineament map is shown in Figure 1.
No dataset source or extraction method is provided (manual digitization, automatic extraction, etc.).

This makes interpretation of Figure 3(o) impossible.

(d) Geological and soil map accuracy concerns:

A key issue is the distribution of Qs (Quaternary Sediments). In Figure 1(b), Qs is shown near 27°15'N, 85°50'E, but additional Qs units occur around 27°10'N, 86°05'E (Bhimasthan). However, Figure 3(h) does not show high susceptibility in the Bhimasthan area. This mismatch indicates the geological map used may be too coarse or inaccurate, which could response to the near-zero Pearson correlation for geology in Table 3 and the model filtered out geology factor (Line 466). Similar concerns apply to the soil map, particularly regarding its resolution, accuracy, and consistency with other datasets.

(e) Unclear spatial and temporal consistency:

Since both landslide inventory and environmental factors lack precise dates, it is unclear whether:

Landslides occurred before or after the input dataset dates
The environmental conditions at the time of failure were accurately represented

This temporal ambiguity reduces confidence in model training and validation.

4. Results, Discussion and Conclusion

After the aforementioned issues are resolved through stronger supporting evidence and clearer methodological explanations, the key findings (Lines 458, 484, and 508) will be substantially strengthened. It is also recommended that the authors clearly articulate the implications for further applications (Line 509).

Author Response

Dear Reviewer:

The response to reviewer comments/suggestions are as in attached PDF file.

Thank you

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report (Previous Reviewer 4)

Comments and Suggestions for Authors

Thanks a lot for authors' effort, they have addressed all my issues.

Author Response

Reviewer comments:

Reviewer 1: Quality of English: Improvement suggested

Author Response: Thanks very much for the sugggestions to improve the langauge. We have tried our best to imrpove the language and corrected the text substantially. We hope it has become quite easy to read now.

Thank you.

Reviewer 2 Report (New Reviewer)

Comments and Suggestions for Authors

The authors have done a commendable job in addressing the previous critiques and strengthening the manuscript, which now presents a much more solid and compelling study.

Author Response

Reviewer 2 Comments:

1. Does the introduction provide sufficient background and include all relevant references? >> Can be improved

Author Response: Thanks very much. Some minor corrections have been incorporated and some more reference papers have been added.

2. Is the research design appropriate? >> Can be improved.

Author Response: Minor change has been incorporated.

3. Are the methods adequately described? >> Must be improved

Author Response: It has been improved and some minor revision has also been adopted.

4. Are the results clearly presented? >> Must be improved

Author Response: The result section has clearly reviewed and improved.

5. Are the conclusions supported by the results? >> Must be improved

Author Response: The conclusion text has been modified as per the modification of the result and discussion.

6. Are all figures and tables clear and well-presented? >> Must be improved.

Author Response: Figure 4, Figure 5, Figure 7, Figure 8, and Figure 8 have been modified and the text, color and other necessary changes have been made.

Thanks very much.

Reviewer 3 Report (New Reviewer)

Comments and Suggestions for Authors

1. Section 2.2 Data and Software Employed

The derivation of secondary datasets should be described more clearly. For example, Sentinel-2A imagery from the 2019–2020 dry season was used to compute vegetation and built-up indices (MNDVI, ARVI, NDBI) (Line 162), and river networks were derived from the DEM (Line 168). Please clarify the specific workflow used to extract the drainage network (e.g., flow accumulation thresholds), as this directly influences the distance to river factor and thus the overall model results. It is recommended that these procedures be documented in detail, for example in an appendix.

2. Data sources and citations

Please clearly declare the data sources, temporal coverage, and spatial resolution of all input datasets (e.g., DEM, satellite imagery). In particular:

Line 147: DMG Nepal (2020), which the geological data source is not cited. Please specify the source (e.g., DMG Nepal, year, map scale/version) and note that Figure 1 appears inconsistent with the official 1:1,000,000 geological map available from DMG Nepal.

(https://dmgnepal.gov.np/en/pages/general-geology-4128).

Line 168: OpenStreetMap (2020) is used but without the citation or data website. Please check and complete similar missing citations throughout the manuscript.

3. Lineaments vs. geological structures (Line 166, Figure 3o)

The manuscript states that lineaments were semi-automatically extracted from DEM-derived hillshade and refined manually [31]. However, the variable “Distance to lineaments” shown in Figure 3(o) appears to represent distance to a mapped major geological structure (a thrust) within the Nuwakot Group (Na), rather than DEM-extracted lineaments. These features are conceptually different.

If mapped thrusts or faults were used, the factor should be renamed to “distance to major geological structures”, and the data source should be clearly specified. In addition, the orientation of the northern thrust appears inconsistent with regional tectonic interpretations. Please verify this using authoritative sources (e.g., DMG geological maps or http://dx.doi.org/10.1016/j.tecto.2013.06.006).

4. Justification of causative factor selection

In response to earlier comments, the revised manuscript states that 15 causative factors were selected based on literature review, regional relevance, and data availability (Line 200). However, this explanation remains insufficient. Please provide appropriate citations and clarify why these specific factors are suitable for the study area, and how this selection strategy can be transferred or adapted to other regions.

5. Grammar and style (Line 708)

Grammar correction suggested: “Consequently, practitioners are encouraged to complement statistical model selection with domain expertise and field-based validation to ensure both predictive accuracy and mechanistic credibility.”

6. Author information

The fourth author’s name should be given in full.

7. Statistical notation consistency

Please ensure consistent notation for Pearson correlation coefficients throughout the manuscript:

Use r consistently (e.g., Line 359, Line 423: r = 0.316).
Replace “Pearson” or “Pearson coefficients” with “Pearson correlation coefficient (r)” where appropriate (Lines 435, 665).
In Figure 5(a), the x-axis title should be “Absolute Pearson correlation coefficient.”
Symbols p and r should be italicized throughout.

8. Weights vs. correlation values (Line 469)

The statement “Curvature (0.940) and Aspect (0.932)” appears to refer to model weights rather than correlation coefficients. Please clarify this by using notation such as Curvature (w = 0.940) and Aspect (w = 0.932).

9. Redundant abbreviations

Once an abbreviation has been defined, it does not need to be redefined or repeatedly expanded (e.g., LSMs in Line 494).

10. Figure–statistics consistency

The legend colors in Figures 6, 7, and 9 are not consistent. In particular, Figure 6 (IG-LR) visually suggests a larger area classified as moderate to very high susceptibility, whereas Figure 7 reports relatively low percentages (29.0%, 18.9%, and 7.2%). Please verify the classification, legends, and statistics. Additionally, Line 569 states that the correlation-LR combination performed best; this should be clearly supported by the figures and quantitative results.

11. Line 631: The interpretation of WO–VIF as suitable for high-sensitivity applications is reasonable given its high recall. However, please clarify that “sensitivity” (Lines 486 and 633) refers to recall (true positive rate) rather than a formal sensitivity analysis, to avoid methodological ambiguity. In addition, since each model applies different feature-selection strategies to the same set of 15 causative factors, it would be helpful to briefly clarify that the reported “sensitivity” reflects model performance (recall) rather than sensitivity to individual input variables.

Author Response

Author response to the reviewer comments is as attached.

Author Response File: Author Response.pdf

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

Comments and Suggestions for Authors

The main idea of the under-review manuscript is the implementation of feature selection methods to improve model efficiency and produce stronger forecasts, eliminating in parallel the process of multicollinearity in landslide susceptibility mapping.

My general comment is that this manuscript is associated with the proposed themes of the Special Issue entitled "Advances in Remote Sensing and GIS for Natural Hazards Monitoring and Management”. The manuscript is well designed (e.g., Introduction, Geological Setting, Material and Methods, Results, Discussion, and Conclusions), but on the other hand, many of the issues described, in my opinion, are already known in the broader landslide scientific literature and community.

As a result, I expect the authors to add a more distinguishing writing for the analysis of their idea, so as to add value to their work and make them different from the already knowledge about the estimation and generation of a landslide susceptibility mapping procedure. To this end, reading carefully the manuscript, some points need to be further clarified, rewritten and analyzed. Thus, for this manuscript being able to be published, the following remarks should be answered satisfactorily by the authors.

Major remarks:

The Discussion Section should be further analyzed by comparing the results with the (possible) similar findings from other researchers. Furthermore, the authors should examine possible drawbacks of the feature selection methods when they implement them in their case study area.
I would like the authors to explain and analyze better the last two paragraphs of the Conclusion Section.

Minor remarks:

Line 163: Make clearer (more visible) the contents of Figure 3.
Line 432: Enlarge the contents of Figure 4. As it is depicted now, it is hardly readable.
I would like to see more recent references. The majority of them are from 2016 and downwards.

Reviewer 2 Report

Comments and Suggestions for Authors

This study primarily utilizes VIF and IG for factor selection and evaluates landslide susceptibility through models like LR, serving as a case study. The following key issues exist:

Unclear research background in the abstract, including a lack of explanation for why this study was conducted and the existing scientific questions in current landslide susceptibility methods.
Incomplete research progress in landslide susceptibility evaluation methods in the introduction. While it is undeniable that methods like LR face collinearity issues among factors, solutions such as tolerance and VIF-based judgments or principal component analysis methods have already been considered during model construction. Additionally, machine learning-based models have achieved significant success in landslide susceptibility evaluation, particularly in addressing collinearity.
VIF has long been used for collinearity diagnosis, which is a fundamental criterion. Where is the improvement in this paper? IG is generally applied to assess factor importance—how does this study address collinearity?
In the list of factors, such as roads, while they can alter surface hydrological processes, generally speaking, only those distributed on or near the landslide body (at the toe of the slope) will significantly affect the stability of the landslide. Otherwise, they will not have a critical impact on the stability of the landslide and should not be considered as variables. In this study, the maximum distance between landslides and roads reaches 13,606 meters, making it difficult to explain the role of roads in landslide stability from a theoretical perspective. Therefore, in the statistical data of this study, what is the distance between most landslides and roads or rivers? Finally, it is worth noting that in most articles, these two factors are used as indicators for susceptibility evaluation, but in fact, from a mechanistic perspective, this approach has serious flaws.
The discussion should emphasize collinearity resolution methods and whether factor selection must entirely eliminate collinearity. Some collinearity between factors (e.g., rainfall and lithology) may be mechanistically reasonable for landslide mechanisms. Excluding one via data-driven algorithms could be inappropriate.
Overly complex conclusions; key data features and limitations need refinement. The authors highlight VIF and IG-based factor selection as the study’s main innovation—this requires cautious summarization without exaggeration.

Reviewer 3 Report

Comments and Suggestions for Authors

This manuscript implemented Weighted Overlay, Linear Regression and Logistic Regression methods combined with feature evaluation techniques for landslide susceptibility mapping. After reviewing this manuscript, I recommend rejection due to the limited novelty of the approach. The use of statistical models and simple information gain feature evaluation methods for landslide susceptibility prediction is already well established in the literature. A simple search using keywords such as "landslide susceptibility Logistic Regression Information Gain" on platforms like Google Scholar yields numerous relevant publications. If we check the following list references, we can find that this integrated technique has already been used in landslide susceptibility mapping. As a result, the manuscript's primary contributions are difficult to distinguish, and it does not offer sufficient methodological innovation or advancement over existing studies.

[1] Chen W, Peng J, Hong H, et al. Landslide susceptibility modelling using GIS-based machine learning techniques for Chongren County, Jiangxi Province, China[J]. Science of the total environment, 2018, 626: 1121-1135.

[2] Du G, Zhang Y, Iqbal J, et al. Landslide susceptibility mapping using an integrated model of information value method and logistic regression in the Bailongjiang watershed, Gansu Province, China[J]. Journal of Mountain Science, 2017, 14(2): 249-268.

[3] Abedini M, Ghasemian B, Shirzadi A, et al. A comparative study of support vector machine and logistic model tree classifiers for shallow landslide susceptibility modeling[J]. Environmental Earth Sciences, 2019, 78(18): 560.

[4] Chen T, Niu R, Jia X. A comparison of information value and logistic regression models in landslide susceptibility mapping by using GIS[J]. Environmental Earth Sciences, 2016, 75(10): 867.

[5] Li X, Chong J, Lu Y, et al. Application of information gain in the selection of factors for regional slope stability evaluation[J]. Bulletin of Engineering Geology and the Environment, 2022, 81(11): 470.

Reviewer 4 Report

Comments and Suggestions for Authors

This manuscript presents a comparative assessment of landslide susceptibility mapping (LSM) in Sindhuli District, Nepal, using three conventional models (Weighted Overlay, Multiple Linear Regression, Logistic Regression) combined with three feature selection methods (Correlation Analysis, Variance Inflation Factor, Information Gain). The study collects a comprehensive dataset (DEM, soil, geology, rainfall, remote sensing) and systematically evaluates the predictive performance of different model–feature selection combinations. The results are clearly presented, and the recommendations for model selection under different conditions are practically valuable. However, the manuscript requires improvements in terms of scientific contribution, methodological justification, and result interpretation before it can be considered for publication. I therefore recommend major revision.

Major Comments

1.The study is mainly based on traditional statistical models (WO, MLR, LR). While the combination with feature selection methods is useful, the level of novelty is somewhat limited. Recent advances in machine learning and deep learning (e.g., RF, XGBoost, SVM, ANN) are not discussed or compared, which weakens the positioning of the paper. The authors should clarify more explicitly how their work advances the field beyond existing literature.

2.The manuscript applies correlation analysis, VIF, and IG, but it is unclear why these three methods were chosen over other widely used approaches (e.g., PCA, LASSO). A stronger justification of the selection criteria and their suitability for LSM is needed.

3.The performance results (accuracy, precision, recall, AUC) are reported, but the interpretation is superficial. For example, why WO+VIF achieves very high recall while LR+correlation analysis provides a more balanced performance is not sufficiently explained. A deeper discussion of the mechanisms behind these differences would strengthen the paper.

4.No sensitivity or robustness analysis is presented. Given the variability of landslide data, the effect of sample imbalance, random data splits, or different thresholds on model performance should be addressed to enhance the credibility of the conclusions.

5.Several figures (e.g., Figure 3, Figure 4) are dense and hard to interpret, and some tables (e.g., Tables 2, 3, 7) contain excessive detail without clear emphasis. The presentation should be improved by simplifying figures, highlighting key results, or using graphical summaries for performance comparison.

Minor Comments

1.The manuscript would benefit from careful editing to reduce redundancy and improve readability. Several sentences are overly long or repetitive.

2.The reference list is dominated by older sources. More recent studies (past 3–5 years) on feature selection and landslide modeling should be cited to better situate this work in the current research context.

3.Performance metrics (e.g., AUC-ROC, F1-score) are written inconsistently across the manuscript. Please standardize terminology and formatting throughout.

4.Although the authors state that data will be available upon request, sharing core datasets or code (e.g., through a public repository) would significantly improve reproducibility.

5.The conclusion repeats parts of the discussion and could be more concise. It should emphasize the main contributions, practical implications, and limitations rather than restating results.