Machine Learning Models for the Spatial Prediction of Gully Erosion Susceptibility in the Piraí Drainage Basin, Paraíba Do Sul Middle Valley, Southeast Brazil

Round 1

Reviewer 1 Report

P3 Line 102

         Figure 1, and all figures and tables, enhancing their

 self-explanatory nature

P4 Line 130

         How to identify the characteristic of the gully with the depth > 0.5m?

Line 164,

         What is VIF and Tol?

P9 for equation (12), is that correct description?

P14 Figure 5, where are a), b), c), and d)?

The work thinks the lineament is the most important factor to influence the gully erosion. Please describe what the nature of lineament is, what characteristics and indicators can illustrate lineament?

quite good, needs a little improvement.

Author Response

Dear, Reviewer 1

Thank you for your suggestions and comments to contribute to improve the article. See below the responses to the suggested comments:

Comment 1: P3 Line 102 - Figure 1, and all figures and tables, enhancing their self-explanatory nature

Response 1:  All figure and table labels have been modified to indicate their self-explanatory nature.

Comment 2: P4 Line 130 - How to identify the characteristic of the gully with the depth > 0.5m?

Response 2:  The requested explanation were incorporated in highlighted, in red, on P4 line 133.

Comments 3: P4 Line 164 - What is VIF and Tol?

Response 3: The requested explanations were incorporated highlighted, in red, on P4 lines 167,168 and 169.

Comment 4: P9 -  for equation (12), is that correct description?

Response 4:  The description of the equation is corrected, highlighted in red in line 404.

Comment 5: P14 Figure 5 -  where are a), b), c), and d)?

Response 5: Corrected

Comment 6:  The work thinks the lineament is the most important factor to influence the gully erosion. Please describe what the nature of lineament is, what characteristics and indicators can illustrate lineament?

Response 6: The explanation was improved between lines 97, 98 and 99, outlined in red.

Reviewer 2 Report

I think the study effectively addresses the global issue of soil erosion, focusing on gully erosion, with a clear explanation of the research objective, comparing machine learning models for susceptibility mapping. However, expanding the discussion on the rationale behind choosing specific models like CART, XGBoost, RF, and SVM would strengthen the methodology. The comparison of model performances using AUROC is well-presented, though a more detailed explanation of why XGBoost and Random Forest performed similarly, and the factors contributing to differences with SVM and CART, would add depth. I guess clarifying the interpretation of false positives in XGBoost versus Random Forest would further enhance the analysis. The identification of key controlling factors, such as lineaments, land use, slope, elevation, and rainfall, is valuable, but more discussion on how these factors interact and influence gully erosion, along with a brief literature review on similar studies, would provide better context. The replicability of the methodology is a strong point, but a more detailed description of how it can be adapted for other regions, along with potential limitations like data availability, would improve the practical relevance. The hypothesis validation is well done, though considering alternative approaches for testing and addressing potential biases in the sample set (n=159) would be beneficial. Adding detailed visual comparisons of spatial prediction maps could enhance reader understanding. Lastly, the conclusion could be expanded to explore broader implications, such as integrating models into land management strategies, and include recommendations for future research, such as incorporating more environmental variables or exploring new machine learning algorithms.

The language is generally clear, but minor revisions could improve the flow. I suggest a thorough review to catch any overlooked issues.

Author Response

Dear Reviewer 2

Thank you for your suggestions and comments to contribute to improve the article. See below the responses to the suggested comments:

Comment 1: However, expanding the discussion on the rationale behind choosing specific models like CART, XGBoost, RF, and SVM would strengthen the methodology.

Response 1: The explanations of this request were incorporated and can be found on page 7, in item 2.2.4. Machine Learning Models, highlighted in red, from lines 291 to 299.

Comment 2: The comparison of model performances using AUROC is well-presented, though a more detailed explanation of why XGBoost and Random Forest performed similarly, and the factors contributing to differences with SVM and CART, would add depth. 

Response 2: Explanations about these requests were incorporated and highlighted in red on page 18, from lines 609 to 615 and lines 621 to 623.

Comment 3: I guess clarifying the interpretation of false positives in XGBoost versus Random Forest would further enhance the analysis. 

Response 3: The explanation of these issues has been improved and highlighted in red, between pages 18 and 19, from lines 624 to 647.

Comment 4: The identification of key controlling factors, such as lineaments, land use, slope, elevation, and rainfall, is valuable, but more discussion on how these factors interact and influence gully erosion, along with a brief literature review on similar studies, would provide better context. 

Response 4: The explanation of this question was present in the article, from line 121 to 126.

“The settlement of Paraíba do Sul middle valley has been characterized by several economic cycles, such as coffee growing and its subsequent replacement by dairy farming in the XIXth century, promoting soil depletion and accelerated erosion on the slopes, due to changes in the regional hydroclimatic dynamics [24]. Concavity's structural control, associated with the subsurface hydrological dynamics and soil use and management, favour the occurrence of erosion processes [19]”.

Comment 5: The replicability of the methodology is a strong point, but a more detailed description of how it can be adapted for other regions, along with potential limitations like data availability, would improve the practical relevance.

Explanations about these questions were incorporated and can be found on page 7, highlighted in red from lines 270 to 275 and on page 19, from lines 648 to 654.

Comment 6: The hypothesis validation is well done, though considering alternative approaches for testing and addressing potential biases in the sample set (n=159) would be beneficial. Adding detailed visual comparisons of spatial prediction maps could enhance reader understanding. 

Response 6: The explanation of these questions were incorporated and can be found on page 19, lines 655 to 658.

Comment 7: Lastly, the conclusion could be expanded to explore broader implications, such as integrating models into land management strategies, and include recommendations for future research, such as incorporating more environmental variables or exploring new machine learning algorithms.

Response 7: As indicated, these suggestions were incorporated in the conclusion of the article, highlighted in red on lines 742 to 749.