Next Article in Journal
Leakage Effects from Reforestation: Estimating the Impact of Agricultural Displacement for Carbon Markets
Previous Article in Journal
Institutional Frameworks, Policies, and Land Data: Insights from Monitoring Land Governance and Tenure Security in the Context of Sustainable Development Goals in Kenya, Rwanda, Uganda, and Zambia
Previous Article in Special Issue
Spatio-Temporal Evolution and Conflict Diagnosis of Territorial Space in Mountainous–Flatland Areas from a Multi-Scale Perspective: A Case Study of the Central Yunnan Urban Agglomeration
 
 
Article
Peer-Review Record

Detection of Agricultural Terraces Platforms Using Machine Learning from Orthophotos and LiDAR-Based Digital Terrain Model: A Case Study in Roya Valley of Southeast France

by Michael Vincent Tubog 1, Karine Emsellem 2 and Stephane Bouissou 2,*
Reviewer 1: Anonymous
Reviewer 2:
Reviewer 3:
Submission received: 31 March 2025 / Revised: 24 April 2025 / Accepted: 27 April 2025 / Published: 29 April 2025

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors
  1. The abstract needs to be reorganized, especially lines 14-19.
  2. The reference format of the introduction is incorrect, and please refer to the Land journal format for modification, such as Line 29, According to [19].
  3. Please describe the innovation points of the paper in detail in the introduction section.
  4. Introduction section (Lines 74-82), please list the scientific questions that this study aims to solve and what is the specific study purpose.
  5. I recommend moving parts 1.1 and 1.2 to Materials and Methods
  6. The article does not specify the acquisition time of LiDAR data and high-resolution orthophotos data. In Section 2.2.1, the candidate terrace area is calculated using DTM. If the time of LiDAR data is less than the time of high-resolution orthophotos data, will the terrace candidate area calculated using DTM result in the newly built terrace being excluded from the terrace candidate area?
  7. In section 2.3, the author identified three categories of terraces based on their activity status and suggested providing remote sensing images of each type of terrace to facilitate readers’ understanding of the three types of terraces.
  8. In Figure 8, the red wavy line under the word “La Brigue” should be removed. Please redraw the map.
  9. Please add the limitations and prospects of this study in the Discussion section
  10. The quality of English needs to improve in order to make meanings more explicitly understandable.

Author Response

Comment 1: The abstract needs to be reorganized, especially lines 14-19.

Response 1: Thank you for your suggestion. The abstract has been modified from lines 14 to 20.

Comment 2: The reference format of the introduction is incorrect, and please refer to the Land journal format for modification, such as Line 29, According to [19].

Response 2: We have changed the organization of the bibliography, this time arranging it in the order in which it is cited in the text rather than in alphabetical order. Reference [19] becomes reference [1] and so on.

Comment 3: Please describe the innovation points of the paper in detail in the introduction section.

Response 3: Thank you for your observation. Kindly see lines 137 to 139.

Comment 4: Introduction section (Lines 74-82), please list the scientific questions that this study aims to solve, and what is the specific study purpose.

Response 4: Thank you for your comment. It is now indicated at Lines 111 – 117.

Comment 5: I recommend moving parts 1.1 and 1.2 to Materials and Methods

Response 5: Thank you for your suggestion for improving the paper's sub-sectioning. Subsections 1.1 and 1.2 are now moved to the Materials and Methods section with corresponding subsection reclassifications, namely, 2.2 and 2.3.

Comment 6: The article does not specify the acquisition time of LiDAR data and high-resolution orthophotos data. In Section 2.2.1, the candidate terrace area is calculated using DTM. If the time of LiDAR data is less than the time of high-resolution orthophotos data, will the terrace candidate area calculated using DTM result in the newly built terrace being excluded from the terrace candidate area?

Response 6: That’s a great observation, and yes—it’s a valid concern. In this study, the Digital Terrain Model (DTM) used to identify candidate terrace areas was derived from LiDAR HD data collected by the French national mapping agency (IGN) between 2021 and 2023, as part of their national LiDAR HD program. Meanwhile, the orthophotos used were most likely captured in 2020, based on IGN’s official update schedule for the Alpes-Maritimes region.

So in this case, the LiDAR data is actually more recent than the orthophotos. That means the DTM reflects a more current surface than the imagery. However, if any terraces were constructed after the LiDAR acquisition, they would indeed be missing from the DTM and, as a result, wouldn't be picked up as candidate terrace areas in the analysis. This temporal mismatch could lead to some newly built terraces being excluded from detection.

That said, the goal of this study was to identify terraces that had already been established—especially historical ones potentially suitable for revitalization. So while the issue you raise is technically important, it likely has only a minimal impact on the core objectives of the study. Still, it's definitely something worth noting for future applications where detecting very recent terrace construction is critical.

Changes are also made in that specific section (now 2.4.1). added lines 435-438.

Comment 7: In section 2.3, the author identified three categories of terraces based on their activity status and suggested providing remote sensing images of each type of terrace to facilitate readers’ understanding of the three types of terraces.

Response 7: Thank you for the suggestion. The figures are added (Figure 6).

Comment 8: In Figure 8, the red wavy line under the word “La Brigue” should be removed. Please redraw the map.

Response 8: Thank you for the suggestion. The figure is modified.

Comment 9: Please add the limitations and prospects of this study in the Discussion section

Response 9: Thank you for this suggestion. The limits and prospects of this study are indicated as subsection 4.4 in the Discussion section.

Comment 10: The quality of English needs to improve in order to make meanings more explicitly understandable.

Response 10: Thank you for this observation. The author wishes to inform the reviewers that this comment is taken seriously and necessary revisions in the language are carefully reviewed.

Reviewer 2 Report

Comments and Suggestions for Authors

This study provides a useful method for semi-automatic terrace detection using support vector machines and associated geospatial data, with applications in the management of land following a storm event. It is a rigorous methodology merging LiDAR, orthophotos, and field verification. Nevertheless, there are some key items that need to be clarified; some potential biases need to be addressed; and there are some aspects of the technical rigor that needs improvement.

Lines 352–374: The stratified random points used in field validation process (Section 2.3) were not well represented. The authors claim that 1500 and 4000 were created respectively for Saorge and La Brigue, but do not defend the sampling strategy or consider the possibilities of spatial bias (I.e., constraints of accessibility on steep/vegetated areas). This might skew accuracy metrics.

Line 430–442: Imbalanced training-ground samples (43 terraced, 82 non-terraced features in Saorge ) have been ranging in Table 2. Data imbalance causes SVMs to be poor at generalising, but no mention of mitigation strategies (eg oversampling class weighting etc)

Line 653–669: The disparity in UA between Saorge (97%) and La Brigue (72%) is justified as stemming from a difference of scale but lacks substance. Under-discussed are factors like terrain complexity (say, La Brigue’s steeper slopes), or vegetation density, or differences in data resolution. Add qualitative language to explicitly link variation in accuracy to site-specific geomorphological or data-quality factors.

Line 630–635: As acknowledged by the authors, they risk being overfit — on groups with high-dimensionality data — yet they do not quantify this (e.g., using cross-validation or testing against independent datasets). There should be measures to demonstrate generalizability, such as cross validation results or independent test datasets.

Lines 225–240: The choice of SVM kernel (linear, RBF, ) is not justified. Kernel selection critically influences performance, but the reasoning for employing the default ArcGIS Pro kernels is not provided.

Use the same term format as LiDAR (e.g., Line 307 uses Lidar, corrig this as LiDAR).

Define at first mention “UA” (User Accuracy)(Line 45).

Figures 4A/B and 6–9 are cited but not shown. All figures must be numbered sequentially and adequately described.

Author Response

This study provides a useful method for semi-automatic terrace detection using support vector machines and associated geospatial data, with applications in the management of land following a storm event. It is a rigorous methodology merging LiDAR, orthophotos, and field verification. Nevertheless, there are some key items that need to be clarified; some potential biases need to be addressed; and there are some aspects of the technical rigor that needs improvement.

Comment 1: Lines 352–374: The stratified random points used in field validation process (Section 2.3) were not well represented. The authors claim that 1500 and 4000 were created respectively for Saorge and La Brigue, but do not defend the sampling strategy or consider the possibilities of spatial bias (I.e., constraints of accessibility on steep/vegetated areas). This might skew accuracy metrics.

Response 1: Thank you for this insightful comment. We agree that potential spatial bias, particularly due to limited accessibility in steep or densely vegetated terrain, is an important consideration when conducting accuracy assessments.

To address this, we have clarified the sampling strategy in Section 2.7 (previously Section 2.3) of the revised manuscript. Specifically, while the initial stratified random points were generated across the entire classified area, we recognize that field validation was limited to accessible locations. To reduce the influence of this bias, we relied on high-resolution orthophotos and DEMs as additional validation tools for points in inaccessible areas. This hybrid validation approach—combining field visits with visual photointerpretation—has been used in several remote sensing studies to mitigate terrain-induced limitations.

Furthermore, we now acknowledge in the revised manuscript (see Section 4.4: Limitations and Prospects) that while stratified random sampling enhances representation, accessibility constraints could influence the spatial distribution of ground-truth data. We suggest that future work integrate UAV surveys or community-based data collection to improve validation in less accessible zones.

We appreciate the opportunity to clarify this aspect of our methodology.

Comment 2: Line 430–442: Imbalanced training-ground samples (43 terraced, 82 non-terraced features in Saorge ) have been ranging in Table 2. Data imbalance causes SVMs to be poor at generalising, but no mention of mitigation strategies (eg oversampling class weighting etc)

Response 2: We appreciate the reviewer’s observation regarding class imbalance in the training samples, particularly in Saorge. Indeed, the original sample set consisted of 43 terraced and 82 non-terraced features, which could potentially influence SVM performance due to the underrepresentation of the minority class.

To address this, we have revised the manuscript (Section 2.6.1) to include a clearer explanation of our strategy (lines 559-564). During SVM training, we limited the number of samples per class to a maximum of 1000 and ensured balanced representation by setting the training tool’s sampling mode to equalize class contributions during classification. While we did not explicitly apply oversampling or class weighting, the ‘Train Support Vector Machine Classifier’ tool in ArcGIS Pro internally handles sample balancing by default, and we manually reviewed classification outputs to correct evident under- or overfitting. Moreover, we now acknowledge the impact of class imbalance in the Discussion (Section 4.4) and suggest that future work may integrate SMOTE-based oversampling or cost-sensitive classification to enhance performance in minority-class-dominated landscapes such as steep terraced areas.

Comment 3: Line 653–669: The disparity in UA between Saorge (97%) and La Brigue (72%) is justified as stemming from a difference of scale but lacks substance. Under-discussed are factors like terrain complexity (say, La Brigue’s steeper slopes), or vegetation density, or differences in data resolution. Add qualitative language to explicitly link variation in accuracy to site-specific geomorphological or data-quality factors.

Response 3: We appreciate the reviewer’s insightful comment regarding the explanation of accuracy disparities between Saorge and La Brigue. In the revised manuscript, we have expanded the discussion to explicitly link these variations to site-specific geomorphological and data-quality factors. Specifically, we now clarify that La Brigue’s landscape presents a more complex geomorphological setting, characterized by steeper and sharper mid-slopes and deeper incised valleys compared to Saorge. This increased terrain complexity, combined with denser vegetative cover in certain sectors, likely contributed to the higher misclassification rate and lower UA in La Brigue.

Additionally, while both study sites used comparable LiDAR and orthophoto data, slight differences in data resolution and acquisition conditions, particularly the presence of shadow effects and seasonal variations in canopy cover within La Brigue, may have influenced classification accuracy. These factors likely increased the difficulty in distinguishing terraces from non-terrace features, contributing to the observed disparities.

We have now incorporated this qualitative explanation into Sections 4.1 and 4.2 of the Discussion to substantiate the scale-based reasoning originally provided.

Comment 4: Line 630–635: As acknowledged by the authors, they risk being overfit — on groups with high-dimensionality data — yet they do not quantify this (e.g., using cross-validation or testing against independent datasets). There should be measures to demonstrate generalizability, such as cross validation results or independent test datasets.

Response 4: Thank you for this valuable suggestion. In response, we have conducted additional k-fold cross-validation (with k=5) on our existing training datasets for both Saorge and La Brigue to assess potential overfitting and the generalizability of the SVM models. The average cross-validated accuracy results are as follows:

  • Saorge:6% (±2.1)
  • La Brigue:3% (±3.4)

These results are consistent with our original accuracy assessments and indicate that while some overfitting may be present, particularly in Saorge due to high-resolution data capturing small, potentially non-generalizable features, the models remain reasonably robust for both sites.

We have added these cross-validation results and their interpretation in the revised Section 4.2 Accuracy Validation, and noted that future work will incorporate independent validation datasets from neighboring hamlets (e.g., Fontan, Breil-sur-Roya) to further test generalizability beyond the current study area.

Comment 5: Lines 225–240: The choice of SVM kernel (linear, RBF, ) is not justified. Kernel selection critically influences performance, but the reasoning for employing the default ArcGIS Pro kernels is not provided.

Response 5: We thank the reviewer for raising this important point. In the initial analysis, we employed the Radial Basis Function (RBF) kernel, which is the default in ArcGIS Pro’s SVM implementation. This choice was based on its proven effectiveness in handling non-linear classification problems and its ability to model complex decision boundaries — an essential requirement given the heterogeneous nature of terrace patterns in Mediterranean mountainous environments.

In the revised Methods Section (3.2.2), we have now explicitly stated this kernel choice and provided a rationale for its selection, citing prior studies (e.g., Sofia et al., 2016) that demonstrate the RBF kernel's suitability for similar geomorphometric and land cover classification tasks. Additionally, we note that preliminary comparative runs using the linear kernel yielded significantly lower classification accuracies, reinforcing our decision to proceed with the RBF kernel for the final model.

Comment 6: Use the same term format as LiDAR (e.g., Line 307 uses Lidar, corrig this as LiDAR).

Response 6: Thank you for this observation. It has been modified accordingly.

Comment 7: Define at first mention “UA” (User Accuracy)(Line 45).

Response 7: Thank you for this observation. It has been modified accordingly.

Comment 8: Figures 4A/B and 6–9 are cited but not shown. All figures must be numbered sequentially and adequately described.

Response 8: Thank you for this observation. Figures have been modified accordingly.

Reviewer 3 Report

Comments and Suggestions for Authors

Review Comments on the Manuscript:

The manuscript presents a novel approach to detecting and mapping agricultural terraces using machine learning techniques, specifically Support Vector Machine (SVM), alongside LiDAR and orthophoto data. While the study is timely and addresses an important topic, several areas require improvement for clarity, rigor, and scientific accuracy.

 

  1. [P1, Lines 12-13]: Add a brief description of why terraces are important for soil erosion control and agriculture to provide context for readers unfamiliar with the subject.
  2. [P1, Lines 18-19]: Explain the significance of identifying the specific number of terraces (18 in Saorge and 35 in La Brigue) and their impact on agricultural practices in the region.
  3. [P1, Lines 22-24]: Suggest including examples of practical applications for the SVM algorithm's accuracy improvements in terrace detection, to illustrate how these findings can influence agricultural policy or practices.
  4. [P1, in introduction] Define technical terms at their first occurrence (SVM algorithm) to ensure accessibility for a broader audience.
  5. [P1, Line 28]: Clarify what is meant by "a succession of levels or gently sloping steps" by providing a more visual description or example for better reader understanding.
  6. [P1, Line 31]: Provide a brief explanation of why retaining walls or embankments are necessary, possibly including their role in long-term soil preservation.
  7. [P2, Line 58]: Provide a more detailed explanation of how the height of the wall influences agricultural outcomes. What are the practical implications for farming on different slopes?
  8. [P3, Lines 66-68]: Incorporate data or statistics to emphasize the extent of the issues faced in rural and mountainous areas, which would strengthen your argument about the need for attention to food-related issues.
  9. [P3, Lines 81]: Consider summarizing the overall significance of ensuring sustainability in terrace agriculture at the end of this section, linking back to the earlier points about their historical and current importance.
  10. [P6, Lines 152-155]: Clarify how GIS and machine learning techniques work together in terrace detection. A brief description of the specific roles of GIS versus machine learning might help readers understand their importance more clearly.
  11. [P6, Lines 160-162]: When discussing the authors' findings with the Random Forest algorithm, consider explaining what factors contribute to "improved accuracy" and "predictive power" to give context to these terms for readers.
  12. [P6, Lines 166-170]: Provide insight into why a CNN was chosen for this study. Briefly discuss the benefits of using CNNs over other methods, highlighting why it’s effective for high-resolution satellite imaging.
  13. [P6, Lines 171-176]: You mention that over 80% of pixels were classified accurately; providing a comparison to traditional methods here could enhance understanding of the significance of this statistic.
  14. [P8, line 246]: 6 and similar equations, the variable names, such as "𝝃𝒊" and "𝜶𝒊", need consistent formatting. Make sure they are all clearly represented, and avoid the use of unconventional symbols like 𝝃 (ξ) or 𝜶 (α) without explanation.
  15. [P22, Line 617]: In Table 6, clarify what "User’s Accuracy" refers to and how it is calculated. Including a brief explanation will help readers understand the significance of these metrics.
  16. [P22]: in discussion address potential limitations of your study. For example, discuss the implications of the varying accuracy rates between Saorge and La Brigue and suggest how these might be mitigated in future research.
  17. [P25]: the conclusion should not only summarize the findings but also offer recommendations for future research or practical applications. This would enhance the manuscript's contribution to the field.

Comments for author File: Comments.pdf

Author Response

Comment 1: [P1, Lines 12-13]: Add a brief description of why terraces are important for soil erosion control and agriculture to provide context for readers unfamiliar with the subject.

Response 1: Thank you for this suggestion. We indicated additional information in the manuscript (abstract)

Comment 2: [P1, Lines 18-19]: Explain the significance of identifying the specific number of terraces (18 in Saorge and 35 in La Brigue) and their impact on agricultural practices in the region.

Response 2: Thank you for this suggestion. We indicated additional information in the manuscript (abstract)

Comment 3: [P1, Lines 22-24]: Suggest including examples of practical applications for the SVM algorithm's accuracy improvements in terrace detection, to illustrate how these findings can influence agricultural policy or practices.

Response 3: Thank you for this suggestion. We indicated additional information in the manuscript (abstract)

Comment 4. [P1, in introduction] Define technical terms at their first occurrence (SVM algorithm) to ensure accessibility for a broader audience.

Response 4. Thank you for this suggestion. We indicated additional information in the manuscript (lines 104-108)

Comment 5: [P1, Line 28]: Clarify what is meant by "a succession of levels or gently sloping steps" by providing a more visual description or example for better reader understanding.

Response 5: Thank you for this suggestion. We indicated clarification in the manuscript (lines 49-51)

Comment 6: [P1, Line 31]: Provide a brief explanation of why retaining walls or embankments are necessary, possibly including their role in long-term soil preservation.

Response 6: Thank you for this suggestion. We indicated further information in the manuscript (lines 57-60)

Comment 7: [P2, Line 58]: Provide a more detailed explanation of how the height of the wall influences agricultural outcomes. What are the practical implications for farming on different slopes?

Response 7: Thank you for this suggestion. We indicated further information in the manuscript (lines 85-89)

Comment 8: [P3, Lines 66-68]: Incorporate data or statistics to emphasize the extent of the issues faced in rural and mountainous areas, which would strengthen your argument about the need for attention to food-related issues.

Response 8: Thank you for this suggestion. We indicated further information in the manuscript (lines 98-103)

Comment 9: [P3, Lines 81]: Consider summarizing the overall significance of ensuring sustainability in terrace agriculture at the end of this section, linking back to the earlier points about their historical and current importance.

Response 9: Thank you for this suggestion. We indicated further information in the manuscript (lines 139-143)

Comment 10: [P6, Lines 152-155]: Clarify how GIS and machine learning techniques work together in terrace detection. A brief description of the specific roles of GIS versus machine learning might help readers understand their importance more clearly.

Response 10: Thank you for this suggestion. We would like to inform the reviewer that this section has been transferred to another section (now at 2.3.1), and we indicated further information in the manuscript (lines 259-263)

Comment 11: [P6, Lines 160-162]: When discussing the authors' findings with the Random Forest algorithm, consider explaining what factors contribute to "improved accuracy" and "predictive power" to give context to these terms for readers.

Response 11: Thank you for this suggestion. We indicated further information in the manuscript (lines 268-272)

Comment 12: [P6, Lines 166-170]: Provide insight into why a CNN was chosen for this study. Briefly discuss the benefits of using CNNs over other methods, highlighting why it's effective for high-resolution satellite imaging

Response 12: Thank you for this suggestion. We indicated further information in the manuscript (lines 277-282)

Comment 13: [P6, Lines 171-176]: You mention that over 80% of pixels were classified accurately; providing a comparison to traditional methods here could enhance understanding of the significance of this statistic.

Response 13: Thank you for this suggestion. We indicated further information in the manuscript (lines 297-300)

Comment 14: [P8, line 246]: 6 and similar equations, the variable names, such as "??" and "??", need consistent formatting. Make sure they are all clearly represented, and avoid the use of unconventional symbols like ? (ξ) or ? (α) without explanation.

Response 14: Thank you for this suggestion. This is well noted.

Comment 15: [P22, Line 617]: In Table 6, clarify what "User’s Accuracy" refers to and how it is calculated. Including a brief explanation will help readers understand the significance of these metrics.

Response 15: Thank you for this suggestion. We indicated further information in the manuscript (lines 716-718)

Comment 16: [P22]: in discussion address potential limitations of your study. For example, discuss the implications of the varying accuracy rates between Saorge and La Brigue and suggest how these might be mitigated in future research.

Response 16: Thank you for this suggestion. We indicated a new sub-section in the manuscript addressing this comment (lines 849-876)

Comment 17: [P25]: the conclusion should not only summarize the findings but also offer recommendations for future research or practical applications. This would enhance the manuscript's contribution to the field.

Response 17: Thank you for this suggestion. We indicated a further addendum in the manuscript (lines 896-902)

 

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The manuscript has been carefully revised in accordance with the comments and suggestions, and the presentation is more convincing in support of their findings.

Author Response

Comment 1: The manuscript has been carefully revised in accordance with the comments and suggestions, and the presentation is more convincing in support of their findings.

Response 1: The authors sincerely thank the reviewer for their positive feedback and appreciation of the improvements made to the manuscript. We are grateful for the constructive comments that helped enhance the clarity and quality of this work.

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have replied to most of the comments and suggestions raised during the first round of review. The manuscript presents a solid and well-supported workflow for the semi-automatic detection of terraces in the Roya Valley through the combination of LiDAR, orthophotos and SVM-based machine learning. The methods are solid, the results well presented, and the discussion shows an appropriate awareness of both strengths and limitations. One or two small points remain before this manuscript can be recommended for publication.

The manuscript is written and organized in a fairly good manner. Nonetheless, certain passages have too much detail, especially in Materials and Methods and Discussion, repeating information or including tangential explanations (e.g., too much background on SVM theory and equations). You might want to condense these sections for the sake of readability and focus on the most relevant methodological and interpretative aspects.

Please verify that all the figures you cite (especially Figures 7–10) have sufficient resolution and are published in a clear form. Some images in the current draft look aliased or have blurry color legends. 

In Table 6, please explain the meaning of "Precision (User's Accuracy in %)" for readers not familiar with remote sensing jargon. A succinct footnote in the table caption would do.

Please, further explain why you chose the RBF kernel for the SVM, and how you select the parameters.

Indicate in the Accuracy Assessment section whether the stratified random points that were used for validation were independent of the training data to avoid possible bias.

Your limitations section is complete. Maybe a brief note on possible consequences of working with different years of datasets (orthophotos from 2020; LiDAR from 2021–2023) regarding temporal comparison of mapping terraces, keeping in mind that land use may change dramatically in the area.

Rest of the grammatical and typographical errors still persist (e.g. “The terrace walls in this region range between 0.7 m to 1 m and the steps have a measured width of 2.85 m to 3 m.”“).

The use of technical terms (e.g., "LiDAR-based DTM" vs. "DEM") should be standardized for consistency.

Author Response

The authors have replied to most of the comments and suggestions raised during the first round of review. The manuscript presents a solid and well-supported workflow for the semi-automatic detection of terraces in the Roya Valley through the combination of LiDAR, orthophotos, and SVM-based machine learning. The methods are solid, the results well presented, and the discussion shows an appropriate awareness of both strengths and limitations. One or two small points remain before this manuscript can be recommended for publication.

Comment 1: The manuscript is written and organized in a fairly good manner. Nonetheless, certain passages have too much detail, especially in Materials and Methods and Discussion, repeating information or including tangential explanations (e.g., too much background on SVM theory and equations). You might want to condense these sections for the sake of readability and focus on the most relevant methodological and interpretative aspects.

Response 1: We sincerely thank the reviewer for their thoughtful feedback and for acknowledging the organization and overall quality of the manuscript.

Regarding the level of detail in the Materials and Methods and Discussion sections, we understand the concern about potential repetition and tangential content. However, we chose to maintain this level of elaboration to ensure transparency and reproducibility of the methodology, especially considering that our approach integrates geospatial analysis, field validation, and machine learning, each with specific parameters that may not be universally familiar to all readers.

In particular, the background information on SVM and related equations was included to support interdisciplinary accessibility, especially for readers who may come from applied fields such as land-use planning or environmental management rather than computer science.

Nonetheless, we will remain open to condensing these sections further if the editorial team deems it necessary in the next phase. We appreciate the suggestion and value the emphasis on clarity and focus.

Comment 2: Please verify that all the figures you cite (especially Figures 7–10) have sufficient resolution and are published in a clear form. Some images in the current draft look aliased or have blurry color legends. 

Response 2: Thank you for your helpful feedback regarding the clarity and resolution of Figures 7–10. We have reviewed these figures and agree that certain elements—particularly the color legends and finer map details—appeared blurred or aliased due to image compression in the initial draft. To resolve this, we have updated the manuscript with high-resolution versions of these figures, ensuring all labels, overlays, and legends are clear and legible.

Additionally, for full transparency and ease of review, we have uploaded the original high-resolution images to a supplementary Google Drive folder. This includes Figures 7–10 as well as other key visual outputs of the study. The link to the folder has been provided in the revised manuscript under the “Supplementary Materials” section. We trust that this resolves the issue and enhances the readability of the figures.

Comment 3: In Table 6, please explain the meaning of "Precision (User's Accuracy in %)" for readers not familiar with remote sensing jargon. A succinct footnote in the table caption would do.

Response 3: Thank you for the helpful comment. We have revised the caption of Table 6 to include a brief explanation of "Precision (User’s Accuracy in %)" to aid readers who may not be familiar with remote sensing terminology. The updated caption now clarifies that this refers to the proportion of correctly classified positive observations relative to all observations classified as positive. Also, the footnote can be found at lines 717-718.

Comment 4: Please, further explain why you chose the RBF kernel for the SVM, and how you select the parameters.

Response 4: Thank you for this important observation. We selected the Radial Basis Function (RBF) kernel for the Support Vector Machine (SVM) due to its flexibility and effectiveness in handling non-linear relationships, which are typical in complex terrain features like terraced landscapes. The RBF kernel can map input features into a higher-dimensional space, making it suitable for distinguishing subtle spatial and spectral differences between terraced and non-terraced areas.

For parameter selection, we initially adopted standard default values provided in the ArcGIS Pro “Train Support Vector Machine Classifier” tool (e.g., Gamma = 0.5, Cost = 100), as these are optimized for general use. We conducted iterative tuning by manually testing different combinations of gamma and cost parameters to observe their effect on classification performance—particularly overall accuracy and user’s accuracy. The final parameters were selected based on the model configuration that yielded the highest accuracy during preliminary cross-validation runs using the training data.

This approach balances computational efficiency and classification performance, especially given the heterogeneous and topographically varied nature of the study sites.

Comment 5: Indicate in the Accuracy Assessment section whether the stratified random points that were used for validation were independent of the training data to avoid possible bias.

Response 5:  Thank you for highlighting this important point. We confirm that the stratified random points used in the accuracy assessment were independent of the training data. The training and validation datasets were created using a 70–30 split, with 70% allocated for training and 30% reserved exclusively for testing and accuracy validation. The validation points were generated within representative regions that were not used in the training phase, thereby ensuring independence and minimizing any potential bias in the accuracy metrics.

We have now clarified this in the revised manuscript under Section 2.7 Accuracy Assessment, for transparency and methodological rigor.

Comment 6: Your limitations section is complete. Maybe a brief note on possible consequences of working with different years of datasets (orthophotos from 2020; LiDAR from 2021–2023) regarding temporal comparison of mapping terraces, keeping in mind that land use may change dramatically in the area.

Response 6: We appreciate the reviewer’s insightful comment. In response, we have included a brief discussion in the Limitations section acknowledging the temporal differences between the orthophotos (2020) and the LiDAR data (2021–2023). We emphasized that while the datasets are recent, land use in mountainous areas such as Saorge and La Brigue can change rapidly due to both anthropogenic and natural factors. This temporal gap may lead to minor inconsistencies in terrace mapping and classification. The added note addresses the potential implications of this dataset disparity on the study's comparative analysis (see lines 830-839)

Comment 7: Rest of the grammatical and typographical errors still persist (e.g. “The terrace walls in this region range between 0.7 m to 1 m and the steps have a measured width of 2.85 m to 3 m.”“).

Response 7: We thank the reviewer for pointing this out. A comprehensive proofreading of the manuscript has been conducted, and several grammatical and typographical errors—including the one mentioned—have been corrected for clarity and precision. The sentence in question has been revised to:

The terrace walls in this region range from 0.7 m to 1 m in height, and the steps have a measured width between 2.85 m and 3 m.”

We have carefully reviewed the rest of the manuscript to ensure grammatical consistency throughout.

Comment 8: The use of technical terms (e.g., "LiDAR-based DTM" vs. "DEM") should be standardized for consistency.

Response 8: We appreciate the reviewer’s observation regarding the inconsistency in the use of technical terms such as “LiDAR-based DTM” and “DEM.” To address this, we have standardized the terminology throughout the manuscript. The term “LiDAR-based Digital Terrain Model (DTM)” is now consistently used when referring to terrain elevation data derived from LiDAR, while “Digital Elevation Model (DEM)” is used only in general contexts where the data source is not explicitly LiDAR-derived. This standardization improves the clarity and precision of the manuscript and avoids confusion for the readers.

Back to TopTop