Importance of Spectral Information, Seasonality, and Topography on Land Cover Classification of Tropical Land Cover Mapping

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

See attached file.

Comments for author File: Comments.pdf

Author Response

Comment 1: In Sections 2.3.1 and 2.3.2, the authors should focus on reporting the methods applied in the study rather than delving into technical details. For instance, while mentioning the function “ee.Classifier.smileRandomForest” provides clarity on the specific technique used for the random forest implementation, it may not be informative for readers unfamiliar with RF. Similarly, although I appreciate the authors for including detailed technical information such as function names and R package names, these details may not be sufficient for readers to understand the broader methodology. The primary concern for readers is the method itself. For example, the workings of Recursive Feature Elimination (RFE), which is the core method in the manuscript, may still be unclear to those who are not well-versed in this technique. Therefore, it is essential for the authors to provide a clear and concise explanation of the methods used, including how they are applied and their relevance to the study, rather than focusing solely on technical specifications.

Response 1: Thank you for the helpful and constructive feedback. In response, we revised Sections 2.3.1 (“Random Forest Classifier and Variable Importance”) and 2.3.2 (“Variable Selection”) to provide clearer, more concise explanations of the methods, emphasizing their conceptual foundations and relevance to the study while minimizing technical and software-specific details.

• In Section “Random Forest Classifier and Variable Importance”, we added a description of how Random Forest improves classification accuracy through ensemble learning and its suitability for high-dimensional remote sensing data (Lines 292–296).
• We clarified the use of Gini impurity for ranking variable importance (Lines 302–305).
• The model configuration was rephrased to emphasize methodological intent rather than software implementation (Lines 306–308).
• In Section “Variable Selection”, we added a concise explanation of the Recursive Feature Elimination (RFE) process, highlighting its iterative nature and role in selecting informative variables (Lines 321–324).

 

Comment 2: If the authors had carefully examined Reference [85], they might have discovered that Equation (8) in Foody’s 2020 paper published in Remote Sensing of Environment is a best practice for testing the statistical significance of differences in overall accuracy. This is true even though Reference [85] itself did not support the usage of kappa coefficient. Additionally, it should be noted that this test should be conducted as a one-sided test.

Response 2: Thank you for this valuable suggestion. we have revised the statistical analysis by adopting the equations in Foody (2020, Eq. 7 and 8), using overall accuracy and its standard error instead of the Kappa coefficient. All Z-statistics are now interpreted using a one-sided test. Corresponding updates have been made to Equation 5 (Line 351–358), the Results section (“3.2 Impact of Spectral Indices, Bi-Seasonal Differences, and Topography on Accuracy in Land Cover Classification”, Lines 431–454), and Figure 6 (Line 455).

 

Comment 3: I have noticed that the reference list requires refinement. There are a total of 137 references cited, but some of them are not properly cited or are missing. For example, on page 9, line 306, a reference related to the R language is cited, but the specific reference for the R package caret is missing. I recommend replacing the current citation [81] with the correct reference for the R package caret. I suggest that the authors carefully review and refine the entire reference list to ensure that all citations are accurate and complete. This will enhance the readability of the manuscript.

Response 3: Thank you for your suggestion. We have made the following revisions to the manuscript:

• The original citation [80], referring to the R caret package, was moved to replace citation [81] to ensure accurate citation (Line 328).
• Reference [40] in the Study area was updated and corrected in the text to reflect the appropriate source (Line 134).
• The citations [78, 79], previously included in the original version (Line 291), were removed as part of the paragraph revision and are no longer required (Line 308).

Reviewer 2 Report

Comments and Suggestions for Authors

This paper proposes a land cover mapping method for subtropical forests, analyzing the roles of spectral information, seasonal information, and topographic features in distinguishing different land cover types. The article also explores the effectiveness of variable selection methods in reducing data redundancy and improving model efficiency. This work is comprehensive, and the writing is clear. I only have some minor comments:

1. Although the article explains the importance of topographic variables and some spectral indices, a deeper discussion could be provided on how these variables reflect the specific ecological environment characteristics and land use patterns in the Kulun region. For example, why is NDTI so important for distinguishing agricultural land from forests? In what ways is the sensitivity of the red-edge band reflected in the differentiation of vegetation types?  

2. Although the IPCC classification standard is mentioned, a more specific analysis could be provided on how the definitions of land cover types in ESA and SERVIR differ from those in this study, and how these differences have led to the final accuracy assessment results.

Author Response

Comment 1: Although the article explains the importance of topographic variables and some spectral indices, a deeper discussion could be provided on how these variables reflect the specific ecological environment characteristics and land use patterns in the Kulun region. For example, why is NDTI so important for distinguishing agricultural land from forests? In what ways is the sensitivity of the red-edge band reflected in the differentiation of vegetation types?

Response 1: Thank you for the suggestion. In response, we have revised Section 4.1.2 (Variable importance) to expand our discussion of how topographic variables and key spectral indices reflect the ecological and land use characteristics of the Kulen region. Specifically:

• We clarified the role of elevation by noting that forested areas in Kulen’s protected zones are located at higher elevations, which receive greater rainfall and humidity. These conditions create favorable ecological environments that support dense evergreen and semi-evergreen forest ecosystems. (Lines 602-605)
• We expanded the explanation of NDTI by highlighting that its sensitivity to soil exposure and non-photosynthetic vegetation makes it particularly effective for distinguishing agricultural areas such as croplands, paddy fields, and plantations. These land uses often involve more frequent disturbance and exposed soil compared to undisturbed or regenerating forests, which tend to maintain stable canopy cover (Lines 615-626)
• We also strengthened the discussion of red-edge bands by explaining their sensitivity to chlorophyll content and leaf structure. This improves the differentiation of vegetation types that vary in species composition, canopy density, and physiological traits, which is especially important in a landscape with diverse vegetation-related land cover types. (Lines 632-636)

 

Comment 2: Although the IPCC classification standard is mentioned, a more specific analysis could be provided on how the definitions of land cover types in ESA and SERVIR differ from those in this study, and how these differences have led to the final accuracy assessment results.

Response 2: Thank you for the suggestion. In the revised manuscript:

• We have revised the manuscript to explicitly acknowledge the limitation that differences in original class definitions and the use of a one-time reclassification (Table A3) may have contributed to discrepancies in accuracy results (Lines 739-746).
• We have provided detail class definition for our product in Table A1. (Line 830)
• We have added a map showing locations of the reference polygons, overlay on a PlanetScope image from March 2021 in supplementary Figure S1.

Reviewer 3 Report

Comments and Suggestions for Authors

The authors develop a method to map forest-agricultural mosaics in Cambodia.  They use Sentinel imagery to generate bands related to vegetation, and use topographic variables and seasonal difference variables as input.

The paper is very well written and clear, with good figures.  It’s not terribly novel, but is very well done and they do show how to classify land cover in a difficult region with thorough and sophisticated methods and field data.  It can be published with minor revisions---please see pdf for detailed comments.

Clarification is needed on how the seasonal and annual maps were generated.  For the annual maps, is it one image for each year?  If so, is the pixel value the average of all pixels for that year, or the max NDVI, or something else?  How many annual maps were there? 

Comments for author File: Comments.pdf

Comments on the Quality of English Language

Very good...see pdf for suggestions.

Author Response

Comment 1: Clarification is needed on how the seasonal and annual maps were generated. For the annual maps, is it one image for each year? If so, is the pixel value the average of all pixels for that year, or the max NDVI, or something else? How many annual maps were there?

Response 1: Thank you for pointing out that we were unclear on this point. In the revised manuscript (Section: Sentinel-2 Dataset), we now explicitly state that the three median composite images (one annual, one for the rainy season, and one for the dry season) were all generated for the land cover classification in 2021. Each composite was produced by applying a per-pixel median across all images in the respective time periods. We also clarify that all spectral indices, including NDVI, were calculated from these median composite images (Lines 202–206).

 

Comment 2: Line 26: The journal may prefer uncrewed aerial vehicles (UAVs)… check to be consistent with their preference

Response 2: Thank you for the suggestion. We have replaced all instances of “drone” with “uncrewed aerial vehicle” for consistency with scientific terminology (Lines 27, 166, 246,251,265, 267,794).

 

Comment 3: Line 31: Tillage? The maps are of forest or also of agricultural lands? Clarify here.

Response 3: Thank you for the comment. We have clarified in the revised abstract that the land cover classification includes both forest and agricultural land types. This helps explain the inclusion of tillage-related indices in the variable set (Line 20-24).

 

Comment 4: At line 70, what is bi-seasonal? At line 71, what is a “full year approach”?

Response 4: Thank you for pointing out that we were unclear here. We have revised the paragraph to clarify that “bi-seasonal composite images” refer to the rainy and dry season composite images, and “full-year approach” refers to annual composites derived from imagery spanning the entire year. (line 72–73, 192–196).

 

Comment 5: At Line 77: Why “may”?

Response 5: Thank you for pointing this out. We have removed the word “may” from the sentence at Line 77 to improve clarity and avoid confusion. (Line 79)

 

Comment 6: line 89: Grasslands consist of a few classes? Reorganize the sentence for clarity.

Response 6: Thank you for pointing this out. We have revised the sentence for clarity by removing the unclear phrase and now refer to urban areas, croplands, and grasslands as examples of relatively homogeneous landscape where RFE has been effectively applied (Lines 90-91).

 

Comment 7: line 94-95: How do interactions complicate classification? Maybe you mean heterogeneity of the mosaic? What is a “high computing platform”?

Response 7: Thank you for pointing this out. we have revised the sentences at lines 93-99 as follows: “However, its application remains underexplored in complex tropical forest landscapes characterized by diverse LC classes and spatial heterogeneity resulting from mixed rural and forest land uses. Bridging this gap necessitates integrating RFE with robust classifiers and scalable geospatial computing platforms, such as cloud-based or high-performance systems, capable of handling the complex patterns and high variability typical of such environments, thus enhancing the accuracy and reliability of LC mapping in tropical landscapes [26].” (Lines 94-100).

 

Comment 8: line 101, add the word “The”.

Response 8: The article “The” was added in the revised text (Line 103).

 

Comment 9: At line 122, “Southeast Asian tropical forest area that was one dominated by pristine forests.” But is it now a mosaic of agriculture and forest? Or of secondary forest? It’s confusing because you say it’s a “tropical forest area”.

Response 9: Thank you for your comment. We have revised the sentence to clarify that the study area is no longer dominated by pristine forest. It is now characterized by a heterogeneous mosaic of forest and agricultural land resulting from extensive human modification (Lines: 124-126).

 

Comment 10: At line 160, remove the word “In this study,”.

Response 10: I followed your comment to remove the word “In this study,” in the revised manuscript (Line: 183).

 

Comment 11: At line 169, “bi-seasonal” (dry and wet season)?

Response 11: Thank you for pointing this out. We have modified the sentence by replace bi-seasonal to dry and rainy seasons in the revised manuscript: “We generated one median composite image each for annual, dry, and rainy seasons for LC classification in 2021, using Sentinel-2 Level-2A surface reflectance data.” (Lines: 192-193)

 

Comment 12: So the annual image is just one image per year? Is the pixel value in the annual image the mean of that pixel location for all images for a given year? Or is the max NDVI (commonly used for compositing).

Response 12: Thank you for pointing this out. See response to comment 1 above.

 

Comment 13: Modification or Modified Normalized Difference Water Index?

Response 13: Thank you for noticing this, it was modified that was meant, that is the “Modified Normalized Difference Water Index” (Line 217).

 

Comment 14: At line 222, why? Were Cashews particularly important or difficult to map?

Respond 14: Thank you for the helpful comment. We have revised the manuscript to clarify that cashew plantations are particularly important due to their rapid expansion, and they are difficult to map due to limited existing reference data and difficulty in distinguishing them from forests using UAV or satellite imagery. These revisions were made in both the introduction and methods sections (Lines: 249-252).

 

Comment 15: Were the numbers of points proportional to the area fraction of each cover in the study area?

Response 15: Thank you for the comment. The number of GPS points collected from field observations was not proportional to the area fraction of each land cover class. Instead, sampling efforts prioritized land cover types that were difficult to delineate from UAV or satellite imagery or lacked existing reference data, such as cashew plantations. This clarification has been added to the revised manuscript (Lines: 249-254). For classification, the final training dataset was generated by randomly sampling up to 1,000 points per class from the reference polygons prepared from GCP, UAV, and satellite imagery. A map of the reference polygons is now included in Figure S1 in the supplementary.

 

Comment 16: At line 242, suggest changing “with” to “, which has”. At line 256, suggest changing “split 70%” to “split into 70%”.

Response 16: We accepted the reviewer’s suggestions and revised them in the revise manuscript (Lines: 272, 286)

 

Comment 17: At line 258 (Section 2.3 Data Analyses), would it make more sense to put this overview paragraphs first in the methods section, and then the details on each variable (as in the current section 2.2.1)?

Response 17: Thank you for this suggestion. We have revised the structure of the Methods section by moving the overview paragraph from Section 2.3 (Data Analyses) to the beginning of the methodology (now Section 2.2.1). This reorganization provides readers with a clearer big-picture understanding of the overall workflow before presenting the detailed descriptions of input variables and processing steps (Section 2.2.1, Line 160).

 

Comment 18: Suggest to remove “As a result,” at line 308; “,” at line 309 and 310.

Response 18: Thank you for your suggestion. We have revised the text accordingly by removing “As a result” at line 308 and the commas at lines 309 and 310, as recommended (Lines: 328, 329, 330).

 

Comment 19: At line 456, the sentence “Further optimization, RFEvar-Hyper model was slightly improved…” is not grammatically correct. “Following optimization”?

Response 19: Thank you for pointing this out. We have revised the sentence at line 456 to improve grammar and clarity. The updated sentence now reads: "Following further optimization, the RFEvar-Hyper model was slightly improved and established as the most accurate and efficient model for LC classification in this study." (Line 478).

 

Comment 20: At line 470, add the phrase “…of the area as “other”.”

Response 20: We accepted your suggestions and added the phrase to the revised manuscript (line 493).

 

Comment 21: At line 548, the phrase “restricted range”, you mean “small fractional coverage”?

Response 21: Thank you for your comment. We clarified that “restricted range of LC classes” referred to the limited number and types of land cover classes specific to the study area, rather than to small fractional coverage. This clarification has been incorporated into the revised paragraph (Line 571).

 

Comment 22: At line 537, replace the word “others” with “other variables”

Response 22: Thank you for the suggestion. We have revised the sentence replacing the word “others” with “other variables” for clarity (Line 597).

 

Comment 23: At line 614, “Unlikely, tcAngleBG” ?

Response 23: Thank you for noticing. We have revised the sentence at line 614 to replace the incorrect phrasing “Unlikely, tcAngleBG” with “While tcAngleBG...”, to improve grammatical accuracy and clearly convey the intended contrast (Lines: 649-652).

 

Comment 24: At line 618, suggest to remove “Additionally,”. At line 644, suggest changing “improves” to “improve”. At line 662, suggest changing “forest land estimates” to “forest land cover estimates”. At line 665, suggest adding the word “used” into the sentence. At line 683, suggest deleting the word “Additionally”. At line 746, suggest changing from “locally derived modes” to “locally-derived models”.

Response 24: Thank you for the helpful editorial suggestions. We have accepted all proposed changes and revised the manuscript accordingly. Specifically:

• “Additionally,” was removed at lines 618 and 683 to improve sentence flow (now lines 653 and 718).
• “Improves” was corrected to “improve” at line 644 (now line 679).
• “Forest land estimates” was revised to “forest land cover estimates” at line 662 (now line 697).
• The word “used” was added for clarity at line 665 (now line 700).
• “Locally derived modes” was corrected to “locally-derived models” at line 746 (now line 697).

Reviewer 4 Report

Comments and Suggestions for Authors

The authors submitted a well written and an interesting manuscript dealing with the improvement of land cover mapping using remote sensing images (Sentinel-2 and Shuttle Radar Topography Mission datasets) and machine-learning techniques. The methodology is well described and the conclusions are supported by the results. Given the current changes in the climate systems, this study would help researchers working to find solutions for challenges resulting from tropical forest deforestation and land use changes. However, the manuscript should be revised before it could be considered for publication. Below are some comments and suggestions to improve the overall quality of the manuscript:

in the section of Materials and Methods, it would be interesting to combine paragraphs lines 157-359 into the subsection of Methodology in with you could include the descriptions of the datasets used and the methods you used to analyse them.  In addition, it would be interest to provide additional information related to references such as orthomosaic on which you could indicate the location of (Reference polygons ) training and validation samples (subsets). You could also provide a table with the descriptions of twelve LC classes you considered in this current study. In addition, in the section of Methodology, it is not so clear how the authors integrated the slope and the aspect to improve the accuracy of LC classification. They could consider this subject during the revision process. 

 

Author Response

Comment 1: in the section of Materials and Methods, it would be interesting to combine paragraphs lines 157-359 into the subsection of Methodology in with you could include the descriptions of the datasets used and the methods you used to analyse them. In addition, it would be interest to provide additional information related to references such as orthomosaic on which you could indicate the location of (Reference polygons) training and validation samples (subsets). You could also provide a table with the descriptions of twelve LC classes you considered in this current study. In addition, in the section of Methodology, it is not so clear how the authors integrated the slope and the aspect to improve the accuracy of LC classification. They could consider this subject during the revision process.

Response 1: Thank you for the valuable suggestions. In the revised manuscript:

• we have revised the structure of the Materials and Methods section by combining the content from lines 157–359 into a single, unified subsection titled “2.2 Methodology.” In the new subsection now we includes both “2.2.1 Data sources” and “2.2.2 Data Analyses” to descriptions of the datasets and the analytical methods used. We believe this integration improves the logical flow and helps readers better understand the overall workflow of the study (Lines 160–380).
• We have added a map showing locations of the reference polygons, overlay on a PlanetScope image from March 2021 in supplementary Figure S1.
• We provide a table describing the 12 land cover (LC) classes (Appendix Table A1), and we refer to it at Line 830)
• We have revised the topographic data section to clarify the integration of slope and aspect, adding the sentence: “Topographic variables, including elevation, slope, and aspect, were included in the Random Forest classification model as part of the full feature set, together with spectral bands, spectral indices, and bi-seasonal differences.” (Lines 224–226, 337).