Characterizing Role of Spatial Features in Improving Mangrove Classification—A Case Study over the Mesoamerican Reef Region

Round 1

Reviewer 1 Report (Previous Reviewer 1)

Comments and Suggestions for Authors

The authors has now clarified the conceptual confusion that previously existed in the manuscript, and by revising the title, has effectively avoided the earlier conceptual misinterpretation. The quality of the manuscript has been significantly improved. Evaluating the effectiveness of various non-spectral spatial features in enhancing mangrove classification model performance is of considerable research value. The author's discussion on the applicability of incorporating spatial attributes into mangrove classification is insightful, and the study’s findings are thought-provoking. The following suggestions are provided for the author's consideration:

-In Line 158, "2019-20" should be revised to "2019–2020".

-In Table 1, change "30-m" to "30 m", and "10-m" to "10 m" in the Spatial-1 column.

-In Line 277, change "3.3" to "3.2"; in Line 302, change "3.4" to "3.3"; and in Line 315, change "3.3" to "3.4".

Author Response

Open Review

Comments and Suggestions for Authors

The authors has now clarified the conceptual confusion that previously existed in the manuscript, and by revising the title, has effectively avoided the earlier conceptual misinterpretation. The quality of the manuscript has been significantly improved. Evaluating the effectiveness of various non-spectral spatial features in enhancing mangrove classification model performance is of considerable research value. The author's discussion on the applicability of incorporating spatial attributes into mangrove classification is insightful, and the study’s findings are thought-provoking. The following suggestions are provided for the author's consideration:

Response: We are sincerely grateful to the reviewer for helping us significantly improve the manuscript and avoid potential conceptual misinterpretations. We acknowledge that previous rounds of review have led to substantial improvements in the manuscript. It is truly encouraging to know that the reviewer finds the work both useful and insightful.

 

-In Line 158, "2019-20" should be revised to "2019–2020".

Response: We thank the reviewer for this suggestion. The line now reads as

“The extent maps were developed for the composite period of 2019-2020.”

 

-In Table 1, change "30-m" to "30 m", and "10-m" to "10 m" in the Spatial-1 column.

Response: Thanks for noticing these typos. These have now been corrected.

-In Line 277, change "3.3" to "3.2"; in Line 302, change "3.4" to "3.3"; and in Line 315, change "3.3" to "3.4".

Response: We apologise for these mistakes. These have now been corrected in place.

Author Response File: Author Response.docx

Reviewer 2 Report (New Reviewer)

Comments and Suggestions for Authors

1. Figure 2: When comparing the importance of different sub-features under various spatial feature conditions, the current layout of four subplots provides limited comparability. It is recommended to merge the four subplots into a single plot and use different color codings for Non-spatial, Spatial-1, Spatial-2, and Spatial-3 to enable stronger visual comparability.

 

2. Section 2.2.3 – Leave-one-out strategy for model calibration: Please revise the terms "leave-one-out approach" and "leave-one-out iteration" to more accurate terminology. The standard "Leave-One-Out" (LOO) method refers to leaving out a single sample point in each iteration. However, the method described in the manuscript actually leaves out an entire spatial grid cell (containing multiple points) in each iteration, which is a standard approach for evaluating a model’s robustness to spatial positional bias (geographical bias) and for mitigating the effects of spatial autocorrelation. Therefore, please adopt the appropriate terminology for spatial cross-validation to avoid confusion. Clearly state that each iteration involves "leaving out all points within one grid cell".

 

3. Furthermore, note that calculating feature importance (based on 500 trees) in each iteration of SLOO (Spatial Leave-One-Out) — a total of 116 iterations — may incur high computational costs. The purpose and post-processing of these results are not clearly described. How are the SLOO validation results used to achieve the goal of assessing the impact of geographical bias?

 

4. Figure 8: Sub-figure labels (a, b, c, d) are missing. In Figure 8, Spatial-3 exhibits a peak in spatial correlation at a distance of 600 km — please explain the reason for this phenomenon. Additionally, for the distance-related variation in spatial correlation shown in Figures 7 and 8, please include a corresponding analysis.

 

5. Section: Role of spatial features in improving classifications (first paragraph): The logical structure of the paragraph needs further improvement. Please standardize the result interpretation, eliminate contradictory statements, and link the discussion to Table 2. Clearly distinguish the modeling roles of spatial positional features versus spatial environmental features. Use concrete examples to explain the mechanism behind error reduction. Also, remove or clearly define the concept of "localized decision trees".

 

6. Discussion section: Consider adding figures or diagrams to support the discussion if necessary. In addition, if possible, please reorganize the methods section to include a schematic of the technical workflow to illustrate the overall research framework.

Author Response

We are grateful to the reviewer for their constructive feedback. We have addressed each comment in the attached word document. We have provided portions of the manuscript with figures wherever appropriate.  Please also find the modifications in the manuscript highlighted in yellow. 

Author Response File: Author Response.docx

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

Incorporating spatiality into remote sensing classification is inherently a significant area of research, as it provides us with data information beyond spectral details, aiding in a deeper exploration of the data's informational content. However, some descriptions in the text appear to be ambiguous.

Firstly, it's essential to clarify that spatiality and spatial autocorrelation are fundamentally different. Spatiality refers to the spatial attributes of data, such as geographic coordinates and distances to other geographic features mentioned in the text, which all fall under spatiality. On the other hand, spatial autocorrelation describes the statistical relationship between the values of a particular spatial attribute, i.e., whether the values of variables at neighboring locations are similar or related. While spatial data may inherently have correlations, determining whether these are spatial autocorrelations requires specific computational methods. When calculating such spatial autocorrelations, it is typically necessary to construct a spatial weight matrix that describes the relationship between each point and its neighboring points, rather than simply describing the correlation between the values at two spatial points.

Based on the above understanding, I revisit the descriptions in the author's article, as the followings.

1. The spatial attributes used in the text, such as geographic coordinates, distance, and elevation, are indeed spatial properties of the data. Although the authors suggest in Figure 2 that they analyzed spatial autocorrelation, in the remote sensing classification process, spatial autocorrelation itself was not involved in the calculations.
2. The correlation used in Figure 2 seems to be Pearson correlation? In fact, calculating spatial autocorrelation requires specialized methods (such as Moran's I, Geary's C). If the authors used these methods, they should be described in the methods section.
3. Inverse Distance Weighting only considers distance-weighted interpolation and is not strictly based on spatial autocorrelation. Why not consider Kriging interpolation, which more accurately accounts for spatial autocorrelation?
4. Based on the third point above, considering the four feature sets used by the authors in the text, the article does not first calculate whether the data has spatial autocorrelation, nor is spatial autocorrelation applied in the classification process. Consequently, the emphasis on spatial autocorrelation in the title does not align with the core theme and focus of the study.

 

Based on the above content, I suggest the authors consider the differences between spatiality and spatial autocorrelation and reorganize the descriptions in the article and change the title. Additionally, there are some suggestions for the author's reference.

1. Consider comparative results with other datasets (Figures 5 and 6).
2. Reorganize the discussion section by introducing subsections.
3. Revise the ambiguous statements in Lines 57–64.
4. The longitude, latitude, and altitude in spatial feature set 1 do not reflect Euclidean distance features, whereas proximity to coast, rivers, and streams in spatial feature set 2 do. Please check Lines 239.
5. From the results of the article, it seems that remote sensing classification incorporating spatiality is more suitable for local processing rather than global. It would be excellent if the authors could specify the applicable conditions for remote sensing classification incorporating spatiality.

It is hoped that the results of this research can be applied to larger areas of mangrove mapping and other remote sensing classification applications in the future.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

General comment:

Each paragraph is unnecessarily separated from the others by one blank line.

 

Comments to the text:

1. Page 1, line 40: Mangrove forests are crucial coastal wetland ecosystems located in tropical and subtropical regions, …
2. Page 2, line 44-45: Remote sensing – based mapping, espectally at the global scale, has been instrumental -> has been crucial..
3. Page 2,. Line 70-71: Appropriately integrated spatial autocorrelation can offer valuable context, which can help capture spatial patterns…
4. Page 2, line 91-92: Unnecessary space between [18,22,33] and dot.
5. Page 2, line 94: Environmental conditions like climate…
6. Page 3, line 140-141: We also tested the robustness of additional features…
7. Page 4, line 158-159: Please correct the description of tall and medium canopies: currently the text says that medium canopies are 2 to 5 m tall, while tall canopies are over 30 m tall.
8. Page 4, figure 1: Check whether the background map used does not violate the author rights and licenses of Bing.
9. Page 4, line 167-168: …as well as for the natural protection against increasing threats of tropical cyclones…
10. Page 4, line 170: under threat from human impact (e.g. coastal development).
11. Page 4, line 176: …, and at an altitude less than 100 m a.s.l.
12. Page 5, line 204: … to capture a range of other land cover types?
13. Page 8, line 280-294: Some of text parts (especially numbers) are in italics: e.g. 5,932, [84].
14. Page 8, line 310-311: “Could be”? Maybe “is”? Is this result tested for statistical significance?
15. Page 16, line 484: Double space between “Wadouc et al.” and “[31]”.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Although the author has responded to my comments one by one, unfortunately, the author has not adequately addressed the issues I raised. I have decided to reject the manuscript.

I agree with the idea that spatial attributes can be incorporated into model construction. However, I cannot agree with the author's statement in the abstract that " This study evaluates strategies to incorporate spatial features in binary mangrove classification account for effect of spatial autocorrelation in model building. "

To reiterate, spatial autocorrelation refers to the statistical relationship between the value of a spatial attribute in a geographic space and the values in its neighboring areas. I understand that the authors mention spatial attributes being influenced by spatial autocorrelation, and in their response, they cite six references, including soil mapping using spatial dependency to model continuous surfaces, introducing spatial features into models to predict two spatial datasets, simulating forest carbon stock maps, and estimating forest biomass. These examples all involve the simulation of continuous variables using spatial attributes, which is suitable for modeling continuous unknown variables (as also mentioned in the literature review section of the manuscript, such as soil properties, pollutants, population, and socioeconomic parameters). This is fundamentally different from the binary land cover classification addressed in the authors' current study.

The focus of the paper (according to the author's explanation) is "account for spatial autocorrelation in regional mangrove mapping" (as stated in the title). The author claims that" These features help capture the spatial structure within a landscape without explicitly calculating spatial autocorrelation during model building. (Lins-488)" and therefore suggests that latitude, elevation, and distance to rivers (Spatial-1 and Spatial-2) can be considered implicit factors reflecting spatial autocorrelation. I do not agree with this equivalence. These factors inherently have spatial properties, and using spatial features to explain spatial autocorrelation requires a complete logical proof process.

Spatial autocorrelation in this paper is only superficially reflected in the IDW interpolation (Spatial-3). However, the method used by the author—“generating interpolated median mangrove spectral reflectance surfaces for each Sentinel-2 band using all mangrove training pixels;……applying IDW interpolation, defining the feature set as the difference between the actual median reflectance values and these interpolated surfaces. (Lins-294~304)”—has significant flaws and contains logical errors. Remote sensing land cover classification inherently relies on spectral information, and IDW interpolation is essentially a spatial smoothing method, which cannot directly capture spatial autocorrelation. Simply put, IDW interpolation was developed precisely because of spatial autocorrelation, not because its use implies spatial autocorrelation has been considered. When actual land cover spectral reflectance values are available, using IDW interpolation is an error-expanding action and does not effectively capture spatial dependence. The author’s claims that the method “explain similarities in the spectral reflectance of nearby pixels. (In coverletter)” and “to account for effect of spatial autocorrelation in model building. (Lins-16~17)” are not well substantiated, and the results obtained are not sufficiently convincing.