Mapping the Extent of Mangrove Ecosystem Degradation by Integrating an Ecological Conceptual Model with Satellite Data

Round 1

Reviewer 1 Report

This manuscript details the application of a new degradation model, based on satellite imagery data products, to detect and map changes in mangrove ecosystem state. The authors present a well justified and structured arguement for the new model and demosntrate its use well within two example case studies. Therefore I believe this manuscript warrants publications subject to a few minor clarifications and edits detailed below. 

It would be useful for the authors to state up front in the introduction whether the new model presented herein is designed soley for use in mangrove ecosystems or whether it has wider applications. I see no reason for it to be limited soley to mangroves other than the two sites chosen to assess the model are mangrove ecosystems, and it may be beneficial to say that potential applications exist beyond this ecosystem.

Following the development of the conceptual model, the authors present a simplified version base on the components relevant to the chosen case-studies (lines 179 - 184). Does this conceptual model therefore only apply to your two sites? What about other degradaton factors that may be dominant in other mangrove settings (i.e. marine and near coastal dredging, sand mining (e.g. in cambodian mangroves)), industrial development and flow regulation (e.g Red River, Vietnam mangroves). More clarity on the scope of application is needed. 

Line 211: How are you defining natural variability? For example, do you account for annual variations in phenology that may bias your analysis? How does the selection of imagery from Google Earth (line 218 - 225) bias your initial data collection in terms of the availability of high resolution imagery and the time of year that that data is available?

With respect to the Rakhine data sets (line 239  - 245), was only one year used in this classification and if so how representative is this? How can you be sure that degradation is real and you are not identifying the natural state of the mangroves as degraded if you dont do the temporal classificaiton as you do for Shark River?

The two different case studies eventually use different data inputs as the Shark River dataset used just the landsat imagery. Therefore, is it strictly the same model used for both case studies or two seperate models? Does this also mean that you can just use landsat imagery to reach the same conclusions about the degradation of mangrove ecosystems? It would be useful for the reader if the authors could explain in more details the limitations of the analysis that has been conducted, especially with regards to the needs of the input data and the level of comparibility between sites where different data sets may be available.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 2 Report

General comments:

The motivation of the current work is good. However, my main concern is that neither remote sensing methods nor models developed in the study are novel. The authors just employed the Random Forests (RF) algorithm with some features derived from multisource earth observation datasets to map the extent of mangrove ecosystems degradation in the two case studies in Myanmar and Everglades National Park, USA. The overall accuracies (OA) are relatively low, ranging from 77.6% to 79.1% and the Kappa coefficients are low, especially for degraded class (0.559). It reflects the misclassifications are substantial. It is unclear how the authors split the datasets into the training and testing sets. The current version has lots of shortcomings that make it unsuitable for consideration for possible publication in Remote Sensing. The authors may address all my comments and suggestions to revise their work. The paper may be suitable for other ecological conservation journals.

Specific comments. 

• Title: It should add the names of the two study sites and remote sensing datasets.
• Abstract: The main results should provide more details such as the Kappa coefficients and the most important features for mapping mangrove degradation. What is the scientific contribution of this work to the remote sensing community? The authors have to underline where the innovative aspect of their work 

1. Introduction
   • The authors should write this section in a much punchier and more reflective way, underlining better the topic importance and also aspects related to the current scientific gaps.
   • The authors did not review enough the current remote sensing and machine learning techniques to justify why the authors just employed the RF algorithm for mapping mangrove forests. 
2. Methodology:
   • Did the authors consider the tidal level when classifying the three classes (Intact, Degraded and Collapsed) mangrove types as the tidal definitely affect the obtained classification results. More detailed information of tidal and dates of date acquisitions are required.
   • Authors should also consider open mangrove canopy as it somehow is relatively similar to degraded mangrove and explain more with relevant references.
   • I suggest adding a Table describe the final training and the testing datasets. 
   • Lines 239- 249. It is unclear how the authors split the datasets into the training and testing sets or did the authors use the K-fold cross-validation?
   • It is unclear how the authors optimised the hyperparameters of the RF algorithm as the classification results are low in terms of Kappa coefficients and OA.
   • It should add more standard metrics such as precision, recall, and F1 score to justify the results. 
   •  
3. Results
   • Maps (Figures 3 and 4) should be improved in terms of spatial resolution and standards. Authors should add the coordinate systems in maps.
   • Fonts size and types should be used consistently. The font in Tables 7 and 9 is different from the font in the main texts. 
   • The variable importance should present in Figures with a percentage of variables contribution. 
   • Why the total relative importance in table 7 is greater than 100% while this number is much lower than 100 in Table 9. It sounds like very strange values???
   • Why the authors only employed the 7 features (Tables 6 & 7) for the Myanmar mangrove and only 5 features (Table 9) for the mangrove forests in the Everglades National park?
   • As the results are low, I suggest including all features from Landsat 8 OLI (7 multispectral bands) and spectral indices such as NDVI, NDWI, NDMI plus HH, HV, ration HH/HV, HV/HH, GLCL computed from HH and from HV to increase the number of features for the final model. Then the authors may apply feature selection either by feature importance or other metaheuristic optimisation for feature selection. 
   • Why in Table 5 authors classified three classes for the mangrove Myanmar but only two classes in Table 8 for the Everglades? 
4. Discussion
   • This part is weak. What is the novelty of this work. 
   • More critical reflection is required more analysis (and deeper) of the results and stronger links to the relevant (contemporary) literature as part of this section.
5. Why no conclusion remarks were drawn in this research???

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 3 Report

Overall, I believe this paper makes a significant contribution to the literature by developing methods for processing remote sensing data into accessible products for assessing ecosystem degradation. The methods are well developed and explained, and the paper is well written. Pending response to a couple general concerns, this paper will be acceptable for publication.

 

First, it is not explicitly clear what “degradation” is measuring in the training set. The introduction (Lines 65-67) states that this can mean changes to species assemblage and abundance, canopy cover, or productivity, and Table 1 relates various ecological links to remotely sensed variables. Also, these definitions make the “degradation” concept concrete, but they are somewhat buried in this paragraph and it would be helpful to the reader to be more explicit about what exactly is being modeled closer to the purpose statement at the end of the introduction. The Methods section on the training set development lays out rules for site selection in the “degraded” class, but relying on a visual assessment of partially cleared mangrove forests to develop this may not account for all of the various ecological links or proposed variable mechanisms. My main question here, then, is if different ecological links such as reduced branch density or stunted growth/leaf number capture degradation through different measurable traits, how can you be certain that those traits are captured in the training dataset and thus give accurate variable importance responses in your models?

 

My last and primary concern regards the Discussion section. Overall, the section does not synthesize the paper’s results by interpreting the different datasets’ performances and model variable importances. Rather, the bulk of this section is devoted to discussing errors and proposing future work. More focus should be given to what your results actually denote. The Discussion paragraph structure works fine as is, starting with the brief paragraphs on the study site results and working through error sources and future work. However, because this is paper focused on methods development and applications of various satellite datasets, I suggest adding a section that dissects in greater depth the different variables, what they indicate, and their relative importances in the models. This would feed into your future development section with more concrete analysis of your results that also indicate which datasets would be most valuable to build on.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

Dear authors,

Thanks for the effort to revise the manuscript. The quality of the revised version has improved, but still has some major points, which are needed to be addressed before it can be further processed for possible publications in Remote Sensing.

Specific comments:

1. Abstract: needs to be improved. Make it punchier and underline better the unique contribution of this work (which is not currently seen well), and please provide the most important conclusions of your findings only in the abstract (e.g. what is really the take-home message of your work?)
2.  Innovation of the research study and unique contribution to the remote sensing community is weak since several such studies have been conducted in the past. Authors need to underline much better where the innovative aspect of their work lies to 
3. Introduction: Some relevant review papers related to the current topics should be reviewed and cited 
   • Remote sensing of mangrove ecosystems: A review
   • A Review of Remote Sensing Approaches for Monitoring Blue Carbon Ecosystems: Mangroves, Seagrasses, and Salt Marshes during 2010 - 2018
4. Method:  I would recommend providing a bit more statistical evidence strengthening the aspect of the statistical significance of the results, demonstrate in your methods and analysis of results that this aspect has been taken under consideration and examined deeply such as other standard criteria such as: F1-score, Precision and Recall beside overall accuracy and Kappa.
5.  Results & discussion: provide more information on the error sources and on the statistical significance of your results. Also, how your algorithm's parameterization settings have affected the obtained results, and can it be possibly compared to other ML techniques?
6.  Discussion and Conclusions: can become more quantitative. What are the study findings implications that need to be discussed as well either in the discussion or in the conclusions sections? 

Author Response

Response: We thank the reviewer for their comments and have addressed them point-by-point below. With their suggestions, we believe the manuscript is clearer and better highlights the contribution of our work.

Please see the attachment for point-by-point responses.

Author Response File: Author Response.docx