Integrating a UAV-Derived DEM in Object-Based Image Analysis Increases Habitat Classification Accuracy on Coral Reefs

Round 1

Reviewer 1 Report

Important results and practical value

Integrating geomorphometric variables into an algorithm Random Forest 497 Continuously increasing overall accuracy, noting increases of up to 11.2%

- the UAV-derived DEM, and derivatives thereof, clearly contain 499 habitat differentiating information and can be used to increase accuracy of automatic hab- 500 itat classification.

Author Response

Dear reviewer, 

We are happy to see that in general you seem content with the paper.
However, it is not fully clear what you mean with the two specific comments as they are repeats directly from the manuscript.
Therefore no further revisions were made.

Kind regards, Brian Nieuwenhuis on behalf of all co-authors

Reviewer 2 Report

Review of remotesensing-1827100 “Combining a UAV-derived DEM and orthomosaic in Object Based Image Analysis enhances accuracy of automated habitat classification on shallow coral reefs”

General comment:

Very interesting paper. The hypotheses, methodology, results, discussion and conclusion are well presented. The manuscript needs some minor changes before being published. Details are listed below:

Introduction

Line 38: “Especially” repetitive, please use another connector word.

Materials and Methods

Line 148: please, add “second” or “s’ after 640

Line 248: based on the description previously done (Lines 241-243), I would suggest adding “(1-3.5m deep)” after reef slope

Lines 249-252: It is not clear the difference between deep and shallow reef slope if it is compared with the description done in Lines 241-243. Please, clarify and rephrase accordingly

Line 266: please add full name for the acronym “ESP2”

Results

Lines 361-362: I would suggest adding the specific references used the figure (e.g. 2a) for “classifiers trained on both geomorphometric and spectral features”

Line 387: please, replace with “Area 1”

Discussion

The new workflow seems promising for habitat mapping of very shallow reef areas, tested in this study. What is the max depth for such habitat mapping approach? I would suggest including a brief discussion on it.

Author Response

Dear reviewer,

 

Thank you for taking the time to review our manuscript.

I am happy to hear that you liked our manuscript.

Your suggestions were appreciated and have been changed in the revised version.

Since they were mostly small textual edits, I assume it suffices to say that we incorporated all of them exactly as suggested.

With regard to your suggestion to add some discussion on the depth limitations I can also be short.

As you suggested we have added a paragraph in the discussion on the depth limitations of our methods.

 

Kind regards, Brian Nieuwenhuis on behalf of all co-authors

Reviewer 3 Report

1.       The title is too long and needs to be revised.

2.       The authors are suggested to specify their new contributions in this paper. They must also mention the disadvantage and drawbacks of the existing methods and clarify how their proposed Algorithm and model can overcome them.

3.       Regarding the UAV underwater images, please introduce the relevant references: Integrating QDWD with Pattern Distinctness and Local Contrast for Underwater Saliency Detection, Journal of Visual Communication and Image Representation, 2018.

4.       Currently, comparative evaluation is weak.

As a journal paper, the proposed method also lacks of comparisons to other state-of-the-art algorithms (e.g. published after 2020).

5.         In the experiments, please add a new subsection for discussing comparisons between different state-of-the-art methods

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

The manuscript remotesensing-1827100-peer-review-v1 analyzes three shallow coral zones through thematic maps combining orthomosaics and DEMs from a UAV based on OBIA and Random Forest approaches. The results indicate an increase in global accuracy when both products are combined. The analysis is correct, and I only have some minor but pertinent suggestions.

LN 26 and so forth: I would not consider the Phantom-4 RTK as a consumer-grade UAV. Consumer-grade UAVs are traditional, out-of-the-box, relatively inexpensive UAVs. I was very interested in using a consumer-grade UAV in corals because of all the limitations regarding this type of platform on shallow ecosystems. Still, I lost interest when I realized the authors used an RTK system. Please eliminate the "consumer-grade" statement throughout the manuscript.

LN 58: Specify that the resolution is "spatial." For instance, this UAV sensor has a meager spectral resolution.

LN 87-90: Authors claimed that this study is the first to address the synergy between UAV-derived DEM and OBIA. However, the authors mention that other remote sensing studies have performed this synergy in different ecosystems. Please briefly include the environment of references 40-42.

LN 143: How much is a minimal tide in this region?

LN 201-202: I understand this reference is new, but Sona et al. (2014) mention the opposite. Honestly, I have tried both Pix4D and Agisoft since 2014, and Pix4D always resulted in worse results, sometimes not being able to align images, while Agisoft worked with no problem. Moreover, Pix4D is way more expensive than Agisoft, although I will circle back to this issue. Please add more information about the use of Pix4D rather than Agisoft.

Figure 5: What happened in the lower right section of the classification? I believe the discussion does not include this explanation.

Figure 6: Please add more explanation about the reasons why Area 1, which has more waves, presented higher accuracy compared with Area 2.

Discussion:

LN 487: Everything depends on the number of thematic classes. Please include the number of classes between both studies. Simple compassion between two studies regarding the same approach is not correct when using different thematic classes.

I do not generally perform UAV flights at such low altitudes and with a high longitudinal overlap because we lose signal fast, the flight time is prolonged, and the SfM algorithm requires more images, resulting in additional computational time. Please mention why the authors made these flights at 40 m and 90% longitudinal overshoot. Generally, the lateral overlap must be more meaningful than the longitudinal one.

The authors mapped a total area of circa 30 ha. This situation is the main limitation of using UAVs in corals and other environments. From my point of view, 30 ha is a minimal area if we consider doing a national study or even frequent monitoring. Please include more discussion about the limitations of the mapped area. Moreover, the mapped area is shallow (4 m). This characteristic is another limitation for many applications worldwide. Please add more discussion about this.

I understand using the Phantom-4 RTK, which has many applications; however, this UAV platform is obsolete. In fact, you can no longer buy batteries for this system (which is problematic for many colleges and me). Please include this in the discussion.

The authors have not mentioned the cost of doing this study. For instance, how many national institutions could afford this equipment and software in the Caribbean? Very few. Please include the hardware and software costs (US Dollars) so that other authors know the initial investment.

References:

Sona et al., 2014. Earth Sci Inform 7, 97-107. http://dx.doi.org/10.1007/s12145-013-0142-2

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 5 Report

Review of the manuscript "Combining a UAV-derived DEM and orthomosaic in Object-Based Image Analysis enhances accuracy of automated habitat classification on shallow coral reefs"

by Nieuwenhuis, Marchese, Marco Casartelli, Andrea Sabino, van der Meij and Benzoni.   

Note: The supplementary data/ appendix is missing, the related file is the letter to the editor. So review is based only in the manuscript.  OOHH Nevermind.. the apendix was at the bottom of the document.

The paper is very well written and easy to follow, and authors must be proud of a work well done.

In general the implementation of this research is very sound and all methods and procedures have been given attention. I do not have any concerns or further suggestions to improve the manuscript except for the following:

Figures 6 & 7 please add at the top of each bar the actual accuracy value, no need to refer the reader to an appendix.

lines 577-579 Although this is partially correct as the spatial resolution of the maps is very high, the classification scheme of 5 very broad classes still can't represent the inherent heterogeneity of shallow reef systems, and combined with the sometimes modest accuracy of particular classes (i.e. macroalgae, coral framework) in some areas, this sentence might be misleading, i suggest to rephrase it.

line 603 should read "high spatial resolution habitat maps"... 

lines 604-605 should read "would be of support for ecosystem-based management decision making." The fact that management will be effective or not does not have to do with the maps.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 6 Report

The topic is interesting and valuable, and a lot of work has been done by the authors. This work and the results were also described in this paper carefully. However, the following questions need to be answered before the paper can be suggested to published in Remote Sensing.

1. The title said that “… automated habitat classification …”, but the abstract said “… can greatly enhance the accuracy of semi-automatic habitat classification …”.

2. Line 186: “In each survey area, at least two 30-metre transects were conducted, starting from the middle UAV GCP (Fig. 1)” For study on coral reef ecosystem, 50- to 100-metre transects was generally used, but in this paper, 30-metre transects was used. What is the rationale for this?

3. It is very difficult to take underwater photos. Although the authors introduced the GoPro camera array in the supplementary file. But how the divers to make sure the photos satisfy the requirements of the SfM, e.g. forward overlap? How the divers to make sure the photos cover the GCPs?

4. SfM is based on the principle that the principal point, the image point, and the object are collinear. However, for the UAV images, as the refractivity of the water, the principal point, the image point, and the object are no longer collinear. How to solve this problem in this paper?

5. Random Forest algorithm was used to classify the images after the segmentation. There are other machine learning algorithms that could potentially be applied and compared to the prediction/classification performance of Random Forest algorithm, for example, neural network, support vector machine and regression. Why did not used them?

6. How to dissolve the possible overfitting problem for the Random Forest algorithm?

7. The capability to determine relative importance of classification features was regarded as an advantage for choosing Random Forest algorithm to do the classification. For this reason, I suggest provided some analysis on the relative importance of classification features.

8. I suggest the authors to added a conclusion section.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

Without intriguing insights into the problem and the solution, I do not think this paper provides much value to the community. Another issue is that the authors fail to provide indepth analysis for the state of the art to explain why they do not work.

Most of the time, the authors simply recite what other authors have done without relating to why they are not good enough and how the proposal is different and why the proposal should work.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 6 Report

The authors have revised the first draft of the manuscript and most of the questions were properly answered. In my opinion, the paper has been significantly improved. The only thing is the text editing need to be further checked and corrected by the authors themselves to make sure the texts and the typesetting to satisfy the requirements of the publication in the journal. As a result, I suggest to consider to accept it for publishing in Remote Sensing.

Author Response

Please see the attachment

Author Response File: Author Response.pdf