An Improved Machine Learning-Based Method for Unsupervised Characterisation for Coral Reef Monitoring in Earth Observation Time-Series Data
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThis paper proposes an improved machine learning approach to automated coral reef monitoring using satellite imagery, which has certain engineering reference value. However, the following problems still exist:
1> In the title, "an improved machine learning approach" is mentioned, but in the method section, there is no indication of what the improvements are. Should the author consider modifying the title to "coral reef monitoring" instead of "shallow water marine environments"? Currently, it seems that the title does not match the content of the article.
2> In the line 4 of the Abstract, it is recommended to add an explanation of what the role and purpose of “…a machine learning-based cloud removal technique using XGBoost”.
3> The keywords used are not accurate enough and the quantity is rather small. It is recommended to make modifications.
4> In the Introduction, it is necessary to supplement what the existing means of coral reef monitoring are and what their advantages compared to Earth Observation (EO) are. It is recommended to add the paper's structural arrangement at the end of this part to enhance readers' overall understanding of the paper.
5> It is recommended that the author only write about the machine learning algorithms relevant to this paper in the section "2. Background on machine learning in remote sensing", and introduce the algorithm principles and improvement ideas in detail.
6> It is recommended that the author rename Part 2 as DATA and Part 3 as method. The current content of Part 2 can be incorporated into the new Part 3.
7>For the “Sentinel-2 Data” part, it is recommended to add example images for comparative analysis, especially example images of different data segments, to analyze the composition of the original images and what the expected goals of this paper are.
8> There are many problems with the layout of the tables. For example, Table 1 should be in the “Sentinel-2 Data” section. Is it necessary to add citations to the table titles? And there are also issues with Table 2. The author is requested to check the whole paper.
9> There is no explanation for Fig. 2 and Fig. 3 in the paper. Especially for Fig. 3, it is recommended to conduct a quantitative analysis of the results of different algorithms.
10> Line 222 what "providing valuable information" specifically refers to ?
11> In Line 265, why is the "Padding" set to 0? Could you please provide the comparison results before and after the padding?
12> It is recommended that the author add the implementation principle and process of “Otsu’s thresholding method” for the images in this paper. Diagrams or formulas can be used for analysis.
13> What is the highlighted ellipse in Fig. 5(B)? There is a significant difference compared with Fig. 5(A). Was there any processing in between? Or it would be better to perform a local magnification and then make a comparison and explanation.
14> It is recommended that the results of the paper be quantitatively analyzed, such as in terms of gray level, geometry, similarity, etc. If such analysis is not possible, effective comparison can also be made by referring to the actual shooting situations obtained by unmanned aerial vehicles.
15> Part 5 and Part 6 can be combined and summarized in a more concise and straightforward way.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThis paper proposes a processing workflow for shallow-water marine environment, which combines PCA, DBSCAN clustering, XGBoost and other algorithms such as Otsu and Kmeans. However, I have difficulties in understanding the main purpose of this study and following the proposed method.
1. Title and Abstract: The main purpose of this paper is unclear. In the title, it seems to develop a method for characterisation of shallow water marine environments. However, in the abstract, it targets at the challenges of image quality assessment and correction for automatic coral reef monitoring. Furthermore, the propose method is also confusing. It says that the method uses PCA with clustering for image selection and quality evaluation and XGBoost for cloud removal, and the workflow involves depth correction and Keamns clarification. What is the exact workflow?
2. Introduction: In lines 87-98, this study have three aims to address these limitations in satellite imagery analysis for coral reef monitoring. The aims in Title and Abstract should be consistent with these in Introduction. It is suggested to change [1],[2],[3] to be (1), (2) and (3), in order to distinguish from the reference citations.
3. Method: The proposed method is questionable in the following aspects.
(1) In Figure 2, it is unclear how the lower resolution imagery are processed from 320m pixel resolution down to the 10m scale.
(2) In section 3.2 filtering scenes, the first step aims to find the most suitable images (lines 176-177). Why should the suitable images be the images that are broadly consistent in brightness and colour balance? Only visible bands have the characteristics of colour. The bands of Sentinel-2A and 2B in Table 1 cover a wide range of wavelengths.
(3) From your description, it seems the PCA and DBSCAN are used for initial image selection. However, it is confusing what the images refer to. Are they some of the Sentinel-2 Spectral bands? Or are they Sentinel-2 images share similar features from a large scale of images captured over times? Or are they similar pixels/points in the Sentinel-2 images?
(4) In lines 178-179, PCA was applied to Preview Image Files (PVI). Please explain whether the PCA is used for dimension reduction (e.g., 3 bands to 1 dimension) or for extracting eigenvectors.In lines 186-189, what does it mean “in this case, deep and shallow areas of water as well as potential outliers missed in the previous processing”?
(5) In section 3.3, how is the combination of SLIC segmentation, PCA and DBSCAN related to image colour correction?
(6) In line 269, a new PCA algorithm was trained on the superpixels. However, PCA is unsupervised march learning algorithm, and it does not require to train.
4. Results:
(1) Experiments were carried out on only one area. The generalization performance of the proposed framework can not be verified well.
(2) In the title of Figure 7, “Images on the right after stretching”. Stretching operation is not mentioned in the entire article.
(3) The method lacks comparisons with other methods.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsIn this paper, the authors propose the overall architectural design of unsupervised machine learning for improving the image quality of marine scenes, in which there are many finely designed steps that can fully utilize the data characteristics and method advantages. Here, I have several comments:
1. The results of the methods are only visualized, and there is no quantitative presentation and analysis of the results, and no comparison with existing methods, which I think is insufficient.
2. There are many steps in the design, and each step has some parameters, which will result in the accumulation of errors, and the parameters will have a decisive impact on the performance, which I have not seen in the analysis.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThere are too many keywords. It is recommended that the author retain 4 to 5 keywords that are more closely related to the article.
Reviewer 2 Report
Comments and Suggestions for AuthorsThe authors have carefully addressed my concerns. I have no further comments.
Reviewer 3 Report
Comments and Suggestions for AuthorsThe authors have addressed my questions and resolved my concerns through careful revisions. The quality of the manuscript has been significantly improved. I recommend accepting it for publication.
