A Deep Learning Method Based on Two-Stage CNN Framework for Recognition of Chinese Reservoirs with Sentinel-2 Images
Round 1
Reviewer 1 Report
This is a very interesting manuscript and I believe your work is a great contribution to the current national reservoir catalog in China. However, this manuscript needs some work before I would consider it ready for publication. Great job on your hard work! Some parts are hard to understand though, so I suggest enlisting someone with full professional proficiency in English to help edit and clarify the manuscript. I believe the wording/ grammar of the paper can be improved for clarity.
Specifically,
1. Need to provide more details, i.e. the size of the images, 36 images of 28 provinces, and then 2000? I would also suggest listing all the Sentinel 2A&B images in this study (i.e., original image capture date, cloud condition, if necessary, etc.) either those were downloaded from ESA or USGS as a table in the Appendix or supplemented documents.
2. Need a clearer definition of “samples,” i.e., line 181. And the caption for Fig 2, "where denotes an 100km×100km sampled 191 area. and XXX (how many) in total. Generally, figures should be able to stand alone and be clear to the reader without the main text. Please make sure all the captions are concise yet descriptive, acronyms/abbreviations are spelled out, individual sections such as subpanels in a multifigure figure are explained, letters labeling subpanels are consistently in the same position (within and across figures), legends explain variables, and the layout is consistent. Any text in and around figures should be large enough to read or be omitted.
3. Line 264-273 (Fig 6) Appropriate citations regarding the LeNet would be good enough. This section is a little redundant. The same issue is with Fig. 10 and related content, given these are established study flows and not the major focus of this research.
4. Line 230, “regarded.” Line 231 “which aimed.” Please be consistent with the tense throughout the manuscript.
5. Line 235, “a” dam… line 236, which “makes” it…
6. Table 3, what is the input data for the FasterCNN, needs more explanation and clarification.
7. Line 544, This will be a major contribution and very valuable to the scientific community. Considering the dataset will be publicly available (assumed to be labeled with annotation to be exact), I would like to have the access to the dataset for a better review/ understanding of the manuscript.
8. Could you please also provide PR and ROC curve figures?
9. The downsampling in the training set would make the data distribution different from the test set. It would be great to evaluate if the downsampling was skipped in the training set and compare the PR and ROC curves.
10. Further explanation will be needed for Figure 11: after fusing the feature map from the two ResNet50, how did you separate the ROI pooling for RGB and NIR?
11. Fig 13. The resolution of the images needs some improvement: For the red box, please zoom-in for readers with better details revealed.
Author Response
Please see the attachment.
Author Response File: 
Author Response.docx
Reviewer 2 Report
The authors showed several methods related to water segmentation, which is a complex problem. The aim is recognize reservoirs as an object detection task with deep neural networks in remote sensing image. Authors tested NIR RCNN and showed that some good results (Reservoirs and dams dectection achieved 80% of average accuracy). However, some issues are described below:
Abstract:
L18, p.1 - What is NIR RCNN?
Introduction:
L45-54 - Long paragraph, please, reformulated it into two paragraphs.
L10-111 - What the authors named "fine-grained"? It could be more specific, such as "a method which use 10m-resolution images"... Define "fine-grained"
L130 - Itenize the last contribution "4) Formed a Chinese reservoir landscape in 2020 with.."
An additional issue here: Since they have a Chinese reservoir dataset from Ministry of Water Resource of PRC (including this source within the text), why the authors want improve such database? They commented trhough the introduction, however, they do not specify details of PRC - Is an outdated dataset? Is it a private dataset? What are the main reasons to update this cited source? Maybe it was described in lines 201-206 in page 5
2.Material and Methods
L. 149, p.3 - Topic 2.1: I've missed an aditional figure to represent the description in this topic (related to relieves locations, maybe a topographic or altimetric map)
L.155, p.4 - Please rewrite this sentence, it is a bit unusual write in this way "13 bands image with the onboard MSI sensors"
L.163,p.4- I do not agree with this sentence. Users can proccess Sentinel L1C images using SNAP and Sen2Cor Plugin, but also, users can also download the image directly from Scihub, in some cases, in L2A...
L.167 - The authors used RGB from L2A, NIR from L2A or both from L2A?
Figure 2, p.5- Reformulated the entire image. What is the scale used here? What is the context (surroundings?). What are the images used to validate and to calibrate the RCNN?
L.194-197, p.5- What are the size of large and medium-sized reservoirs and dams in China?
L.222-225, p.7 - I do not understand the entire paragraph.
L.234-p.7 - Earth instead of earth
L.310-P.10 - Define FPN and RPN
Althought the efforts to create a mannual dataset and improve water system detections, the results are not well described. A lot of details are provided in methodology and results are summed up, the same occured in conclusion.
In addittion I did not understand very well the concept of " 4756 positive samples and 58746 negative samples in the constructed dataset (training set and test set) - Line 222, p.7" and relations of images and water system detections.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
Please have the manuscript proofread/edited by someone with full professional proficiency in English one more round.
Author Response
Dear reviewer:
Thank you again for your sincerely comments on my manuscript. Following with your comments, I have revised some sentences in the manuscript, and also changed the expression of some vocabulary to make it easier to understand. Please find my revisions in the re-submitted files.
