Swin-UperNet: A Semantic Segmentation Model for Mangroves and Spartina alterniflora Loisel Based on UperNet
Round 1
Reviewer 1 Report
The manuscript presents a study on segmentation of mangrove and spartina alterniflora Loisel using a semantic segmentation model based on UperNet. The formal aspects of the paper are proper. The paper is well prepared and well-organized, it brings valuable results. However, some problems need to be improved before it can be published.
1. 29 references are not enough. I suggest authors add a paragraph about semantic segmentation methods and its applications in remote sensing images. Introduction of UperNet, its advantages and disadvantages, and its successful applications in other area are necessary to convince readers.
2. Line 59-61: Sentence “Many scholars have proposed different deep learning methods …“ is not clear. Any segmentation method has fault and omit errors in its result, this should not be the flaw of pervious methods. Please discuss more details about flaws and shortages in previous studies.
3. Section 2.2 GF-1 and 2.3 GF-6 should be a subsection of 2.1 Data. Temporal resolution of GF-1 and GF-6 needs to be listed out. A table of selected GF-1 and GF-6 images with their dates is suggested. Number and spatial coverage of GF-1 and GF-6 image scenes you collected should be clarified.
4. In Data Preprocessing section, details about how you decide the locations of the cropped 480X480 images should be discussed as most area in your Figure 1 image is land. Maybe a flowchart of generating experimental dataset and some example images in the final dataset are great for readers to better understand this section.
5. Also, the 800 480X480 images seems a small number considering the time span of 5 years and scene size of 12078X11125. Please add discussions about the process of deciding the number of 800, and convince readers it’s a reasonable number for this experiment with evidences.
6. In Data concatenation module, 7 channels were used. My question is why blue band is not included here. According to USGS, blue band is useful for “Bathymetric mapping, distinguishing soil from vegetation and deciduous from coniferous vegetation”.( What are the best Landsat spectral bands for use in my research? | U.S. Geological Survey (usgs.gov) )As mangrove is deciduous vegetation and spartina alterniflora Loisel is coniferous vegetation, utilizing blue band in your study may be a good choice.
7. Start from line 244, please check the numbers of mIoU in this paragraph, it’s not consistent with the result in table 2.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
This paper proposes a novel segmentation model based on UperNet. I believe the model is technically sound and practically useful. I got the following opinions for the authors consideration.
1. The authors details the model without giving the rationale first. It’s recommended that author should explain the reason for designing the model. Why design the model like this? What’s your consideration to choose specific modules? Please give the rationale/logic reason behind model. It’s important to give your reason, so that readers can buy your idea and be convinced that your model is not designed randomly. I recommend you add more relevant discussion in the abstract and introduction sections.
2. I suggest the authors mention the paper’s usefulness in broader context, as the last sentence of abstract. What’s the contribution of your work to greater picture? Please consider.
3. The section for relevant works is short. Please expand your literature review in the introduction. The current reviewed papers are too little to support a research paper.
4. I recommend the authors to cite the following relevant works:
Unsupervised Change Detection Using Fuzzy Topology-Based Majority Voting
Novel Multiscale Decision Fusion Approach to Unsupervised Change Detection for High-Resolution Images
Novel Higher-Order Clique Conditional Random Field to Unsupervised Change Detection for Remote Sensing Images
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
In this work, the authors proposed a semantic segmentation model for mangrove and spartina alterniflora loisel based on Upernet. The model has a good novelty and the results are compared with the existing models. There is one suggestion to the authors- to study the effects on more than 7 channels as in their work -"7 channels were used to enhance spectral information for mangrove and spartina alterniflora loisel segmentation"
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
The authors’ response has made most of my concerns clear. But still some small questions left unclarified.
1. Maybe I didn’t express myself clear in last comment. By “details about how you decide the locations of cropped images”, I mean did you decide the locations manually or is there any automatic detection method or some sort of land mask you used in deciding the locations of cropped images. If its manually selected, authors should give out a distribution map of all selected locations, if it’s automatically decided, method description (maybe with flow chart) should be added.
2. Is Figure 2 covers all of your study area or its just an example for image cropping? Any way it’s hard for readers to discover your study area without a clear geographic map.
3. In line 158-160 “The experimental dataset was generated by flipped…… with random multiplicity”, I assume any one of the operations would double the size of your experimental dataset, so could you please explain in your manuscript (section 4.1) what’s the final images in your experimental dataset? I guess it contains (augmented cropped images, flipped augmented cropped images, scaled augmented cropped images, scaled flipped augmented cropped images), whose size is (800, 800*3, 800, 800*3).
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
The authors have addressed my concerns
Author Response
Thanks for your crucial comments.