Triangle Water Index (TWI): An Advanced Approach for More Accurate Detection and Delineation of Water Surfaces in Sentinel-2 Data
Round 1
Reviewer 1 Report
The manuscript (ID: remote sensing-1961371) conducts interesting work in extracting surface water from satellite imagery. It proposed a Triangle Water Index (TWI) based on Sentinel-2 data and tested the accuracy and robustness of the TWI method using Sentinel-2 data of several water bodies. The results confirm the TWI is effective to improve the accuracy of extracting water bodies, especially in mountainous regions with highly rugged terrains, and areas covered by snow and ice. The data and methods as well as its well-written, organized and comprehensive style contribute to the robustness of this study. Therefore, this study fits the scope of this journal and will be of interest to its readers.
I recommend this work for publication after some revisions.
(1) Line 171: There is an error in formula (1).
(2) Line 217: Please add an abscissa or abscissa description for Figure 3.
(3) The format, font, and font size of the text used in the formulas, figures, and tables should be unified. For example, there are two formats of brackets in formula (18), and there are texts with different font sizes in Figure 4. There are still similar errors. Please check the author carefully
(4) Figure 2 shows the reflectance spectra of 11 water bodies in Figure 1. Is there only one sample per water area (i.e., the reflection spectrum on one pixel) taken as the reflection spectrum? Is the sample representative?
(5) In Section 3.6, “There are 1581 sampling points that were collected based on different types of water bodies that were selected for accurate assessment by a stratified random sampling method. The classification of sample points is based on currently available data. Each point is evaluated and labeled using Google Earth images and maps”. Are Google Earth images and maps consistent on the time with Sentinel-2 images used for water extraction? If not, is there a similar situation: a point is a body of water on Google Earth images and maps, but when the Sentinel-2 satellite transits, the point changes to another land type due to time changes?
Author Response
Dear Reviewer
Please see the attachment,nameed "Response to Referee1".
Author Response File: 
Author Response.docx
Reviewer 2 Report
This paper presents the development of a new water body index, the triangle water index (TWI). It is based on different wavelength bands of sentinel-2 data and on the triangle similarity theorem. The study presents a comparative study with other well-known indices and claims that the TWI achieves better performances in the identification of small water bodies within low albedo or mountain regions covered by snow. The paper is well written, and the introduction and description of the methods are mostly clear, although some points need revision. The description of the results and conclusions is very synthetic and omits needed explanations. The main weakness of the work is that it does not relate the construction of the method (namely the selection of sample regions) with the water body characteristics it claims the index is designed for (low-albedo, mountain terrains, snow cover). The paper lacks a discussion section with a more detailed comparison between TWI and the other indices. The reference to the use of TWI to detect inundation and droughts at the end of the conclusions is out of place (a picture is included in Appendix A without further information). Specific points that the authors should address:
1- Please justify the image selection in line 136. Why did you select the images on the dates specified in Table 3? What was the reason, or at least the criterion?
2- Equation (1) seems to be wrong. The numerator and denominator are the same.
3- Section 3.3 - A better justification for the regions presented in Figure 1 is needed. What are the characteristics of the sample regions in terms of your claim that makes the TWI a better index (low albedo, rough terrain, snow cover, and so on)?
4- Line 213-214 – state the reason for excluding bands B1, B8A, and B10.
5- In Table 2 it would help if there was another column specifying the sensitivity of each one of these bands to the water body characteristics.
6- In section 3.3 a paragraph should be included describing what water body groups I and II represent.
7- In Figure 3 what is the x-axis?
8- Line 241 – “In Figures 2 and 3 surface reflectance values of the type I water body shows the highest mean valued in B3 and B5..”. Really? It seems to me that the ones having higher values are water 11 and 9 belonging to group II. Please explain better or correct.
9- Line 307 – Why do you sum the TWI1 and TWI2 at the final? Please give a justification.
10- Line 346 – The classification accuracy lacks a more detailed explanation. For instance, explain what kappa coefficients and confusion matrices are.
11- Line 365 – Comparative maps are exported based on mean values of Fig. 5 and Fig. 6. This needs to be explained. The whole 4.1 section needs more clarification.
12- Line 396 – This is repeated (the Otsu method is used to calculate a threshold) several times but is never explained how.
13- Line 468 – How do we know that the thresholds are stable if you never show them?
14- Lines 496 to the end – Something that is not explained or described should not be in the conclusions. This part of the work should either be omitted or presented in another section.
Author Response
Dear Reviewer:
Please see the attachment
Author Response File: Author Response.docx
Reviewer 3 Report
Major Remark
From the manuscript, the newly proposed water index (TWI) is constructed by two preliminary indices (TWI=TWI1+TWI2) and that's because you've divided all 11 water samples into two types. One possible major issue is that the sample sites are too concentrated (all located in Xinxiang City, China) and the number seems to be too small to be representative enough. A better approach would be to consider different properties of the water sample, such as water covered by snow and ice, salt content (saline or freshwater), suspended matter (clear or turbid) and some other hydrologic attributes. Moreover, it would be better if the sample sites were geographically uniform. As for the TWI index, which is simply added by TWI1 and TWI2, it would be more appropriately if we gave them the right weights. Or why you make the same weights? Please clarify this.
Minor Remark
1. p1 l28 missing % after 97.
2. There are some typos in this manuscript, including but not limited to line 189, 209, 383 and Table3. The city’s name is Xinxiang.
3. p2 l45 Using visual interpreters could be more appropriately.
4. p2 l46 Currently widely used are... This sentence is not properly expressed.
5. p2 l86 Maybe you could shift the focus from maintaining water resources to the key of this paper.
6. From Table 2, the spatial resolution of Band 5 and 6 (originally 20 meters) have been resampled to 10 meters as you described in order to be consistent with Band 2, what are the specific scientific reasons and why not apply same resample process to Band 12 which TWI also considered? Maybe you could present the original data information.
7. Formula (1) has a fundamental error: the denominator should be the sum of two terms, i.e. .
8. Figure 1: Please make sure the size of Sentinel-2A image of water body samples 1 to 11 is the same. And there is no explanation of mumber 208.
9. Figure 2: The beginning of the reflectance should be 0 in B8 bands.
10 Figure 4: The font sizes are not uniform.
11. The legend should contain all the content of the figure. For example, apart from water and non-water, there are still some unspecified regions in Figure 5.
12. For same content of the figures, it should be represented by the same legend (Weihui City Boundary versus weishihui in Figure A.2).
13. It should be style unified in References, and errors should be corrected (Reference 18 and so on).
Recommendation
1. It is recommended to expand the scope of your literature review. For example, you mentioned several classical machine learning models and pointed out the disadvantages caused by their complex algorithm, but you seem to ignore the convenience of semantic segmentation by convolutional neural network (CNN). I suggest you refer to but not limited to this paper, which has advantages of both effective performance and lightweight structure. (Liu W, et al. LaeNet: A Novel Lightweight Multitask CNN for Automatically Extracting Lake Area and Shoreline from Remote Sensing Images. Remote Sensing. 2021; 13(1):56. https://doi.org/10.3390/rs13010056)
2. It is recommended that all the formula in this manuscript be consistent in style, including the font size, font thickness and whether italicized or not.
3. It is recommended that the drawing style be consistent through all the manuscript (arrange the size and position of the scale and compass properly).
4. Some grammatical errors should to be corrected and English language and style may need minor polished.
Author Response
Dear Reciever:
Please see the attachment
Author Response File: Author Response.docx
