Identifying Dynamic Changes in Water Surface Using Sentinel-1 Data Based on Genetic Algorithm and Machine Learning Techniques

Round 1

Reviewer 1 Report

Comments

• For more clarification, it is preferable to make an illustrative chart or schematic of the steps in which the research was carried out, including all the methods used.
• In Figure 1: Please add the coordinates for Sites A, B, C, and D
• Please replace Fig 3 (a, b, c) with a more clear plot, there are many black dots around the text and lines in this figure. Doesn’t look professional
• In Figures 5-8, please add the coordinates.
• Please define the symbols (TP, TN, Pe,…. ) in the equations immediately after the equation.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Please see attached file.

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

1. The revised paper is basically unreadable since all the changes are included directly in the manuscript.  The journal should require a finalized version of the manuscript along with a version showing the changes.  As it stands, the revised version of the paper that I am reading is a mess. Perhaps I missed a clean version, in which case this isn’t an issue.
2. The authors keep mentioning irrelevant information about the software they used. For example, why is it important that they used GDAL v. 2.2.2 and Python v3.6?  This is equivalent to saying they used Microsoft Word to write the manuscript.   I guess this is fine, as this approach seems to be prevalent among articles dealing with remote sensing in general.  I would much prefer to see the authors actually describe their methods instead of listing the software they used to avoid this “black box” approach to data analysis.
3. The authors clarify their use of 3x3 grids in the revised manuscript; however, in my opinion their approach to “eliminate the impact of data imbalance” actually magnifies the issue. For example, if there are 2,118 water samples and 54,928 non-water samples in Site A, then why take 2,000 random samples from each?  All this means is that the water grids are drastically over sampled (94.4% are used in the analysis) while the non-water grids are undersampled (3.6% are used in the analysis).  This seems like a fairly substantial imbalance in the data.
4. The authors still maintain that choosing “pure” samples means the results are realistic; however, choosing samples without any boundary effects (which is inherently non-random) means the underlying data are already biased against areas with the same boundary effects. This is clearly shown in the results whereby areas along the land-water interface are not always properly identified.  The authors provide several references in Response 15 showing that pure sampling is practical, so although I still have some issue with this approach, the literature justifies the method.
5. In Response 11, the authors state that the optimal combination of hyper-parameters varies with data sets. This is exactly my point, and the reason why I suggested testing this approach using other data (not subsets of the same data).  It’s no real surprise that different hyper-parameters are produced using different optimization methods, but what is interesting is to see if the resulting ML models are applicable to other Sentinal swaths.  In this research the models are only applicable to a single swath of Sentinal data; therefore, the results aren’t wrong they’re just highly limited.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf