Evaluation of SAR and Optical Data for Flood Delineation Using Supervised and Unsupervised Classification
Round 1
Reviewer 1 Report
The paper is well-structured and easy to understand. The methods are also clearly described in the text. But there are also some parts that need to be explained more in the discussion as affecting the implementation of the methods, with the data and the specific classification method you used, as well as the results you got.
1. Make sure that first time you used an abbreviation, it is presented properly (with its full meaning in the text)
2. Section 3.2 Flood maps derived from unsupervised methods: Otsu and CThS methods
2.1 Figure 6. It is difficult to see exactly the border or boundary of the flooding when overlain on the true colour images. Will it be possible to increase the transparency of the background image or use a brighter blue colour (instead of black) to represent water, so that they can easily be distinguishable?
2.2 Figure 6. The results show the differences of floods mapped using the two different methods, but difficult to see exactly where they differed. A combination map showing the extents of flooding generated from both methods (overlain with each other) can give a better overview where they differed from. You can use two different colours when the two results are overlain.
2.3 Figure 6 – title - indicate the two different areas where the flooding happened in 2017 and 2020.
2.4 How big are difference in sizes between the two methods at the two study areas? – by this you can quantify your results better than by just providing visual results which can be difficult to know how big the differences are between the two methods.
2.5 Make sure that these are described in the results
3. Section 3.3 Flood maps with supervised methods: Random Forest classification
3.1 Same comments as in 3.2 a-d
4. Section 3.4 Evaluation of flood delineation
4.1 Do not forget to mention the number of samples again here in this part to have a clearer view of how big the values are when referring to percent.
4.2 Table 5. Aside from the years (2017 and 2020), include the study area where the study was applied to make it clearer that you are referring to two different flood events at different areas. This differences in the study areas have also implications in your results (refer to succeeding questions)
4.3 Lines 547-548- you mentioned: “The complexity of landscape units of the region of the event 2020 led to a less accurate water delineation using RF classifier than the event 2017.” – expound more on this. How and in what way does the complexity of the landscape (or as I would say the sites) have affected the results? This is important as it shows the heterogeneity in the results based on the two study areas.
4.4 How can the method used further increase the inaccuracy with complex landscape? Important part of the discussion.
4.5 Disparities between the two study areas are larger when SAR is used than MS. Why? – discussion
4.6 Why does SAR have higher disparity between the two study areas than MS? – discussion
4.7 Table 6. Lower RMSe for 2017 than 2020. How can this be explained? The same thing as with using SAR vs MS and among the different methods - – discussion
Author Response
Please see the attachment. Thank you.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
This is a very interesting manuscript that elaborates on using the adjusted DCT-PLS method to reconstruct the Sentinel-2 image in the case of cloud contamination. A minor revision is required to enhance the quality of this manuscript.
My concerns are listed as follows,
(1) The introduction section should be better organized in appropriate sequences. Firstly, the scientific importance/meaning of this study; Secondly, based on the literature review, what are the major research gaps between previous studies? Thirdly, what should be done to fill these gaps, and what are the research goals of this study?
(2) Why only two cases were selected for this study? What is the scientific importance of these two study areas?
(3) There are several popular machine learning-based classifiers such as the SVM, AdaBoost, RF, and deep learning algorithms. In section 2.3.3, why only the RF classifier was employed? A parallel comparison between different classifiers is expected.
(4) In Tabel 2, giving that the different backscattering intensity of S-1 VV and VH bands, it seems only one band with higher backscattering intensity should be used to reduce the complexity and the dimensionality of the dataset.
(5) Discussion section containing the new findings and limitations of this study must be elaborated.
(6) The conclusion needs to be refined to be more concise. It is somewhat too long.
(7)Overall citations are somewhat out of date and the latest references should be added.
Author Response
Please see the attachment. Thank you.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
The authors' effort in improving the quality of this manuscript is highly appreciated. I think this manuscript merit publication in its present form.
