A Novel Fully Automated Mapping of the Flood Extent on SAR Images Using a Supervised Classifier

Round 1

Reviewer 1 Report

The authors present an iterative training algorithm to classify flooded waters on flood SAR imagery based on supervised classification of pre-flood SAR and Sentinel-2 imagery.

In principle I like this paper. It proposes (yet) another method for rapid SAR flood classification but it deserves publication in my opinion as it adds to existing scientific literature on methods to classify flooding on satellite imagery.

The paper is very well written and is scientifically sound. I would suggest adding the following to the paper:

- Please be careful when using the term "fully automated" when the training is supervised in the sense that there is a selection of wet and dry pixels for the learning algorithm. In that sense I do not really agree that there is no operator intervention if the training classifier is supervised (line 274 for example). I suggest replacing fully automated throughout the manuscript and the title with quasi-automated for example;

- Please also be aware that for most flood mapping CSK and TerraSAR or not free and available to most users and applications, these two sensors do not provide open data. Only under Copernicus and NASA are data really truly open. I do however see the point and value in the paper of showing the Tewkesbury demo but really prefer the Myanmar case of S1 and S2. This restriction on data access should be clearly noted;

- there should also be a small paragraph somewhere that shows more clearly the value of the method proposed over existing fully automated methods as for instance proposed by Matgen et al (2011) included in the reference list. That would be of value to this paper I think.

Author Response

Dear reviewer,

Please see the attached word document for the responses to the points addressed.

Thanks.

Best regards,

Hakim Benoudjit.

Author Response File: Author Response.docx

Reviewer 2 Report

(- Line 126) Review the research papers only related to the technique proposed by the author and describe each pros and cons.
(Line 136-144) The novelty suggested by this paper is that NDWI is used as to construct a training dataset in supervised classification. However, studies that distinguish between flood and non-flood by learning have already been done.
  - Potential and Limitations of Open Satellite Data for Flood Mapping (Notti et al, Remote Sensing, 2018)
  - A Neural Network Approach to Flood Mapping Using Satellite Imagery Skakun, Computing and Informatics, 2010
(Line 147-161) The experimental period is 2007, which is the 2016 image, 9 years after the optical image used in the training. It is necessary to confirm the no-change of the study area for 9 years.
(Line 174-222) If it is not original content, author(s) can reduce the description. The readers are likely to be confused.
(Line 195) There are many classifiers, but there is no clear reason for choosing SGD.
(Line 234) Write the definition of "regularization item".
(Line 300-345) Is it effective to train only for water and land? How about adding a vegetation class?
(Line 300-353) In 3.3.2 Extraction of the training dataset, you separated the pre-flood image into water and land classes using NDWI and trained the model using SGD. Then, in 3.3.3 Classification of the post-flood SAR image, you classified the post-flood image as flood and non-flood classes using the trained classifier. However, since the model is trained as water and land, the result of the post-flood image should be classified as water and land. If you want to obtain classification results in flood and non-flood, you should train with flood and non-flood instead of water and land. Furthermore, if you have decided to change of class (Land –> Water) as flood, is it uncertain whether it is due to actual change or flood effects?
(Line 386-408) Authors do not have to describe the method of accuracy analysis. This can be described shortly by adding reference.
(Line 505-520) The processing time has been described as being very short, and a reference has to be added that this amount of time is helpful for prioritizing tasks during rescue operations.

Author Response

Dear reviewers,

Please see the attached word document showing the points addressed.

Thanks.

Best regards,

The authors.

Author Response File: Author Response.docx

Reviewer 3 Report

Dear Editor/Authors

The manuscript entitled "A Novel Fully-Automated Mapping of the Flood Extent on SAR Images Using a Supervised Classifier", by A. Benoudjit and R. Guida, presents an excellent work with some aspect of significant novelty. The manuscript is well written and clearly organized. However, I have some comments and suggestions to improve the current version of the manuscript. In general, the manuscript should be acceptable for publication but some minor correction must be done prior to publication. 

Firstly, as a research paper, this submission needs to critically assess work previously carried out in the field. Although this has been done to a limited extent in the introduction, some key points are missed. Perhaps the most significant articles in the research field, similar worldwide researches etc.

Secondly, it is necessary to add the results of the accuracy assessment of the proposed model of the automatic classification to the abstract, because only in line 12 states that the applied algorithm offer a good compromise between computation time and precision.

In conclusion, the authors could describe and suggest the possibility of applying the proposed classification methodology to other radar and optical satellite images.

Best regards

Author Response

Dear reviewer,

Please see attached the addressed points in a word document.

Thanks.

Best regards,

The authors.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

Everything tbe reviewer pointed out was complemented.