Round 1

Reviewer 1 Report

Reviewer

 

MDPI Remote Sensing

 

Manuscript Number: ID 1451301

New Manuscript Number: ID 1539939 – revised version of ID 1451301

 

Title: Sentinel-1 spatiotemporal simulation using Convolutional LSTM Flood Mapping

 

 

As requested, I have reviewed the revised version of the above-titled paper for potential publication in the Remote Sensing – MDPI Journal. I divided my comments in the sections presented as follows.

 

 

 

Contribution

 

This paper proposes the analysis of Sentinel 1 SAR images (10-m spatial resolution and 12 days temporal revisit) for evaluation of flood impact focusing on change detection algorithms.

 

In this sense, the authors raise the question that errors derived in such type of approach are then expected due to the fact that not all land cover changes are flood induced. The authors propose the use of historical SAR images to improve the detection of flood-induced changes using the Long-Short Term Memory (LSTM) method (e.g. Recurrent Neural Network – RNN).

 

More specifically, the authors use the Convolutional Log-Short Memory (ConvLSTM) based on deep learning spatiotemporal simulation network for producing a synthetic image using Sentinel 1 intensity time series in order to capture both spatial and temporal dependencies present in the datasets.

 

The objective is to produce a flood proxy mapping framework using tools provided by big data cloud and Google Earth Engine. The proposed experiment for evaluating  pairs of images pre /post- flood event in order to evaluate the corresponding impact included 4 models, namely: (i) Baseline model 1 (B1) – uses the last pre-event image before the flood event; (ii) Baseline model 2 (B2) uses the historical mean as the pre-event image; (iii) RNN LSTM model (L1); (iv) ConvLSTM  model (L2); The last two options propose to generate the synthetic image as the pre-event image for change detection.

 

Just to mention, the authors also refer to previous studies using a texture-based Bayesian probability approach for flood mapping employing the normalized difference sigma-naught index (NDSI) and Shannon´s Entropy of NDSI (SNDSI) (e.g Ulloa et al. , 2020 – MDPI Remote Sensing – case study: Mozambique). They mention the use of such references to evaluate some aspects related to the present work.

 

Three case studies have been studied in this work: (i) Australia – cyclone Debbi on March 28, 2017; (ii) Mozambique – cyclone Idai on March 14, 2019; (iii) Brumadinho – Minas Gerais state, Brazil – collapses of an iron ore tailing dam on January 25, 2019. In the case of Brumadinho, Sentinel 1 Gamma naught radiometric corrected data from Alaska Satellite Facility (ASF) was used, while Sentinel 1 Ground Range Detected (GRD) sigma naught data from Google Earth Engine was used for Australia and Mozambique. The authors justify that option due to radiometric effects due to the terrain, since that issue is more relevant in the case of Brumadinho-Brazil.

 

The authors discuss some differences among the three case studies including the potential effect of land cover (Figure 9 for Australia and Mozambique) and  topography for the case of Brumadinho based on VV and VH polarization. For evaluating the performance, the authors used Sentinel-2 derived water mask, while the affected zone due to the dam break in Brumadinho was manually digitized from 4-m spatial resolution Planet Labs images.

 

They claim that the research results are able to support the evidence of the good performance of the proposed  methodology. In addition, the authors foresee the extended use of the adopted approach.

 

I found the manuscript has an interesting goal to be pursued and presents the conditions to be published in the Remote Sensing MDPI Journal.

Please, see further comments in the attached file. 

 

Comments for author File: Comments.pdf

Author Response

We would like to take this opportunity to thank the reviewer once more for taking the time to review our manuscript. His insightful and thorough comments and suggestions haven significantly improved the quality of our manuscript. Please find our replies in the attached document, highlighted in red font. 

 

Author Response File: Author Response.docx

Reviewer 2 Report

I appreciate the authors' hard work for the improvement and correction of their work. At present I have no specific objections for the acceptance of the manuscript for publication. The minor additional tuning of the text and figure description might upgrade their product.

Author Response

We would like to thank the reviewer for taking the time to review our manuscript. Please find our replies in the attached document. 

Author Response File: Author Response.docx

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

Reviewer

 

Remote Sensing

 

Manuscript Number: ID 1451301

 

Title: Sentinel-1 spatiotemporal simulation using Convolutional LSTM Flood Mapping

 

 

As requested, I have reviewed the above-titled paper for potential publication in the Remote Sensing – MDPI Journal. I divided my comments in the sections presented as follows.

 

 

 

Contribution

 

This paper proposes the analysis of Sentinel 1 SAR images (10-m spatial resolution and 12 days temporal revisit) for evaluation of flood impact focusing on change detection algorithms.

 

In this sense, the authors raise the question that errors derived in such type of approach are then expected due to the fact that not all land cover changes are flood induced. The authors propose the use of historical SAR images to improve the detection of flood-induced changes using the Long-Short Term Memory (LSTM) method (e.g. Recurrent Neural Network – RNN).

 

More specifically, the authors use the Convolutional Log-Short Memory (ConvLSTM) based on deep learning spatiotemporal simulation network for producing a synthetic image using Sentinel 1 intensity time series in order to capture both spatial and temporal dependencies present in the datasets.

 

The objective is to produce a flood proxy mapping framework using tools provided by big data cloud and Google Earth Engine. The proposed experiment for evaluating  pairs of images pre /post- flood event in order to evaluate the corresponding impact included 4 models, namely: (i) Baseline model 1 (B1) – uses the latest pre-event image before the flood event; (ii) Baseline model 2 (B2) uses the historical mean as the pre-event image; (iii) RNN LSTM model (L1); (iv) ConvLSTM  model (L2); The last two options propose to generate the synthetic image as the pre-event image for change detection.

 

Just to mention, the authors also refer to previous studies using a texture-based Bayesian probability approach for flood mapping employing the normalized difference sigma-naught index (NDSI) and Shannon´s Entropy of NDSI (SNDSI) (e.g Ulloa et al. , 2020 – MDPI Remote Sensing – case study: Mozambique). They mention the use of such references to evaluate some aspects related to the present work.

 

Three case studies have been studied in this work: (i) Australia – cyclone Debbi on March 28, 2017; (ii) Mozambique – cyclone Idai on March 14, 2019; (iii) Brumadinho – Minas Gerais state, Brazil – collapses of an iron ore tailing dam on January 25, 2019. In the case of Brumadinho, Sentinel 1 Gamma naught radiometric corrected data from Alaska Satellite Facility (ASF) was used, while Sentinel 1 Ground Range Detected (GRD) sigma naught data from Google Earth Engine was used for Australia and Mozambique. The authors justify that option due to radiometric effects due to the terrain, since that issue is more relevant in the case of Brumadinho-Brazil.

 

The authors discuss some differences among the three case studies including the potential effect of land cover (Figure 9 for Australia and Mozambique) and  topography for the case of Brumadinho based on VV and VH polarization. For evaluating the performance, the authors used Sentinel-2 derived water mask, while the affected zone due to the dam break in Brumadinho was manually digitized from 4-m spatial resolution Planet Labs images.

 

They claim that the research results are able to support the evidence of the good performance of the proposed  methodology. In addition, the authors foresee the extended use of the adopted approach.

 

I found the manuscript has an interesting goal to be pursued, however the text should be revised to provide the reader with more details with respect to the methodology used and the corresponding workflow of the work for achieving the results shown in the manuscript.  Assumptions should be more clearly presented along the manuscript.

 

There are some gaps along the text with respect to the methodological approach and evaluation of the datasets used. Statistical analysis is poorly presented. To use only Kappa for evaluating results for such type of work is not much. There is a requirement to present more in-depth statistical analysis. There is also a need to produce more results and including more Figures and Tables to better explain and justify the methodology used. Some further comments are going to be raised in the next section with respect to the points raised.

 

In general, the text is well written and organized. Just a couple of mistypes (for example, please use km² instead of Km² remain in the manuscript. In this sense, the manuscript is generally fine to read.  

 

Therefore, the more specific comments and questions regarding the manuscript paper are going to be presented jointly with the evaluation of the contribution of the manuscript in the next section. They are provided below with more details with respect to the methodological approach. I think it would be interesting to have some feedback from the authors and also to have fully revised version of the text in order to better refer to the points I will raise in the next paragraphs. That might also lead to explore or reflect about different scenarios still not well and thoroughly explored by the authors in the proposed manuscript but that deserves attention.

Please, see further comments in the attached file.

Comments for author File: Comments.pdf

Reviewer 2 Report

Please, consult my enclose.

Review of the article titled as “Sentinel-1 spatiotemporal simulations ~ “.

 

General mark:

There exist many unordered features in the main text. From such a mismatch, I am very reluctant to assign any reliable credits to their analysis. My best recommendation is that the authors withdraw this article and rewrite and resubmit again.

 

Some details:

Line 162: typo

on March 28^th, 2017 at approximately  >> on March 28^th, 2017 at approximately

 

Line 166-167:

the recorded accumulated precipitation >> the recorded accumulation of precipitation

 

Line 262: typo

ConvLSM >> ConvLSTM

 

Line 271-302

Subsection “5.1.1 Australia Flooding”, it is very confusing and also somewhat dubious in the analysis of figures 2 and 3 and the explanation of the figures in the main text seems to be the wrong figure numbers. The fact makes me hard to believe the construction of the figures themselves.

 

Line 450

Long-Short Term Memory and --- >> Long Short-Term Memory ?

Long-Short Term Memory or Long Short-Term Memory? Which one is the expression you want to use?