Mapping Pluvial Flood-Induced Damages with Multi-Sensor Optical Remote Sensing: A Transferable Approach

Round 1

Reviewer 1 Report

The paper entitled "Mapping pluvial flood-induced damages with multi-sensor optical remote sensing: a transferable approach using change detection, very high spatial resolution and machine learning" shows an interesting use of ML techniques (RF and CNN) on high-resolution images to detect pluvial floods. The manuscript is well-written and well-articulated. Some suggestions:

- INTRODUCTION: There is no mention to physically-based models for the space-time prediction of rainfall-induced shallow landlides. Please, include some recent efforts as: https://doi.org/10.1016/j.compgeo.2022.105175 https://doi.org/10.5194/nhess-18-1919-2018 

-CONCLUSION: Lines 781-792 should be shortened for a more concise paragraph

 

Author Response

The paper entitled "Mapping pluvial flood-induced damages with multi-sensor optical remote sensing: a transferable approach using change detection, very high spatial resolution and machine learning" shows an interesting use of ML techniques (RF and CNN) on high-resolution images to detect pluvial floods. The manuscript is well-written and well-articulated.

Thank you for this comment.

Some suggestions:

• INTRODUCTION: There is no mention to physically-based models for the space-time prediction of rainfall-induced shallow landlides. Please, include some recent efforts as: https://doi.org/10.1016/j.compgeo.2022.105175; https://doi.org/10.5194/nhess-18-1919-2018.

This is a good and relevant comment and addition to the introduction, thank you. References were added to the manuscript.

• CONCLUSION: Lines 781-792 should be shortened for a more concise paragraph

The beginning of the conclusion section was shortened for better concision.

Reviewer 2 Report

The manuscript entitled "Mapping pluvial flood-induced damages with multi-sensor optical remote sensing: a transferable approach using change detection, very high spatial resolution and machine learning", presents an interesting work with a great aspect of significant novelty. The manuscript is well-written and clearly organized. The text is very clear and provides a unique and easy approach to the comprehensive framework for mapping pluvial flood-induced damages.

Although I am sure that this process has been carried out rigorously and the results will be of interest to all scientific community, I am convinced by the applicability of the study to subsequent studies, given that it deals with pluvial flood-induced damages and does explicitly make broader contributions to the assessment field.

This study introduces a brand-new automatic remote sensing technique termed FuSVIPR for Fusion of Sentinel-2 & very high resolution imaging for pluvial runoff. With regard to pluvial floods that are not close to active streams, this strategy is intended to close the identification gap for flood-related deteriorations. The novelty of this study is not in the use of these two approaches, which are frequently used in Machine learning/Deep learning classification of remote sensing data, but in the application of these approaches to precisely identify pluvial flood-induced deteriorations. The main point of this study's originality is its goal of great generalizability, or the classifiers' excellent performance on instances and locations for which they were not taught.

All citations used in the article are appropriate and were created under the publisher's guidance.

I would like to propose some small suggestions to the author(s) of the article.


- The title of the article is very long, if possible try to shorten it.

- The article does not directly address several instances of pluvial flood-induced impacts, such as soil erosion, gullies, landslides, and mudflows located farther away from the stream. The damage caused by the aforementioned pluvial flood is not depicted in the images of the research results. It would be better to present them in various colors on article figures with a clear legend.

- Try to further describe the applied machine learning methods used in the research.

- Although the conclusion section is in line with the data and arguments made, it includes many superfluous phrases detailing pluvial floods. We should only discuss the outcomes and inferences that can be made from those outcomes.

Author Response

The manuscript entitled "Mapping pluvial flood-induced damages with multi-sensor optical remote sensing: a transferable approach using change detection, very high spatial resolution and machine learning", presents an interesting work with a great aspect of significant novelty. The manuscript is well-written and clearly organized. The text is very clear and provides a unique and easy approach to the comprehensive framework for mapping pluvial flood-induced damages.

Thank you for this comment on the quality of organization and redaction.

Although I am sure that this process has been carried out rigorously and the results will be of interest to all scientific community, I am convinced by the applicability of the study to subsequent studies, given that it deals with pluvial flood-induced damages and does explicitly make broader contributions to the assessment field.

This study introduces a brand-new automatic remote sensing technique termed FuSVIPR for Fusion of Sentinel-2 & very high resolution imaging for pluvial runoff. With regard to pluvial floods that are not close to active streams, this strategy is intended to close the identification gap for flood-related deteriorations. The novelty of this study is not in the use of these two approaches, which are frequently used in Machine learning/Deep learning classification of remote sensing data, but in the application of these approaches to precisely identify pluvial flood-induced deteriorations. The main point of this study's originality is its goal of great generalizability, or the classifiers' excellent performance on instances and locations for which they were not taught.

Exactly, the main point is the generalizability to other areas and pluvial flood events, and also the possibility to use of different types of sensors for the VHR post event image (not only Pléiades satellite images).

All citations used in the article are appropriate and were created under the publisher's guidance.

I would like to propose some small suggestions to the author(s) of the article.

• The title of the article is very long, if possible try to shorten it.

We shortened it a little bit. Thank you!

• The article does not directly address several instances of pluvial flood-induced impacts, such as soil erosion, gullies, landslides, and mudflows located farther away from the stream. The damage caused by the aforementioned pluvial flood is not depicted in the images of the research results. It would be better to present them in various colors on article figures with a clear legend.

We are not sure to understand your comment fully. The FuSVIPR method specifically identifies these kinds of damages. However, it does not distinguish between the different types, i.e. it only indicates “damage” or “no damage” in the final maps.

So we are guessing that you mean that we do not provide a map indicating “this is a landslide”, “this is an eroded soil” etc.? In this case, we added a mention for this in the beginning of section 2.4.:

“Hence, in this work, the developed classifiers were not designed to distinguish the different types of deteriorations associated to intense overland flow (landslides, mud deposits, debris flows, eroded soils, gullies etc.) in a multi-class framework. They only indicated the presence or absence of one of these PF-induced damages.”

• Try to further describe the applied machine learning methods used in the research.

Thank you for this comment. We understand your request. The U-net algorithm is already fully described in 2.4.2. For the Random Forest algorithm in 2.4.1, it is less developed but references were added. As the manuscript is already quite lengthy, we chose to limit the description for Random Forest as it is already used a lot in the literature.

• Although the conclusion section is in line with the data and arguments made, it includes many superfluous phrases detailing pluvial floods. We should only discuss the outcomes and inferences that can be made from those outcomes.

Indeed we shortened the beginning of the conclusion to remove superfluous phrases that are already exposed in the introduction. Thank you.

Reviewer 3 Report

 

Happy to read this article “Mapping pluvial flood-induced damages with multi-sensor optical remote sensing: a transferable approach using change detection, very high spatial resolution and machine learning”. I appreciate the authors for making a very good effort to map the pluvial flood damages. But I think still there is a need for significant changes for the improvement of the manuscript. I recommend this paper should be accepted, after the suggested revisions.    

Abstract: It is not a good practice to write abbreviations in this section. Also, the first few sentences should be summarized in such a way as to be specific. I suggest please revise your abstract significantly from lines (13-27), as it is really hard to understand. There is a serious scientific writing issue in the entire abstract section.

Introduction: I would suggest that you can clarify your problem statement. It is a mix of literature and certainly no coherence at all. I did not see what is significant in your work in this section. In other words, it is a literature review only. A lack of scientific writing also exists in this section too.

Materials and methods

Why you did not use Radar imagery? Radar imagery may be useful to estimate multisector damages with reliable accuracy in the post-flood instance.

Line 202+203

The Images used for the Durban event are obtained from a different source as used for Aude 1+2 and Alpes Mar, the results may show inconsistency due to the use of different sources as a single source may be more beneficial.

Line # 224-227+ 230

The use of top-of-atmosphere TOP radiance for the ultimate production of TOP reflectance depends on solar variability, this can limit the data collection for appropriate mapping and image production and need a clear sky for the accurate source of imagery, as you have to use optical imagery.

Line # 245

Calibration techniques use highly controlled conditions for accurate and precise outcomes and need well-equipped techniques

Line 303-305

I don’t think, such information is important to mention here.

Table 1

The satellite data used for Mars-Alps is different from that used for other events, this will result in different data outcomes, and also the use of spatial resolution varies. The source of satellite data and spatial resolution used should be the same for all events for the effective study if possible. 

Line # 309-315

The ground truth data and validation used for all events vary from each other that require an extensive process to detect accurate results.

Line # 317-318+320

In the Durban event case, no photo-interpretation was carried out and ground truth points were not directly taken by an appropriate method.

Line # 341

The two supervised learning pixel-based approaches were used and it is also suggested that the soil type detection technique should be included which is lacking and it is important as the damages were studied related to land i.e. soil erosion, land sliding etc.

Results

Line # 492

Both the RF and U-net classifiers were trained only on the Aude 1 learning sample provides limited information and other events' damages will not be properly evaluated according to the objective. All damages should be properly evaluated, this section needs significant improvements.  

Line # 530

Using the SPCD method results in higher uncertainty in change detection, as  in the spring season which may lead to inconsistency of data, results, and accuracy.

Line # 547-548

S-2 change information and model training performances were lower for both methods showing the inconsistency of your results

Line # 562

The RH data is comparatively better than U-net in terms of ADR-based detection rates, RH results obtained will not provide accurate outcomes and will not support decision-making later for implementation of your results.

Line # 564-565

The addition of SWIR change information did not seem to significantly improve the performance of either RF or U-net, showing high contrast of data obtained from the RF or U-Net approach.

Line # 586

Cross-validation results show less accurate outcomes

Line # 607

Using U-net for object identification gives inverse results.

Figure # 9: Line # 639

Overflowing damages are detected by the FuSVIPR method (RF only) with no account for accurate assessment of final thematic images

Line 653-655

The RF and ADR methods used make no sense in obtaining authentic information

Line # 658-659

The methods i.e. U-Net and RF used are not satisfactory in obtaining information

Line # 691-692

Final performances from ML and DL classifiers roughly exceeded 75% detection rates, showing low detection rate as compared to other studies.

Line # 797

Satisfactory detection rates (~ 75%) by using FuSVIPR are not satisfactory as compared to other studies

 

Conclusion:

In this section, the scientific contributions should be exhibited, rather than the repetition of results.

General views

1.      Complex nature of remote sensing techniques used in the study

2.      Accuracy assessment does not show the authenticity of results as compared to other studies

3.      Accuracy methods and remote sensing techniques used show less accuracy

4.      Final data obtained is not satisfactory to supporting objectives effectively

5.      Land or soil texture is important to study flood-related damages, this section is not considered in this study 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 3 Report

I am happy with current changes. indeed, the accuracy of the results are not satisfactory. Also, the scientific contribution of this work is not clear to me. 

 

Author Response

Thank you very much for taking the time to review our work. There are some good comments and feedbacks for improvement of the manuscript. However, your review report is globally concerning, as it seems that you didn’t understand multiple key aspects of the developed method.

We still agreed to address your main suggestion at the end of the discussion subsection 4.2:

“The PF damage detection rates reached by FuSVIPR (~ 75%) turned out to be lower than what other related studies on flood damages and rainfall-induced landslides have found (~ 85% to 90%, [29,30]). This however has to be alleviated considering that our results are valid on four different events with a classifier trained on only one of the events. Indeed, most classification methods featuring higher detection rates (85% and more) are usually tested on the same event (and thus the same remote sensing images) that they were trained on. In particular, for instance, the SPCD method had also been found to reach up to 90% detection rates when trained and validated on only one event [32] before these accuracies were updated and lowered to ~ 70% after performing transferability assessment over two additional events [27].”

We believe this paragraph answers your comments regarding the performance of our method.

Also, we shortened the title as suggested.

Yours sincerely,

The Authors