Flood Detection in Polarimetric SAR Data Using Deformable Convolutional Vision Model

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The study is very interesting and the authors did good work. My comments as follows:

In the Introduction, Line 89, the authors mentioned the use of Deep Learning in other engineering fields. However, there are other studies used DL in flood forecasting as well. The authors are encouraged to include such studies. To name few: A review of hybrid deep learning applications for streamflow forecasting, Predicting streamflow in Peninsular Malaysia using support vector machine and deep learning algorithms.

The authors mentioned the use of Transformer but has not include related studies where TNN was applied in hydrology. Please refer to A novel application of transformer neural network (TNN) for estimating pan evaporation rate.

"Figure 3. The structure of Feature Extraction Module." Not clear, please improve. 

"Figure 4. The structure of Deformable Convolution v3." Why not the Input and Output features have not been shown in the Figure. 

The reviewer feels that the Methodology section is very lengthy. 

2.3. Experimental Metrics, The authors can refer to other studies to show that these metrics are widely used. Maybe authors can refer to Streamflow classification by employing various machine learning models for peninsular Malaysia

"Figure 10. (a)" what does Bg means?

" AdamW algorithm", can the authors show some studies prove this is better than other optimizers. 

"floating-point operations" can you please add to the manuscript how it calculated. 

"Figure 12. The extract results of FWSARNet in the testing region." Not really clear to the reader why the authors include the figure here. 

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Firstly, I want to thank the authors for their efforts in making their manuscript. Globally, the detection of flood extent presents a challenging area of research, particularly in arid and/or urban regions. Introducing effective methods to realize accurate spatial extraction of flood boundaries has gained considerable attention from many researchers worldwide. This interest is driven by the widespread negative impacts of floods, especially in the context of escalating climatic changes over the last few decades.

``The manuscript titled " Flood Detection in Polarimetric SAR Data Using Deformable Convolutional Vision Model" aims to introduce an advanced flood mapping-based deformable convolutional vision model using Sentinel-1 Synthetic Aperture Radar (SAR) imagery as its primary data source. The authors used the SAR datasets from the ETCI 2021 flood dataset to train and evaluate the performance of their developed FWSARNet model. In addition, they compared the intensity values between two polarization channels (i.e., |VV| and |VH|) to improve the accuracy and effectiveness of their FWSARNet algorithm for flood detection. Furthermore, they utilized multi-level feature map fusion in the FWSARNet model's decoder part to enhance the extraction of detailed flood spatial data from SAR images. For the evaluation of their proposed model, the authors employed various semantic segmentation metrics, including Intersection over Union (IoU), mean Intersection over Union (mIoU), F1-score, precision (P), and recall (R). These metrics evaluated the effectiveness of the FWSARNet model in accurately detecting flood-prone regions in SAR imagery.

I have some concerns, comments, and suggestions concerning the manuscript:

Regarding the Introduction Section:

·       In lines 41-44, please clarify whether these flood events occurred globally.

·     In line 72, please provide definitions for the abbreviations (NOAA GMU, SNPP/VIIRS).

·  I highly recommend incorporating more comprehensive details about applying various machine learning techniques, particularly deep learning models, in mapping flood-prone regions.

·    It would be beneficial to provide a more in-depth insight into the main reasons for selecting the deformable convolutional vision model for detecting flooded areas over other deep learning models.

·  The presence of floodwater can lead to a decrease in backscatter intensity in both VV and VH channels. It would be valuable to cite additional information on how the |VV|/|VH| ratio can be effectively utilized to improve flood boundary detection.

 Regarding the Methods Section:

· I recommend including a flowchart that comprehensively overviews the entire methodology.

· For the evaluation metrics, Precision, Recall, and F1 Score are commonly utilized for assessing classification problems, while Intersection over Union (IoU) and Mean Intersection over Union (mIoU) are particularly relevant for evaluating object detection and semantic segmentation tasks. It would be beneficial to provide more detailed explanations on why these specific metrics were chosen to assess the performance of the FWSARNet model.

· Could you explain how you addressed the substantial computational requirements and managed numerous parameters while utilizing the deformable convolutional vision model to detect flood-prone areas?

·   In lines 505-506, you mentioned that the findings of the FWSARNet model were compared to other state-of-the-art models applied to the ETCI2021 dataset, resulting in significant performance improvements. To strengthen your statement, consider citing the published papers or reports that have utilized the ETCI2021 dataset for mapping floods.

 Regarding the Experiments Section:

·       Could you consider adding latitude and longitude information to the maps in Figures 7, 8, and 9?

Regarding the Discussion Section:

·  It would be valuable to engage in a more comprehensive discussion of the major limitations encountered in your study, such as the substantial computational requirements and numerous parameters in the deformable convolutional vision model, the presence of striping artifacts in some of the ETCI2021 datasets, and the insufficient availability of multi-temporal SAR images over several years for accurate flood boundary detection.

· While you conducted a generalization experiment in Hebei Province, it is notable that the robustness and performance of your FWSARNet model were not evaluated using other SAR datasets from diverse flood-prone regions worldwide, especially in challenging arid and/or urban regions severely affected by significant flood events. Could you elaborate on your expectations regarding the advantages and disadvantages of using your model to detect floods in arid and/or urban regions?

·   I highly recommend relocating the paragraph in lines 504-518 to the Discussion section.

 

·    I recommend moving the “4.2. Optimizer analysis” section in lines 545-606 to the Results section.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The manuscript is well written and structured in flow. I have following comments to further improve its quality:

1.       Table 1 demonstrates a reasonably low accuracy for ViT even lower than ResNet which does not seems to be right. Can authors subjectively expand on this behaviour within the abstract and justify this performance.

2.       The training results could have been compared in many other ways as well including training loss plots, etc

3.       It would be nice to clearly discuss the limitations of the proposed model and corresponding future research directions.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors did improve the paper. I am inclined to accept it.