Sentinel-1 SAR Images and Deep Learning for Water Body Mapping
Round 1
Reviewer 1 Report
Summary/Contribution: This research describes an approach for categorizing flooded areas using synthetic aperture satellite photos, the U-NET neural network, and the ArcGIS platform. Los Rios, a region of Tabasco, Mexico, is the focus of the research. Despite the minimal number of training samples, the results suggest that U-NET performs well. Its accuracy rose as training data and epochs increased.
Comments/Suggestions:
1. I would like to request that you consider summarizing the related work section in tabular form as a way to better identify the limitations of the related works and emphasize the originality of your proposed approach.
2. How does the concentration of rainfall in certain periods affect the hydrology of the region, and what are the implications for water management and infrastructure?
3. What other technical factors were taken into account in the selection and processing of the SAR images, such as incidence angle or spatial resolution?
4. How does the methodology of the research compare to other studies that have used SAR images for mapping flood extent or vegetation?
5. To improve the study's quality and effect, I recommend including a paragraph on formal methodologies for AI-based technique verification. Formal approaches, which use mathematical models and logic to validate system correctness, are becoming more important in the development and validation of AI-based techniques.
6. Some important sources to this topic that the authors should consider include:
a. https://ieeexplore.ieee.org/abstract/document/9842406
b. https://incose.onlinelibrary.wiley.com/doi/abs/10.1002/inst.12434
7. Explain how the generation of RGB and binary layers improves the accuracy and resolution of the images, and provide examples of potential applications of these layers beyond training the deep learning model.
8. Describe how the 2020 flood map generated by SINAPROC complements the binary layer and training samples, and provide information about potential limitations or sources of error associated with the 2020 flood map.
9. What are some potential sources of error or bias associated with using SAR images from a specific time period, and how were they addressed in the research?
10. How were the image chips and labels captured in the training process, and what were the criteria for inclusion or exclusion?
11. Provide information on how the parameters of epochs, batch size, and chip size were selected and adjusted in the ArcGIS platform, and describe the potential effects of these parameters on the accuracy and reliability of the trained model.
12. Describe the methodology of using object segmentation and classification in satellite imagery for mapping flood extent and vegetation, and provide a comparison with other approaches such as object-based analysis or traditional image classification.
13. What are the main limitations or sources of error associated with using the same deep learning model for repeated classification of flooding in the same geographical area?
14. How were the training parameters established, and what were the potential effects of these parameters on the training and validation loss?
15. How does the methodology of using deep learning methods for pixel classification in water body mapping compare to other approaches, such as object-based analysis or traditional image classification?
Must be improved
Author Response
Response to comments/suggestions. Reviewer 1
1. I would like to request that you consider summarizing the related work section in tabular form as a way to better identify the limitations of the related works and emphasize the originality of your proposed approach.
Observation attended
Table 1 was added with important data from different related works. New references were also added.
2. How does the concentration of rainfall in certain periods affect the hydrology of the region, and what are the implications for water management and infrastructure?
Observation attended
Added two paragraphs in “introduction” section 3. What other technical factors were taken into account in the selection and processing of the SAR images, such as incidence angle or spatial resolution?
Observation attended
4. How does the methodology of the research compare to other studies that have used SAR images for mapping flood extent or vegetation?.
It is difficult to make a comparison because the study areas are very different. Each zone has a different morphological structure. Although the methodologies use the same types of images from the same satellite and the same type of DL, the results can be very different.
5. To improve the study's quality and effect, I recommend including a paragraph on formal methodologies for AI-based technique verification. Formal approaches, which use mathematical models and logic to validate system correctness, are becoming more important in the development and validation of AI-based techniques.
Observation attended
Added U-Net description
“Model evaluation" section added 6. Some important sources to this topic that the authors should consider include:
Several references were considered
7. Explain how the generation of RGB and binary layers improves the accuracy and resolution of the images, and provide examples of potential applications of these layers beyond training the deep learning model.
Observation attended
8. Describe how the 2020 flood map generated by SINAPROC complements the binary layer and training samples, and provide information about potential limitations or sources of error associated with the 2020 flood map.
Observation attended 9. What are some potential sources of error or bias associated with using SAR images from a specific time period, and how were they addressed in the research?
The following paragraph was added in conclusions:
However, SAR images may contain some errors that can influence flood detection, such as: 1) image artifacts such as false edges or discontinuities caused by the acquisition process and image processing; 2) terrain topography that can affect the backscattering of SAR waves and generate false positives or false negatives that can influence flood detection using SAR imagery.
10. How were the image chips and labels captured in the training process, and what were the criteria for inclusion or exclusion?
Observation attended
11. Provide information on how the parameters of epochs, batch size, and chip size were selected and adjusted in the ArcGIS platform, and describe the potential effects of these parameters on the accuracy and reliability of the trained model.
Observation attended 12. What are the main limitations or sources of error associated with using the same deep learning model for repeated classification of flooding in the same geographical area?
Observation attended
The following paragraph was added in conclusions:
On the other hand, using the same DL model repeatedly to the same study area may be subject to limitations and errors. This is because floods can alter the topography and terrain characteristics. Therefore, if the model is trained with pre-flood images and used to classify post-flood images, it may not capture terrain changes and new features that may emerge. Another issue is that floods can vary in magnitude and extent over time. If the model is trained on historical data and applied to more recent imagery, there may be significant differences in flood conditions. This can lead to a lack of model adaptability and decreased classification accuracy. On the other hand, if the training data used for the model has biases or limitations, such as limited coverage of flood events or lack of diversity in lighting conditions and scale, the model may be unable to generalize producing incorrect or biased results.
13. How were the training parameters established, and what were the potential effects of these parameters on the training and validation loss?
Observation attended
Comments on the Quality of English Language
Author Response File: Author Response.pdf
Reviewer 2 Report
The paper describes the application of a Deep Learning algorithm to Sentinel 1 data for the automatic detection of flooded areas. Automatic flood detection is a challenging issue because flood scenarios are typical examples of complex situations in which different factors have to be considered to provide accurate and robust interpretation of the situation on the ground. Moreover, deep learning techniques have shown good excellent classification performances in many applicative fields, and also in the analysis of remotely sensed images.
The paper is clear and well written. For all these reason I recommend the acceptance after minor revisions.
In particular, I recommend the to consider the following suggestions:
· - to clarify as the authors face the double bounce problem in the preprocessing step of thresholding;
· - to specify in table 2 the number of images available and considered in each mentioned period;
· - to explain if the training patches are partially overlapped or not;
· - to correct some typos (as the wrong reference to section 1 in line 78 or the repetition in line 204).
Author Response
Response to comments/suggestions. Reviewer 2
The paper describes the application of a Deep Learning algorithm to Sentinel 1 data for the automatic detection of flooded areas. Automatic flood detection is a challenging issue because flood scenarios are typical examples of complex situations in which different factors have to be considered to provide accurate and robust interpretation of the situation on the ground. Moreover, deep learning techniques have shown good excellent classification performances in many applicative fields, and also in the analysis of remotely sensed images.
The paper is clear and well written. For all these reason I recommend the acceptance after minor revisions.
In particular, I recommend the to consider the following suggestions:
1. to clarify as the authors face the double bounce problem in the preprocessing step of thresholding;
Observation attended
Sección 3, Linea 297
In urban areas and flooded vegetation, double bounce backscatter dominates, they form right angles in the direction of the radar and the signal bounces twice reflecting most of the energy back to the radar.
2. to specify in table 2 the number of images available and considered in each mentioned period;
Observation attended
A table of available and selected images has been added (see next page).
3. to explain if the training patches are partially overlapped or not;
No, the aim was to avoid overlapping.
4. to correct some typos (as the wrong reference to section 1 in line 78 or the repetition in line 204).
Observation attended
Fixed errors in the references
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
The authors responded to my comments and suggestions. Good luck.
Can be improved.