Stochastic properties of coastal flooding events &ndash; Part 1: CNN-based semantic segmentation for water detection

Kang, Byungho; Feagin, Rusty A.; Huff, Thomas; Duran Vinent, Orencio

doi:https://doi.org/10.5194/egusphere-2023-231

Preprints

https://doi.org/10.5194/egusphere-2023-231

Preprints

03 Apr 2023

| 03 Apr 2023

Stochastic properties of coastal flooding events – Part 1: CNN-based semantic segmentation for water detection

Byungho Kang, Rusty A. Feagin, Thomas Huff, and Orencio Duran Vinent

Abstract. The frequency and intensity of coastal flooding is expected to accelerate in low-elevation coastal areas due to sea level rise. Coastal flooding due to wave runup affects coastal ecosystems and infrastructure, however it can be difficult to monitor in remote and vulnerable areas. Here we use a camera-based system to monitor wave runup as part of the after-storm recovery of an eroded beach on the Texas coast. We analyze high-temporal resolution images of the beach using Convolutional Neural Network (CNN)-based semantic segmentation to study the stochastic properties of runup-driven flooding events. In the first part of this work, we focus on the application of semantic segmentation to identify water and runup events. We train and validate a CNN with over 500 manually classified images, and introduce a post-processing method to reduce false positives. We find that the accuracy of CNN predictions of water pixels is around 90% and strongly depend on the number and diversity of images used for training.

Received: 11 Feb 2023 – Discussion started: 03 Apr 2023

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this preprint. The responsibility to include appropriate place names lies with the authors.

Download & links

Preprint (PDF, 10782 KB)

Notice on discussion status
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
Preprint (10782 KB)

Download & links

The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.

Journal article(s) based on this preprint

03 Jan 2024

Stochastic properties of coastal flooding events – Part 1: convolutional-neural-network-based semantic segmentation for water detection

Byungho Kang, Rusty A. Feagin, Thomas Huff, and Orencio Durán Vinent

Earth Surf. Dynam., 12, 1–10, https://doi.org/10.5194/esurf-12-1-2024,https://doi.org/10.5194/esurf-12-1-2024, 2024

Short summary

Byungho Kang, Rusty A. Feagin, Thomas Huff, and Orencio Duran Vinent

Interactive discussion

Status: closed

RC1:
'Comment on egusphere-2023-231', Anonymous Referee #1, 11 May 2023

The manuscript “Stochastic properties of coastal flooding events – Part 1: CNN-based semantic segmentation for water detection” describes the development and accuracy of a Convolutional Neural Network (CNN) model that automatically detects water on the beach. The authors describe the training methodology and develop a post-processing method based on the classification probability (e.g., water, background, or sky) that improves the accuracy of the model’s predictions. I find the paper to be well-written and concise.
This is a two-part submission, where Part 2 (also submitted to ESurf) uses the output from the CNN model developed here to evaluate flooding probabilities. I can see why the authors split the paper into two, as I believe aspects of the methodology are novel and there is a detailed validation process that would make one paper very long. However, I believe for this manuscript to be a standalone paper, the authors should contextualize the novelty of the developed algorithm (e.g., is this more automated and potentially faster than other classifying/ML models? Is the post-processing method unique?) and highlight the main contribution of the work to the field. Following on, it is hard to tell if this system was developed solely for validating the probabilistic analysis of Rinaldo et al., 2021 and what is described in Part 2 (Kang et al., in review). If that is the case, perhaps this is better as one paper. However, if the authors plan to use this technique in future applications, a discussion of the applicability/strengths to other sites could also help this manuscript stand alone. Some specific suggestions are found below.
General Comments:
The introduction section is missing material that contextualizes the broad use and long history of using optical remote sensing for coastal-based field work. The way that Lines 27 – 35 reads suggests that camera-based systems are new tools for determining wave runup/shoreline position. Remote sensing wave runup using cameras (i.e., optical remote sensing) has been around for over 20 years now (see Holman and Stanley, 2007) and the authors should at least mention these in their introduction to camera systems for context. Time stacks of wave runup from cameras (e.g., how the Stockdon et al., 2006 R2% wave parameterization was generated) have long been manually digitized or QA/QCed, although there are also detection methods using edge detection algorithms, intensity thresholds, and other filters (https://github.com/Coastal-Imaging-Research-Network/CIRN-Quantitative-Coastal-Imaging-Toolbox). Recently, the US Geological Survey has indeed also moved towards leveraging the strengths of machine learning (Palmsten et al., 2020). I think this information can strengthen the authors’ position that automated methods for classification are important.
Following on, other types of camera systems use geometric corrections and ground control points to measure elevations/water levels/wave runup. Can the authors explain and contextualize the strengths of a system that is only measuring the percent of beach flooded? Is this tool intended to be for long-term monitoring/deployment? Are the authors planning on using it for beach-flood validation in other locations?
Terminology: The authors motivate their work discussing “flooding due to wave runup” and understanding the properties of “runup-driven wave events.” I just want to make the point that the water levels the authors capture do indeed have runup in them, but they’re total water levels (runup + still water level) as the authors are not removing the still water level signal, so the flooding is generated by other components related to the still water level (e.g., storm surge, sea level anomaly) not just the wave runup.
Is the intent of this system to only look at low level events? I find the use of the terminology “coastal flooding” (in the title) a little broad and misleading, as it seems this is really looking at “beach flooding.”
Figure 9 displays a decreasing trend in the mean absolute percentage of error for different flooding % images and the number of training images. If the accuracy keeps decreasing, why not see where it “bottoms out”/becomes stable, i.e., where an increase in the number of training images or ratio of flooded images in training set doesn’t affect the MAPE anymore? Maybe the authors also want to include a plot here that shows the increase in computation time relative to variation in # training images to show why a certain value is chosen, rather than the maximum decrease in MAPE.
How computationally efficient is this entire effort?
Specific Lines:
Line 17 – 19: I am confused by this citation, as the Kang et al. paper’s (Part 2 of this work) focus is not this.
Figures 5 and 8 – open circles with closed circles plotted on top of them make the open circles hard to see.
Label Fig 8 and 9 with a,b,c; a,b, respectively instead of using “left” or “right.”
I didn’t see any data availability statement or a location where codes could be found – please see the journal’s Data Policy.
References:
Holman, R. A., & Stanley, J. (2007). The history and technical capabilities of Argus. Coastal engineering, 54(6-7), 477-491.
Palmsten, M., Birchler, J. J., Brown, J. A., Harrison, S. R., Sherwood, C. R., & Erikson, L. H. (2020, December). Next-generation tool for digitizing wave runup. In AGU Fall Meeting Abstracts (Vol. 2020, pp. EP051-10). https://ui.adsabs.harvard.edu/abs/2020AGUFMEP051..10P/abstract

Citation: https://doi.org/10.5194/egusphere-2023-231-RC1
- AC1: 'Reply on RC1', Orencio Duran Vinent, 28 Jul 2023
  
  The manuscript “Stochastic properties of coastal flooding events – Part 1: CNN-based semantic segmentation for water detection” describes the development and accuracy of a Convolutional Neural Network (CNN) model that automatically detects water on the beach. The authors describe the training methodology and develop a post-processing method based on the classification probability (e.g., water, background, or sky) that improves the accuracy of the model’s predictions. I find the paper to be well-written and concise.
  This is a two-part submission, where Part 2 (also submitted to ESurf) uses the output from the CNN model developed here to evaluate flooding probabilities. I can see why the authors split the paper into two, as I believe aspects of the methodology are novel and there is a detailed validation process that would make one paper very long. However, I believe for this manuscript to be a standalone paper, the authors should contextualize the novelty of the developed algorithm (e.g., is this more automated and potentially faster than other classifying/ML models? Is the post-processing method unique?) and highlight the main contribution of the work to the field. Following on, it is hard to tell if this system was developed solely for validating the probabilistic analysis of Rinaldo et al., 2021 and what is described in Part 2 (Kang et al., in review). If that is the case, perhaps this is better as one paper. However, if the authors plan to use this technique in future applications, a discussion of the applicability/strengths to other sites could also help this manuscript stand alone. Some specific suggestions are found below.
  Reply) Thank you for your helpful comments and suggestions. Indeed, we find that using CNN-based image segmentation is faster than other methods for such a large number of images (>50,000) and don’t require calibration or feature-extraction pre-processing. In total, we manually labelled only a 10% of the images (using a MATLAB function and a few days of work), and the trained algorithm was able to predict the other 90% in about two hours. Also, our error-minimization method is unique. Our main objective was to test the suitability of CNN-based image segmentation to handle complex coastal imagery and accurately identify water pixels in arbitrary locations, with the goal of expanding our ability to detect and predict water overtopping. Of course, this work stands alone and it is not tied to the validation of Rinaldo et al., 2021, which is exactly the reason of submitting a two-parts paper.
  General Comments:
  The introduction section is missing material that contextualizes the broad use and long history of using optical remote sensing for coastal-based field work. The way that Lines 27 – 35 reads suggests that camera-based systems are new tools for determining wave runup/shoreline position. Remote sensing wave runup using cameras (i.e., optical remote sensing) has been around for over 20 years now (see Holman and Stanley, 2007) and the authors should at least mention these in their introduction to camera systems for context. Time stacks of wave runup from cameras (e.g., how the Stockdon et al., 2006 R2% wave parameterization was generated) have long been manually digitized or QA/QCed, although there are also detection methods using edge detection algorithms, intensity thresholds, and other filters (https://github.com/Coastal-Imaging-Research-Network/CIRN-Quantitative-Coastal-Imaging-Toolbox). Recently, the US Geological Survey has indeed also moved towards leveraging the strengths of machine learning (Palmsten et al., 2020). I think this information can strengthen the authors’ position that automated methods for classification are important.
  Reply) We acknowledge the previous use of optical remote sensing in coastal-based fieldwork and the importance of providing a comprehensive historical context in our introduction. We will add this to the introduction in the revised manuscript.
  Following on, other types of camera systems use geometric corrections and ground control points to measure elevations/water levels/wave runup. Can the authors explain and contextualize the strengths of a system that is only measuring the percent of beach flooded? Is this tool intended to be for long-term monitoring/deployment? Are the authors planning on using it for beach-flood validation in other locations?
  Reply) Our goal in this work was to test the method for water detection in complex high-resolution imagery (arbitrary light and weather conditions) and not necessarily to properly estimate flooded area. We found that approximating the water area fraction based on the water-pixel fraction is enough for a statistical analysis of beach overtopping events, although of course it does not provide actual water area or elevation. This work can be easily expanded and improved in the future by combining it with hydrostatic pressure sensors at the beach and adding photogrammetry methods.
  Terminology: The authors motivate their work discussing “flooding due to wave runup” and understanding the properties of “runup-driven wave events.” I just want to make the point that the water levels the authors capture do indeed have runup in them, but they’re total water levels (runup + still water level) as the authors are not removing the still water level signal, so the flooding is generated by other components related to the still water level (e.g., storm surge, sea level anomaly) not just the wave runup.
  Reply) We agree and will correct it in the revise manuscript.
  Is the intent of this system to only look at low level events? I find the use of the terminology “coastal flooding” (in the title) a little broad and misleading, as it seems this is really looking at “beach flooding.”
  Reply) The method is general and can thus be applied to arbitrary large overtopping (flooding) events, as long as the camera system can withstand the storm. It just happens that most of the events we captured in this location consisted of beach flooding, howerver those same events can easily lead to nuisance flooding in lower-elevation areas.
  Figure 9 displays a decreasing trend in the mean absolute percentage of error for different flooding % images and the number of training images. If the accuracy keeps decreasing, why not see where it “bottoms out”/becomes stable, i.e., where an increase in the number of training images or ratio of flooded images in training set doesn’t affect the MAPE anymore? Maybe the authors also want to include a plot here that shows the increase in computation time relative to variation in # training images to show why a certain value is chosen, rather than the maximum decrease in MAPE.
  Reply) We will extend Figure 9 to include more data points, aiming to illustrate where the MAPE reaches stability, in the revised version. Furthermore, it only took about three hours to train and run the CNN model with the maximum number of training images we had (500 images) so we didn’t evaluate the computation time as function of the size of the training set.
  How computationally efficient is this entire effort?
  Reply) Roughly three hours to train and run the CNN model for 51,000 images using an external GPU (NVIDIA GTX 1660TI with 6 GB GDDR6 memory).
  Specific Lines:
  Line 17 – 19: I am confused by this citation, as the Kang et al. paper’s (Part 2 of this work) focus is not this.
  Reply) The referee is correct. We will change it in the revised version.
  Figures 5 and 8 – open circles with closed circles plotted on top of them make the open circles hard to see.
  Reply) We agree, we will adjust these figures.
  Label Fig 8 and 9 with a,b,c; a,b, respectively instead of using “left” or “right.”
  Reply) We will change it in the revise version.
  I didn’t see any data availability statement or a location where codes could be found – please see the journal’s Data Policy.
  Reply) We will upload the data to the Texas Data Repository (TDR) and update the information in the revised version.
  References:
  Holman, R. A., & Stanley, J. (2007). The history and technical capabilities of Argus. Coastal engineering, 54(6-7), 477-491.
  Palmsten, M., Birchler, J. J., Brown, J. A., Harrison, S. R., Sherwood, C. R., & Erikson, L. H. (2020, December). Next-generation tool for digitizing wave runup. In AGU Fall Meeting Abstracts (Vol. 2020, pp. EP051-10). https://ui.adsabs.harvard.edu/abs/2020AGUFMEP051..10P/abstract
  
  Citation: https://doi.org/10.5194/egusphere-2023-231-AC1
RC2:
'Comment on egusphere-2023-231', Anonymous Referee #2, 22 Jun 2023
General comments
The authors present an application of a Convolutional Neural Network to the segmentation of water from coastal imagery. The performance of the model and its sensitivity are quite thoroughly explored.
The general approach and description of the CNN, its training and validation are fairly clear and easy to follow. Something I had expected but is not mentioned in the current manuscript (or part 2) is the mapping of pixels from the image to real-world coordinates. Given the eventual purpose of the segmentation, mapping to real-world coordinates seems like a very useful thing to do as it would result in the actual height the waves have reached instead of just identifying which pixels contain water. And if it is not possible or feasible, it would be interesting to the reader to explain why.
In the manuscript, the term flooding is used a lot in a confusing way, at least for this reviewer. The introduction mentions flooding causing damage to coastal communities and infrastructure, which I then interpret as flooding of the hinterland (say landwards of the beach or dune crest). Going on the camera views shown in Fig. 2, the camera view contains the beach up to the start of vegetation, so from the images one can detect whether the beach itself is submerged (flooded). Furthermore the manuscript also mentions ‘flooding events driven by wave runup’, which is confusing from a semantics standpoint, as strictly speaking as soon as it leads to flooding (of the hinterland), it is by definition wave overtopping instead of wave run-up. In short, the authors should make more explicit and consistent what they mean with the term flooding, and what is of interest considering the context, namely high-frequency nuisance flooding that leads to damage to infrastructure and coastal communities.
Specific comments
L011: “Coastal flooding can cause significant damage to coastal ecosystems…” depending on your viewpoint, one could argue that coastal flooding is – or should be – a natural part of that very ecosystem (especially in the case of barrier island systems). So I do not think this is a very strong argument.

L029: “to monitor coastal flooding” – I think it’s more accurate to state that you are monitoring local wave runup (or actually total water level). The coastal flooding itself is not monitored with this setup, right?

L090: “…weights in the cost of accuracy.” should be “…weights at the cost of accuracy.”

L115: ‘images’ should be ‘image’

Fig 6: I don’t see a superimposed red line in the center panels mentioned in the caption. (only red pixels in the right panels).

L236: The automatic segmentation of water in coastal imagery itself will not directly improve prediction of coastal flooding/overtopping, just the monitoring of wave run-up in a specific location. Indirectly, that data can be used in a way that does improve prediction as well, but this does not directly follow from (part 1 of) the manuscript.
Citation: https://doi.org/10.5194/egusphere-2023-231-RC2
- AC2: 'Reply on RC2', Orencio Duran Vinent, 28 Jul 2023
  
  General comments
  The authors present an application of a Convolutional Neural Network to the segmentation of water from coastal imagery. The performance of the model and its sensitivity are quite thoroughly explored.
  The general approach and description of the CNN, its training and validation are fairly clear and easy to follow. Something I had expected but is not mentioned in the current manuscript (or part 2) is the mapping of pixels from the image to real-world coordinates. Given the eventual purpose of the segmentation, mapping to real-world coordinates seems like a very useful thing to do as it would result in the actual height the waves have reached instead of just identifying which pixels contain water. And if it is not possible or feasible, it would be interesting to the reader to explain why.
  Reply) We thank the reviewer for her helpful comments and suggestions. Our goal in this work was to test the method for water detection in complex high-resolution imagery (arbitrary light and weather conditions) and not necessarily to properly estimate flooded area. We found that approximating the water area fraction by the water-pixel fraction is enough for a statistical analysis of beach overtopping events, although of course it does not provide actual water area or elevation. We agree that our work can be easily expanded and improved in the future by combining it with hydrostatic pressure sensors at the beach and adding photogrammetry methods. We will comment on this in the discussion section of the revised manuscript.
  In the manuscript, the term flooding is used a lot in a confusing way, at least for this reviewer. The introduction mentions flooding causing damage to coastal communities and infrastructure, which I then interpret as flooding of the hinterland (say landwards of the beach or dune crest). Going on the camera views shown in Fig. 2, the camera view contains the beach up to the start of vegetation, so from the images one can detect whether the beach itself is submerged (flooded). Furthermore the manuscript also mentions ‘flooding events driven by wave runup’, which is confusing from a semantics standpoint, as strictly speaking as soon as it leads to flooding (of the hinterland), it is by definition wave overtopping instead of wave run-up. In short, the authors should make more explicit and consistent what they mean with the term flooding, and what is of interest considering the context, namely high-frequency nuisance flooding that leads to damage to infrastructure and coastal communities.
  Reply) We agree with the referee and will address this point in detail in the revised version.
  Specific comments
  1. L011: “Coastal flooding can cause significant damage to coastal ecosystems…” depending on your viewpoint, one could argue that coastal flooding is – or should be – a natural part of that very ecosystem (especially in the case of barrier island systems). So I do not think this is a very strong argument.
  Reply) Here we are referring to the harm more frequent coastal flooding can have on the salt-intolerant vegetation that is part of the coastal dune ecosystem. We will clarify this point in the revised version.
  2. L029: “to monitor coastal flooding” – I think it’s more accurate to state that you are monitoring local wave runup (or actually total water level). The coastal flooding itself is not monitored with this setup, right?
  Reply) Yes, we agree and will correct it in the revised version.
  3. L090: “…weights in the cost of accuracy.” should be “…weights at the cost of accuracy.”
  Reply) We will correct it in the revised version.
  4. L115: ‘images’ should be ‘image’
  Reply) We will correct it in the revised version.
  5. Fig 6: I don’t see a superimposed red line in the center panels mentioned in the caption. (only red pixels in the right panels).
  Reply) We will revise the caption to accurately reflect the content of the figure.
  L236: The automatic segmentation of water in coastal imagery itself will not directly improve prediction of coastal flooding/overtopping, just the monitoring of wave run-up in a specific location. Indirectly, that data can be used in a way that does improve prediction as well, but this does not directly follow from (part 1 of) the manuscript.
  Reply) Yes, we agree and will correct it in the revised version.
  
  Citation: https://doi.org/10.5194/egusphere-2023-231-AC2

Interactive discussion

Status: closed

RC1:
'Comment on egusphere-2023-231', Anonymous Referee #1, 11 May 2023

The manuscript “Stochastic properties of coastal flooding events – Part 1: CNN-based semantic segmentation for water detection” describes the development and accuracy of a Convolutional Neural Network (CNN) model that automatically detects water on the beach. The authors describe the training methodology and develop a post-processing method based on the classification probability (e.g., water, background, or sky) that improves the accuracy of the model’s predictions. I find the paper to be well-written and concise.
This is a two-part submission, where Part 2 (also submitted to ESurf) uses the output from the CNN model developed here to evaluate flooding probabilities. I can see why the authors split the paper into two, as I believe aspects of the methodology are novel and there is a detailed validation process that would make one paper very long. However, I believe for this manuscript to be a standalone paper, the authors should contextualize the novelty of the developed algorithm (e.g., is this more automated and potentially faster than other classifying/ML models? Is the post-processing method unique?) and highlight the main contribution of the work to the field. Following on, it is hard to tell if this system was developed solely for validating the probabilistic analysis of Rinaldo et al., 2021 and what is described in Part 2 (Kang et al., in review). If that is the case, perhaps this is better as one paper. However, if the authors plan to use this technique in future applications, a discussion of the applicability/strengths to other sites could also help this manuscript stand alone. Some specific suggestions are found below.
General Comments:
The introduction section is missing material that contextualizes the broad use and long history of using optical remote sensing for coastal-based field work. The way that Lines 27 – 35 reads suggests that camera-based systems are new tools for determining wave runup/shoreline position. Remote sensing wave runup using cameras (i.e., optical remote sensing) has been around for over 20 years now (see Holman and Stanley, 2007) and the authors should at least mention these in their introduction to camera systems for context. Time stacks of wave runup from cameras (e.g., how the Stockdon et al., 2006 R2% wave parameterization was generated) have long been manually digitized or QA/QCed, although there are also detection methods using edge detection algorithms, intensity thresholds, and other filters (https://github.com/Coastal-Imaging-Research-Network/CIRN-Quantitative-Coastal-Imaging-Toolbox). Recently, the US Geological Survey has indeed also moved towards leveraging the strengths of machine learning (Palmsten et al., 2020). I think this information can strengthen the authors’ position that automated methods for classification are important.
Following on, other types of camera systems use geometric corrections and ground control points to measure elevations/water levels/wave runup. Can the authors explain and contextualize the strengths of a system that is only measuring the percent of beach flooded? Is this tool intended to be for long-term monitoring/deployment? Are the authors planning on using it for beach-flood validation in other locations?
Terminology: The authors motivate their work discussing “flooding due to wave runup” and understanding the properties of “runup-driven wave events.” I just want to make the point that the water levels the authors capture do indeed have runup in them, but they’re total water levels (runup + still water level) as the authors are not removing the still water level signal, so the flooding is generated by other components related to the still water level (e.g., storm surge, sea level anomaly) not just the wave runup.
Is the intent of this system to only look at low level events? I find the use of the terminology “coastal flooding” (in the title) a little broad and misleading, as it seems this is really looking at “beach flooding.”
Figure 9 displays a decreasing trend in the mean absolute percentage of error for different flooding % images and the number of training images. If the accuracy keeps decreasing, why not see where it “bottoms out”/becomes stable, i.e., where an increase in the number of training images or ratio of flooded images in training set doesn’t affect the MAPE anymore? Maybe the authors also want to include a plot here that shows the increase in computation time relative to variation in # training images to show why a certain value is chosen, rather than the maximum decrease in MAPE.
How computationally efficient is this entire effort?
Specific Lines:
Line 17 – 19: I am confused by this citation, as the Kang et al. paper’s (Part 2 of this work) focus is not this.
Figures 5 and 8 – open circles with closed circles plotted on top of them make the open circles hard to see.
Label Fig 8 and 9 with a,b,c; a,b, respectively instead of using “left” or “right.”
I didn’t see any data availability statement or a location where codes could be found – please see the journal’s Data Policy.
References:
Holman, R. A., & Stanley, J. (2007). The history and technical capabilities of Argus. Coastal engineering, 54(6-7), 477-491.
Palmsten, M., Birchler, J. J., Brown, J. A., Harrison, S. R., Sherwood, C. R., & Erikson, L. H. (2020, December). Next-generation tool for digitizing wave runup. In AGU Fall Meeting Abstracts (Vol. 2020, pp. EP051-10). https://ui.adsabs.harvard.edu/abs/2020AGUFMEP051..10P/abstract

Citation: https://doi.org/10.5194/egusphere-2023-231-RC1
- AC1: 'Reply on RC1', Orencio Duran Vinent, 28 Jul 2023
  
  The manuscript “Stochastic properties of coastal flooding events – Part 1: CNN-based semantic segmentation for water detection” describes the development and accuracy of a Convolutional Neural Network (CNN) model that automatically detects water on the beach. The authors describe the training methodology and develop a post-processing method based on the classification probability (e.g., water, background, or sky) that improves the accuracy of the model’s predictions. I find the paper to be well-written and concise.
  This is a two-part submission, where Part 2 (also submitted to ESurf) uses the output from the CNN model developed here to evaluate flooding probabilities. I can see why the authors split the paper into two, as I believe aspects of the methodology are novel and there is a detailed validation process that would make one paper very long. However, I believe for this manuscript to be a standalone paper, the authors should contextualize the novelty of the developed algorithm (e.g., is this more automated and potentially faster than other classifying/ML models? Is the post-processing method unique?) and highlight the main contribution of the work to the field. Following on, it is hard to tell if this system was developed solely for validating the probabilistic analysis of Rinaldo et al., 2021 and what is described in Part 2 (Kang et al., in review). If that is the case, perhaps this is better as one paper. However, if the authors plan to use this technique in future applications, a discussion of the applicability/strengths to other sites could also help this manuscript stand alone. Some specific suggestions are found below.
  Reply) Thank you for your helpful comments and suggestions. Indeed, we find that using CNN-based image segmentation is faster than other methods for such a large number of images (>50,000) and don’t require calibration or feature-extraction pre-processing. In total, we manually labelled only a 10% of the images (using a MATLAB function and a few days of work), and the trained algorithm was able to predict the other 90% in about two hours. Also, our error-minimization method is unique. Our main objective was to test the suitability of CNN-based image segmentation to handle complex coastal imagery and accurately identify water pixels in arbitrary locations, with the goal of expanding our ability to detect and predict water overtopping. Of course, this work stands alone and it is not tied to the validation of Rinaldo et al., 2021, which is exactly the reason of submitting a two-parts paper.
  General Comments:
  The introduction section is missing material that contextualizes the broad use and long history of using optical remote sensing for coastal-based field work. The way that Lines 27 – 35 reads suggests that camera-based systems are new tools for determining wave runup/shoreline position. Remote sensing wave runup using cameras (i.e., optical remote sensing) has been around for over 20 years now (see Holman and Stanley, 2007) and the authors should at least mention these in their introduction to camera systems for context. Time stacks of wave runup from cameras (e.g., how the Stockdon et al., 2006 R2% wave parameterization was generated) have long been manually digitized or QA/QCed, although there are also detection methods using edge detection algorithms, intensity thresholds, and other filters (https://github.com/Coastal-Imaging-Research-Network/CIRN-Quantitative-Coastal-Imaging-Toolbox). Recently, the US Geological Survey has indeed also moved towards leveraging the strengths of machine learning (Palmsten et al., 2020). I think this information can strengthen the authors’ position that automated methods for classification are important.
  Reply) We acknowledge the previous use of optical remote sensing in coastal-based fieldwork and the importance of providing a comprehensive historical context in our introduction. We will add this to the introduction in the revised manuscript.
  Following on, other types of camera systems use geometric corrections and ground control points to measure elevations/water levels/wave runup. Can the authors explain and contextualize the strengths of a system that is only measuring the percent of beach flooded? Is this tool intended to be for long-term monitoring/deployment? Are the authors planning on using it for beach-flood validation in other locations?
  Reply) Our goal in this work was to test the method for water detection in complex high-resolution imagery (arbitrary light and weather conditions) and not necessarily to properly estimate flooded area. We found that approximating the water area fraction based on the water-pixel fraction is enough for a statistical analysis of beach overtopping events, although of course it does not provide actual water area or elevation. This work can be easily expanded and improved in the future by combining it with hydrostatic pressure sensors at the beach and adding photogrammetry methods.
  Terminology: The authors motivate their work discussing “flooding due to wave runup” and understanding the properties of “runup-driven wave events.” I just want to make the point that the water levels the authors capture do indeed have runup in them, but they’re total water levels (runup + still water level) as the authors are not removing the still water level signal, so the flooding is generated by other components related to the still water level (e.g., storm surge, sea level anomaly) not just the wave runup.
  Reply) We agree and will correct it in the revise manuscript.
  Is the intent of this system to only look at low level events? I find the use of the terminology “coastal flooding” (in the title) a little broad and misleading, as it seems this is really looking at “beach flooding.”
  Reply) The method is general and can thus be applied to arbitrary large overtopping (flooding) events, as long as the camera system can withstand the storm. It just happens that most of the events we captured in this location consisted of beach flooding, howerver those same events can easily lead to nuisance flooding in lower-elevation areas.
  Figure 9 displays a decreasing trend in the mean absolute percentage of error for different flooding % images and the number of training images. If the accuracy keeps decreasing, why not see where it “bottoms out”/becomes stable, i.e., where an increase in the number of training images or ratio of flooded images in training set doesn’t affect the MAPE anymore? Maybe the authors also want to include a plot here that shows the increase in computation time relative to variation in # training images to show why a certain value is chosen, rather than the maximum decrease in MAPE.
  Reply) We will extend Figure 9 to include more data points, aiming to illustrate where the MAPE reaches stability, in the revised version. Furthermore, it only took about three hours to train and run the CNN model with the maximum number of training images we had (500 images) so we didn’t evaluate the computation time as function of the size of the training set.
  How computationally efficient is this entire effort?
  Reply) Roughly three hours to train and run the CNN model for 51,000 images using an external GPU (NVIDIA GTX 1660TI with 6 GB GDDR6 memory).
  Specific Lines:
  Line 17 – 19: I am confused by this citation, as the Kang et al. paper’s (Part 2 of this work) focus is not this.
  Reply) The referee is correct. We will change it in the revised version.
  Figures 5 and 8 – open circles with closed circles plotted on top of them make the open circles hard to see.
  Reply) We agree, we will adjust these figures.
  Label Fig 8 and 9 with a,b,c; a,b, respectively instead of using “left” or “right.”
  Reply) We will change it in the revise version.
  I didn’t see any data availability statement or a location where codes could be found – please see the journal’s Data Policy.
  Reply) We will upload the data to the Texas Data Repository (TDR) and update the information in the revised version.
  References:
  Holman, R. A., & Stanley, J. (2007). The history and technical capabilities of Argus. Coastal engineering, 54(6-7), 477-491.
  Palmsten, M., Birchler, J. J., Brown, J. A., Harrison, S. R., Sherwood, C. R., & Erikson, L. H. (2020, December). Next-generation tool for digitizing wave runup. In AGU Fall Meeting Abstracts (Vol. 2020, pp. EP051-10). https://ui.adsabs.harvard.edu/abs/2020AGUFMEP051..10P/abstract
  
  Citation: https://doi.org/10.5194/egusphere-2023-231-AC1
RC2:
'Comment on egusphere-2023-231', Anonymous Referee #2, 22 Jun 2023
General comments
The authors present an application of a Convolutional Neural Network to the segmentation of water from coastal imagery. The performance of the model and its sensitivity are quite thoroughly explored.
The general approach and description of the CNN, its training and validation are fairly clear and easy to follow. Something I had expected but is not mentioned in the current manuscript (or part 2) is the mapping of pixels from the image to real-world coordinates. Given the eventual purpose of the segmentation, mapping to real-world coordinates seems like a very useful thing to do as it would result in the actual height the waves have reached instead of just identifying which pixels contain water. And if it is not possible or feasible, it would be interesting to the reader to explain why.
In the manuscript, the term flooding is used a lot in a confusing way, at least for this reviewer. The introduction mentions flooding causing damage to coastal communities and infrastructure, which I then interpret as flooding of the hinterland (say landwards of the beach or dune crest). Going on the camera views shown in Fig. 2, the camera view contains the beach up to the start of vegetation, so from the images one can detect whether the beach itself is submerged (flooded). Furthermore the manuscript also mentions ‘flooding events driven by wave runup’, which is confusing from a semantics standpoint, as strictly speaking as soon as it leads to flooding (of the hinterland), it is by definition wave overtopping instead of wave run-up. In short, the authors should make more explicit and consistent what they mean with the term flooding, and what is of interest considering the context, namely high-frequency nuisance flooding that leads to damage to infrastructure and coastal communities.
Specific comments
L011: “Coastal flooding can cause significant damage to coastal ecosystems…” depending on your viewpoint, one could argue that coastal flooding is – or should be – a natural part of that very ecosystem (especially in the case of barrier island systems). So I do not think this is a very strong argument.

L029: “to monitor coastal flooding” – I think it’s more accurate to state that you are monitoring local wave runup (or actually total water level). The coastal flooding itself is not monitored with this setup, right?

L090: “…weights in the cost of accuracy.” should be “…weights at the cost of accuracy.”

L115: ‘images’ should be ‘image’

Fig 6: I don’t see a superimposed red line in the center panels mentioned in the caption. (only red pixels in the right panels).

L236: The automatic segmentation of water in coastal imagery itself will not directly improve prediction of coastal flooding/overtopping, just the monitoring of wave run-up in a specific location. Indirectly, that data can be used in a way that does improve prediction as well, but this does not directly follow from (part 1 of) the manuscript.
Citation: https://doi.org/10.5194/egusphere-2023-231-RC2
- AC2: 'Reply on RC2', Orencio Duran Vinent, 28 Jul 2023
  
  General comments
  The authors present an application of a Convolutional Neural Network to the segmentation of water from coastal imagery. The performance of the model and its sensitivity are quite thoroughly explored.
  The general approach and description of the CNN, its training and validation are fairly clear and easy to follow. Something I had expected but is not mentioned in the current manuscript (or part 2) is the mapping of pixels from the image to real-world coordinates. Given the eventual purpose of the segmentation, mapping to real-world coordinates seems like a very useful thing to do as it would result in the actual height the waves have reached instead of just identifying which pixels contain water. And if it is not possible or feasible, it would be interesting to the reader to explain why.
  Reply) We thank the reviewer for her helpful comments and suggestions. Our goal in this work was to test the method for water detection in complex high-resolution imagery (arbitrary light and weather conditions) and not necessarily to properly estimate flooded area. We found that approximating the water area fraction by the water-pixel fraction is enough for a statistical analysis of beach overtopping events, although of course it does not provide actual water area or elevation. We agree that our work can be easily expanded and improved in the future by combining it with hydrostatic pressure sensors at the beach and adding photogrammetry methods. We will comment on this in the discussion section of the revised manuscript.
  In the manuscript, the term flooding is used a lot in a confusing way, at least for this reviewer. The introduction mentions flooding causing damage to coastal communities and infrastructure, which I then interpret as flooding of the hinterland (say landwards of the beach or dune crest). Going on the camera views shown in Fig. 2, the camera view contains the beach up to the start of vegetation, so from the images one can detect whether the beach itself is submerged (flooded). Furthermore the manuscript also mentions ‘flooding events driven by wave runup’, which is confusing from a semantics standpoint, as strictly speaking as soon as it leads to flooding (of the hinterland), it is by definition wave overtopping instead of wave run-up. In short, the authors should make more explicit and consistent what they mean with the term flooding, and what is of interest considering the context, namely high-frequency nuisance flooding that leads to damage to infrastructure and coastal communities.
  Reply) We agree with the referee and will address this point in detail in the revised version.
  Specific comments
  1. L011: “Coastal flooding can cause significant damage to coastal ecosystems…” depending on your viewpoint, one could argue that coastal flooding is – or should be – a natural part of that very ecosystem (especially in the case of barrier island systems). So I do not think this is a very strong argument.
  Reply) Here we are referring to the harm more frequent coastal flooding can have on the salt-intolerant vegetation that is part of the coastal dune ecosystem. We will clarify this point in the revised version.
  2. L029: “to monitor coastal flooding” – I think it’s more accurate to state that you are monitoring local wave runup (or actually total water level). The coastal flooding itself is not monitored with this setup, right?
  Reply) Yes, we agree and will correct it in the revised version.
  3. L090: “…weights in the cost of accuracy.” should be “…weights at the cost of accuracy.”
  Reply) We will correct it in the revised version.
  4. L115: ‘images’ should be ‘image’
  Reply) We will correct it in the revised version.
  5. Fig 6: I don’t see a superimposed red line in the center panels mentioned in the caption. (only red pixels in the right panels).
  Reply) We will revise the caption to accurately reflect the content of the figure.
  L236: The automatic segmentation of water in coastal imagery itself will not directly improve prediction of coastal flooding/overtopping, just the monitoring of wave run-up in a specific location. Indirectly, that data can be used in a way that does improve prediction as well, but this does not directly follow from (part 1 of) the manuscript.
  Reply) Yes, we agree and will correct it in the revised version.
  
  Citation: https://doi.org/10.5194/egusphere-2023-231-AC2

Peer review completion

AR: Author's response | RR: Referee report | ED: Editor decision | EF: Editorial file upload

AR by Orencio Duran Vinent on behalf of the Authors (18 Aug 2023) Author's response Author's tracked changes Manuscript

ED: Referee Nomination & Report Request started (05 Sep 2023) by Sagy Cohen

RR by Anonymous Referee #2 (15 Sep 2023)

RR by Anonymous Referee #1 (24 Oct 2023)

Suggestions for revision or reasons for rejection

I am pleased with the authors responses to my original review. I have a few remaining questions/corrections, which range from technical to minor.

Relating to the extended analysis of the MAPE of the accuracy ratio of CNNS trained with different fractions of flooding (Figure 9A): Did the authors test the MAPE accuracy with a proportion of more than 26% of flooding images? There is also a decline in accuracy between 10 and 15%, so I’m wondering if the decline in accuracy the authors note continues past 26% or not?

Following on, I believe that the authors see the lowest MAPE when 23% of the training set was composed of flooding images. It looks like the validation set was 21 flooding image and 70 normal images (line 215) making the percentage of flooded images in the validation set also 23%. Is this a coincidence? I suggest testing a different proportion of flooding images to ensure the best accuracy ratio proportion for training is independent of the percentage of flooded images of validation data used for validating/testing.

What are the limitations of this approach? For example, if a camera moves, would the accuracy of the predictions be at risk?

Line by Line:
Line 240: Maybe I’m misunderstanding something, but isn’t this statement contradictory? Accuracy of CNN predictions independent of amount of water in images, yet it was better for images with a larger fraction of water? This would tell me it might be dependent, because more water = more fraction of water?
Line 252: is precise the right word to use here? Considering its just pixels from the images, I’m not sure it would be considered a precise observation of beach or back-beach overtopping. How is overtopping defined based on the imagery?
Line 253: I suggest editing this sentence to say, “This method has the potential to enhance….” Since it hasn’t been combined with photogrammetry this is untested, so I would leave it as it has the potential, rather than it can enhance…

Figures
Figure 2: I suggest the authors label the figure to show which photos are considered flooding and which are considered normal.
Figure 3: I suggest the authors label “top row” and “bottom row” rather than just “top” and “bottom”.

Hide

ED: Publish subject to minor revisions (review by editor) (26 Oct 2023) by Sagy Cohen

AR by Orencio Duran Vinent on behalf of the Authors (07 Nov 2023) Author's response Author's tracked changes Manuscript

ED: Publish as is (08 Nov 2023) by Sagy Cohen

ED: Publish as is (14 Nov 2023) by Niels Hovius (Editor)

AR by Orencio Duran Vinent on behalf of the Authors (16 Nov 2023) Manuscript

Journal article(s) based on this preprint

03 Jan 2024

Stochastic properties of coastal flooding events – Part 1: convolutional-neural-network-based semantic segmentation for water detection

Byungho Kang, Rusty A. Feagin, Thomas Huff, and Orencio Durán Vinent

Earth Surf. Dynam., 12, 1–10, https://doi.org/10.5194/esurf-12-1-2024,https://doi.org/10.5194/esurf-12-1-2024, 2024

Short summary

Byungho Kang, Rusty A. Feagin, Thomas Huff, and Orencio Duran Vinent

Viewed

Total article views: 562 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
386	153	23	562	18	18

HTML: 386
PDF: 153
XML: 23
Total: 562
BibTeX: 18
EndNote: 18

Views and downloads (calculated since 03 Apr 2023)

Month	HTML	PDF	XML	Total
Apr 2023	128	40	6	174
May 2023	58	11	2	71
Jun 2023	49	16	2	67
Jul 2023	47	16	5	68
Aug 2023	33	8	2	43
Sep 2023	24	13	2	39
Oct 2023	14	17	2	33
Nov 2023	14	7	0	21
Dec 2023	18	20	2	40
Jan 2024	1	5	0	6
Feb 2024	0
Mar 2024	0
Apr 2024	0
May 2024	0
Jun 2024	0
Jul 2024	0
Aug 2024	0
Sep 2024	0

Cumulative views and downloads (calculated since 03 Apr 2023)

Month	HTML	PDF	XML	Total
Apr 2023	128	40	6	174
May 2023	58	11	2	71
Jun 2023	49	16	2	67
Jul 2023	47	16	5	68
Aug 2023	33	8	2	43
Sep 2023	24	13	2	39
Oct 2023	14	17	2	33
Nov 2023	14	7	0	21
Dec 2023	18	20	2	40
Jan 2024	1	5	0	6
Feb 2024	0
Mar 2024	0
Apr 2024	0
May 2024	0
Jun 2024	0
Jul 2024	0
Aug 2024	0
Sep 2024	0

Viewed (geographical distribution)

Total article views: 562 (including HTML, PDF, and XML) Thereof 562 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 02 Sep 2024

Download

The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.

Preprint (10782 KB)
Metadata XML

Short summary

Coastal flooding can cause significant damage to coastal ecosystems, infrastructure and communities, and is expected to increase in frequency with the acceleration of sea level rise. In order to respond to it, it is crucial to measure and model their frequency and intensity. Here, we show Deep-Learning techniques can be successfully used to automatically detect flooding events from complex coastal imagery, opening the way to real-time monitoring and data acquisition for model development.


Total:	0
HTML:	0
PDF:	0
XML:	0