Image-Based Classification of Double-Barred Beach States Using a Convolutional Neural Network and Transfer Learning
Round 1
Reviewer 1 Report
The authors present an extension of the work by Ellenson et al (2020) to use convolution neural networks to identify and classify double barred beach morphology from time averaged images of the breaking wave intensity patterns. The analysis and descriptions of the work are particularly clear, and I have few if any suggestions. I appreciate the detailed, but not overly esoteric, description of the CNN used here. One result of this work, that I'm a little curious about, is that it seems obvious to me that a doubled barred data set would need to be included in the training imagery to be able to allow for classification of a double barred system. Thus, it's surprising to me that there was any (though admittedly little) skill in classifying a novel data set using a system trained only on single barred beach images. I have only these few minor suggestions: line: 102 - change compared to comparable 59 - there are better references on in situ morphology measurement techniques than the Argus (remote sensing) references listed. 635 - change watercolour to two words, otherwise your referring to the painting medium.Author Response
We thank reviewer 1 for the kind remarks and the suggestions for improving the manuscript. Below, the reviewer’s comments are given in italic. Accordingly, italic mentions of line numbers refer to those in the original manuscript, other mentions of line numbers refer to those in the revised manuscript.
The authors present an extension of the work by Ellenson et al (2020) to use convolution neural networks to identify and classify double barred beach morphology from time averaged images of the breaking wave intensity patterns. The analysis and descriptions of the work are particularly clear, and I have few if any suggestions. I appreciate the detailed, but not overly esoteric, description of the CNN used here.
> We thank Reviewer 1 for the kind remarks.
One result of this work, that I'm a little curious about, is that it seems obvious to me that a doubled barred data set would need to be included in the training imagery to be able to allow for classification of a double barred system. Thus, it's surprising to me that there was any (though admittedly little) skill in classifying a novel data set using a system trained only on single barred beach images.
> This is indeed an interesting observation. Although the model trained on images from single-barred beaches recognizes features in the double-barred images, the correspondence is not yet enough to provide an accurate classification. We reflect upon this in the discussion section (lines 575-590) and state that the misclassification of the inner bar may be due to its proximity to the shoreline.
I have only these few minor suggestions:
line: 102 - change compared to comparable
> Done.
59 - there are better references on in situ morphology measurement techniques than the Argus (remote sensing) references listed.
> Thanks for pointing this out. We now cite Gallagher et al. (1998) and Almar et al. (2010) in line 61, where the difficulties of acquiring in-situ field measurements are explicitly mentioned.
635 - change watercolour to two words, otherwise your referring to the painting medium.
> Done.
Reviewer 2 Report
-The paper should be interesting ;;;
-it is a good idea to add a block diagram of the proposed research (step by step);;;;;;
-it is a good idea to add more photos of measurements;;
-What is the result of the analysis?;;
-figures should have high quality;;;
-labels of figures should be bigger;;;; - Figure A1. T; Figure A2. T etc.
-please add photos of the application of the proposed research, 2-3 photos (if any) ;;;
-what will society have from the paper?;;
-Please compare the proposed method with other approaches;;
-references should be from the web of science 2020-2022 (50% of all references, 30 references at least);;;
-Conclusion: point out what have you done;;;;
-please add some sentences about future work;;;
Author Response
We thank reviewer 2 for the suggestions for improving the manuscript. Below, the reviewer’s comments are given in italic. Accordingly, italic mentions of line numbers refer to those in the original manuscript, other mentions of line numbers refer to those in the revised manuscript.
-The paper should be interesting ;;;
> We feel the manuscript provides a thorough presentation of our research, and that it is of significant interest to the (coastal) remote sensing community and the readers of the MDPI journal “Remote Sensing”.
-it is a good idea to add a block diagram of the proposed research (step by step);;;;;;
> We thank the reviewer for this suggestion. We have now added a diagram at the beginning of Section 3 that provides a schematic overview of the methodological steps taken. In addition to the list of the methodological steps taken, also at the beginning of Section 3 (lines 219-224), we now include the suggested block diagram.
-it is a good idea to add more photos of measurements;;
> It is not clear to us what measurements the reviewer means here. We used three sets of daily video images, form three different field sites, and we provide examples of these images in Figure 3. Moreover, in Figure 5, we demonstrate how we further process these images to increase the amount of data for training the model.
-What is the result of the analysis?;;
> The main results, elaborately discussed in the sections 5 and 6 and summarized in section 7 (conclusions), are that (1) a CNN trained with images from single-barred beaches shows poor performance when classifying double-barred beach states, (2) transfer learning, where limited data from a double-barred beach is added to the single-barred model, allows for the training of a well-performing model for classifying double-barred beach states and (3) including outer-bar labels in the transfer learning has a larger impact on the resulting model performance than when labels from the inner bar are included.
-figures should have high quality;;;
> We have revised all figures to ensure their quality and improved where necessary.
-labels of figures should be bigger;;;; - Figure A1. T; Figure A2. T etc.
> We thank the reviewer for pointing this out. We have revised all figures and increased the font size where necessary. Also, we have limited the number of panels in figure A3 from 10 to 6.
-please add photos of the application of the proposed research, 2-3 photos (if any) ;;;
> The application of the methodology described in our paper results in the classification of bar states from single- and double-barred beaches, and therefore photos of its application are not applicable here. We show figures of the results in figures 6-13. Moreover, we elaborate on the processing of the data in section 3, where we included photos of the technique used to manipulate the data for training the model.
-what will society have from the paper?;;
> The research will allow for the automated classification of bar-beach states, and will benefit the understanding of nearshore morphodynamics and management of coastal zones as larger amounts of data can be processed more easily. We provide this societal context in the first paragraph of the introduction section (section 1; lines 28-36).
-Please compare the proposed method with other approaches;;
> We reflect upon the different approaches used to classify beach states in Lines 69-90. Here, we discuss that deep-learning techniques provide an innovative way to automate the classification of sandbar morphology, as opposed to approaches based on human-designed rules, as done previously. A direct quantitative comparison with such approaches is at the core of our research, as images and classifications (labels) are used to train and validate the CNN. Moreover, we build upon the work of Ellenson et al (2020) and directly compare our findings with their work.
-references should be from the web of science 2020-2022 (50% of all references, 30 references at least);;;
> We have ensured that all relevant previous work is cited and we place our work within the context of recent work, most notably that of Ellenson et al (2020).
-Conclusion: point out what have you done;;;;
> In our view, we provide a complete summary of our approach, findings and discussion thereof in the Conclusions section, which can be read separately from the main body of the manuscript.
-please add some sentences about future work;;;
> In lines 657-662 we explicitly address what future research, building upon our work, could look like: more data form different locations, and the applications of object detection and object tracking.
Reviewer 3 Report
Dear authors,
Firstly, I’d like to acknowledge the work you put in the different experiments and in the paper. Your work is a good add-on to previous work and a step forward to better automatic beach classification.
Therefore, I want to recommend your paper to be accepted with minor revisions (listed below) as I do think some clarifications are needed in the description of the different models/experiments and in the discussion.
Best regards
Abstract
- From the first line I think there is a confusion that continues all along the paper. You here talk about subtidal sandbars. However, you say in the discussion that the difficulty of differentiate the shoreline and the sandbar can be an issue for sandbar recognition. Those bars are intertidal bars, not subtidal. As it is usually the case for TBR states (and LTT).
Introduction
- L30 & L37: remove subtidal; same comment as above.
- L46: rephrase, sandbars don’t always present every states. They can however evolve from one state to another depending on hydrodynamic conditions
- L53: seaward-most outer bar; landward-most inner bar
- L57: what do you call in-situ measurements? Video images are in-situ methods.
- L69 to 77: I think you can tell a bit more about those methods/models, their accuracy, how do they deal with images issues such as sun reflexion, water levels etc
Field sites and datasets
- From L134, a lot of work as been done on Narrabeen and you could use more references. L152: storms are usually defined using the POT method based on both Hs and duration thresholds (Harley, 2017)
- L162: rephrase
- L180: remove “on the order of”
- L185: remove “(due to high tide or low waves)
- L198: as for
- L200: unclear. Averaged over which period of time?
- L206: to fit the best…
Methods
- L237: output “from” the previous processes
- L261: their weights
- L265: not always “available”
- You should here give a bit more details about the choice of the settings. Are they all based on previous work?
- Do you think that the geometric transformations (augmentations) can be a reason for some misclassifications that occurred?
Setup
- I think this part describing the two experiments need more clarifications. It’s a bit confusing and the difference between the two experiments is actually better described in the conclusions (difference between CombinedSINGLE and CombinedINNER/OUTER). A diagram for the second experiment could help.
Moreover, does using datasets covering the same periods of time play a role in the CombinedSINGLE computation (as you prepare chronologically the data beforehand)?
- L373: prefer increasing to varying
- Table 6: describe TP, FN, FP and TN in the caption
Results
- Table 7: what the 0.78 value at the bottom?
- L440: not lower as the maximum value is the same. Bigger range or lower min value
- Figure 7b: TBR and LBT are classified as RBB, but true RBB are always well classified. Do you know why?
- L460: “relatively often”? rephrase
- L477: prefer biggest to greatest.
- Figure 9a deserves more description even if the results are not showing a good classification
- L499 to 503: I understand what you mean but the sentence is unclear/very heavy. Rephrase
- L511: in the case of the OUTER-CNN/ L515: INNER-CNN
- L512: not slight. 50%! It’s significant, not slight
- L522: reaching instead of achieving
- L523: remove ranging
- L533: only tested
- Fig 13b: Do you think the higher values of recall values can be linked to the lower options of beach states? You don’t talk about that point in the discussion. Do you think that the number of options (beach states) has an influence on the results and the accuracy of the models?
Discussion
- In general, I think that the discussion could use a bit more work. First, when you are coming back to results, you must help the reader and specify which figures you are referring to. So far, it is very hard to follow. Moreover, the paragraph L613 to 624 is confusing and need re-writing.
- L539: remove “notable is that”
- L543 & 544: corresponding to
- L571: That can explain why the inner bar is not well identified but it doesn’t explain why the outer bar is a problem. Indeed, the outer bar is subtidal (low tide or hight tide) and not linked to the shoreline. Any explanation comes to mind?
- L576: does the model need more data or better methods to identify the structures and differentiate the inner bar from the shoreline?
- L588: You said imbalanced data must be the explanation for the results. However, bias is observed for LTT and TBR states. If TBR can be explain by a 62% frequency, the LTT represents only 13% (less than RBB). How can you explain the bias in classifying LTT then? Moreover, for the training data of the outer bar, LBT is biased while its percentage is very close to TBR. Do you think the imbalance can really explain the bias here too? I am not really convinced here.
- L594: comparable to
- L600: this is consistent with your previous results showing there are more difficulties to identify the inner bar compared to the outer one
- L633 to 638: You are pointing at different reasons, but which are the reasons in your specific case that can explain that “non positive effect” on the model? As in the description of the data, you said that you limited your studied area to avoid artificial reef etc…
Conclusions
- The conclusion could better explain the main messages from the paper
- L663: one from both
- L680: “equal amount […] from each location”. If you use 50% Gold Coast, you then use 25% Narrabeen and 25% Duck.. therefore, not an equal amount of each location. Rephrase.
Author Response
We thank reviewer 3 for the kind remarks and highly appreciate the suggestions for further improving the manuscript and clarifying our work. Below, the reviewer’s comments are given in italic. Accordingly, italic mentions of line numbers refer to those in the original manuscript, other mentions of line numbers refer to those in the revised manuscript.
Dear authors,
Firstly, I’d like to acknowledge the work you put in the different experiments and in the paper. Your work is a good add-on to previous work and a step forward to better automatic beach classification.
Therefore, I want to recommend your paper to be accepted with minor revisions (listed below) as I do think some clarifications are needed in the description of the different models/experiments and in the discussion.
> We thank reviewer 3 for the kind words and appreciate the suggestions for further improvement. Below, we address each of the raised concerns.
Abstract
- From the first line I think there is a confusion that continues all along the paper. You here talk about subtidal sandbars. However, you say in the discussion that the difficulty of differentiate the shoreline and the sandbar can be an issue for sandbar recognition. Those bars are intertidal bars, not subtidal. As it is usually the case for TBR states (and LTT).
> The reviewer raises a valid point. Although we consider the inner bar at the Gold Coast as a subtidal bar, the Low Tide Terrace (LTT) state is inherently intertidal. Sandbars that originate in the intertidal zone (i.e. intertidal bars; slip-face ridges and ridge and runnel systems) are not included in our analysis and instead we focus on what we call subtidal (or submerged) bars. To prevent this misunderstanding, we changed “subtidal” into “nearshore” in the first line of the abstract, and stress that we focus on subtidal sandbars in particular (line 30).
Introduction
- L30 & L37: remove subtidal; same comment as above.
> We adapted line 30 by stating that “...sandbars, in particular subtidal sandbars,...”
L37: removed “subtidal”
- L46: rephrase, sandbars don’t always present every states. They can however evolve from one state to another depending on hydrodynamic conditions
> Thank you for pointing this out. We now adapted this (line 45-48) accordingly: “In general, during low-energetic accretionary conditions, sandbars advance sequentially in downstate direction, whereas sandbar morphology may jump to a higher state during high-energetic erosional sequences.”
- L53: seaward-most outer bar; landward-most inner bar
> Done.
- L57: what do you call in-situ measurements? Video images are in-situ methods.
> We define in-situ measurements as measurements that require that the instrumentation be located directly at the point of interest and in contact with the subject of interest (water, sand, bed level). In contrast, remote sensors are located some distance away from the subject of interest; in our case the video cameras. We now clarify what me mean by in-situ measurements by stating in line 57: “In-situ measurements, where hydrodynamics, sediment transport and bed levels are measured directly in the field,…”
- L69 to 77: I think you can tell a bit more about those methods/models, their accuracy, how do they deal with images issues such as sun reflexion, water levels etc
> These methods were mainly meant to illustrate the application of machine learning techniques on nearshore video imagery. However, an important aspect of our method is the deep learning aspect in which no human tuning is required and focusses on the automated processing and classification of the images. Hence, we thought the image processing methods in these studies not to be relevant to our study.
Field sites and datasets
- From L134, a lot of work as been done on Narrabeen and you could use more references. L152: storms are usually defined using the POT method based on both Hs and duration thresholds (Harley, 2017)
> We thank reviewer 3 for the suggestions and we included references to: Splinter et al (2018), Harley et al (2011) & Harley (2017).
- L162: rephrase
> We have changed this line (now line 166) to “Furthermore, the nearshore predominantly consists of quartz sand with a median grain size of 0.225mm, and exhibits an average slope of approximately 0.02.”
- L180: remove “on the order of”
> Done
- L185: remove “(due to high tide or low waves)
> Done
- L198: as for
> Done
- L200: unclear. Averaged over which period of time?
> One image of the Gold Coast is an average of 600 images taken at low-tide period with a frequency of 1 Hz. So, this is done over a time frame of 10 minutes. To clarify we now say “..time-averaging..” in line 202
- L206: to fit the best…
> Done
Methods
- L237: output “from” the previous processes
> Done
- L261: their weights
> Done.
- L265: not always “available”
> Done
- You should here give a bit more details about the choice of the settings. Are they all based on previous work?
> In line 282, after “several adjustments were made to the training settings as used by Ellenson et al (2020), we added “these adjustments were made empirically.”
- Do you think that the geometric transformations (augmentations) can be a reason for some misclassifications that occurred?
> The applied transformations do not change the aspect ratio of the image or sandbar morphology, and only concern a rotation, shift or changed lighting. As such, the morphology remains identical. Any misclassifications of the method depend on variability within the originally acquired images and not the following transformations. To clarify this, we now added “morphology” in line 321-322: “..while simultaneously preserving the morphology and labels.”
Setup
- I think this part describing the two experiments need more clarifications. It’s a bit confusing and the difference between the two experiments is actually better described in the conclusions (difference between CombinedSINGLE and CombinedINNER/OUTER). A diagram for the second experiment could help.
> We have rephrased certain parts, and we think it will be clearer now. We focused on expressing the differences between the experiments. L359: ‘In the second experiment, we assessed the performance of models trained using not only single-barred data like in Experiment 1, but also double-barred data.’ & L370: ‘The training data of Duck and Narrabeen combined, as used for training the CombinedSINGLE-CNN was supplemented with data of the Gold Coast.’
Additionally, we have referred more often to the training table in which the experiments, the corresponding models and the locations on which they have been trained are clearly shown.
- Moreover, does using datasets covering the same periods of time play a role in the CombinedSINGLE computation (as you prepare chronologically the data beforehand)?
> Good point. The datasets are prepared separately from each other. Therefore, this does not play a role.
- L373: prefer increasing to varying
> Done
- Table 6: describe TP, FN, FP and TN in the caption
> We now include the description of these abbreviations in the caption.
Results
- Table 7: what the 0.78 value at the bottom?
> This value is for the case where the CombinedSINGLE-CNN is tested on the combined testdata of Duck and Narrabeen.
- L440: not lower as the maximum value is the same. Bigger range or lower min value
> Done
- Figure 7b: TBR and LBT are classified as RBB, but true RBB are always well classified. Do you know why?
> As it is hard to say on what specific features a CNN trained, we cannot say for sure why RBB is well classified. The most likely explanation is what is stated on L571-572: “Due to the lack of a consistent optical signature of the shoreline at Narrabeen, the separation between shoreline and sandbar is less evident...” Hence, it could be that the features of the RBB state at Duck are best recognized.
- L460: “relatively often”? rephrase
> Done
- L477: prefer biggest to greatest.
> Done
- Figure 9a deserves more description even if the results are not showing a good classification
> This is now Figure 10a; we have more elaborately described the corresponding findings in lines 481-489.
- L499 to 503: I understand what you mean but the sentence is unclear/very heavy. Rephrase
> Rephrased
- L511: in the case of the OUTER-CNN/ L515: INNER-CNN
> Done
- L512: not slight. 50%! It’s significant, not slight
> Line 527 now states a “significant bias”
- L522: reaching instead of achieving
> Done
- L523: remove ranging
> Done
- L533: only tested
> Done
- Fig 13b: Do you think the higher values of recall values can be linked to the lower options of beach states? You don’t talk about that point in the discussion. Do you think that the number of options (beach states) has an influence on the results and the accuracy of the models?
> We adjusted the figures so it might be clearer this way. In Figure 12b the confusion matrix is shown for the self-test of the OUTER-CNN, so the CNN trained with only the data of the outer bar. This bar exhibits only 3/5 states, hence there are no true R and LTT images. As this CNN is only trained with three classes it is only able to classify three classes. On the other hand, in Figure 14b (Figure 13b) we see the confusion matrix of the model trained with data of Duck, Narrabeen and the Gold Coast with the outer bar labels. As Duck and Narrabeen exhibit all 5/5 states, this model is still able to classify all five states. Hence, the R and LTT are no NaN values, but are shown as 0. As an answer to your question: The option of the number of beach states is the same, so in this case it has no influence on the recall values.
Discussion
- In general, I think that the discussion could use a bit more work. First, when you are coming back to results, you must help the reader and specify which figures you are referring to. So far, it is very hard to follow. Moreover, the paragraph L613 to 624 is confusing and need re-writing.
> References to figures have now been added to the text. Additionally, we tried to clarify the intention of lines 613 to 624 by adding a short sentence, introducing the aim of the paragraph: “Depending on the dataset used for training, the bias towards certain beach states” (L 630-631).
varied.
- L539: remove “notable is that”
> Done
- L543 & 544: corresponding to
> Done
- L571: That can explain why the inner bar is not well identified but it doesn’t explain why the outer bar is a problem. Indeed, the outer bar is subtidal (low tide or hight tide) and not linked to the shoreline. Any explanation comes to mind?
> The largest margin of error for the Gold Coast is probably since when trying to classify the outer bar, the whole image is taken into consideration and not only the outer bar. Hence, when an image is classified as outer bar it cannot be classified as the inner bar, and vice versa. When classifying the Gold Coast the test data is similar for both the inner and outer bar cases. We see F1 scores of 0.32 and 0.55 corresponding to the classification of the inner and outer bar, meaning that approximately 0.32% is classified as inner bar and therefore is limiting the number of images in which the outer bar can correctly be classified.
- L576: does the model need more data or better methods to identify the structures and differentiate the inner bar from the shoreline?
> Yes, both aspects mentioned here will result in improved results and, as such, form logical next steps. We mention this is the final paragraph of the discussion section (L657-663), where directions for future research are mentioned.
- L588: You said imbalanced data must be the explanation for the results. However, bias is observed for LTT and TBR states. If TBR can be explain by a 62% frequency, the LTT represents only 13% (less than RBB). How can you explain the bias in classifying LTT then? Moreover, for the training data of the outer bar, LBT is biased while its percentage is very close to TBR. Do you think the imbalance can really explain the bias here too? I am not really convinced here.
> There is probably a small misunderstanding here. As described on L593: ‘... R/LTT/TBR/RBB/LBT being 13%/62%/18%/2%/5%’. TBR is not explained by a 62% frequency, but LTT is (representing 62% and not 13%).
- L594: comparable to
> Done
- L600: this is consistent with your previous results showing there are more difficulties to identify the inner bar compared to the outer one
> Sentence added in which this is mentioned (L616-617): “This is consistent with our previous observations showing that there are more difficulties to identify the inner bar compared to the outer bar.”.
- L633 to 638: You are pointing at different reasons, but which are the reasons in your specific case that can explain that “non positive effect” on the model? As in the description of the data, you said that you limited your studied area to avoid artificial reef etc…
> This sentence indeed a bit confusing, as we mix up our own double-barred site and the addition of potential future sites. We now clarify this paragraph (L645-655), including the comment “Reasons for this may be the site-specific features, such as tidal range, number of cameras, and wave climate. A majority of images coming from one specific site could result in the model training on such a specific feature.”.
Conclusions
- The conclusion could better explain the main messages from the paper
> We now start the conclusions section with a paragraph that summarizes our main findings: “The main findings of our work are that (1) a CNN trained with images from single-barred beaches shows poor performance when classifying double-barred beach states, (2) transfer learning, where limited data from a double-barred beach is added to the single-barred model, allows for the training of a well-performing model for classifying double-barred beach states and (3) including outer-bar labels in the transfer learning has a larger impact on the resulting model performance than when labels from the inner bar are included.”
- L663: one from both
> Done
- L680: “equal amount […] from each location”. If you use 50% Gold Coast, you then use 25% Narrabeen and 25% Duck.. therefore, not an equal amount of each location. Rephrase.
> Rephrased, now mentioning: ‘... with an equal amount of the total training data...’
Reviewer 4 Report
General comments:
The paper shows results that are extended from the previous similar article published in 2020: 10.3390/rs12233953. The methods and datasets are similar to the previous study (like it was mentioned in line 106).
The discussion seems to be like interpretation or results, however, it should be enlarged covering at least the following topics:
- the implication of your study to a wider scientific audience
- limitations of your study
- recommendations for future research
I suggest using for example the following references to enrich the discussion:
- 10.2112/JCOASTRES-D-20-00150.1
- 10.1016/j.csr.2020.104213
Specific comments:
line 77: suggested additional reference: 10.1016/j.enggeo.2022.106615 - utilisation of ML to classify the seafloor and extract sandbars from airborne Lidar bathymetry.
line 190: can you provide more details of the cameras used in the Argus system?
Author Response
We thank reviewer 4 for the kind remarks and the suggestions for improving the manuscript. Below, the reviewer’s comments are given in italic. Accordingly, italic mentions of line numbers refer to those in the original manuscript, other mentions of line numbers refer to those in the revised manuscript.
The discussion seems to be like interpretation or results, however, it should be enlarged covering at least the following topics:
- the implication of your study to a wider scientific audience
> We chose to refrain from a wider scientific interpretation of our results in the discussion and to focus on the application of our technique to beach states in particular. Instead, we chose to provide an elaborate and systematic description of our methodology, from which a wider scientific audience may take lessons out of our application of transfer learning to image classification.
- limitations of your study
> We thoroughly discuss the limitations of our study throughout the discussion section, such as the causes of misclassifications, the expression of the beach state morphology, and the processing of the available data.
- recommendations for future research
> Our explicit recommendations for future research (lines 657-662) include: (1) collection of more data, (2) object detection and (3) object tracking.
I suggest using for example the following references to enrich the discussion:
- 10.2112/JCOASTRES-D-20-00150.1
- 10.1016/j.csr.2020.104213
> Thank you for the suggestions. We chose not to add these to the discussion section, as these studies use other techniques that do not directly relate to our approach.
Specific comments:
line 77: suggested additional reference: 10.1016/j.enggeo.2022.106615 - utilisation of ML to classify the seafloor and extract sandbars from airborne Lidar bathymetry.
> We thank review 4 the suggestion, and we have added the suggested reference.
line 190: can you provide more details of the cameras used in the Argus system?
> We left detailed specifications of the Argus cameras out of the paper on purpose and provide the relevant references. We now refer to these references explicitly as “For details on the Argus camera system see [ref]”, in lines 181 and 201 in Section 2.2 .
Reviewer 5 Report
Dear authors,
The reviewer is satisfied with the manuscript. It describes well the data, method, and results. Please consider the following in preparing the final manuscript.
The reviewer supposes that the authors are interested in monitoring and understanding the coastal processes, behaviors, and others. Please describe in a proper place how the method described in this work can contribute to these. Further, what is the required F1 score for properly analyzing coastal processes and behaviors?
Best regards,
Author Response
We thank reviewer 5 for the kind remarks and the suggestions for improving the manuscript. Below, the reviewer’s comments are given in italic. Accordingly, italic mentions of line numbers refer to those in the original manuscript, other mentions of line numbers refer to those in the revised manuscript.
Dear authors,
The reviewer is satisfied with the manuscript. It describes well the data, method, and results. Please consider the following in preparing the final manuscript.
> We thank the reviewer for the positive feedback!
The reviewer supposes that the authors are interested in monitoring and understanding the coastal processes, behaviors, and others. Please describe in a proper place how the method described in this work can contribute to these.
> We added a sentence in the conclusion that as our work shows that single and double-barred beaches can be classified using a CNN and transfer learning, this is a step forward to better automatic beach state classification (line 681).
Further, what is the required F1 score for properly analyzing coastal processes and behaviors?
> We thank reviewer 5 for the valid point. As we did not find an F1 score which is defined as ‘properly’ for analyzing coastal processes and behaviors, we analyzed everything relative to previous work. We defined a proper F1 score when it was comparable to or better than previous work.
Round 2
Reviewer 2 Report
-figures should be better quality
-references to figures, formulas, and tables in the text should be added;;
-it is a good idea to add some more photos of sensors/applications of the proposed research;;
Author Response
We thank Reviewer 2 for the additional comments. Please find our responses below, we hope this sets all remaining concerns at rest.
-figures should be better quality
> We have adjusted figures 3, 5 and 6 to improve their quality.
-references to figures, formulas, and tables in the text should be added;;
> We ensured that we refer to all figures, formulas and tables in the text. After the first round of reviews we added additional references to figures in the discussion section. Additionally, we have added references on L195 & 215. However, unless specific instances of additional references are provided, we are unaware where else these could be added otherwise.
-it is a good idea to add some more photos of sensors/applications of the proposed research;;
> We are not sure what the reviewer wishes to see here. We provide links to further details about the cameras in the methods section (which include photos of the camera systems). Otherwise, the application of our research concerns deep learning algorithms that predict the beach state within a provided video image.
With kind regards,
Stan Oerlemans
