A Transformer Model for Coastline Prediction in Weitou Bay, China

Round 1

Reviewer 1 Report

Article summary

Delineating the position of coastlines automatically, quickly, and accurately is important for tracking long-term expansion and erosion trends necessary to implement appropriate coastal management strategies. Whilst numerous methods for delineating coastlines exist, their wide use is hampered by issues including data availability, computational power, and applicability for long term analysis. To improve coastal position detection, the authors assess the performance of two machine learning algorithms (support vector regression and long short-term memory) and a transformer model against coastline delineations derived from an image classifier with tidal and manual corrections using several uncertainty metrics. The authors conclude that transformer models yielded the highest accuracy scores, although undeveloped coastlines were often delineated more accurately than developed coasts. 

 

General comments

The manuscript is generally well written, and contributes to the important challenge of automatically delineating coastlines from aerial images accurately. I have four main concerns with this work:

1.     The text in the discussion section largely reads as an extension to the results and so fails to “discuss” the results. The authors miss opportunities to demonstrate how their findings impact the field. A reworking of the discussion section is required.

2.     The abstract (line 16) and conclusion (lines 477-478) are misleading in suggesting that this work presents a method to predict future coastline change. No evidence is presented that any of the models can actually predict coastal change into the future. From reading the abstract, I expected to see models attempt to predict coastal change from 2010 to 2020, with a biannual assessment of their performance to the actual position of the coastline (i.e., using a hindcast modelling approach). Whilst there is a temporal element to this work (with images taken at two periods per year between 2010 to 2020; table 1), and whilst IWE+TC were generated for each image (line 196) and all data was apparently used to train the transformer model (356 -370), only the performance of each inference method for the first quarter of 2020 are assessed. I believe the authors should clarify that this manuscript has developed a method to delineate coastlines accurately from aerial images - not predict future coastline positions as is presently done, and clarify why analysing multiple image years was necessary.

3.     The instantaneous waterline extraction and tidal correction (IWE+TC) method is used to delineate the “real coastline”, with some additional “manual correction” (line 331). Two things are unclear to me: 

a.     a typical approach of assessing the accuracy of inferred methods (transformer, SVR, and LSTM here) is to compare their performance against direct observations (i.e. “ground-truthing”). Is there a reason no GPS coordinates were taken along at least part of the coastline against which to assess performance? And what steps were taken to ensure the IWE+TC method fairly represents the coast? Inset of Fig. 6 does imply a visual inspection of the results was done to confirm accuracy. Please clarify.

b.     as a means to delineate coastlines, isn’t IWE+TC better than transformer, SVR, and LSTM models, given that it was selected as the “real coastline”? Please clarify.

4.     There is no description of how SVR and LSTM models were trained nor what their purpose was. Could the authors please include this in the methods section. 

I also include a number of specific (and generally minor) suggestions below. Given the scale of the changes outlined above, I recommend major revisions of this work be done before being accepted for publication.

 

Specific comments

Lines 16 & 18: “Coastline data” – please be specific. Do you mean “Location”?

Line 20: Without much context, the line “with three evaluation indices called Correct, Complete and Quality” is unhelpful. I suggest removing.

Lines 20-21: What is meant by “Compared with those 20 of the traditional methods”? Please clarify or simply delete.

Lines 22-23: Without explanation of what “Correct, Complete, and Quality” mean, this is unhelpful. Please clarify.

Line 28: Please add a final discussion style sentence at the end of the abstract to highlight take-home messages of the paper.

Line 40: “destructive disasters” is a little vague, please clarify.

Line 77: “In general” is unnecessary, please delete.

Line 79: Delete “it”

Line 88: “Furthermore” is unnecessary. Please delete. Please also add a return to start this as a new paragraph.

Line 130: There should be a small addition to the final paragraph (line 130), which needs a connecting sentence – something like “Having identified potentially suitable methods for delineating coastline positions, ”

Line 132: Please clarify what is meant by “feasibility”. Does this relate to computing time?

Lines 139-142: This sentence is difficult to understand, please revise (ideally as two separate sentences).

Figure 2: For propriety, please only include identifier prefixes at the bottom of each image if referred to in the figure caption / elsewhere in the text.

Line 163: “are” should be “were”. Please use past tense phrasing throughout.

Line 163: Please reference “the USGS archives”, either in the reference list or using a web link.

Lines 164-165: Please clarify what is meant by “first quarter” and “third quarter”. I presume you mean quarters of the year? 

Table 1: Since “Datum”, “Projection”, “Resolution”, “Path/row”, and “File type” are consistent across each image, please remove these from Table 1 and instead provide a sentence summarising this information in the text.

Line 199: Replace “, and this” with “. This”

Line 248: Please ensure “m” is in italics

Line 274: Please remove “Specifically”

Line 299: Should k be italicised, and I am unclear what is meant by “a stride of size 1”. Please clarify.

Line 307: Should “extracted” read “observed”?

Line 311: Please delete “, and a method”

Line 331: Please remove “This method is extended to this article:”

Lines 356-370: This should appear in the methods section.

Line 367: “Moreover” is not required. Please delete.

Line 371: In the two following paragraphs, the results could do with some deeper insight. E.g., it would be beneficial to state, in metres, how large the uncertainty can be in the areas highlighted in the Fig. 7 inset.

Lines 374-375 & 386-388 & 421-423: As these are sentences are interpreting the results, please move to the discussion.

Figure 7: I think it’s slightly misleading to use the term “Real coastline”, given that this was delineated using the instantaneous waterline extraction + tidal correction method. I would therefore use the identifier “IWE+TC (real coastline)” instead. 

Figure 8: Requires a scale bar and a faded basemap. The lines are quite thick as well, I’d rather see them thinned.

Lines 403-409: Please move to the methods section.

Line 427: “is” should be deleted. 

Line 493: What is meant by “tidal factors”? Please clarify.

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Overall, the paper is interesting and well-structured and well written in my opinion. There are only some minor issues I would refer the authors to:

 

1. Key words should not start with capital letters, for instance Coastline prediction.

2. Spaces should be checked throughout the paper, for instance:

 nature[1, 2] should have a space before the bracket and that is the case throughout the whole text.

3. References are not written according to the instructions

4. The questions that should be raised by the paper for instance in the Conclusion (that should be a bit stronger) are:

- who is it useful for;

- what are the weaknesses of the research;

- what are the recommendations and implications of the research;

- how could/should futures studies improve the model etc.

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Line 145 coastline type

In the description of the research area, it would be necessary to add oceanographic characteristics and parameters, especially tidal parameters (when the tide occurs, whether it has diurnal, semi-diurnal, or mixed characteristics) that significantly affect the coastline. In the paper, the authors describe only the tourist characteristics of the area they work on.

Mandatory information in Table 1 is the recording time and the current state of tides in the observed area.

In the workflow chart in Figure 3, the pre-processing of the data needs to be elaborated because radiometric and atmospheric correction and cloud masking are missing. It would be good to describe each of these corrections in more detail. Equally, the tidal correction is not well described and requires a more detailed explanation and description.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf