Deep Learning of Sea Surface Temperature Patterns to Identify Ocean Extremes

Round 1

Reviewer 1 Report

Introduction section is poor. Please improve it.

Please provide better literature review (use new an update papers).

Provide a graphical abstract for better understanding.

Don not put some figures and tables continuously, each tables and figures must be discussed separately.

Figure 12 needs more explanation.

Some of figures’ caption are too ling (e.g. Figure 12), make shorter.

Conclusions must be extended, no need repeat again the result in conclusion section.

Author Response

Thanks for your careful reading of our manuscript and your helpful comments and criticisms. We comment on each below and detail the resultant revisions to the manuscript. Introduction section is poor. Please improve it. -- We found it difficult to write this Introduction in part because the intended audience ranges from ML to SST experts. Indeed, the other reviewers have apparently found it up to standard which probably reflects their differing backgrounds. In any case, we welcome specific suggestions to further improve it. Please provide better literature review (use new an update papers). -- This is a good suggestion. We have added a sentence referencing Ma et al. 2019 who provided a meta-analysis of the literature on ML applied to remote sensing as well as a number of references of the use of machine learning associated with sea surface temperature fields and how our work differs from these. Provide a graphical abstract for better understanding. -- This is an excellent recommendation. We now include a graphical abstract. Don not put some figures and tables continuously, each tables and figures must be discussed separately. -- Sorry, we don't understand this comment. The Table in the Appendix is not near a Figure. Maybe the reviewer is concerned that Figure 11 is on its own page? Or that Figures 8 and 9 fill a page. In any case, we have made a number of changes for the figures to be better embedded in the text. Figure 12 needs more explanation. -- We have added a few clarifying points to the figure caption. This has made it a bit longer in contradiction with the next comment by the reviewer. The main text, meanwhile, has several paragraphs dedicated to introducing and discussing this figure. We did not add additional text. Some of figures’ caption are too ling (e.g. Figure 12), make shorter. -- As per the reviewer's request above, we *increased* the caption for Figure 12. We have, however, reduced the caption for Figure 8. We attempted the same for Figure 9, but found all of the text required to faithfully describe it. Conclusions must be extended, no need repeat again the result in conclusion section. -- We believe the current text matches well the guidelines outlined for the Conclusion section except for acronyms in the Instructions for Authors (https://0-www-mdpi-com.brum.beds.ac.uk/journal/remotesensing/instructions#preparation). We have expanded the acronyms and added additional text to make this section more self-contained.

Reviewer 2 Report

This is an interesting work in Oceanography community. The study designed and developed a ULMO model which is a deep learning model for the identification if Sea Surface Temperature (SST) in the ocean. The model was trained using the data from MODIS satellite images and the data were clearly filtered with the cloud cover threshold. The model also able to retrieve SST patterns from the masked/cloud portion. The patterns learned by ULMO can be able to understand the geostrophic currents and biogeochemical processes within the ocean. 

Author Response

-- We appreciate the reviewer's positive remarks.

Reviewer 3 Report

This is a very nice paper on the growing trend of application of machine learning algorithms on large oceanographic data sets. In this case, CNN is applied to SST to identify anomalous patterns. I think that it is good paper, carefully written and it is based on a solid idea, with much future potential and interest to others in this rapidly-advancing and useful field. So I recommend acceptance.

I have just two curiosity driven questions:

1) Can you really do something you could not before? Cornillion has been dealing with SST for ever and I am wondering whether CNN really identified something new that he hasn’t seen before? Or it is just another method to do the same, i.e. what other methods can do, maybe less painfully or more elegantly, but the result is not a big surprise for the trained scientist’s eyes? Is there a net big advantage to using this for oceanographic data sets?

2) If you have submesoscale features, is overfitting during training an issue? Have you encountered overfitting due to small features on the training set that gave bad results on the test case? 

Author Response

This is a very nice paper on the growing trend of application of machine learning algorithms on large oceanographic data sets. In this case, CNN is applied to SST to identify anomalous patterns. I think that it is good paper, carefully written and it is based on a solid idea, with much future potential and interest to others in this rapidly-advancing and useful field. So I recommend acceptance.

-- Thanks for your positive comments.  We respond to your
  two questions in turn, below.

I have just two curiosity driven questions:

1) Can you really do something you could not before? Cornillon has been dealing with SST for ever and I am wondering whether CNN really identified something new that he hasn’t seen before? Or it is just another method to do the same, i.e. what other methods can do, maybe less painfully or more elegantly, but the result is not a big surprise for the trained scientist’s eyes? Is there a net big advantage to using this for oceanographic data sets?

-- We have added text (lines 284-288) in the revised manuscript detailing the uniqueness of the anomaly fields ULMO has found - yup, Cornillon has been looking at these data forever :-) We have also added a paragraph (lines 341-352) emphasizing that, in addition to finding what we see as unique cutouts, the algorithm has also found similarities in the structure of cutouts between widely separated regions in the ocean. We are particularly excited about new insights these findings may offer and we are exploring this in the context of a similar analysis of the llc4320 OGCM output.

2) If you have submesoscale features, is overfitting during training an issue? Have you encountered overfitting due to small features on the training set that gave bad results on the test case? 

-- We believe the reviewer is concerned that the autoencoder may not capture
 all of the fine-structure (submesoscale) features in the data.  We agree.
 And, we have developed new pre-processing approaches to accentuate those.
 An earlier draft of the manuscript included this analysis, but we have since
 decided to defer its presentation to a future manuscript.  Kindly stay tuned.

Round 2

Reviewer 1 Report

Dear authors, thanks for your updated version. This manuscript still needs some corrections:

Introduction needs better literature review.

As I mention before, a graphical abstract is required for this study.

Discussion part is missed! Please add a suitable discussion for this study.

Author Response

Thanks again for your review.  Our responses are below.

 

Introduction needs better literature review.
  ++ In the resubmission we included references to 6 papers on ML and
  remote sensing and also the excellent review article by Ma et al. 2019.  We welcome the Reviewer to propose specific additional 
  articles and we will gladly include them.

As I mention before, a graphical abstract is required for this study.
  ++ We provided a graphical abstract in the previous resubmission.  Not sure how to share it with you, but please look for it.

Discussion part is missed! Please add a suitable discussion for this study.
  ++ Thanks for pointing this out.  We chose to combine Results and Discussion but did not properly label the Section.  We have done so now.