Semantic Segmentation of Metoceanic Processes Using SAR Observations and Deep Learning

Round 1

Reviewer 1 Report

This  paper used  SAR data and machine learning techniques for studying the semantic segmentation of ten metoceanic processes . They concluded that a fully-supervised model outperforms any tested weakly-supervised algorithm. Also, they found out that adding more segmentation examples in the training set would further increase the precision of the predictions. The topic is relevant to the satellite and physical oceanography community. However, the paper is very concise and needs to be expanded more to clearly address the topic. I recommend theis paper for publishing in the Remote Sensing after the authors address the following minor to major issues:

• Although “deep learning” is an appealing part of the paper title and also a keyword to this paper, the methodology behind it has less been discussed. The type of neural network that was used, number of layers, computational efforts, and training data are the most relevant ones. If the convolutional neural network is finally used, please include more details.
• Add a schematic(figure) showing the structure of the neural network in this paper.
• Please include more details about the Dice Index (concept, formulation, and implications in this research)
• Equation1: define all parameters (some of the parameters have not been defined)
• Please include more details about the ground-trough data. Are they in situ measurements or spectral data from other satellite sensors?
• Please include example figures of different phenomena inferred based on the segmentation approach.
• In Figure 3, the table is empty. Is it supposed to be like this or something is missing?
• Please describe how the results from this research are used for quantification of oceanic phenomena? In this paper the results were discussed qualitatively. It is important to address quantification of the ten oceanic phenomena discussed in the paper.
• Ln 97-98: “The distribution of these phenomena over the globe, as contained in the TenGeoP-SARwv dataset, is depicted in the figure 1” : what do you mean by “distribution”? Is it number of data acquired by satellite or the quantity of the parameters itself? This sentence is confusing when you see the caption for figure1.

Author Response

Thank you for your comments. They proved to be highly beneficial to this paper. Please see the attachment PDF file that we uploaded. We hope it will answer the different points that you raised.

Author Response File: Author Response.pdf

Reviewer 2 Report

In the submitted manuscript, studies have been carried out about the semantic segmentation of ten geophysical phenomena, including both oceanic and meteorologic features using SAR images and deep learning. Two paradigms of semantic segmentation, namely fully-supervised framework and weakly-supervised framework, are investigated and evaluated based on the TenGeoP-SARwv dataset. 

The manuscript technically sounds and the discussed case could be of interest to the potential readers of this manuscript. However, I think the authors should prepare their manuscript more carefully and I have some concerns and comments which are suggested as below:

1. There are no examples of output on elements of the test set for each segmentation method in Figure 3? Please check the manuscript carefully. 
2. In table 1, the unit m is missing for 100.px-1, 200.px-1 and 340.px-1.
3. In line 96, the Pure Ocean Swell is one of the ten metoceanic processes studied in your work. However, in the rest of the manuscript, the Pure Ocean Wave is used. Please check it to ensure that it is consistent. 
4. In the paragraph beginning from line 251, I think you miss mentioning Pure Ocean Waves (POW). In table 2, the supervised framework gives the highest Dice index of 47.9%, compared with the other methods.  
5. In Table 2, the Dice index was chosen to evaluate the segmentation results. It will be better for readers to understand and analyze the results if a brief introduction about the dice index could add to the manuscript. 

Author Response

Thank you for your comments. They proved to be highly beneficial to this paper. Please see the attachment PDF file that we uploaded. We hope it will answer the different points that you raised.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The authors addressed all my comments.

Reviewer 2 Report

This reviewer thinks that  the revision can be accepted.