Ship Detection in Sentinel 2 Multi-Spectral Images with Self-Supervised Learning
Round 1
Reviewer 1 Report
This paper presents an application of deep learning methods to the problem of ship detection from multispectral images taken from a satellite (Sentinel 2). A toolchain for this is presented, involving the U-Net architecture with custom modifications and parameter tweaking. By incorporating self-supervised learning, good results are shown even with a small number of labeled datasets available for training.
In general I found this work to be well written and interesting. I have some comments/suggestions for the authors:
- In Section 1, you should more immediately explain what the "Sentinel 2" (which is in the paper title) is. You refer to a "Copernicus Sentinel" in line 35, but it's not clear if this is Sentinel-2, or if there is also a Sentinel-1. Perhaps include a picture of the satellite and/or a table of its capabilities within Section 1?
- Line 58: "the increasing number of datasets publicly available" I think?
- In the opening paragraph, you talk about ship detection; however, could say a few words about whether ship identification would be possible as well? Maybe this is not doable with commercial satellite imagery due to the low resolution, but are you aware of anyone doing this? See my last comment below
- Line 175: "As a first step we rasterize the ground truth polygons .. into a binary mask" - I don't understand the word "rasterize" here, is this done by hand? For me "rasterization" involves the process of converting a vector image into a pixelized image, yet here I think you're doing a segmentation?
- Line 256: you mention that you run your pertaining on a multi-GPU machine; yet on lines 124 and 131, you mention your architecture running well on a "desktop GPU" and "single desktop GPU". So do you need a multi-GPU at some point, or am I missing something?
- Line 353: "scratch achieved better F1
- Line 360: a qualitative analysis of their results
- Line 361: Some piers and
- Line 371: I think "spareness" of data sets, not frugality
- Line 388: to fine-tune its weights further,
- Line 408: I think you can remove the word "thoroughly"
- Line 433: you mention "SAR methods" here - could you add a reference or explain a bit more? These seem to give much better performance than your system if I understand correctly
Author Response
Please see the attachment
Author Response File: 
Author Response.docx
Reviewer 2 Report
The authors presented a method to train a ship detector in Sentinel-2 images using self-supervised learning. Their method plugs in an SSL-trained backbone in a U-NET architecture. Their work shows that there is room for improvement although the direction towards this goal remains unclear. The manuscript is well written and should be of great interest to the readers. More clearly results. All figures could be bigger. They are not clear.
Author Response
Please see the attachment
Author Response File: Author Response.docx
Reviewer 3 Report
The paper is overall well-written and clear. It is well-organized and balanced. The approach is interesting and results are encouraging. I have only the following minor comments before acceptation:
- The first question regards the performance metrics used. Why did the authors not consider accuracy?
- More discussion about the overall computational complexity could be somehow useful to the readers.
Author Response
Please see the attachment
Author Response File: Author Response.docx
