Ship Detection with Deep Learning in Optical Remote-Sensing Images: A Survey of Challenges and Advances

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

In this paper,the author studied methods for automatic recognition of ships in images and precise classification and localization.And the author focuses on studying methods such as ships operating in complex marine environments, insufficient discriminative features, large scale variations, dense and rotational distributions, large aspect ratios, and imbalances between positive and negative samples.The author provides the advantages and disadvantages of these methods.This article makes a good review of the relevant algorithms of the ship detection, which is conducive to the future research of the ship detection algorithm, so I'd like to recommend the paper.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Dear authors,

in general, I find your review titled "Ship Detection with Deep Learning..." comprehensive and useful. It is necessary to explicitly state that it concerns only ship detection in nadir imagery (if not, please point to examples of applying the described methods to oblique imagery, - I have probably overlooked them in the paper). Further, the methods described concern not only ship detection, but also their classification – this is worth mentioning explicitly.

A couple of minor comments:

Line 49: the abbreviation SAR has not been explained before use

Figure 1: The orange exert in the large image and in the zoomed-in image have different content

line 166 – line 176 are the repetition of lines 157-166

line 304 – probably you meant “occluded” ships?

4.1 Datasets: “smaller” ship in your terminology only refers to the pixel size of the object, not to its actual size: please state it explicitly in the text, because a ship occupying 100 pixels in a 16m resolution image (LEVIR) is in effect larger than a ship occupying 1000 pixels in a 0.5m resolution image (DOTA).

Table 7 – what does the column “Category” mean? Are this classes of ships in the dataset?

Tables 8 and 9 - what is the motivation between summarizing the performance differences as one value in Table 8 but then demonstrating different performance for different object classes in Table 9? This makes the comparison of algorithm performance on different datasets very difficult. Also, the order of algorithms is different, they are grouped in one table and given as a single list in another one - what is the reason for this inconsistency?

Lines 989-990: If your intention is to compare the feature extraction capabilities between CNN and Transformer, why are you using different input imagery (Figure 20 and Figure 21 demonstrate different inputs)? Wouldn't the use of the same input be more reasonable? 

Wish you all the best with your further research!

Comments on the Quality of English Language

Please make sure to get a proof-reader before publishing: there are some minor grammar improvements needed.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

#Review

Article: remotesensing-2879607

Title: Ship Detection with Deep Learning in Optical Remote Sensing Images: A Survey for Challenges and Advances

        I.          General Appreciation

This manuscript presents a review of ship detection in optical remote sensing images (SDORSI). A detailed analysis of some methods is carried out and both weaknesses and potential are analyzed. The authors consider a set of data that they tested for two methods in order to obtain improvements in terms of accuracy in detecting this type of objects, respectively, CNN (Convolutional Neural Networks) and Transformer methods.

The article addresses a relevant topic in the field of object detection from optical images of Remote Sensing. The manuscript is robust and is well supported with recent bibliographic references, although too long. Presents several methods in detail and only two of them are used. The idea underlying the title of the article is interesting, so the description of all these methods could be summarized, presenting tables where limitations and advantages of each of them could be compared and analysed and maintain the use of the two methods based on manual feature extraction, CNN and Transformer. Therefore, It is suggested that chapter 3 be reorganized to make it less extensive.

It is understood that the article will be accepted for publication, after the reorganization of chapter 3 according to the indications already indicated.

Comments on the Quality of English Language

Minor editing of English language required.

Author Response

Please see the attachment

Author Response File: Author Response.pdf