Deep Hash Remote-Sensing Image Retrieval Assisted by Semantic Cues
Round 1
Reviewer 1 Report
The English writing is very peculiar and hard to understand, e.g., ‘We propose a move to deep hash remote sensing retrieval method’. Also, the papers lack scientific vocabulary, e.g., depth network instead of the deep network. The paper also misses scientific formal writing (e.g., by the way).
What is semantic information referring to? Never defined in the paper. Also pay attention to how this is used in the paper. Some sentences are quite misleading and hard to understand (e.g., the semantic information used by the image itself to identify image categories is also helpful for representing image retrieval.)
Pay attention to the mathematical notation. Some letters have different meanings. The index i indicates both the image (xi) and the probability of each class (pi). Of course, this makes the text hard to understand.
What is metric learning? From this sentence (and the rest of the section) it looks like a contrastive loss: ‘However, learned by metric learning, the features in the embedding of the positive class are clustered together while the features of the negative class are farther away from other classes and the boundary.’ Anyway, I could not understand how it is implemented in the proposed method.
Moreover, the innovation of the proposed method is extremely limited, and the scientific soundness of the method is not convincing. Hence, considering also the strong effort required for reading the paper given the English writing and the confusion of the structure, I cannot recommend the publication.
Minor comments:
Introduction.
· The structure is confusing. A single concept is repeated several times, and the flow of information is not clear and logical.
· The description of the novel contribution is not clear and mixed with the state of the art. Should be rewritten.
· Remote sensing is not mainly about Earth's surface: there is planetary, there is also subsurface, etc.
· Improve Fig 1 with the inputs and the outputs. It is not clear the flow of the algorithm.
Methodology.
· I cannot understand how Fig 3 ‘shows that the optimized network with classification loss can clearly distinguish different classes of images, but the distance between different classes of images and 288 the boundary is small and the distribution between similar images in embedding space is 289 also discrete.’
· Why the Hamming space and not the feature space?
· What is the proxy set?
Experiments:
· Missing experiments with other contrastive learning methods (I suggest looking at Begum et al.)
· Missing experiments with more complex and big datasets to evaluate also the retrieval time, like BigEarth.
Author Response
Please see the attachment.
Author Response File: 
Author Response.docx
Reviewer 2 Report
This article focused on deep hash remote sensing image retrieval, can significantly improve the researchers work on the rapid increase in the number of remote sensing images. The subject of this paper is insteresting to the readers. But there are several suggestions:
1. The English of this paper need a native language speaker to improve the manuscript, such as some syntax, style, and phrasing problems.
2. The label of figure 2, 3, 4, 7, 8 and tables 1 to 4 is too simple, must be detailed to explain the main results of the figure or table want to convey.
3. The conclusions of this paper is too simple, must be improved and add more discusions about the further work.
All in all, this paper need a major revision.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 3 Report
Review for "Deep Hash Remote Sensing Image Retrieval Assisted by Semantic Cues"
The manuscript is generally good and the topic seems to present very interesting results for readers. I suggest a "Major" revision before possible consideration of the application in remote sensing. My comments are listed as:
1. Although the manuscript is generally well-written, a language check by a professional native speaker or an editing agency is needed to fix some syntax, style, and phrasing problems.
2. The introduction needs to be improved and further discussion is needed. Also, the introduction required adding newly references. And subtitles should be added to the introduction and to related work section.
3. Please check the references as there are many references without Doi. Please check them carefully.
4. I think it is necessary to add a discussion separately of the results and increase it
5. L23: what is UCMD
6. add different keywords and remove (hash; remote sensing; image retrieval)
I look forward to seeing a better version of the manuscript.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 4 Report
The paper is dedicated to deep hash remote sensing image retrieval. Authors have demonstrated that adding the classification label as a kind of similarity semantic cue can provide improved the retrieval effectiveness.
They validated their method on two commonly used remote sensing retrieval datasets and showed the superiority of the method. The paper is mostly written well. The results obtained are new, interesting, and valuable for the field. The results are clear, but their discussion should be provided in section Discussion, which should be organized in the paper. Section Materials and Methods, and section Results also should be organized. The paper also needs some other corrections before its publication in the journal.
Corrections suggested.
1. The paper should have the following structure: Introduction, Materials and Methods, Results, Discussion, and Conclusions. Please, reorganize the paper in accordance with the required structure: https://www.mdpi.com/journal/remotesensing/instructions.
2. Please, add a blank between the word and reference number in the text on page 1.
3. Please, add an empty line before Algorithm 1 caption.
4. Line 347. ‘the ⨁ represents’ should be ‘⨁ represents’.
5. Line 362. Please, add a blank between a value and its dimension.
6. Please, add an empty line after Table 1, 2 and 3.
7. Please, indicate in conclusion where the obtained results can be used.
8. Please, prepare all the references exactly in accordance with the journal template, abbreviate journal names, and add missing volumes, pages, and DOIs. You may use any recently published paper in the journal as an example.
So, the paper needs minor revision.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
The authors strongly improved the manuscript and addressed most of my comments. Fig 3 still misses a color legend and indications inside the figure on the axes labels. Moreover, there are several misleading sentences that should be improved (e.g., Current deep metric learning methods only care about whether two samples have the same label (positive or negative), but not about which class each sample specifically belongs to.).
Author Response
Please see the attachment
Author Response File: Author Response.docx
Reviewer 2 Report
This paper has been revised and can be accepted.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 3 Report
Accept
Author Response
Please see the attachment.
Author Response File: Author Response.docx
