Semantic Labeling in Remote Sensing Corpora Using Feature Fusion-Based Enhanced Global Convolutional Network with High-Resolution Representations and Depthwise Atrous Convolution
Round 1
Reviewer 1 Report
No comments.
Author Response
We thank the reviewer and editor for the comments and suggestions. The paper has already been fully revised by follows the feedback and advice via https://drive.google.com/open?id=1X9DmChYy9uElqeu8_zNad9n8F9obv6xk
- We thank the reviewer for his/her positive scores on our manuscript.
Author Response File: Author Response.pdf
Reviewer 2 Report
First of all, I would like to thank the Remote Sensing Editorial Committee for the possibility of reviewing this work, as well as the authors themselves for their interest in having this journal as a means of disseminating the results of their research.
The presented paper addresses a topic of the greatest interest, such as the application of methodologies based on the application of artificial intelligence techniques for the automatic classification of images, both captured from satellites and from airborne platforms. The work presents an adequate structure and length. Furthermore, it is important to note that it has an important analysis of current methodologies with an adequate number of updated bibliographic references.
Its most notable contribution is the design and implementation of a new deep learning architecture, called HR-GCN-FF-DA, which is explained in a very adequate level of detail in section 3. In addition, the work includes real cases of application using reference data tests, both in terms of satellite images and aerial images, showing the results compared to the application of other previous methods. The results show the good results derived from the proposed method, including parameters for performance analysis and validation of classifications.
Finally, make reference to the high scientific rigor of both the discussion of the results carried out, as well as the own conclusions derived from the work.
In this sense, I recommend the publication of the work in its current format.
Just indicate a small error that should be corrected on line 168, when referring to the pixel sizes of the high-resolution Landsat and aerial images, indicating that they are 30m2 / pixel and 9cm2 / pixel when in fact they would be 900m2 = 30mx30m and 81cm2 / pixel = 9cmx9cm.
Author Response
We thank the reviewer and editor for the comments and suggestions. The paper has already been fully revised by follows the feedback and advice via https://drive.google.com/open?id=1X9DmChYy9uElqeu8_zNad9n8F9obv6xk
We thank the reviewer for his/her positive comments. Regarding the small error of this work, we have already revised it at point 4.
Thanks to the reviewer’s comment. We have revised the number in the last paragraph of Section 3.1 to be 900 m2 = 30mx30m and 81 cm2 / pixel = 9cmx9cm.
Author Response File: Author Response.pdf
Reviewer 3 Report
The article presents a number of improvements to deep learning image classification in a continued effort by this group of authors. This is the second round, after publishing a very similar article [16] in Remote Sensing in 2018, which was also enhancing CNN and which they call their baseline now. (The references have two more of their papers, [4] and [5], which however are not cited in the text of the manuscript - they should be removed).
Indeed, by applying three further refinements (HR, FF and DA), better results are obtained, compared to the previous article, using the same input datasets. So it seems the authors are still on track. So far so good.
What is not good is that as a reader I do not get anything useful out of the article. It does not improve my understanding in deep learning, a main argumentation is missing, the article is fragmented and meandering. It gives bits and pieces of information, apparently coming from many different places, and it is unclear whether those are cited, expected, observed, concluded, or just advertised. As compared to the previous article, large pieces are re-formulated (but remain essentially the same).
It is unclear what has been done to build the system and to conduct the experiments. Was there existing software involved? Have new programs been written? How do those algorithms look? What is the workflow? Could I ever do something like this myself? How to go about it? Or is it that kind of rocket science that is out of reach to ordinary mortals? I have that Vaihingen dataset on my computer - I should be able to reproduce the result, shouldn't I?
From "Author Contributions" it appears that four out of five did not do anything. Perhaps all five should sit together and figure out a better story - one from which the readers could also benefit a little. This would be a resubmission then (not a revision).
Author Response
We thank the reviewer and editor for the comments and suggestions. The paper has already been fully revised by follows the feedback and advice via https://drive.google.com/open?id=1X9DmChYy9uElqeu8_zNad9n8F9obv6xk
(1)
We really apologize for the mistake in our reference. We have revised and removed the unused references.
(2)
In this work, we aim to improve the performance of our previous work by proposing three major improvements in our model: HR, FF, and DA.
They are specifically designed to overcome a common problem in the semantic segmentation that cannot perform well on rare classes, e.g., river, pineapples, low vegetation, etc.
As shown in the experimental results, we have succeeded in our goal by obtaining an accuracy beyond 90% on almost all classes. Also, our proposed model is the winner and outperforms all baselines including our previous work.
Furthermore, there is one more data set called “Landsat-8w3c corpus”, which is not included in our previous work. We agree that this point may not clearly state in the paper, so we have revised the paper to show this point in the first paragraph of Section 4.1.
(3)
According to the reviewer’s comments, we have provided more details and revised the paper structure in order to show related deep learning concepts and better explain our proposed ideas. Therefore, there are many major revisions throughout the whole paper, especially in Introduction and Section 3 (3.2, 3.3, and 3.4).
(4)
We totally understand the reviewer’s concern. Our paper is not a kind of rocket science, so please don't be disrespectful.
This research work is under the non-disclosure agreement (NDA), especially any outcomes of the private data sets provided by Geo-Informatics and Space Technology Development Agency (Public Organization), Thailand. Moreover, it is still on-going research under collaborations between many parties.
Therefore, we really cannot disclose the code yet. Anyway, we can still provide the pre-trained weights of the model for the ISPRS Vaihingen data set as in the link below.
https://github.com/kaopanboonyuen/SemanticLabelingOnRemoteSensingCorpora
Furthermore, we confident that we have tried our best to provide all details of our models along with hyperparameters to ensure that it can be reimplemented by other researchers, especially in the deep learning community.
Hope you understand our situation and accept our resolution.
(5)
About this point, we would like to declare that each author really plays an important role in the paper. Without a contribution to everyone, this paper would have not been accomplished.
There is a little mistake with this section for the first version of the article. And we have rewritten in this section to describe the responsibility of each author.
Moreover, we would like to explain that this work is a result of continuing research for many years. Also, we have shown many experiments in order to support all proposed ideas in the paper, so it is not just a story, but it is scientific research.
I hope you understand our effort that we have put into this work. We really wish to contribute and share our scientific ideas to researchers, especially in the remote sensing domain.
Author Response File: Author Response.pdf
Round 2
Reviewer 3 Report
Dear Authors,
Thank you for submitting a revised version while paying attention to my earlier comments. I had concerns about the lack of clarity and detail in the first version and I apologize for formulating those concerns a bit harshly.
The situation has partly improved, and for the remainder it appears to be a consequence of an NDA between the authors and a third party. I wonder whether this should then be explained in the text, and also whether this is considered acceptable from MDPI's point of view -- I gladly leave those decisions to MDPI, however. From my side I can go for 'accept'.