Effective Training of Deep Convolutional Neural Networks for Hyperspectral Image Classification through Artificial Labeling

Round 1

Reviewer 1 Report

I thank the authors for their efforts to improve the previous version of the text, incorporating each and every one of the considerations indicated by the reviewers. I believe that the motivation and conclusions of the article are now much more clearly justified.

However, I believe that the structure is still somewhat confusing, so below are some changes that may help to improve its understanding.

Point 2.4 is not well understood because it is shown in the section on materials and methods, as it has a similar structure to the different experiments in the section on results. In any case, it is very strange to reference the use of a database that will be described later in section 3.1.1. and a network architecture that is also presented later.

Figure 2 should try to be shown before section 2.5 so that it can be viewed near its reference. I understand that in the current format of the text it may not be possible, but perhaps for the final publication, especially if it is not in paper format, if possible.

In section 3, not only are the results described, but also the experiments conducted, perhaps it would be more appropriate to call it Experiments and Results or to move the description of the experiments to the previous section.

Table 1 should be shown before starting with experiment 2 and not after to avoid confusion.

Figures 3 and 4 should also be shown before experiment 2 and separate in a table the values of OA, AA and k. In any case, a clearer explanation of the conclusions drawn from your study should be given in the last paragraph of experiment 1 and not simply say "The results confirm stated hypothesis and the validity of the proposed approach".

An illustrative figure of experiment 2 (in the style of figure 5) would be very useful to understand well the grid density and the number of strips of table 2.

In Tables 2 and 3, the results of AA and k are missing as in Table 1 and Figures 3 and 4.

In the caption of figure 6, I think the final (see text) is left over.

The results and conclusions related to figure 7 are very interesting. I think that they should be highlighted to give them the importance that I think they have.

Author Response

Dear Reviewer,

We wish to thank you for your insightful comments. We tried to answer them to the best of
our ability and hope that you will find the resulting revised article satisfying. You can find our
answers to your specific comments in the attached PDF.

Once again, thank you for your constructive comments.

Most sincerely,
Wojciech Masarczyk, Przemysław Głomb, Bartosz Grabowski, Mateusz Ostaszewski
Institute of Theoretical and Applied Informatics, Polish Academy of Sciences

Author Response File: Author Response.pdf

Reviewer 2 Report

Authors in this paper presented effective training method of deep convolutional neural networks for hyperspectral image classification. Authors compared different deep convolutional neural networks training methods. Article is well-written, but structure of this manuscript should be changed. Methodology is mixed with result.

Some more comments are given below:

Abstract:

The main results of the study could be presented (some values, accuracy etc.).

Introduction:

The introduction should be improved, first bearing in mind, that the manuscript is submitted to remote sensing journal there is no need to say for remote sensing community that the Hyperspectral imaging is.

In introduction it is missing more concentration on problem. Not clear that is the aim of this study. It could be wider commented what result achieved by other researchers in their research not just mentioned that they did such investigation.

Methods:

Information about research object is missing. Used methods could be explained more detail.

Line 213. What was k number in your case?

Results:

Some parts ( 3.1.1. Description, 3.2.1. Description, 3.3.1. Description) of result paragraph should be moved to Materials and methods paragraph.

Line 319. Need more detail explanation why you choose n=15

Line 314. I think most part of this should be in methodology.

Line 324-330. I think you need add here numbers to base your statement.

Line 330. I think you must give more explanation. It is not enough to say that results confirm stated hypothesis and you must find this in figures.

Conclusions should be more specific and based on the results of the study.

Author Response

Dear Reviewer,

We wish to thank you for your insightful comments. We tried to answer them to the best of
our ability and hope that you will find the resulting revised article satisfying. You can find our
answers to your specific comments in the attached PDF.

Once again, thank you for your constructive comments.

Most sincerely,
Wojciech Masarczyk, Przemysław Głomb, Bartosz Grabowski, Mateusz Ostaszewski
Institute of Theoretical and Applied Informatics, Polish Academy of Sciences

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

After this third review, the authors have considered almost all the issues that have been indicated to them, substantially improving the structure of the text and some aspects that improve understanding.

I still think that the text would be improved if the results of AA and k were incorporated into tables 2 and 3, although I understand that it would have a very high temporary cost in order to incorporate it into the article now.

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

In this paper, the authors proposed a new pre-trained method before training DL models using existing training samples. This idea is interesting, however, some key points haven't been addressed appropriately in the manuscript. The followings are my comments:

1) Basically, English writing (academic writing) needs to be addressed extensively in the revised manuscript. 

2) Introduction section needs to rewrote due to the following reasons: (1) literature review is not exhausted where the discussions on HSI classification, transfer learning (if you think the term "transfer learning" really matters in your manuscript), DL methods, and transfer learning-related DL methods are not provided in a detailed manner; (2) no need to create a separate sub-section 1.1.

3) Method section (section 2) is too short and too simple. It is really really difficult for the readers to understand why you design such a model, why it can work, and how it can be implemented. I think section 3.1 can be removed into section 2 including adding more detailed illustrations of your proposed model.

4) Regarding the proposed method itself. I'm confused about the "pre-train" step since you mentioned it as an "unsupervised" way. How can it been "pre-trained" without any label information utilized? And I don't know why it is effective since only those patches split from the whole HSI were used directly 

5) Within the Results section, more methods should be included for comparison. Other different quantitative indices (e.g. average accuracy, Kappa) and qualitative results (e.g. classification maps) should be incorporated as well. 

In general, the quality of this manuscript can and should be improved by considering the aforementioned comments, and most parts of this paper need to be rewritten before the acceptance for further publication.

Reviewer 2 Report

This article describes a strategy of transfer learning that uses unsupervised pre-training step without label information. This strategy is applied to hyperspectral classification tasks.

Transfer learning is a machine learning method where a model developed for a task is reused as the starting point for a model on a second task.
Thus, they propose a simple unsupervised method for assigning artificial labels to hyperspectral data which can be used for pretraining a neural network. After the authors evaluate the proposed methodology with different deep neural network in different dataset.

The contribution seems specified in the “Introduction” part. The number of references related to keywords in this part is 27.
In Related works part there is 6 (33) specific references. This sub-section is separate like in other papers, for example:

Remote Sens. 2019, 11, 1342; doi:10.3390/rs11111342

I find the paper to be interesting, the procedure is well planned and well described. The authors describe the proposed method in the lines 126 – 152. Details of step 1, generating artificial labels and motivation are in lines 132-152. It is possible that the content of section 3.2 can be introduced into section 2.

In the proposed method, the authors generate artificial labels by a simple segmentation algorithm based on the local homogeneity of spectral characteristics. Then, they establish spectral labels and they don´t need the supervision for expert (unsupervised method).

It seems that they have experience with the characteristics of land cover classes in hyperspectral imaging and then they propose artificial labeling.

In order to improve the document, the classification maps should be incorporated in the results.

Overall Accuracy analyzes the quality of results. It is possible to add others metrics like Kappa index. The discussion of results is correct and the conclusions are compatible with the results.

In addition, please the authors should provide more details about computation time of the strategy proposed.
I think that the authors have more data and information, but they seemed unnecessary to show it

 The contribution seems specified in the “Related works” part and I consider it a valuable contribution to the Remote Sensing journal.

Reviewer 3 Report

This is an interesting article on transfer learning. Although I think that, given the amazing results, the results obtained should be better explained and justified.

In general, I think that it is very important for its credibility that the possible generalization of the results obtained to other networks should be well justified.

I will now indicate some specific issues to be reviewed:

Figure 1 I think should be shown after its reference before point 2.2

I think that point 2.2 should be extended to justify more clearly the motivation for the proposed form of artificial labelling, since it is really surprising that such pre-labelling helps to improve, as it seems from the results.

In point 3.1, it is not clear which reference corresponds to which architecture, the first being with reference 10 or with reference 35?

Surely the improvement has much to do with the specific networks proposed, so I think an effort should be made to try to justify why this initial misleading pre-labelling is improved in some more than in others.

In point 3.2.1, I think the parameters used n, n1, n2 should be justified. why is the average subtracted?

Table 1, I think should appear after point 3.2.2 and not before.

In point 3.3.1, it is not clear because only the IP is used and not the two datasets. On the other hand, it is not clear how to divide by vertical stripes, if the number of stripes is 2, does it mean that there are only 2 classes located to the right and left of a single vertical line?

I think table 2 should be shown after point 3.3.2 and not before.

In 3.4.1, why is it divided by the standard deviation in addition to subtracting the mean?

In figure 2 I think that it is missing to indicate what each image is with correlative letters and to indicate it in the caption.

I think that Table 3 should be shown after point 3.4.2 and not before.

I think that experiment 3 is very interesting and its results should be better explained so that they really support the results of the previous experiments.

I think that, in the discussion section, the advantages of the proposed algorithm (line 310-313) are not correctly justified. Being in general very confusing.

The conclusions are too general and I think they should also show the concrete conclusions of the study carried out.