Combining Spectral Unmixing and 3D/2D Dense Networks with Early-Exiting Strategy for Hyperspectral Image Classification

Round 1

Reviewer 1 Report

The paper is written well, in my opinion. I want to emphasise on an explanation of why VCA is used in this paper. VCA unmixing, authored ~15 years ago, is surpassed by several methods in recent years, like, R-CoNMF (doi: 10.1109/WHISPERS.2015.8075385) or regression-based unmixing (doi: 10.3390/rs11151792, doi: 10.3390/jimaging5110085). A synthetic experiment showing the effect of added noise is recommended.

Author Response

We would like to express our gratitude to the editor and reviewers for their invaluable comments. We have carefully considered and addressed all of their concerns in revising this work. The following presents a summary of our response (which is given in italic) to the reviewers’ comments (shown in bold), on a point-by-point basis. Please see the revised manuscript for actual changes made, as highlighted.

Author Response File: Author Response.docx

Reviewer 2 Report

This paper proposed a 3D/2D dense network that integrates 3D convolutions and 2D convolutions to perform spatial-spectral processing. In the end, this method is more capable of handling spectral-spatial features, while containing fewer parameters compared with the conventional 3D-based convolutions, and further boosts the network performance with limited training samples.

 

My first comment is about the literature review, I think that there are more papers about densenet and HSI, please reference it.

 

The second commend, I suggest to add two datasets more challenges.

1.- Pavia with fixed train/test data.

2.- Houston 2018.

Both datasets are avaiables on http://dase.grss-ieee.org/

Author Response

We would like to express our gratitude to the editor and reviewers for their invaluable comments. We have carefully considered and addressed all of their concerns in revising this work. The following presents a summary of our response (which is given in italic) to the reviewers’ comments (shown in bold), on a point-by-point basis. Please see the revised manuscript for actual changes made, as highlighted.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

accept in present form

Author Response

We would like to express our gratitude to the reviewers for their invaluable comments.

Reviewer 2 Report

To be honest, I am a little disappointed with the review in the sense that I think the authors should have made better efforts, both in the search for references and in the experimentation.

For example, concerning the references, without diminishing the importance of the proposed works, I believe that there are articles more cited than those considered by the authors and with more impact that could give more quality to the proposed work.

With regard to experimentation, I strongly recommend to include Pavia or Indian Pines with fixed train/test data. Also, authors claim that their method needs little amounts of trainining data, but I believe that the experiments do not demonstrate it correctly, I think it is necessary to show the performance of the proposed method when different percentages of training data are used in the disjoin datasets. I also believe that the Houston data should be included in the manuscript.

Author Response

Response to Review Reports

Reviewer #2: For example, concerning the references, without diminishing the importance of the proposed works, I believe that there are articles more cited than those considered by the authors and with more impact that could give more quality to the proposed work.

R: Thank you very much for your invaluable suggestion, we have read some more cited and more impact papers about dense networks and HSI, and have cited them [1][2][3] in the revised manuscript.

 

[1] Paoletti, M.; Haut, J.; Plaza, J.; Plaza, A. Deep learning classifiers for hyperspectral imaging: A review. ISPRS Journal of Photogrammetry and Remote Sensing 2019, 158, 279 – 317.

[2] Wang, W.; Dou, S.; Jiang, Z.; Sun, L. A fast dense spectral–spatial convolution network framework for hyperspectral images classification. Remote Sensing 2018, 10, 1068.

[3] Paoletti, M.E.; Haut, J.M.; Plaza, J.; Plaza, A. Deep&dense convolutional neural network for hyperspectral image classification. Remote Sensing 2018, 10, 1454.

 

# With regard to experimentation, I strongly recommend to include Pavia or Indian Pines with fixed train/test data. Also, authors claim that their method needs little amounts of trainining data, but I believe that the experiments do not demonstrate it correctly, I think it is necessary to show the performance of the proposed method when different percentages of training data are used in the disjoin datasets. I also believe that the Houston data should be included in the manuscript.

R: Thanks very much for your suggestion. Experiments on effect of training samples have been added in Section 4.1. The classification results of Indian Pines with fixed train/test data is also include in Section 4.1.

Experiments on Houston 2018 have been added in Section 4.4.