Version 1
: Received: 4 December 2019 / Approved: 5 December 2019 / Online: 5 December 2019 (04:05:48 CET)
Version 2
: Received: 11 February 2020 / Approved: 12 February 2020 / Online: 12 February 2020 (05:40:08 CET)
Li, R.; Zheng, S.; Duan, C.; Yang, Y.; Wang, X. Classification of Hyperspectral Image Based on Double-Branch Dual-Attention Mechanism Network. Remote Sens.2020, 12, 582.
Li, R.; Zheng, S.; Duan, C.; Yang, Y.; Wang, X. Classification of Hyperspectral Image Based on Double-Branch Dual-Attention Mechanism Network. Remote Sens. 2020, 12, 582.
Li, R.; Zheng, S.; Duan, C.; Yang, Y.; Wang, X. Classification of Hyperspectral Image Based on Double-Branch Dual-Attention Mechanism Network. Remote Sens.2020, 12, 582.
Li, R.; Zheng, S.; Duan, C.; Yang, Y.; Wang, X. Classification of Hyperspectral Image Based on Double-Branch Dual-Attention Mechanism Network. Remote Sens. 2020, 12, 582.
Abstract
In recent years, researchers have paid increasing attention on hyperspectral image (HSI) classification using deep learning methods. To improve the accuracy and reduce the training samples, we propose a double-branch dual-attention mechanism network (DBDA) for HSI classification in this paper. Two branches are designed in DBDA to capture plenty of spectral and spatial features contained in HSI. Furthermore, a channel attention block and a spatial attention block are applied to these two branches respectively, which enables DBDA to refine and optimize the extracted feature maps. A series of experiments on four hyperspectral datasets show that the proposed framework has superior performance to the state-of-the-art algorithm, especially when the training samples are signally lacking.
Computer Science and Mathematics, Artificial Intelligence and Machine Learning
Copyright:
This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Received:
12 February 2020
Commenter:
Li Rui
Commenter's Conflict of Interests:
Author
Comment:
Specifically, the revisions are as followings:
1. We tried our best to modify the language expression throughout our manuscript.
Besides, we used professional English editing service to polish the manuscript.
2. We reconstructed the framework of the paper. There are 6 parts in the revised
manuscript. In the Introduction section, we review the background, significance,
and the present research status of HSI classification. In Related Work section, we
introduce the basic modules used in our framework. In Methodology section,
we illustrate the whole framework of our method. In Experiments Results
sections and Discussion section, we provide the experimental results and make
sufficient analyzation of the experiments. And a conclusion of the entire
manuscript with the direction of future work are presented in Conclusion
section.
3. Introduction section: We re-wrote the Introduction section. In the previous
manuscript, the examples about SAE, DAE, RAE, DBN, CNN, RNN, and GAN
are provided in the following three paragraphs dispersedly, which may cause
confusion. Hence, we delete the sentence “Stacked Autoencoders (SAE), Deep
Autoencoder (DAE), Recursive Autoencoder (RAE), Deep Belief Networks (DBN),
Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), and
Generative Adversarial Networks (GAN) et al. have been successively used in the areas
of HSI classification.” and just provide the examples directly.
4. Related work section: We re-wrote the related work section, and add a brief
introduction of HSI Classification Framework Based on 3D-Cube and 3D-CNN
with Batch Normalization which may help to understand our framework.
Furthermore, we are so sorry about the negligence of error “Error! Reference
source not found.”, and all of the references are checked and corrected.
5. Methodology section: We provide a brief introduction before taking the Indian
Pines dataset as an example, each dimension represents in the matrices are
illustrated, and a methodology flowchart of our framework are added.
6. Experiments Results section: We reconstructed the Experimental Setting,
Section 4.2 and illustrated the input data of the classifier. We just select a certain
number of adjacent pixels centering on the target pixel as a patch, which is
without any other pre-procession. If the target pixel is on the edge of the image, Remote Sens. 2019, 11, x FOR PEER REVIEW 2 of 2
the values of missing adjacent pixels are set as zero. What we fed into the
classifiers are the patches of the image. Furthermore, we provided an explanation
about HSI Classification Framework Based on 3D-Cube in section 2.1, which may
be helpful to understand the input of the framework.
7. Discussion section: We did ablation experiments in the Discussion section to test
the Effectiveness of Attention Mechanism and the Effectiveness of Activation
Mechanism, so we rename the sections as “the Effectiveness of Attention
Mechanism” and “the Effectiveness of Activation Mechanism”, respectively.
What’s more, we added a supplementary experiment about the effectiveness of
activation function in Section 5.3, which supports the conclusion about the
superiority of our framework. We visualized the result of the experiments by
making some charts, and we make more specific explanation of the experiments.
8. Conclusion section: Based on the results and analyses of the experimental results,
we draw a powerful conclusion.
Commenter: Li Rui
Commenter's Conflict of Interests: Author
1. We tried our best to modify the language expression throughout our manuscript.
Besides, we used professional English editing service to polish the manuscript.
2. We reconstructed the framework of the paper. There are 6 parts in the revised
manuscript. In the Introduction section, we review the background, significance,
and the present research status of HSI classification. In Related Work section, we
introduce the basic modules used in our framework. In Methodology section,
we illustrate the whole framework of our method. In Experiments Results
sections and Discussion section, we provide the experimental results and make
sufficient analyzation of the experiments. And a conclusion of the entire
manuscript with the direction of future work are presented in Conclusion
section.
3. Introduction section: We re-wrote the Introduction section. In the previous
manuscript, the examples about SAE, DAE, RAE, DBN, CNN, RNN, and GAN
are provided in the following three paragraphs dispersedly, which may cause
confusion. Hence, we delete the sentence “Stacked Autoencoders (SAE), Deep
Autoencoder (DAE), Recursive Autoencoder (RAE), Deep Belief Networks (DBN),
Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), and
Generative Adversarial Networks (GAN) et al. have been successively used in the areas
of HSI classification.” and just provide the examples directly.
4. Related work section: We re-wrote the related work section, and add a brief
introduction of HSI Classification Framework Based on 3D-Cube and 3D-CNN
with Batch Normalization which may help to understand our framework.
Furthermore, we are so sorry about the negligence of error “Error! Reference
source not found.”, and all of the references are checked and corrected.
5. Methodology section: We provide a brief introduction before taking the Indian
Pines dataset as an example, each dimension represents in the matrices are
illustrated, and a methodology flowchart of our framework are added.
6. Experiments Results section: We reconstructed the Experimental Setting,
Section 4.2 and illustrated the input data of the classifier. We just select a certain
number of adjacent pixels centering on the target pixel as a patch, which is
without any other pre-procession. If the target pixel is on the edge of the image, Remote Sens. 2019, 11, x FOR PEER REVIEW 2 of 2
the values of missing adjacent pixels are set as zero. What we fed into the
classifiers are the patches of the image. Furthermore, we provided an explanation
about HSI Classification Framework Based on 3D-Cube in section 2.1, which may
be helpful to understand the input of the framework.
7. Discussion section: We did ablation experiments in the Discussion section to test
the Effectiveness of Attention Mechanism and the Effectiveness of Activation
Mechanism, so we rename the sections as “the Effectiveness of Attention
Mechanism” and “the Effectiveness of Activation Mechanism”, respectively.
What’s more, we added a supplementary experiment about the effectiveness of
activation function in Section 5.3, which supports the conclusion about the
superiority of our framework. We visualized the result of the experiments by
making some charts, and we make more specific explanation of the experiments.
8. Conclusion section: Based on the results and analyses of the experimental results,
we draw a powerful conclusion.