We Need to Communicate: Communicating Attention Network for Semantic Segmentation of High-Resolution Remote Sensing Images
Round 1
Reviewer 1 Report
In this paper, the authors proposed an attention communication network for remote sensing image semantic segmentation. The title is charming and fully shows the novelty of the proposed method. However my concerns can be listed as follows
1. Are the code and dataset released if the paper is published?
2. More information on dataset should be provided.
The writing of this paper is urgently to improve.
Author Response
Response to Reviewer 1’s comments
In this paper, the authors proposed an attention communication network for remote sensing image semantic segmentation. The title is charming and fully shows the novelty of the proposed method. However my concerns can be listed as follows:
Thank you very much for your valuable and constructive comments, which indeed help us improve our manuscript. We have revised our manuscript according to your comments, and the point-by-point responses to your comments are listed below. Some explanation and clarification about your concerns are provided in this response letter as well. All the changes in the revised manuscript are highlighted with red text. We have also attached the relevant changes in the revised manuscript in this response letter for your convenience in reviewing.
- Are the code and dataset released if the paper is published?
Thank you for your questions regarding the code and the open source of the dataset. We promise that after the publication of the paper, the model and the related code for training will be open-sourced on Github to facilitate researchers to better reproduce the new methods proposed in this paper. In addition, the two datasets we used are open source datasets, provided by ISPRS, which you can find in the two links below.
https://www.isprs.org/education/benchmarks/UrbanSemLab/2d-sem-label-potsdam.aspx
https://www.isprs.org/education/benchmarks/UrbanSemLab/2d-sem-label-vaihingen.aspx
- More information on dataset should be provided.
Thank you for your suggestion on the dataset information. The dataset we utilize is an open source dataset provided by ISPRS, which is widely used in semantic segmentation tasks of remote sensing images. In the manuscript of the article, the area where the dataset is located, the number, size, spatial resolution, and band information of the internal data are presented. In order to present the information of the dataset more fully, we add the image numbers used as training, validation, and testing in the dataset and a general overview of the dataset in figure 5.
The relevant changes made in the manuscript are attached below:
Line 373-375 in Section 3.6:
We employed images numbered 2_13, 2_14, 3_13, 3_14, 4_13, 4_14, 4_15, 5_13, 5_14, 5_15, 6_13, 6_14, 6_15, and 7_13 for testing, images numbered 20_10 for validation, and the remaining 22 images for training.
Line 380-382 in Section 3.6:
We utilized images numbered 2, 4, 6, 8, 10, 12, 14, 16, 20, 22, 24, 27, 29, 31, 33, 35, and 38 for testing, image number 30 for validation, and the remaining 15 images for training.
- The writing of this paper is urgently to improve.
We have edited and polished the English writing of the manuscript to meet the requirements of an academic paper
Reviewer 2 Report
Comments:
The authors propose an communicating attention network (CANet) for the semantic segmentation task of remote sensing images, which uses an communicating mutual attention (CMA) and communicating self-attention (CSA) modules for enhancing the interaction and fusion of different levels of feature maps. Both qualitative and quantitative experimental results demonstrate the effectiveness of CANet. Before it is accepted, I have the following comments.
1) Lines 103-105, Line 110, Layout formatting error.
2) In the related work, some more semantic segmentation methods should be reviewed, such as “Collaborative attention-based heterogeneous gated fusion network for land cover classification, IEEE TGRS” and “Novel Adaptive Region Spectral-Spatial Features for Land Cover Classification with High Spatial Resolution Remotely Sensed Imagery, IEEE TGRS”.
3) Line 255, Equation (6), the subscript of “????????????????????” is too long, suggest abbreviation instead.
4) Line 259, “Where ? is the channel number of the feature map”, “where” should be lowercase and not indented, it's not a separate paragraph.
5) The dataset used in the experiment is too simple with few land cover classes, and it is recommended to test on a complex dataset.
6) The computational complexity of the proposed method is not analyzed.
Author Response
Please see the attachment
Author Response File: 
Author Response.pdf
Reviewer 3 Report
This study aims to improve the semantic segmentation accuracy by implementing different levels of feature communication through the attention module. The method proposed in this study seems to be innovative and the designed experiments can validate it. However, several issues should be addressed before it goes further.
1. The manuscript does not mention the specific image numbers used in the training set, validation set and test set, please add the above information.
2. There are some similarities between the abstract and the conclusion of the manuscript. Please revise the content of the conclusion.
3. The content of the first column of Table 2 is data sets, but the name is the method, please change it.
4. There are some problems with the English expression, please improve the English presentation of the article to make it meet the requirements of academic papers.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
The authors solved all my concerns and the paper can be accepted.
The writing must be improved.
