Layover Detection Using Neural Network Based on Expert Knowledge
Round 1
Reviewer 1 Report
The overall structure of this paper is complete. In addition, the research on 3D Array SAR topography reconstruction by the author is of great significance. Some problems in detail in this paper are as follows:
1) This paper presents a Layover detection using a neural network based on expert knowledge,which has been mentioned many times in the article, but the explanation on how to add expert knowledge or how to use expert knowledge is not sufficient and specific, and I do not think using FFT transform can be called expert knowledge, which is just transformed data into frequency domain.
2) It is suggested to quote more references from the recent two years.
3) The description of the the proposed method is not explicitly. For example it is confusing what the phase convolution is.
4) The author should compare the results with more recent deep learning semantic segmentation methods such as U-Net, U-Net++, Deeplab V3, and DeepLab V3+. The ablation experiment should be performed to show the effectiveness of all the proposed methods in the paper.
5)The grammar should be checked carefully to improve the qulity of the paper.
Author Response
Dear Reviewer:
Thanks for your patient reading and constructive comments. Those comments are all valuable and helpful for revising and improving our paper. We have studied all comments carefully and made conscientious correction. The responses and explanations of revisions related to every point are listed below:
- As the response to the comment “This paper presents a Layover detection using a neural network based on expert knowledge...”, we re-described the relationship between expert knowledge and FFT in detail and the principle how FFT helps to distinguish signals in height and furtherly detect layover. (Section 3.2.1 Theoretical basis and FFT Residual Structure, line 248)
- As the response to the comment “It is suggested to quote more references from the recent two years.”, we added the reference 38-40 and 46. (Section Reference)
- As the response to the comment “The description of the proposed method is not explicitly. For example, it is confusing what the phase convolution is.”, we not only re-described the part of FFT residual structure in comment 1, but also completed the description of phase convolution including the design principle and specific processing. (Section 3.2.2 Phase convolution, line 304)
- As the response to the comment “The author should compare the results with more recent deep learning semantic segmentation methods...”, we conducted the contrast experiments with other recent semantic segmentation methods including your mentioned U-Net, U-Net++, DeepLabV3 and DeepLabV3+. And we analyze the results in detail. We also performed the ablation studies to compare our proposed components with original model and show the effectiveness of two components. (Section 4.3 Comparative experiments with other deep learning methods, line 511; section 4.4 Ablation studies, line 528)
- As the response to the comment “The grammar should be checked carefully to improve the quality of the paper.” we have asked Embellishment agency to polish our paper. Please see if the revised version met the English presentation standard.
Thank you and best regards.
Author Response File: Author Response.docx
Reviewer 2 Report
This paper focuses on the layover detection in 3D Array-SAR topography reconstruction, and proposes a U-Net based method with FFT and phase components to rebuild the terrain. Generally the paper is well-written and the performance is inspiring. Before being accepted, the following suggestions can be considered:
1.The note of Figure 2 is relatively simple, the author can explain more in the note.
2.The author can add ablation studies to show the functions of different components.
Author Response
Dear Reviewer:
Thanks for your patient reading and constructive comments. Those comments are all valuable and helpful for revising and improving our paper. We have studied all comments carefully and made conscientious correction. The responses and explanations of revisions related to every point are listed below:
- As the response to the comment “1. The note of Figure 2 is relatively simple, the author can explain more in the note.”, we enriched the note of Figure 2. (Section 1. Introduction, line 70)
- As the response to the comment “2. The author can add ablation studies to show the functions of different components.”, we conducted the ablation studies to compare our proposed components with original model and show the effectiveness of two components. (Section 4.4 Ablation studies, line 528)
Thank you and best regards.
Author Response File: Author Response.docx
Reviewer 3 Report
In some forms, it seems that other sources have been used, in which case it needs to be referenced.
Abbreviations should be spelled out in full the first time they are used.
Is there a need for image processing on radar images before applying the used method?
The DEM image production section should be explained more.
Author Response
Dear Reviewer:
Thanks for your patient reading and constructive comments. Those comments are all valuable and helpful for revising and improving our paper. We have studied all comments carefully and made conscientious correction. The responses and explanations of revisions related to every point are listed below:
- As the response to the comment “In some forms, it seems that other sources have been used, in which case it needs to be referenced.”, we checked the article and add the reference 33-46, 38-40 and 46. (Section Reference)
- As the response to the comment “Abbreviations should be spelled out in full the first time they are used”, we spelled out FFT in full at the first time. (Section Abstract, line 14)
- As the response to the comment “Is there a need for image processing on radar images before applying the used method?”, we listed the image processing including image registration and phase compensation in Figure 1 and re-emphasized them in the text part. (Section 1. Introduce, line 37)
- As the response to the comment “The DEM image production Section should be explained more.”, we explained explicitly in the place where we only use DEM once. (Section 4.1.1 Simulation Experiment, line 393)
Thank you and best regards.
Author Response File: Author Response.docx
Reviewer 4 Report
This paper proposes a novel layover detection method that combines the complex-valued (cv) neural network and expert knowledge to extract features in the amplitude and phase of multichannel SAR. First, inspired by expert knowledge, an FFT residual convolutional neural network was developed to eliminate the training divergence of the cv-network, deepen networks without extra parameters, and facilitate network learning. Then, another innovative component, phase convolution, was designed to extract phase features of the layover. Subsequently, various cv-neural network components were integrated with FFT residual learning blocks and phase convolution on the skeleton of U-Net. Due to the difficulty of obtaining SAR images marked with layover truths, a simulation was performed to gather the required dataset for training. The experimental results indicated that our approach can efficiently determine the layover area with higher precision and fewer noises.
Author Response
Dear Reviewer:
Thanks for your patient reading and comments. I'm greatly thankful again for your affirmation.
Thank you and best regards.
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
The quality of the paper has been improved, and all my questions have been addressed.