Road Extraction in SAR Images Using Ordinal Regression and Road-Topology Loss

Round 1

Reviewer 1 Report

In this paper, the authors conduct the road detection task and the road centerline extraction task jointly with multitask learning scheme to solve the problem of the road extraction from SAR imagery. In general, the methods in this paper are not well described. Besides, the organization should also be improved. The comments are listed as follows.

 

1. In formula (3), the authors should clearly clarify how to determine the value d_M.

 

2. In line 260, the authors should clearly clarify how to determine the threshold T.

 

3. In this paper, the authors repetitively emphasize that they detect road and extract road centerline simultaneously. This is also the major difference compared to traditional work, as traditional methods mainly include two subtasks, i.e., road detection and road centerline extraction. However, the authors still compare performance based on traditional metrics. These metrics cannot well describe performance of authors’ method, as these metrics are developed for traditional methods based on characteristics of traditional methods. To some degree, this comparison cannot highlight the characteristics of presented method. Therefore, the authors are strongly suggested to use some other metrics which can highlight the characteristics of presented method.

 

4. In Fig. 7, the units of x-axis and y-axis should be labelled.

 

5. This paper is not organized well. The authors’ method detects road and extracts road centerline simultaneously. At this point, their method cannot be segmented into road detection and road centerline extraction steps. However, section 2 still has both steps. This would confuse readers. Therefore, the authors should improve their organization.

 

6. Some words are misspelled.

In line 167: extraction -> extraction

In line 298: blod -> bold

 

The authors should carefully proofread the whole paper to avoid similar issues.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

The scientific content of the paper is very good and the topic is interesting. But I would like to see some more information in the paper before it gets published.

1. What is the source of the ground truth data? i.e, How did the road centerline and road width ground truth data obtained? It needs to be mentioned in detail in the paper.
2. A Google Earth/ Optical image of the study area is expected.
3. The authors have mentioned that they have used the TerraSAR-X spotlight datasets. Which polarization is used? Are you using sigma nought/intensity images? Is there any preprocessing required?
4. In the results section, I would like to see a zoomed view of a small region of the road, like highway. Then the zoomed Google Earth view of that region, the SAR image, the road centerline and the extracted road for that region can be shown. This helps the readers to visually compare the results obtained from the different methods which you have shown in this study.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

The paper addresses an important task of road detection in SAR images and determination of their centers. A new approach based on NN learning is proposed, its novelty is described well. The method is tested, its advantages are demonstrated. Meanwhile, to my opinion, the paper can be improved. Important items are the following: 

1) Can the authors give details about TerraSAR images that have been used, namely, speckle properties, spatial resolution sensed terrain size? 

2) Is there some statistics about road width, do the authors consider only asphalt roads of other types as well? 

3) the total number of images used in experiments are given and it is impressive, the question is are these images for China territory of for territories of different countries? 

4) possible reasons of false detection are mentioned but have You analyzed what are the main reasons of false detection for the designed network? 

5) some references can be added, for example, https://www.tandfonline.com/doi/full/10.1080/22797254.2019.1694447 and some paper of A. Shelestov, N. Kussul and others on crop classification using CNN and Sentinel SAR data 

Some minor comments are: 

1) Line 25: what is IoU metric? 

2) line 40: SAR images instead of SAR 

3) line 259: 1024x1024 is not resolution but image size 

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The reviewer's comments are well addressed. At this point, the reviewer suggests that this paper is accepted.

Reviewer 2 Report

The authors have addressed the comments raised by me and I recommend the manuscript for publication.

Reviewer 3 Report

I am more or less satisfied by the reply and corrections done by the authors. So, I think that the paper can be accepted now.