Deep Learning for Landslide Detection and Segmentation in High-Resolution Optical Images along the Sichuan-Tibet Transportation Corridor

Round 1

Reviewer 1 Report (Previous Reviewer 1)

I am appreciated the authors revising this manuscript.  

Author Response

Tanks very much for your kind work and consideration on publication of our paper. On behalf of my co-uthors, we would like to express our great appreciation to you. Thank you very much for your previous comments and suggestions. Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches.

 

Best regards.

 

Yours sincerely,

Wandong Jiang (on behalf of all authors)

Reviewer 2 Report (Previous Reviewer 2)

The manuscript has been improved significantly after the major revision. All of the problems were solved and in my opinion it is ready for publishing in present form

Author Response

Tanks very much for your kind work and consideration on publication of our paper. On behalf of my co-uthors, we would like to express our great appreciation to you. Thank you very much for your previous comments and suggestions. Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches.

 

Best regards.

 

Yours sincerely,

Wandong Jiang (on behalf of all authors)

Reviewer 3 Report (Previous Reviewer 3)

I noticed that this manuscript is a revised version. The author carefully revised the manuscript according to the previous comments. However, the current version cannot be accepted as it is. My main suggestions are in the discussion section. The current discussion is too weak and deviates from what the discussion should have. The author is suggested to make major modifications to the discussion section. It is necessary to compare with previous work in terms of data, methods, results, etc. In addition, prospective aspects should be included. To sum up, a moderate revision is recommended.

Author Response

Re: Thanks for your constructive comments. As suggested, we have added the following in the discussion and conclusions section in the revised manuscript.

4.3. Features of the TL-Mask R-CNN method

In this paper, deep learning was used to detect landslides along the SSTC, which has positive implications for the safe construction and operation of the SSTC. Firstly, the LRSTTC dataset generated in this study has been made freely available for the public, which can significantly reduce the time of data collection and labeling for other researchers. There have been limited (if any) available geohazard datasets in this study area for deep learning. Secondly, the TL-Mask R-CNN method presented in this paper can be utilized to detect old landslides and ice avalanches with better performance than previously reported landslide detection methods (e.g. Mask R-CNN, Unet, Unet++, and Deeplabv3+). Given that new and old landslides have similar shapes, and landslides and ice avalanches can be seen as slides along slopes with different materials, it is feasible to use transfer learning to realize the detection of different geological. Thirdly, the TL-Mask R-CNN method is able to segment landslides one by one, which appears to be a big challenge for most of the previously published methods.

4.4. Limitations of the TL-Mask R-CNN method

It should be pointed out that the TL-Mask R-CNN method has two major limitations at the moment.

(1) Limited sample size: Deep learning always requires large sample sizes, but the sample size of the LRSTTC is still small. It is believed that the TL-Mask R-CNN meth-od could perform even better with an increasing sample size of the LRSTTC dataset.

(2) Model transferability: Geological and weather conditions vary a lot along the STTC, and the key influencing factors of landslides can be different from one place to another [69], which makes the transferability of the TL-Mask R-CNN method a challenge. To address this issue, it would be desirable, once again, to increase the sample size of the LRSTTC dataset.

5. Conclusions

……

Currently, our approach achieves the migration between three different types of geohazards including new landslides, old landslides, and ice avalanches. Our future research will study the transfer of knowledge in different regions and landscapes, exploring typical feature transfer under the diversity scenarios. How to find these effective features is a key underpinning to transferring learning. In addition, a reasonable combination of geological knowledge is also important rather than simply adding all kinds of geological data to train the model. We will also combine multi-source remote sensing data and geohydrological data to detect landslides and continuously update the LRSTTC dataset.

Author Response File: Author Response.pdf

Reviewer 4 Report (New Reviewer)

Comments for author File: Comments.pdf

Author Response

Thanks for the constructive suggestion, we agree with you on that – in the revised manuscript, we carefully revised it according to your suggestion, and highlighted the changes.

Author Response File: Author Response.pdf

Reviewer 5 Report (New Reviewer)

This study aims to use Mask R-CNN and TL-Mask R-CNN for new and old landslide detection and segmentation. Manuscript is well organized. While the topic has some merit, there is several modifications that are necessary. The authors should consider the following specific comments diligently.

 

1.  Line 28, the wrong use of acronym. the Sichuan-Tibet Transportation Corridor is not the first to appear in abstract. In addition, Line 100, when it first appears in introduction, both the full name and the acronym need to be given.

2.        The introduction section needs strengthening. Authors could briefly discuss the topic of landslide hazard in the introduction part, one of the most debated ones due to its socio-economic consequences. In order to provide a more general overview of this main issue, I suggest to insert some references from international literature.

3.        Line 120, please insert comma thousand separator in the number. Similar problems in some other places in the manuscript.

4.        In the Methods, there is no need to elaborate the specific method content in a very detailed description. Section 2.3 is too long.

5.        Section 3.1. Accuracy Evaluation should be written in the Methods, it is not the result of this study.

6.        In Figure 11, numbers and texts in the maps are difficult to read.

7.        The authors should discuss the potential sources of uncertainties in the manuscript.

8.        English is inaccurate in several sentences. The manuscript would have benefited of English edit by a native speaker.

Author Response

Thanks for the constructive suggestion, we agree with you on that – in the revised manuscript, we have made careful revisions as suggested by the reviewers.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report (Previous Reviewer 3)

Accept

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

 

 The authors effectively used Mask 22 R-CNN and TL-Mask R-CNN for new and old land-23 slide detection and segmentation.   The methodology of using deep learning to do landslide detection can help us enlarge the landslide database along the Sichuan-Tibet transportation corridor.  The structure of the manuscript is well organized. The contents of the manuscript are suitable for publication in “Remote Sensing”.   

The following are the questions for this manuscript:

1.        In introduction, the authors described too many methods used in the landslide detection. The readers cannot know clearly pros and cons of these methods. Please make it more convenient to read.

2.        Many abbreviations do not have their full names in the first appearance, e.g., PRN in line 158; R-CNN in line 101; ROI in line 155.

3.        Line 297, for the samples used for Machine-learning, how do you distinguish the new landslide from the old one?  Do you have some specific index?

4.        In Figure 5, please give the sources and descriptions of the landslides shown in the figure.

 

 

Reviewer 2 Report

Dear authors,

The effort to evaluate the potential offered by remote sensing data, machine learning and cloud computing to classify the changes that occurred in fire-affected areas is interesting and would bring a significant contribution in this field.

The effort to detect landslides using deep learning and high-resolution remote sensing data is interesting and would bring a significant contribution in this field.

Manuscript is perfectly presented, but my suggestion will be to describe used remote sensing data and reasons of using it instead of other remote sensing data. In my opinion it is critical when implementing such approach in practice.

I wish that my comment would be helpful in improving the quality of this research.

Thank you.

Reviewer 3 Report

In this manuscript, the authors focus on the extraction task targeting old landslides. Compared to previous DL based landslide extraction for post-earthquake/rainfall landslides, few scholars have explore the capability of ML in old landslide detection, so I consider the research of this work significant. According to the results, the model designed by the authors is useful in extracting old landslides and avalanches. However, the manuscript' logic in describing the principle part of the method is not optimistic, I could not understand exactly how the TL-MASK-RCNN proposed by the authors works after the first reading, and the Figures provided do not clearly fit with the content of the article to enhance the persuasiveness, so I strongly suggest the authors to spend time to re-write the methodology part. One suggestion: try to describe how the model works, rather than just emphasizing how the parameters of each module in the network are set. Finally, I think it is important for the author to polish the article, either by hiring a professional organization or inviting a professor whose native language is English.

1.       In introduction, when introducing the typical four types of landslide extraction methods, the authors do not seem to distinguish well between pixel-based landslide extraction methods and ML-based landslide extraction algorithms, and I think the authors should select more typical literature related to pixel-based landslide extraction.

2.       In Line 65, there is an extra space here.

3.       you should specify in Figure 1 that 2 represents the enlarged area of 1

4.       It seems that all the formulas listed in the manuscript are not uniform, and the serial number at the end should be centered

5.       Line 156, RCNN or R-CNN?

6.       Line 158 and Line 162, you have used the abbreviation RPN in Line 158

7.       Line 125-149, I suggest that the authors do not need to spend time and space to introduce some basic knowledge about CNNs.

8.       Line 166-180, please draw a brief figure of the structure of ResNet rather than depicting the detailed parameters of the modules in ResNet extensively in the manuscript.

9.       Line 181-188, I don't understand which of the five parts of Mask-R-CNN is related to here. The formulas given also makes it difficult to understand how feature combination network works. Besides, the authors did not mention feature combination network before.

10.   When describing the method, you should use more passive rather than first person.

11.   Line 189-201, you should re-explain how RPN and ROI work, it seems that from your expressions, it is hard to understand what they really are. You'd better draw a structure diagram or flowchart of PRN and ROI.

12.   Line 188, plooing?

13.   The formatting of formula symbols in the content of paragraphs should be adjusted

14.   Line 228-229, what are n, m, x, y, i, j in the formula and how many expressions does Dt represent?

15.   I don't see the difference between Figure 3 and Figure 4? Is it just the import of pre-trainned weights and FT? I suggest the authors to add a figure or sub-figure describing how TL work in Figure 4, making it easy to understand the role of TL for readers.

16.   In Figure 7, you should specify the extent of the landslide.