A Research on Landslides Automatic Extraction Model Based on the Improved Mask R-CNN

Round 1

Reviewer 1 Report

I think authors did an excellent work and the paper is well written and well discusse.

I have some minor changes to advice:

Line 29: “In the mainland of China, there are about 70% of the area are mountainous areas with complex topography, and geological hazards.” It becomes: “In the mainland of China, about 70% of the areas are mountainous areas with complex topography, and geological hazards.”

Line 31: “For example, the 2008 Wenchuan earth-31quake, the 2010 Qinghai Yushu earthquake, the 2013 Sichuan Ya'an earthquake, and the 322017 Sichuan Jiuzhaigou earthquake have caused huge losses to the country and people33[4]. In these earthquakes, the losses of life and property can account for more than half of 34the entire earthquake losses[5].” Ok, it is hazard, but how does it fit with landslides? I understand that you want to consider earthquake-triggered landslides, but you can said it explicitly. Moreover you know that losses of life are not due only by landslides, but also by earthquake by itself. So explain better the issue.

Line 36: “rapid acquisition of hazard information such as the regional distributions, 36numbers, and scales of landslides immediately after the earthquake”. Perhaps it is important not only after the earthquake…I would cut “after the earthquake”.

Page 2, line 82-83: “With the rapid development of high-resolution satellites, high-resolution remote sensing data has increased dramatically. Extracting targets from massive high-resolution data has become the key technology [28].” True, but probably not the right place where to put this sentence. We are talking about OBIA in this point, so it is not the case to interrupt the subject.

Line 126: “The developed model” for this title sounds a bit better

Line 295 Also here, could you insert a heading such as “Application of the model to the Jiuzhaigou County”?

Line 339, Figure 9. For me it is useless.

Line 342-358 I am curious to know the time required for the training steps.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

I suggest a major revision: 1. Draw a flowchart from your work flow that briefly shows the process and in the Discussion section. 2. Compare your results with the results of other researchers. 3. The literature review part may be further improve.  4. In abstract and conclusion, authors need to add some numerical results. 5. Description of the figures should be more complete 6. Improve the English.   7. The conclusion should include more details.  8. Please improve the Conclusion section, this section need to more details and more explain. 9. Please refer to new paper about this field.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The paper “A research on landslides automatic extraction model based on the improved Mask R-CNN” would be very useful to accurately detect earthquake-induced landslides from aerial images since the proposed method worked effectively. However, the current manuscript partially lacked some important explanations and citations to be improved. The specific comments are shown below;

 

Line 48-90

These three methods introduced here were very classical and not strongly related to this study. Since the descriptions seems bit lengthy, could you shrink them appropriately?

 

Line 91

Since deep learning model was not introduced as number four method, please delete the numbering.

 

Line 109

Please do not use abbreviated words (R-CNN, YOLO and SSD) before explaining the detail. And please add citations for the previous methods.

 

Line 111-114

The explanations in the lines seem strange. If the Mask R-CNN model was already proposed in previous researches, the problems described here were almost solved, weren’t they? Please clarify the problems of the previous Mask R-CNN model to be solved in this study.

 

Line 115

Please provide citations for Mask R-CNN model.

 

Line 128

Plead provide a citation to the study by He Kaiming et al.

 

Line 136

What is RoI Align?

 

Line 143, Figure 1

This figure lacks necessary information. What is C, P, and FC? Where is C1 explained in Table 1? Do the outputs of “Classes softmax”, “Boundary box regressor” and “Mask” correspond to Classification, Positioning and Segmentation as shown in Line 227-228? If so, please clearly described here.

If possible, could you visualize the diagram by showing schematics for easier understanding?

 

Line 169, Figure 2

A lot of explanations are necessary also for Figure 2. What is 256-d, the numbers in the figure, and 1x1? What did the sign plus (+) in the figure mean?

 

Line 258-260

This section is too short. The caption of Fig. 5 is not the flowchart for Mask R-CNN. Please add appropriate explanations for them.

 

Line 354

Please add a citation for the official COCO2014 dataset.

 

Line 399-400, Figure 11

Please replace the explanations for (b)-(d) “Experiment one-three results” to appropriately. The caption also needs more explanations for the figure.

 

Lien 403-428

In the learning process, the image was splitted to 256x256 pixels as shown in Fig. 8. I am wondering how the experiment in Fig. 11(b)-(d) obtained? The experiment image was also splitted? Please add more explanations for assessing the experiment.

 

Line 414, 423 and 430, Figure 12-14

These captions also need more explanations for understanding the figures. The scales and legends are also needed.

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

I found this manuscript clear and well-written. The developed method is also innovative and interesting to explore this in other landslide prone areas. however I have some questions and comments:

Line 312: the area of this area! Size of study area. The same for line of 314.

Line 445: which has achieved acceptable results.

In the manuscript I did not find information about the type of landslides in the study area. Also, it is not clear form the text that the developed method can be used for all type of landslides or just specific type of landslides. Another important point that could be discussed is the feasibility of the approach to identify different landslides size, Did they check this? Besides it is worth to discuss the implications of the method, for instance in multi-temporal mapping of landslides which could be created with the current available high resolution remote sensing imageries. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

The author improved manuscript accordingly. However, still needs some corrections:

Provide a better map by high resolution instead of Figure 6.

Deep learning model’s information/parameters should be listed at a table.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

Thanks you for your responses. The authors almost promptly replied my questions. But original text should be modified based on your responses.

Line 122
Concerning the question for RoI Align, I meant that the explanations or references about the RoI Align should be included in the manuscript not only in the response letter.

Line 399, Figure 11
My previous suggestions were that authors should change the explanations of the figure (b)-(d) "Experiment one, two and three results" because these explanations did not tell any information. The reviewer suggests to change to (b) Result of ResNext50, (c) Result of ResNext50+Improved FPN, and (d) ResNext50+Improved FPN+Ledge, for example. 
With the same reason, the captions of Fig. 12-14 should be modified appropriately to show that the results (b)-(d) are corresponding to the results of the techniques shown above.

Author Response

Please see the attachment.

Author Response File: Author Response.docx