Deep Learning Method of Landslide Inventory Map with Imbalanced Samples in Optical Remote Sensing

Round 1

Reviewer 1 Report

Please see the attached file.

Comments for author File: Comments.pdf

Author Response

See the seperated attachment please.

Author Response File: Author Response.docx

Reviewer 2 Report

The manuscript titled “Deep Learning Method of Landslide Inventory Map with Imbalanced Samples in Optical Remote Sensing” is an interest work. The authors adopted machine learning algorithms to landslide detection with remote sensing images in recent decades. The full CNN was adopted as the basic model. While this manuscript is suitable for publication in Remote Sensing, some issues should be addressed before we can consider proceeding with this paper. I therefore recommend a major revision.

 

1.     As a full research article, it is important to state the scientific implication of this research. The whole manuscript describes the machine learning algorithms concerning the landslide inventory mapping, yet it lacks discussions on the scientific importance of their proposed algorithm compared to traditional method.

 

2.     Similarly, it is necessary to add more basic introduction on the landslide inventory mapping and the processing of remote sensing images. This journal focuses on the science and application of remote sensing technology. However, the manuscript is more like a technique report on deep learning method.

 

3.     Regarding the methodology, please specify how the complex background of the historical landslide could affect the detection results. And please discuss how to avoid its effects in detail.

 

4.     In the structure of Full CNN, the convolution features of the first and second down sampling were added to the final prediction graph. Please explain why it could increase the feature level. Does it increase the probability of False Positive or False Negative?

 

5.     The authors should clearly describe the applicability and their future work. As they state, it may not be applicable in areas with less surface vegetation coverage, such as the Loess Plateau. In fact, the Guizhou Province has abundant rainfall and hence it has a high vegetation (forest) coverage of more than 60%. This value is a relatively high level compared to other provinces in China. This algorithm should not become a Guizhou-specific method. Please discuss how this method could be extended to other regions with moderate or low vegetation coverage.

 

6.     Please increase the quality of figures. For example, please add the name of “Fa’er” and “Jichang” to Figure 1, as well as other cities of Guizhou Province. please use a different colour to represent the landslide points in Figure 8. The green colour is difficult to distinguish.

Author Response

See the seperated attachment please.

Author Response File: Author Response.docx

Reviewer 3 Report

 

 Deep Learning Method of Landslide Inventory Map with Im-balanced Samples in Optical Remote Sensing

 

Dear Editor

I have read above article. I have the following major comments on this study. Some parts need to clarify by authors.

Please see detailed comments.

 

 

 Abstract

This part is written so general. You must explain about methodology parts, achieved results, and a conclusion about landslides. You are describing deep learning, and this isn’t suitable.

 

 

 1. Introduction

Any details about research questions and hypothesis?

What is difference of the current study and study of Ghorbanzadeh et al. [33].

 

2. Study Area

Please add coordinate system of the study area in Fig. 1 and as well in text.

770 landslides and 2003 non-landslides data: why do you used from non-equal no-landsides in this study? Any details?

Any details about landslides? Their type? Area? Volume?

Please add climate features of the study area.

 

3.1. Content and process of LIM

The landslide interpreted by geological knowledge was then used as the reference truth value to select the remote sensing classification method. “What is your mean?”

 

Unfortunately, authors are concentrating on method, this is important to used from this method about landslides researches. I cant see anything about explaining landslides. This is a negative point in this study.

Even, authors ran this model on an identified landslides by Ji et al. [28]. This is just a method running.

 

4.1. Landslide sample augmentation

Why 256×256 pixels? What is original resolution of landslides?

 

 

4.3. Comparison of FCN-FLK, SegNet, and U-NET models

What is reason of this comparison? I cant see any details in methodology bout this.

Any discussion?

This is important to compare results of the current study by some previous studies.

 

Author Response

See seperated document, please.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The uploaded manuscript and cover-letter are revised based on the reviewer's comments.

I accept it.

Reviewer 2 Report

I appreciate the efforts of authors in revising the manuscript. My comments were well responded. The article is now ready to be accepted.