Detection of Small Impact Craters via Semantic Segmenting Lunar Point Clouds Using Deep Learning Network
Round 1
Reviewer 1 Report
This paper proposed a deep learning-based algorithm to extract the craters from the moon’s surface. They proposed a new convolutional network including a global feature module to improve the performance. The input of the network is point cloud data generated from DEMs and output is segmented point cloud for craters and non-craters areas. They’ve also applied some post-processing methods to improve the results. My main comments are listed below.
My greatest criticism is that there is a scientific error in this paper. In lines 142-143, the authors mentioned that “converting DEM data to point cloud data does not result in information loss, which often happens when converting DEM data to image data.” My questions are: What is the DEM format (raster or vector)? And how does it generated in Chang’E data?
If the DEM or Digital Elevation Model is a raster-based representation of elevations which is generated from a point cloud data, therefore, it has already lost the information because of the interpolation, and if you convert it to the point cloud again, it doesn’t gain more information!
As I understood from the paper, the DEM is generated from satellite images (Line 27). In this case, first image matching techniques are applied to the satellite images to generate the point cloud, and then the point cloud is converted to the DEM. Therefore, if you want to use a raw point cloud without losing information, you need to use it after image matching techniques and before converting to the DEM. Converting a DEM to the point cloud is not a correct and scientific solution. Please explain more about the format of the DEM to eliminate this misunderstanding.
Line 11: “we provide a new point cloud dataset of impact craters from the Chang’E data”.
What are the resolution, accuracy, point density, scale, and extensions of the new point cloud? In Section 3.1. you need to explain the details of the data.
Line 43:”First, we transform DEM data into point cloud data”.
What are the spatial resolution and vertical accuracy of the DEM? How did you transform a DEM to the point cloud? The method?
Line 53-54:”The method can robustly detect small impact craters even with the disturbances of sunlight”.
There is no proof of this statement in the results section.
In Related Work, the accuracies and quantitative results of other studies are required, so that we can compare them to the current paper.
Line 159-165: You need to add a section as “preprocessing” to explain the coordinate transformations. What are these rules: transformation rule, conversion rules, projection rule? If they are ArcGIS Tools, you need to cite the software.
Line 168: “point cloud data which is converted from DOM data”. DOM or DEM?
Line 169: “In this work, we implement a new network based on pointnet++ architecture”. However, I cannot see the pointnet++ in Figure 3! Which part of the network is based on pointnet?
Line 170:”Global Feature Exchange Network(GFE-Net)” Why did you add a Global Feature Exchange Network? Since the pointnet includes the global feature module. What is the advantage of your global feature module?
Lines 170-171 and the first three lines of Section 3.2.1 are the same! You need to remove one of them.
In Figure 6, you need to present the Ground truth (real point cloud), related DOM and DEM.
Formulas 5-9 need citations.
In Section 4.3., is your network robust to the orientations, density of the point cloud and the input size?
The notations in formulas 5-9 and tables 1-3 should be the same.
Line 240: “In the end, 16384 points are sent to the network for training”. What are the sizes of the training and test datasets?
The English language should be improved. I suggest a nature speaker go over it once more or have a proof reading service correct the English. There are many language/grammatical problems and considerable effort from the authors is needed to increase the quality of the paper. Several typos exist, and the paper must be re-checked. Examples:
- digital elevation map (DEM) >> Digital Elevation Model (DEM)
- In order for the network to learn global feature >> In order to learn global feature for the network
- we propose new network module >> we propose a new network module
- the output new features >> the new output features
- If without the morphological dilation >> without the morphological dilation
- In this work, We implement >> In this work, we implement
Author Response
Please see the attachment.
Author Response File: 
Author Response.docx
Reviewer 2 Report
This paper presents a deep learning pipeline for impact crater detection. The contribution is fair. However, the presentation of the paper requires further editing. Please address the points mentioned below:
- Please expand the acronyms at first mention (e.g., DEM in Abstract).
- It is necessary to add quantitative results to the Abstract.
- When describing the dataset, it is necessary to display several instances from it.
- When describing the several steps of the proposed pipeline, it is important to display illustrative examples for each step, not only for some of them.
- It is important to mention the training/test split sizes in a table.
- In Table 1, please add the references of the comparison baselines.
- It is necessary to polish language mistakes throughout the paper.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 3 Report
The article "Detection of Small Impact Craters Via Semantic Segmenting Lunar Point Clouds using Deep Learning Network" is satisfactorily conceived and presented. Certain parts of the article, such as the introduction and methods, might be shortened, but in general the article should not need further correction.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
No comments. The authors have responded to all my concerns in this version.
Reviewer 2 Report
The comments have been addressed. The paper may now be accepted.
