Building Extraction from Airborne Multi-Spectral LiDAR Point Clouds Based on Graph Geometric Moments Convolutional Neural Networks
Round 1
Reviewer 1 Report
I strongly suggest to revise and resubmit the manuscript according to the following major comments as well as the other comments from the reviewer board.
- The title suggests that the manuscript is about "multispectral point clouds", however I do not see any usage of the multispectral information. It looks like only point clouds are necessary and the multispectral information is never used. Therefore, I highly suggest to change the title.
- The introduction suggests that the building extraction problem is still not solved, however not clearly describing the challenges of the building extraction process on point clouds. What are the specific problems with the previous algorithms and why they could not be solved so far? Please improve the information.
- I find the experiments very limited. The shown rooftops could be easily identified when a height threshold is used to select the rooftop and the ground points. There is not a need for an advanced network structure. Please show results on more complicated scenes; such as buildings at a hill (uneven surface) and buildings which are surrounded by tall trees.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 2 Report
I'd like to give some comments to the manuscript, hopefully make it better.
- Please be very clear with the term "building extraction". I don't really sure when I read the title. Is it "building point-wise classification" or "building vectorization" or "3D building model generation". I recommend to use the specific term in the title. I think it is a binary classification: building or non-building points.
- The authors made comparison with many point-based deep learning network. I appreciate the idea. However, I recommend to compare with PointNet++ instead of PointNet. And also I would like to see PointCNN* in the comparison, as I personally had some experiments of point cloud segmentation with Airborne Laser Scanner data using PointCNN and the results are very good.
- Most readers interest in pictures. Not only in number and table. Please add more results in pictures to make the manuscript looks more informative.
https://arxiv.org/abs/1801.07791
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 3 Report
The authors present a deep learning framework to extract the point cloud of the building. The FPS-KNN is proposed to generate numerous samples with the fixed number of points, and geometric moments have been used in learning point features. The proposed method has been tested in one data set and shows that the proposed method is better than existing frameworks. The authors present an interesting work, but some points should be addressed.
- The authors present interesting results, but it will be good if the results of building extraction for an entire 3.2km2 of the study area present. Moreover, an additional case study must be added because only one case study is not enough to prove the robustness of the proposed method. Particularly, in this paper, data for training and testing must satisfy certain conditions (like the same flight path). Thus, the proposed method must test on a general multispectral LiDAR.
- Several exiting methods (non-deep learning methods) have been used to extract the building with high accuracy (see ISPRS benchmark). It would be good to compare with them.
- It is unclear what features will be computed for points although the authors mention to compute both first and second-order geometric moments.
- The authors address the framework uses both spatial information and three spectral values. How are the spectral values used to compute the features of the points?
- why the knn = 16, 32 and 48 points are chosen.
- the authors should present/discuss a lost function.
Some other comments:
L 403: “effective to fuse the multiscale features”. Please provide supporting information
L 534-548: the authors must present quantitative evaluation rather than addressing general statement like “our model relatively completely recognized the main rooftop . . . only some cracked pieces were recognized”
Figure 7: it is too hard to see where the building is. The building figures (a right column) must be ordered from 1-3.
L 591: “complicated than the commonly used urban area” how can the authors identify it?
L 593: Can the authors give more information to explain the difference between the point-based vs. pixel-based evaluation, and why the point-based evaluation is better than others in this study?
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 4 Report
While the results of the proposed method are promising, I suggest the following improvements:
- As the paper is focused on multi-spectral LiDAR data, it will be useful to highlight the advantages of multi-spectral LiDAR data over traditional LiDAR data, and describe how they were used in the proposed method.
- In Section 4 Methodology, before describing the proposed framework, the rationale/principle of the proposed framework should be explained first.
- The number of testing data (study area) is limited.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
Authors have revised the manuscript considering the reviewer comments.
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.