Multiscale Feature Fusion for the Multistage Denoising of Airborne Single Photon LiDAR
, Yulin Ding 1, Xuekun Yuan 1, Ying Jiang 1, Yigao Jin 1, Xuming Ge 1, Yeting Zhang 2, Jie Chen 3 and Xiaocui Guo 4Round 1
Reviewer 1 Report
This work presents a methodology for point cloud denoising. It is based on the selection of point cloud features in a multi-scale level, which are fed into a Random Forest classifier to detect noisy points. Multi-stage denoising is also considered to improve performance. This work uses 3D point clouds from a Single Photon LiDAR as input, which have higher levels of noise than linear mode LiDAR. The contribution may be of interest due to the limited use of SPL technology. However, the methodological contribution is based on the use of techniques already well established in the field to adapt to a different type of noise, so the interest for the reader may be limited as well.
My main concern is with the way in which the scientific contribution is conceived. This paper demonstrates that the proposed technique performs better than existing techniques (Radius Outlier Removal, Statistical Outlier Removal) for SPL point clouds, but it does not demonstrate whether this same method would achieve similar improvements in non-SPL point clouds. I find this relevant because if this method also overperforms other methods in LML point clouds, this could be presented as a “3D point cloud denoising method” rather than a “single photon LiDAR denoising method”. Thus, I would suggest to apply this method to a non-SPL 3D point cloud, and compare the improvement with respect to other noise removal algorithms.
Furthermore, this method is compared with well-stablished noise removal algorithms, but the state of the art has newer methods also based on Deep Learning or Machine Learning which should be taken into account. Would a DL-based method overperform this method if trained with SPL data?
Finally, the data labeling procedure remains unclear. How was the data labeled for training and testing? Manual labeling of 3D point clouds is an expensive task, and the high level of noise of the SPL point cloud may make it even more difficult.
Other issues:
- - Equations should be referenced in the text: E.g. There is no “Equation 1” reference anywhere in the text (furthermore, in Eq. 1 the covariance matrix is defined as “D” while in the text is named “C”).
- - Quality of some figures may be improved. Figures 7, 8, and 9 are difficult to interpret.
- - There are some minor spelling errors along the manuscript that should be checked.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
This paper proposes a mathine learning method for point cloud noise data removal, it seems intersting and sound. The structure is logical, the figures are of good quality and the historical background has given credit as is appropriate. The references are adequate. One note that the method is feasible for a special region, maybe could not be feasible for the other regions, this is the common quention for ML-based method, it presents well in the trained data benifiting from nonliear fitting, while maybe not common for the other regions. Overall, I think this manuscript can be considered after minor revision if the author could adequately address the comments below.
1. Line 4: SPL firstly appear, add “single photon lidar”;
2. how do you labeling the noise point cloud? detailed information should be clarifiled.
3. tainning, testing, validation performance need to be added for better understanding.
4. how many time and computer consumption, need to be clarified?
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
This reviewer wants to thank the authors for their work improving the manuscript.
My comments have been properly addressed. I find interesting the application of Rand-LA for comparison with the authors' method.
The only minor improvement that I would point is to add Rand-LA results in the mountain area (Table 4) and Water Area (Table 5) as it has been done in Table 3. This will allow the reader to comapre the generalization of authors' method and Rand-LA.
Apart from that, I would recommend to accept this paper.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
