High-Accuracy Filtering of Forest Scenes Based on Full-Waveform LiDAR Data and Hyperspectral Images

Round 1

Reviewer 1 Report

The article describes research on an algorithm for ground filtering for data obtained using full-wave-form LiDAR Data. The entire processing was divided into 3 parts, i.e. data preparation, the use of a neural network, and the use of the CSF algorithm to improve the results. The main part of the description (correctly) concerns the method of data preparation and the application of the neural network (modification of the DGCNN network). Ultimately, I believe that the presented research and results are interesting and worthy of publication. However, I believe that the manuscript is underdeveloped and requires major revision.

Comments:

1. Figures
a) there is simply too many of them - at least figures 6, 11, 12, 16, 22 are in my opinion unnecessary,
b) figure 17 - the division into training and test sets can be presented in one of the other figures. In my opinion, there is no need to make a separate figure for this purpose.
c) The titles of figures: 3, 7, 8, 9, 10, 11, 20, 21 need to be corrected so that the title shows what they represent.
d) Figures 1, 14, 15, 16, 17, 18, 19, 20, 21 should be supplemented with information allowing for their spatial orientation and scale.
e) Figure 13 is of too low quality and illegible, the units of the graphs axes should also be completed.
f) Figures 18, 19, 20 - figures should fit on one page.
g) Figure 14 - whether "Waveform decomposition results" are actually shown or are selected (e.g. first or last) reflections shown. What does the small black rectangle outline in the western part of the figure mean?
h) Figure 2 - "Waveform features" appear twice. Is it supposed to be like that or shouldn't there be "Geometric Features" in one place? Is "Quality evaluation" a part of this workflow ?
i) What does the small black rectangle outline in the western part of the figure mean?
h) if something is outlined like in figures 16, 19, 20 please describe on the figure or in its title what this outline means.

2. Table 4 should be deleted.

3. Reference points:
a) I don't think there's anything you can do about it at this stage, but 23 reference points isn't a lot.
b) The reliability of GNSS RTK measurement results in the forest may raise some doubts. Was the measurement made using the GNSS RTK or GNSS RTN method ?
c) The location of the reference points should be shown in one of the drawings (eg figure 21).

4. Study area and datasets
a) You mention that Mengjiagang Forest Farm has "a total area of 167 million square meters.". I think it's worth converting this value to square kilometers.
b) Whether Figure 1 shows the entire area of Mengjiagang Forest Farm or a fragment used in training and testing the neural network ?
c) Is "point cloud density of about 14 points/m2" before or after waveform decomposition ? I am trying to understand what was the actual area used for training and validation of the model.
d) In my opinion in this section you should write about data actually used for traning and validation. How many points were used ? How many of them were used for training / validation ? How many points was in the ground class ? How reliable was the reference classification ?

5. Methodology
a) "Waveform decomposition", around line 299 - describing the formulas in words instead of giving  equations is quite unusual.
b) "Band selection from hyperspectral image" looking at figure 5 - does it make sense to choose other bands than 63 ?
c) According to strategy described in lines 369 - 375 if n = 80 then 8 features should be selected, if n = 8 then 6 features should be selected, but if n = 125 only 4 features should be selected.
d) I have ambivalent feelings about the breadth of the description of solutions already existing and described in open access articles. Certainly, however, the detailed description of the discrepancies from the models proposed so far is correct.

6. Experimental Results
a) In my opinion the figure(s) presenting the differences between heights of reference points and points chosen with classification algorithms (i.e. errors) would be useful. Evaluating results based on model smoothness is not necessarily objective.
b) Does data in tables 2, 3, 5 concern validation set on training set or entire set ?

7. Discussion

a) lines 717-726 repetition.
b) lines 744-749 in my opinion in the given task unsupervised learning is not a good idea. 
c) line 759 - in my opinion realistically not theoretically,
d) line 727 - there is no figure 25 in the manuscript,
e) line 730 - please rephrase - "the maximum error" in table 5 is not 0.26m,
f) line 732 "left side" -> "west side" ?
g) In my opinion there is no description of the significance of the results. Maybe comparison to results of other algorithms would be helpful ? 

8.  Conclusion 
a) "Conclusion" -> "Conclusions"
b) line 774 - "significantly". Taking into account results I am not sure about that word.
c) please rephrase conclusions because now part of it sounds like an abstract.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

In this paper, the authors a novel filtering method for forest scenes utilizing full-waveform LiDAR data 26 and hyperspectral image data. The structure of this paper is clear; however, it needs major revision and there are some issues that should be addressed as listed in detail as follows:·

1. The paper is interesting, but what interests me more is how to register data for any Lidar and hyperspectral image data. This is crucial for the practical application of both image processing and algorithms.

2. The references regarding GCN for HSIC should be investigated in your introduction, e.g., Unsupervised Self-correlated Learning Smoothy Enhanced Locality Preserving Graph Convolution Embedding Clustering, Multi-scale Receptive Fields: Graph Attention Neural Network, MultiReceptive Field: An Adaptive Path Aggregation Graph Neural Framework, AF2GNN: Graph Convolution with Adaptive Filters and Aggregators.

3. Please explain why band selection is necessary in this way(line 372-375, PP. 10) and whether there are other better options. Please use Formula and experiment to deduce and prove.

4. The Self-attention layer is transformer? Please provide a detailed explanation. If it is a Transformer, this cannot be considered your work. Please introduce it in the relevant work section and mark it in the schematic diagram.

5. How are the graphs constructed in the models proposed by the authors? How are the graphs constructed in the models proposed by the authors? Why use A and what is the motivation of the authors? I think conventional GCN may yield better results.

6. The manuscript should include ablation experiments and compare the computational complexity of different algorithms.

7. I think the language of the manuscript needs further revision.

I think the language of the manuscript needs further revision to improve readability.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

No more comments. The paper can be accepted as presented form.