Sample Plots Forestry Parameters Verification and Updating Using Airborne LiDAR Data
Round 1
Reviewer 1 Report
Sample Plots Forestry Parameters Verification and Updating Using Airborne LiDAR Data
This paper presents a novel method in doing tree crowns segmentation using the a profile-rotating algorithm, and do local tree species classification using generated geometric statistics from tree crowns. The paper also predicted biomass and stock volume from the extracted info of trees in the study area. This study claims that the prediction models can be extended to map large-scale forest resource monitoring with large scale LiDAR data of the same quality. The paper did an accuracy evaluation and concluded that the F1-score of the tree segmentation is >0.95, the accuracy of tree species exceeds 90%, and R2 of tree height, east-west canopy width, north-south canopy width, diameter at breast height, above-ground biomass, and stock volume are 0.893, 0.757, 0.694, 0.840, 0.896 and 0.891, respectively.
Dear editor, after reviewing the paper, I have comments below for the authors:
1. The paper uses a profile-rotating method to extract tree edges from different angles. Based on examples in Figure 6, it seems like traditional methods like watershed or these studies would be sufficient: https://www.sciencedirect.com/science/article/abs/pii/S0924271615001951
https://www.tandfonline.com/doi/full/10.1080/01431161.2021.2018149
Could you please provide a benchmark comparison between your method and the traditional ones? If you are rotating round the point clouds, why not just directly use 3D segmentation methods? It would be better if you can create another figure of point clouds to show the advantage of using profile rotating (e.g points that are wrongly detected as edge points in traditional methods and edge points correctly identified by your method).
2. The tree top finding is based on a local maximum method. It would not work well on deciduous tree species. Again, a 3D segmentation method should be explored (e.g. this one with an self-attention mechanism): https://ieeexplore.ieee.org/abstract/document/9540373
3. There are only seven studied tree species in this research and different simplified fitted shapes are used to model trees. Are these tree crown shape models universally applicable (for other species with similar morphological features or for sub-species classification)? Would this simplification make 3D info loss on more challenging species classification tasks?
4. This paper mentioned the challenge in delineating overlapping canopy trees and overshadowed trees (line 575 and 695). It seems like it is partially resolved in this paper: https://www.sciencedirect.com/science/article/abs/pii/S0924271615002257?casa_token=_ax90xhGph8AAAAA:eFtPI4QvkLrJzOPIAlxoVGOag6VQ0rgBDN9TvmgGM5aeDZiGGVP4jBJmWO8HjXOKPVJhE1RVg34
5. In figure 9, it seems like at least 5 extracted trees are missing from the reference after alignment. Are these errors in the reference data?
6. Shouldn’t formula (14) be a softmax function instead of Sigmoid one?
7. Grid size of 1.5m provides the best balance between under and over segmentation. Is there any justification for this? Would this be universally applicable?
8. Apart from the 7 species, there is an “other” type. It would be good to elaborate more on the species and characteristics in the “other” type.
9. Line 681 and 715 mentions about computation intensity of varies grid sizes and how the proposed method reduces computational burden. I suggest the author add a table quantifying the algo’s computational intensity and improvement.
10. Suggest the author add reference to support the claim in line 704.
Good.
Author Response
Please see the attachment
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Title: parameters measured on sample plots are generally considered to be the most accurate, and they serving for parametrization and calibration of less accurate estimation models. Hence the "..Verification and Updating.." should be replaced by ".. as a base for large scale forest inventories using LiDAR"
Key words: Tree species identifocation is doubled.
Introduction: extensive and reliable, but all devoted to individual tree detection (IDT) methods. A mention about area besed (ABA) methods missing, and should be added.
Line 183: The above ground biomas at breast height has no sense.
Line 188: To verify and update the error information of manually collected sample plots has no sense (see comments for Title).
Some parts of Introduction duplicate Methodology and might be removed: for example lines 196-203.
Line 221: section 6 does not exist.
Methodology: extensive and sometimes repeated. For example Figure 1 might be completed by adding numbers of plots to images, and then Figure 2 is not necessary and can be removed.
Line 242: first time the term "canopy" is introduces for "tree crown". Canopy represent tree layer, not individual tree crown. Hence I strongly recommend use "tree crown", or simply "crown" instead od "canopy" in whole text.
Line 250: how sample plots data was confirmed in the outfield?
Table 1: For average values (means?) a parameter describing variability might be added (std, se, ..). How cumulative stock volume is defined? Please complete.
Fig. 8: (a) - original or normalized point cloud?
Deep believe network (DBN) description is too extensive for purpose of this paper. Some details might be removed.
Figure 15: .. again: reclassification of ground measured data by estimation by model is nonsens..
Equation 15: estimation of stock volume based just on dbh, without tree height is questionable - while we don´t know how stock volume is defined (see also above).
Results: look good and are presented in sufficient detail.
Tables 11 and 12 have shifted reference in the text (10 and 11 respectivelly).
Table 13: in the heading, Collected (in the field) vs. Estimated (by the model) should be labeled, not True as now is written.
Discussion: is not discussion, just summarization of own results. Some confrontation and references to results of other autors should be added.
Especially the success of tree species detection from LiDAR data by suggested DBN method is surprisingly high i.. worthy of more attention.
Author Response
Please see the attachment
Author Response File: 
Author Response.docx
Reviewer 3 Report
General Comments
The manuscript evaluates the performance of a recent approach for tree segmentation and classification from LiDAR data for estimating key forest resource indicators such as forest stock volume, diameter at trough height, and large-scale forest biomass. This is a topic of great interest to RS readers.
All in all the manuscript is well structured, but a strong rearrangement is necessary to make it more accessible and independent.
The work should be an evolution of the article just published by the same authors on RS (https://doi.org/10.3390/rs15020406) but in the current form of the submitted manuscript, the differences and peculiarities of the new work are not well highlighted.
Please describe in detail the objectives and added value of the current work compared to the previous one from the beginning (from the introductory paragraph up to methods and results).
As an example, authors should highlight the concept of matching degree and the use of dominant trees for plot matching as well as the extraction of Forestry parameters.
All the method paragraphs, including steps identical to the previous work, have to be shortened by referring to the published work; whereas, the methodological paragraphs currently present in the Results section have to be moved in the method.
Detailed point-by-point comments are given below.
Specific comments
Abstract
The abstract needs to be updated once all the text has been rearranged.
Line 20: The prediction models of forestry parameters can also be verified … (can or were verified ???)
Introduction
Please reorganize the introductory section by introducing your recent work (https://doi.org/10.3390/rs15020406) and clearly define the objectives of the present study by emphasizing the expected improvements in Lidar-based tree classification with respect to the previous one.
The current goals are a bit confusing: on the one hand, the author proposes Lidar classification to improve in situ sampling; on the other hand, they use in situ sampling to validate the lidar classification. Reading the entire manuscript, the purpose is clear, but the aims of the work should be clear from the introduction.
More generally, the introduction is lacking in recent literature on the subject, please provide an updated picture. Possible suggestions can be: https://doi.org/10.1016/j.rse.2022.113143, https://doi.org/10.3390/rs12030515, https://doi.org/10.3390/f12020131, https://doi.org/10.1016/j.rse.2021.112307)
The reference number should be close to the reference name. For example: line 51, Ouma [6]; line 156, Shrestha & Wynne [36],
Line 221: Section 6 should be Section 5.
Study Area and Data
Figure 1: As the same Lidar data are shown in the next figure (Figure 2), it is more useful to have some zooms of the RGB images instead of the Lidar circles.
Section 2.1.2. Sample Plots Data: it is not clear the origin of in situ data; please better explain if they are collected by the author or they are provided by an external source. Additionally, more information about the methodology of measurements have to be added (in particular for indirectly measured parameters) as well as information on the dates of field campaigns are necessary to understand the temporal shift with respect to the Lidar acquisitions (same year, 1 year, 10 years???).
Lines 250-252: What does “the manually obtained information from the sample plots is confirmed in the outfield” mean? Please clarify the outfield operations to confirm the in situ measurements.
Table 2: Please correct “Attitude of Points”. A part of the misprint of altitude; it is better to refer to “flight height”
Figure 2: Please add the legend for the Lidar data.
Methods
Section 2.2: Much of the methods (filtering, data segmentation, classification, etc.) look similar to your previous work (https://doi.org/10.3390/rs15020406). Please shorten the identical part by referring to the published article and better highlight the aspects that provide new or more detailed operations.
Line 285: (mainly tree points), please specify that mainly tree points is not generic but refers to a peculiarity of your study area.
Lines 294-286: the filtering operation in the pre-processing have also to include the elimination of outliers (noise), i.e. points with unreasonably high or low elevation values that must be removed before the ground identification (see e.g., https://doi.org/10.3390/rs2030833 , https://doi.org/10.3390/f13030380 , https://doi.org/10.1201/9781420051438 ).
Lines 290-295: How do authors determine the threshold values for PTD implementation? Please add a short comment in the text.
Lines 296-303: This paragraph can be shortened in one or two sentences.
Line 301: Please explicit the CNTP acronym. Not all readers of RS are familiar with point clouds.
Line 317: unit means grid?
Lines 320-321: How do authors determine the threshold D for redundant tree top elimination?
Figure 5: About the comments on the left, the first and the second partially describe the same operations. Please revise the text of the boxes. Moreover, the figure should be self-explicative, please describe the variables Ctop, Cntp, and D in the caption.
Figure 6: in the caption, also the T points have to be explained.
Lines 412-413: “the majority of tree species information obtained through field surveys is correct.” What does the “majority” mean ? and, how the authors evaluate the goodness of tree species information from the field surveys. The sentence looks more like a generic comment, please rewrite the whole period 410-414.
Lines 468-470: Please, provide references for pros. of DBN classifier (small scales, small samples).
Line 530 and Table 4: To avoid confusion with the D parameter of segmentation (Line 320), please maintain the DBH also in the equation models.
Table 5: Similarly to Table 4, please preserve the variables AGB and DBH instead of D and M.
Results
Methodological paragraphs (Lines 550-556; Lines 574-585; Lines 563-571) have to be moved to the Method section. Moreover, a dedicated Validation section in the methods is appropriated with details on reference data.
Lines 550-553 and Table 7: How do you evaluate the correct classified ground points? what are the reference data ?
Lines 578-581: the use of F-score, although popular, can generate misleading results on imbalanced datasets, because it fails to consider the ratio between positive and negative elements (https://doi.org/10.1186/s12864-019-6413-7). Thus, it should be coupled with the more reliable Balanced Accuracy (https://doi.org/10.1109/ICPR.2010.764) metric that accounts for the unbalance of class distributions, such as this case of tree classes. BA has also proven useful for Lidar data applications (https://doi.org/10.3390/rs14205127 ) as the smaller class performance is not shaded by the proportional influence of large classes.
Table 10: Please, specify in the caption that the Deviation Values are referred to the tree position and not to the tree height.
Lines 614-615: 1D feature vector classification … What does it mean?
Figure 18: how many points are used for the regression of tree parameters? Please add in the txt or in the caption. Why the authors do not divide the dataset into training and validation samples?
Discussion
Comments on the results of similar approaches (pros vs cons) should be added to improve the discussion. Approaches already cited in the introduction (e.g. https://doi.org/10.14358/pers.78.1.75 , https://doi.org/10.3390/rs4020484) can be compared to support the proposed approach.
English is quite good; there are some typos that can certainly be removed during the rewriting of the text.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Minor revision
Please do more proof reading and grammar error corrections.
Author Response
Please see the attachment
Author Response File: Author Response.docx
Reviewer 3 Report
The revision implemented has greatly improved the manuscript (details and readability) and, following some minor changes listed below, the manuscript can be accepted for publication.
***
Please order the number of references. The added citations are not ordered with the previous ones.
Figure 1: Please add the plot numbers to support the comparison with Lidar circles.
As for accuracy metrics, the authors are right about using BA for Lidar classifications, but it can be easy in tree segmentation as well. The True Negatives can be simply evaluated considering the number of initial TreeTop correctly eliminated by the segmentation procedure (i.e., Initial Ctop-(TP+FP+FN)).
This approach can make the article even more interesting and a reference for a solid accuracy metric in segmentation procedures.
Author Response
Please see the attachment
Author Response File: Author Response.docx
