A Multi-Threshold Segmentation for Tree-Level Parameter Extraction in a Deciduous Forest Using Small-Footprint Airborne LiDAR Data
Round 1
Reviewer 1 Report
Some comments:
location plots should be measured with better accuracy. more reliable results could be received if some trees dontt have heights, those can be estimated with other measured trees segmentation methodology itself is goodAuthor Response
Thanks to the LiDAR data and field data provided by the Department of Forestry at the University of Kentucky, some researches about how to improve the tree segmentation were implemented in our paper. Indeed, more reliable results could be received with more accurate plot locations. Among the field data, only 84 of the 271 plots have on-site measured tree heights data. If we get the description of species code in the 271 plots, we can use regression models of species and DBH to estimate the missing heights. Furthermore, more reliable results could be received with better segmentation algorithm. We are trying to utilize deep learning method to improve the tree segmentation algorithm, we sincerely hope that the reviewer can let us elaborate on the next article.
Thank you again for the guidance, I have learned a lot, which makes the article better.
Reviewer 2 Report
The paper presents a new approaches to three level parameters extraction for forest resource inventory and management Experimental results show that a combination of multi-resolution threshold segmentation based on stratification and cross-layer tree segments merging method can provide a significant performance improvement in individual tree-level forest measurement.
The paper is well written and able to comunicate the information. Figs and Tabs are well represented.
I suggest to better frame the topic talking about ther fields in which the segmentation techniques could play a very important role.
e.g Masonry-stone segmentation which is performed to identify the cracks in historical masonry structures. To this end I suggest to consider the following papers and others in which the identification of the cracks in the masonry construction is very important:-Survey and seismic vulnerability assessment of the Baptistery of San Giovanni in Tumba (Italy). Journal of Cultural Heritage, 26, 64-78. -Automated defect detection and classification in ashlar masonry walls using machine learning. Automation in Construction, 106, 102846
Author Response
Thank you for your suggestion. Indeed, in recent decades, with the development of remote sensing, segmentation technology has played an important role in many fields, including historical building survey, forestry resource inventory, geological disaster survey, mining survey, and digital city construction, etc. I've read the papers you recommended to me. There is some correlation between the crack identification technology of historical stone masonry structure and the tree segmentation technology of forest resources inventory, which is worth further study. I got a lot of inspiration from these papers. However, for our paper, I think it might be better for the framework theme to focus on research on segmentation techniques in the field of forest resource inventory.
Thank you again for your comments and suggestions, which have broadened my horizon and inspired me a lot.
Reviewer 3 Report
It is necessary to explain better in the paper the DSM algorithm implemented (figure 5):
GMX identification, what is this? Is it the maximum apex of the model or of an area? Generation of 8 vertical profiles. What do they mean? Why the number of 8? In the identification of the crown boundary: What is the GAP? Are They isolated areas? Is getting applying a neighborhood algorithm? What is the minimum local detection? Finally it seems that Convex Hull is used to obtain the crown areas, to determine the segments or regions? It is not enough cleary. Is getting a clustering of the different convex hulls obtained in each individual crown?
Figure 6. You must explain how group segmentation is performed.
Combination of vertical segments. Please, must indicate whether the relationship between the regions of the different layers is carried out because there is an intersection in the XY projected regions.
It is not clear the combination criterion of the flowchart in Figure 8. In this flowchart a secondary node is a group chosen from Layer 1?
The paper is very similar to “Hamraz H, Contreras M A, Zhang J. Vertical stratification of forest canopy for segmentation of understory 672 trees within small-footprint airborne LiDAR point clouds. ISPRS Journal of Photogrammetry & Remote 673 Sensing, 2017, 130:385-392.” It is good that authors indicates what is new in the current paper?
Author Response
Please see the attachment
Author Response File: 
Author Response.pdf
Round 2
Reviewer 3 Report
The answers to the questions raised are correct. For my part they can accept the article and proceed to its publication.
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.
Round 1
Reviewer 1 Report
I acknowledge the authors' work in the resubmission of the paper: they have included more graphs and text in this revised version. They have attempted to adress comments made by the reviewers.
Unfortunately, the new information provided often renders the paper redundant making its reading somehow more confusing. For example, Fig. 1, 3 and 6 look alike, see L175 and L213.
Some concepts are announced first in the paper and then later explained: a reader does not know if the authors refer to the same concept or something new (e.g. L 183, L261 and L304).
The results and discussion section contains some methodology information (e.g. L344)
L141: what is the size of the plots?
L156: there are two Lidar datasets. Were both datasets used for the analysis?
L171: Add a Methodology section. The methodology followed can be divided in four main steps (Figure 1). Section 1: Isolating layers, Section 2: Implementing the multi-threshold segmentation, Section 3: combining tree segments", Section 4: Delineating individual trees.
L236: why eight uniformly spaced profiles? Does that mean that layer 1 is systematically segmented into 8 lidar-tree tops (see Fig. 9)?
L319: which datasets were considered? The two lidardatasets? One lidardataset and the field dataset? L143: there is an average error of upto 5m in plots georeferencing. How can you be sure that a lidar tree-top coincides with a given field tree using equation 5?
L338: how were the 23 plots selected? 23/248 overall plots is not representative.
L345-349: redundant, already mentioned in methodology section
L380: Before concluding that the multi-threshold segmenation is better, what is the field reality for this plot?
L393: overstory trees rather than overstoty
L426: Tree matching is based on x,y, and z coordinates. How do you deal with conflicting scenarios where for a given lidar tree-top coincides two field trees (from the stem map)?
L430: Sometimes you use feet, sometimes meters to refer to distance: please select one same unit
L437: Comparison rather than Compared
L445: the paper refers a lot to Hamraz et al study. Yet, these authors have already mentioned a tree segmentation approach according to L111. Clearly state what is the novelty of your paper compared to Hamraz et al study.
L462: what is the Multi-threshold tree segmentation approach? What take-home message should a reader remember about the multi-threshold tree segmentation approach?
Reviewer 2 Report
apologies for my brevity, i am currently out of office but do not want to have you wait.
The manuscript is much improved over the previous version (420735) where I was rev.2. mostly in terms of clarity, readability and overall quality of the figures. however some serious issues remain with the structure and quality of the results.
Results and discussion are mixed together, for clarity please present them in two separate sections.
for the result section: any of the comparisons between datasets or regarding data quality will need actual statistics (including some form of testing of significance of the difference) please define your methods in the method section and report on which differences are significant and at what levels
I still am very much unconviced of the rationale behind your new method. the improved figures (1, 3, 11) show very little basis for using return nr to define layers. I will therefore need to see some comparison to actual ground truth quality data (for example stem counts, basal area, locations at the ground. not just another tree segmentation algorthm) to indicate the relevance of the method.
while you have pointed to practical uses in the revised manuscript, an actual evaluation of your restults in terms of usable forest metrics (so compare results to some form of truth measurement en relation to standard methods) would add some much needed relvance to the presentation of the algorithm.
> especially in fig 11, the chosen plots do not look anything like the shape of a single tree... leading me to believe your segmentation algothim is wrong. in most high density lidar datasets the morphology of trees is usuallly well represented in the point cloud when looked at in profile.
Reviewer 3 Report
This paper is a replication of the method proposed by Hamraz et al. (2016), and provides no new development.
