An Unsupervised Canopy-to-Root Pathing (UCRP) Tree Segmentation Algorithm for Automatic Forest Mapping
Round 1
Reviewer 1 Report
review comments on article "An Unsupervised Canopy-to-Root Pathing (UCRP) Tree Segmentation Algorithm for Automatic Forest Mapping"
This is an interesting approach to segmenting forest point cloud into separate trees. The article follows the naming convent of Wang et al, which is also the state-of-the-art method, the results are compared to.
The introduction section is very comprehensive in historical view, in my opinion, less detail on history and more on recent point cloud methods would be more interesting, but it is not too relevant to this review.
I recommend the article to be published with minor review.
Minor detailed comments:
Chapter 1.3, Page 3, line 98: "laser raging" -> "laser ranging"
Chapter 2.1.1, Page 6, lines 270-272: "Instead of introducing error by attempting to filter segmentation results, we used the point clouds to manually measure the tree positions and heights of all trees in the six plots which originally had been excluded from the provided validation data." Please rephrase so that it is unambiguous (currently "which" could refer to "six plots" rather than to "tree positions"). Add, what you did with the manual measurement, did you remove trees, that were found and matched the manual locations or what?
Chapter 2.1.2, Page 6. line 283: "the 4D compartment of Martell Forest" I guess the 4D compartment is known to your group, to the world it is not, and I could not find info on it on the internet. Please consider introducing 4D compartment*.
Chapter 2.1.3, Page 7, lines 298 - 313: This part contains information that would be better placed in introduction.
Chapter 2.1.3, Page 7, lines 314 - 318: * In my opinion, your text would be clearer if you separated the site introduction from the device and methodology - for example 2.1.2 the Martell Forest compartment, 2.1.3 UAV Lidar datasets 2.1.4 UAV photogrammetric datasets. Currently, there is repetition on site description. Added image of extension of UAV Lidar and photogrammetric areas would be nice. Add description of reference data collection, what is minimum height / dbh for a tree? Or is it relevant to declare tree validity constrictions?
2.2.1. “Point Cloud Normliation” -> Point Cloud Normalization”
Chapter 2.2.6, Page 12, lines 488 - 490: "The algorithm then returns a point cloud with a classification value for all points. Points not included in any tree are given a classification of 0 (Figure 2.i)." Please use "segment label" instead of "classification" for clarity.
A lot is being done about forest point cloud segmentation, tree delineation, species classification, name it..., around the world. What is in my opinion too rarely discussed in publications, is "what is enough?". Especially with different needs for collecting forest data - CO2 budget computations vs timber volume. If purely academically the best segmentation algorithm is searched, the point of why and to what extent is missed.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The manuscript under review focuses on a real interesting and challenging topic, as is the aim for automatizing the process of tree segmentation following cloud point acquisition. The utility of airbonrne-LIDAR, TLS or photogrammetry methods in forest inventory and forest management relies on the capacity of having automatized methods for processing such huge datasets, resulting in accurate estimates of such traits as tree number, tree position, tree height, etc, making no necessary any manual supervision of the process. The work under manuscript presents a new algorithm to deal with this, which is therefore applied over different data sets (UAV photogrammetry, LIDAR and TLS) and type of forests. The manuscript is nicely written, very easy to follow, and constitue a nice piece of work which deserve enough merits as for recommending its publication.
I only have a few minor concerns/comments on the text, that I’ll expose in the next paragraphs, together with some detected minor errors
MAIN COMMENTS
- Sensitivity of the method to the predefined parameters
The proposed algorithm relies on different main parameters directly fixed by the authors: voxel size, minimum height for canopy classification, maximum height for ground classification, merging distance, number of neighbours, etc… It would be interesting to know whether the authors have tested other possible values, or even carrying out a sensitivity analysis to detect which could be the potential effect of a wrong selection of these initial parameters. Of course, I’m not demanding to do this in this manuscript, but at least this is a topic for further research that should be mentioned in the discussion.
- Algorithm implementation
It is not clear under which programming language has been the algorithm coded-implemented, and if it is available for other external users.
MINOR ERRORS
Line 195-196: contribution number (4) is missed
Section 2.1.1. While TLS data have been obtained from an external dataset, I miss some general information on scanning parameters and point processing: number of point per m2, vertical angle scanning, point clouds registration methods, cloud depuration…
Line 255 I suggest “hypsometer” better than “inclinometer"
Line 318-322: Photo-Natural plot is the same plot used for UAV Lidar data set? If so, this should be addressed here
Line 425, eq 2: what are pi and pj? I assume that are the coordinates for the superpoints i and j, but this should be clarified
Line 426-427: what do the authors mean by “L2 norm”?
Tables 5&6, and page 18, fist line: Commission errors have not been previously defined
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Evaluation of An Unsupervised Canopy-to-Root Pathing (UCRP) Tree Segmentation Algorithm for Automatic Forest Mapping
This is a very well written paper on a subject of interest to forestry .
The only thing that bothered me was the assumption that manual estimation of tree parameters from the point cloud were as valid as physically measured Only in the Finnish dataset did it seem that there were any field measurements used. Possibly one can make an argument that point cloud measures can be more precise and accurate than field measurements.
Also the authors maybe a little too confident in the value of this method. Results of a single study may need to be verified by others before some of the claims made in the abstract can be accepted. However, the results do seem quite compelling.
Specific comments
Line 253-256. Wording is confusing trees were located on point cloud manually then description of DBH and height using physical on the ground techniques. What technique was used to assure spatial correspondence of the point cloud stem and the field measurement? Such as total station survey or taped gridding.
Line 291- accuracy was quoted for 50m height but flights were 80-110 meters. maybe some statement as to the degradation of accuracy with altitude cold be added.
Line 295 no mention of any on the ground verification.
Line 312 inappropriate present tense. Use found instead.
Line 319 was the 58x58m plot the same in the photogrammetry as the plot used for airborne lidar.
Line 496 drew inspiration seems a little whimsical for a scientific paper. Used Wang method as best available may be more mundane way to state you are comparing to the best currently available.
Line 496 Wang citations only location on Github not publications that describe the method and its use.
m
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
