Segmentation of Individual Tree Points by Combining Marker-Controlled Watershed Segmentation and Spectral Clustering Optimization

Round 1

Reviewer 1 Report

The manuscript proposes a hybrid method that combines watershed segmentation and spectral clustering for single tree segmentation. The method is straightforward, and the technical pipeline is clearly designed. The author describes their methodology well, and the results are well-organized. Although the proposed methodology is not novel, the study was thoroughly designed, executed, and analyzed. The results are of high quality. Moreover, the manuscript is well-written. Based on this evaluation, I recommend publication pending clarification of minor questions.

Section 1.2: Title Revision Needed

The title of this section is inappropriate and misleading. The summarized methods mentioned in this section are not machine learning-based, but rather point-based methods in contrast to the CHM-based methods. It is possible that the authors used this title to differentiate from the deep learning methods. I suggest revising the title or regrouping the methods mentioned in sections 1.2 and 1.3.

Additionally, the phrase 'shallow machine learning-based methods' on line 154 is inappropriate.

In section 2.1, the relationship between tree crown radius and height was measured in situ. This study utilized two sets of open-source datasets. Were the measurements already included in the datasets or how were they obtained? Additionally, how many samples were used to construct the relationship shown in Figure 2?

Regarding Figure 2, it is unclear why the lower bound was chosen. If the main goal is to correct under-segmentation, wouldn't a larger crown radius (window size) be more appropriate? Using the lower bound could theoretically lead to over-segmentation. 

This issue is also mentioned on line 253. This method is primarily suited for trees with a conical shape, such as conifers. This method is not significantly different from simpler methods such as CHM-only, which are only effective for conifer trees (as mentioned in the introduction). Therefore, it is necessary to discuss how well this method applies to deciduous trees with irregular crown shapes. 

Additionally, please elaborate on what the straightforward histogram method is (L290).

Author Response

Kindly see attached response file.

Author Response File: Author Response.pdf

Reviewer 2 Report

General comments:

The Introduction is structured quite well, with a good overview of current individual segmentation approaches, although it would be valuable to have some additional information (even a simple one liner explanation) of each method covered, rather than simply naming them. Personally, I am not a huge fan of the last paragraph of the introduction (contributions) as I think this information is in the wrong place. I would rather see a clear statement of the studies objectives and expected results.

The methodology section is extensive and mostly clearly presented for the detection phase (although I could not find any mentions of which software were used for the analysis), but I feel some of the actual analysis and the different testing set ups carried out are missing from this section, while they are being presented together with the results in the following section (Performance evaluation). This makes the reading very confusing. I would strongly suggest restructuring these two sections following a more traditional research article style (Materials and Methods (starting with the datasets used, then presenting the theory behind the proposed approach and including the different testing designs), Results, and Discussions). The text is all there already, just needs to be moved around in the different sections. I would also argue that there are too many tables and figures, and some of them could easily be placed in the supplementary materials, although this is not strictly necessary.

The conclusions section reads well and can be left as is.

Minor comments:

L25: Please change “technique” with “technology”.

L33: I believe you meant to say “relevant” not “irrelevant”. Please change it.

L129-131: This sentence is hard to follow, please clarify what you mean.

L190: I would suggest changing “process” to “workflow”.

L214-216: Please change this sentence as follow “The optimal model is then selected based on the best fit to observations. Here, the quadratic model performed best, achieving an R² of 0.56.”

L220: Please remove the adjective “vividly”.

L223: Please remove “the” before “under-segmentation”.

Figure 2: I am a little confused here by your choice of using the 99% lower bound curve for your estimations. From the looks of it, you seem to have negative crown radius for small (<12.5m tall) trees. Did you have any trees in your sample areas that were shorter than this threshold? It might be appropriate to comment on this (unsure if you do mention it later in the manuscript).

L277: Please clarify what does the T in the condition equation refer to?

L290-292: Does this mean that the threshold is manually selected for each patch?

L298-299: I think this sentence is not appropriate. I would argue that under- and over-segmentation might have rather high impacts depending on the type of forest you are working on, being it deciduous or conifer, and having a closed canopy versus open canopy forest as well (irrespective of common growth form). You could easily have over-segmentation problems with conifers in an open canopy forest, where lateral branches extend too much and start growing next to the dominant shoot. I would suggest either softening this statement to allow for more possibilities, or really narrow it down to your specific study area.

L407-408: What do you mean by “and so on”? What other ground refences are included in the dataset? If you are using any other, please mention them here, otherwise I would suggest to simply remove that.

L428-430: How accurate did you manage to be with the manual segmentation to derive the validation dataset? Might be worth adding a comment on this.

Author Response

Kindly see attached response file.

Author Response File: Author Response.pdf

Reviewer 3 Report

This article introduces a method that combines watershed segmentation with spectral clustering algorithms controlled by markers for the identification and segmentation of individual trees. The article is logically and well organized, while it lacks an explanation of the principles behind the marker-controlled watershed algorithm. Totally, it is suitable for publication, and some minor comments are listed follows:

Line225: The article mentions non-dominant trees for the first time. Please elaborate on how to distinguish between dominant trees and non-dominant trees.

Line 335 : The article selects a fixed rotation angle of 45 degrees. Please provide justification for this choice and explain whether relying solely on these specific orientations is reliable.

Author Response

Kindly see attached response file.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

I would like to thank the authors for their time in revising the manuscript. The structure of the manuscript is a lot clearer now, although I would still argue that there is need of some minor refinements. Some parts of the results section belong to the discussions, as you should not have any need of citing references when presenting your results. Also, you should leave the comparison of your results to others in the literature for the discussion section. Please move all those parts into the appropriate section.

Minor comments:

L129-131: The sentence is missing the verb. What are you trying to say here? Am I correct in thinking you meant “is” instead of “as”?

L179: Please change to “obtains”.

Table 1 should be placed at the end of 2.1 Datasets section, not before.

L235: Please change to “The watershed”.

L248: Please change to “issues can still persist”.

L346: Please change to “number of points”.

Thank you and kindest regards

Author Response

Kindly see attached file.

Author Response File: Author Response.pdf