Dear Carrieri et al.,

Your work is important and will likely be very impactful! You demonstrate the effectiveness of using UAV imagery in combination with pre-trained deep learning models for 1) detecting and delineating tree crowns in the ATE, and 2) estimating tree attributes (position and height). Your field dataset is impressive, covering a wide geographic area of the Italian Alps that is representative of heterogeneity at multiple scales and with respect to important climatological, biological, and topoedaphic variables.

However, your work could be elevated with more nuanced discussion of treeline ecology. You make some general statements in the introduction and conclusion about facilitation and competition, but you do not discuss important nuance related to 1) the degree of stress in the system (especially wind-stress), which can lead to a predominance of facilitative interactions; 2) whether any moisture or nutrient limitations are known to exist in your system that would lead to a predominance of competitive interactions; 3) the species composition of your sites and any important biological factors related to the species present, such as relative tolerance of known stressors at the adult and seedling stages, dispersal modality, and growth rates (or any evidence of age-size relationships in your treeline systems); and 4) anything known about the spatial patterns of variables related to the suitability of sites for colonization, such as distribution of soil characteristics, snowpack, or shelter. You need to discuss the relevant ecological literature in your introduction, use it to inform your hypotheses about the spatial analyses you conduct (which are also missing from the introduction), use it to justify the size classes you delineate in your spatial analysis methods, and finally discuss your results within this ecological context in your discussion.

Secondly, you need to revisit the size classes you use for your point-pattern analyses from the standpoint of data reliability. In your discussion, you make it clear that your model’s detection of small trees is biased based on the proximity of those small trees to larger trees. You specifically state that you are more likely to miss the small trees that are closer to large trees. Yet you did a bivariate analysis that relied on accurate detection of small trees to see if they tended to occur close to, or far away from, large trees. Your finding that the small trees tended to be located further from large trees than expected from a relaxed-random distribution could simply reflect the bias in your dataset with respect to the detection of small trees. You absolutely must demonstrate that this is not the case to justify the conclusions you drew based on this analysis. Furthermore, as you discuss the findings of the spatial analyses based on your remote sensing dataset, you must frame these as hypotheses of process based on the observed patterns. There are multiple possibilities that could explain the observed patterns, with competition and facilitation being among them. I listed many papers that you may find useful for adding this nuance.

I want to emphasize that the above does not devalue your study overall, and I very much agree that UAV data fill an important gap in the translation of field data and small-scale processes and patterns at treeline to larger-scale patterns (and potentially processes). Your finding that trees tended to be clustered at scales less than 20 m is very interesting, and seems sound despite the potentially missing small trees. (The opposite result would be less justifiable, given that it could be due to missing trees.) This finding also makes sense based on what is known of processes at treeline (including, but not limited to, facilitation).  However, previous spatial analyses of tree patterns in the ATE, done by Elliot et al. (2010), found spatial randomness. Sindewald et al. (2023 – dissertation publication embargoed until Sept. 2025) also found spatial randomness. I would ask that you elaborate on your definition of individual trees. Many conifer species reproduce clonally at treeline, and the stone pine is dispersed by the European nutcracker, which caches seeds. How did you determine which were individual trees and which were clusters of clonal stems? If your cluster analysis was based on discrete “canopies”, but those canopies are actually different crowns of a single tree, it would explain why you saw such high levels of clustering at smaller scales as well as why your clusters tended to be of similar sizes.

I would also like to note that Grant Elliot’s work comparing spatial patterns across treelines predates your work and compares treelines spanning a greater geographic area (~600 km between the Medicine Bow range and the Sangre de Cristos range). My condolences, but you will need to walk back your claims in your discussion. Instead, compare how your spatial analysis methods differ from Elliot's and, potentially, why your methods would be better as the standard to use for comparison across treelines.

Regarding your model evaluation, I think you need to add clarity to how you divided your data for training, validation, and testing. It sounded like you were using 70% of the data for training and 20% for validation, then later using those same data within your cross-validation. If this is the case, the results of your cross-validation of the model would not be a true test of model generalizability because the model would have already seen those data during training. Usually, researchers either divide their data into training, validation, and testing sets, or they do cross-validation and reserve a geographically distinct dataset for testing. I do not understand why you have done both. You can speculate that your model will generalize well based on the variability represented in your dataset, but if I am correct that the cross-validation folds included training data the model had previously seen, you cannot use that evaluation to draw conclusions about model generalizability. At the beginning of the section, you also state that you tested the effectiveness of training the model with only 3% of your data (given that the model was pre-trained, off-the-shelf), but unless I am mistaken, you do not report the results of this test. (It would be useful to know how this worked out!)

Lastly, I would like to compliment you on how well-written your manuscript was! Your English is very good and surpasses drafts I’ve seen from some native speakers. That said, I did try to assist with grammatical edits throughout. I do insist on the value of the Oxford comma, for example. I also tried to assist with wording whenever your meaning was unclear. There may be cases where I misunderstood your meaning, and in these cases my suggested rewrites may not be correct. Please take them as suggestions and tweak as needed.

Overall, I think your paper is a valuable contribution and I look forward to seeing the next version.

Respectfully,

Laurel Sindewald

Dear Erik and colleagues,

You present a useful and successful approach for mapping spatial patterns in treeline ecotones. In this paper, you focus on tree positions and sizes, but I think the general methodology holds further promise for creating extended vegetation maps and terrain models as explanatory variables (a DTM was created as a necessary step to create the vegetation-height model but was not further used at this point for analysing the tree-distribution patterns, it appears).

The paper is very well written (there are always details, see the yellow markings in the attached pdf and my comments below) and the methodology is generally clear and well presented. The graphs are informative and nicely formatted and the number of figures and tables is adequate. I have just one recommendation that may take some more analyses, and a number of minor recommendations, as listed below, but nothing that cannot be addressed.

Since I took so long to do this review (sorry for that), I can now also refer to the excellent review comments by Laurel Sindewald. I would agree that it needs to be made clearer in the introduction understanding what kind of processes require the mapping of individual trees. However, I understand this paper mainly as a methodological one, and the point pattern analysis as a first application of the results to demonstrate their usefulness. Therefore, I think that the ecological context could be kept quite short and instead the methods could be further validated:

It would be very interesting to see what the inaccuracy in tree detection (of different size classes, as also suggested by Laurel) means for the resulting spatial pattern metrics. This is possibly the most important criterion for deciding whether / for answering what questions the method is good enough: if the resulting point pattern analysis is very sensitive to including or excluding the missed trees (big or small), this would limit the use of the presented method for this analysis, but if it is not, then the method is good enough for such analysis in spite of some inaccuracy. Explaining why clumping / dispersion is an interesting pattern in an ecological sense (e.g. importance of facilitation vs. competition, as also suggested by Laurel) would not hurt in this context.

I am looking forward to seeing the revised version.

With best wishes, Maaike Bader

Detailed comments:

The yellow markings in the pdf indicate tiny spelling, grammar, or punctuation errors, like missing or superfluous words, commas, or hyphens, inconsistencies between verb and subject (e.g. plural vs. singular), etc., or sometimes they mark sentences that have some issue of logic, as explained below in detail.

L11-12 make clear what kind studies you are talking about and what “overall pattern” you would want to model

L14 the “therefore” does not logically lead to conservation…

L16 what gap?

L24 It is not clear what is meant with “tend to progressively occupy safe sites” (and the term “safe sites” does not occur anywhere else in the paper…)

L35-36 Citing Körner here (e.g. Körner & Paulsen 2004) would be appropriate

L42 Mienna et al. did not study carbon sequestration at all. They just mention in the introduction, like many papers do, that treeline dynamics are important for carbon sequestration, but that does not make it a suitable reference for this point.

L43 It is very popular to claim that treelines respond sensitively to climate change, but the interesting thing about them is that some do not respond sensitively at all….

L48-49 this sentence is not clear.

L50-51 The great spatial heterogeneity is actually a great chance in the context of understanding pattern-process relationships, because it allows us to relate many different patterns to many potential drivers. If they all looked the same it would be boring. It is not so clear here what you are aiming at with “hinders case-study observations”. What would you be trying to generalise here?

L51 I recommend “elevation” for height of the terrain above sea level, and “altitude” for height in the air.

L58-59 I recommend being a bit friendlier about field data, since they are still the best quality data there are, also for detecting tree spatial distributions!

L60 The use of remote sensing dates back a lot longer, with aerial imagery available from the start of the 20^th century, and ground-based remote sensing even longer!

L63 long time periods rather than wide time intervals (the latter sounds like a low temporal resolution, which is not what you meant)

L66 Uncrewed or Unmanned?

L68 friendly use = user-friendliness

L72 Although I really like single-tree scales in treeline ecology, and it is useful and even essential for answering some of the questions, this is not the only possible or useful scale in treeline ecology.

L80 & L82 These are not real hypotheses. You could write “We aimed to show that….”

L85-87 Explain why mapping trees at treeline is different.

L97-98 Are you sure there is a gradient from oceanic to continental from west to east? In the Alps the climate gradients usually vary from the outer flanks to the inner valleys, and I have never heard that the area of your eastern study sites should be oceanic. Also, the precipitation data in table 1 do not really reflect such a gradient.

L98-99 What are the implications of this northern and southern air; this is not very clear.

L99 What do these temperature ranges refer to. The study sites? But that does not fit with table 1. And it would be very warm for a treeline to have an annual temperature of 10°C, if the globally consistent SUMMER temperature at climatic treelines is around 6.4°C…

L100 The term “physiography” is unusual and imprecise. What do you mean with it exactly?

L101-102 Here you describe the general regions, but that hardly seems relevant; I would rather expect a description of the treeline sites here

L102 In what sense a “sequence”?

L114 shown here are the entire Alps, not just the Italian Alps, right?

L115 consider explaining abbreviations in figure captions to make them more self-explanatory

Table 1: I suggest sorting the table by longitude (west to east)

Table 1: were the temperatures from CHELSA corrected to correspond with the site elevation?

L121 It is a bit hard to believe that your study sites were really randomly selected. Where there no considerations of e.g. accessibility?

L132-133 how did starting from the middle of the plot affect the spatial resolution if the flight height was fixed relative to the highest point of the plot? And why did you choose a fixed flight height rather than following the terrain?

L134 “shade patterns from clouds”

L135 “respectively, 80 and 80%” -> “both 80%”

L135 “ensure a comprehensive coverage of the surface” -> “to allow calculating robust structure-from-motion outputs” or something like that…

L141 What is a “ground classification”? An extraction of the ground (as in terrain) surface from the point cloud? How does this “normalise” the point cloud. For readers not familiar with this procedure, like me, this is not clear.

L142 Some readers may appreciate another definition (full name) of DSM and CHM here

L143-144 you might want to start this sentence with “in the filed work” (since you were talking about processing the point clouds and suddenly jump back to the field

L160 the “reference data” = “dataset” how were these tiles turned into datasets exactly?

L161 what do you mean by “sites geographical distribution”?

L167-169 It is not so clear to me how the “reference dataset”, created with segment-anything, relates to the ground truth through manual image interpretation. Or how the segment-anything segmentation relates to your trained DL model. Why not just use the segment-anything segmentation directly to predict tree canopies across the study areas, if you assume that that model performs better (since you use it to validate the DL model, L213-214)? In L227 you even mention a “ground-truth crown”…. Is this the manual one or the segment-anything one? This logic needs some more explanation.

L174 Did the larches and the pines and spruces really have similar spectral information? That is quite surprising!

L207 what doe you mean with the “not cross-validated results”? Those where the model was evaluated in the same site where it was trained?

L225 This does not seem correct to me. Sounds to me like the MAE gives more weight to large absolute differences than to small one, but the same weight to the same difference irrespective of the relative differences, i.e. irrespective of the total height of the tree.

L239 Why would a program for point-pattern analysis be grid based? What do you mean by this?

L246 Can you explain how you let the intensity vary across the study area. I.e. what did the null model look like. I guess the ecotonal gradient from forest to no trees was somehow translated into this null model. But how? Since this is important for understanding the outcome of the PPA, it would be good to explain this in the paper.

L281 I would call this an expectation rather than a hypothesis

L298 This Wilcoxon test does not seem to be a very important result. Why did you even want to know whether these accuracies differed significantly? I think you could leave it out and just present the accuracies themselves.

L305 Explain what you mean by "better"

Figure 3 Explain what the orange areas are in b and d.

Figure 4 Explain abbreviation to make the figure self-explanatory. And if you start explaining what the violin plot shows, you might explain it more completely. How does a relative deviation in m work? Might this be the absolute deviation?

L322 & 327 clusterization & clusterized -> clustering & clustered

Figure 6 add information about the spatial scale of the images (i.e. size of the study sites) in the graph or caption.

Table 3 Add: Egli, & Höpke, 2020

L378-379 You might also argue that mapping and monitoring the small trees is particularly important to understand the processes going on in the ecotone. This is actually what I expectd as a logical conclusion after the previous sentence…

L381 First time a test of the leaf-off effect is mentioned. This should be introduced earlier.

L384 First time you mention this test too….

L386 Do you mean that the detection rate was low in the cross-validation because the image looked quite different from the others due to the low sun angle causing long shadows and different colours?

L390 This is a very promising result. At this point, you would usually “conclude” rather than “hypothesise”, although I appreciate the caution with which this statement is made. How about “It therefore appears that ….”? This made me wonder: how did you treat Pinus mugo? As a small tree, or as non-tree? (please mention this somewhere). I suspect that sites with lots of Pinus mugo as a matrix among which trees stick out (or not) could cause particular problems. This may be worth mentioning at this point, as an opening for further research into the question of how important background vegetation is.

L394 To make this logic easier to follow, add as second sentence something like: “and some of the deviations may in fact be due to inaccuracy in the ground control data rather than the UAS images.”

L403 Since this paragraph presents a lot of numbers, it would help if you would repeat the accuracy in m here. I am not sure that these detailed comparisons of accuracies at the cm scale is very exciting for the readers though. More interesting would be to know why the accuracies differ (whereby actually the differences are not huge and all the same order of magnitude), and why yours would be better or worse than other studies, if this can be explained by your method vs the other SfM methods.

L407 Instead of “as high as” a more cautious “not much worse than” may be better, since high-quality laser scanning must reach higher accuracy, and you also give such examples here yourself.

L422 At this detail and extent and for the Alps it may indeed be unprecedented, but I would indeed recommend toning down a little bit in this section. There are studies in the Alps on e.g. patterns of tree recruitment at different sites (e.g. Nicoud et al 2025), and there are multi-site studies about spatial pattern in e.g. the Pyrenees (e.g. cited papers by Camarero, Batllori, Guttierez et al, or not-yet-cited papers by Ameztegui et al 2016, Birre et al 2023).

L426 The pattern is very different between point and polygon metrics, it is clearly not just an effect of  a different range of significance (which, if I understand it correctly, would only result in a wider confidence interval around the null model).

L429 With “systematic effect” you mean “methodological bias” or something like that?

L438 Please explain “active niche selection” for trees (both the “active and the “niche” part)

Page 24 Please cut this long paragraph in two (e.g. at line 439).

L451 This conclusion may need some more thought. How do you imagine competition may work across 50m? May this “repulsion” relate to a large-scale gradient from established forest at one end of the gradient to recruitment at the other end, or is this taken care of in the null model (please see my previous comment about explaining the handling of the gradient in the null model)?

L467-469 Using good training data, VHR images can also distinguish tree species even if they are RGB or monochromatic, since species have different crown shapes. This may be worth mentioning here too (i.e. the investment can be either in better sensors or in better training data, depending on whether one has more money or more time available).

L487 … on the ground or by expensive aerial surveys or lower-resolution, lower-quality and more expensive satellite imagery.

Cited references:

Ameztegui, A., L. Coll, L. Brotons, and J. M. Ninot. 2016. Land-use legacies rather than climate change are driving the recent upward shift of the mountain tree line in the Pyrenees. Global Ecology and Biogeography 25:263-273.

Birre, D., T. Feuillet, R. Lagalis, J. Milian, F. Alexandre, D. Sheeren, R. Serrano-Notivoli, M. Vignal, and M. Y. Bader. 2023. A new method for quantifying treeline-ecotone change based on multiple spatial pattern dimensions. Landscape Ecology 38:779-796.

Egli, S., and M. Höpke, 2020. CNN-Based Tree Species Classification Using High Resolution RGB Image Data from Automated UAV Observations. Remote Sensing, 12(23), 3892. https://doi.org/10.3390/rs12233892

Körner, C., and J. Paulsen. 2004. A world-wide study of high altitude treeline temperatures. Journal of Biogeography 31:713-732.

Nicoud, B., A. Bayle, C. Corona, R. P. Chambard, L. Francon, M. Fructus, M. Bensa, and P. Choler. 2025. Climate, not land-use, drives a recent acceleration of larch expansion at the forest-grassland ecotone in the southern French alps. Science of the Total Environment 959:178326