Thank you for the opportunity to review this manuscript. In this article the authors present an innovative method to incorporate citizen science photographs of trees to segment and classify ten deciduous tree species from aerial images, using a Convolutional Neural Network. The two-step approach of using simple labels of citizen science data to create masks for a segmentation model is innovative and highly relevant. I think that the paper fits well within the scope of this journal and presents an application of an interesting new approach to remote sensing.

The manuscript as a whole is very well structured. Only the first part of the abstract could be shortened significantly.

Comments

1) The first part of the abstract, that presents an overview of the problem could be shortened to make it more concise (try to summarise each section of the manuscript in 1-3 sentences).

2) l. 250 why did you choose EfficientNetV2L over the other tested backbone architectures?

3) l. 261 how much % of the images were assigned NA? Did this influence the model training?

4) Could you explain the term “replacements” (e.g. l. 240)?

5) Do you think the amount of misclassified data could be a problem for the training of the segmentation model? (l. 297-298)

6) 0.22 cm already seems like very high resolution. Many remote sensing studies focus on making high resolution reference data more usable over large areas (i.e. by adapting it to satellite data). You argue for the use of even finer resolution data in the future. What research objectives could be studies using this very high resolution of UAV data? Is there a research gap for very high prediction accuracy over relatively small areas? Could multispectral/hyperspectral sensors be more useful than higher resolution?

Minor comments

l. 29 Please remove the “and” between “data” and “by”

l. 51 “unleash” might not be the right word; “harness” might be better suited

“provided” might be better instead of “given”

l. 56-60 This sentence is not completely clear to me. Maybe you can reformulate it to make it

easier to read.

l. 63 Please remove “similar”, as it is unnecessary

l. 66 Consider combining sentence “[…] costly, as training data […]”

l. 81 Is the training data limited or just costly/time consuming to generate?

l. 89 “platforms”

l. 90/95 “mil” or “M”;

please remove “of”

l. 97 Please remove “The” before “Pl@ntNet”

l. 109 “Ideally, for species mapping applications […]”

l. 115-120 This part might fit better in the Methods section

l. 198 Please remove “Accordingly”

l. 235 “were afterward rasterized”

l. 240-241 What does “sampled with replacement” mean?

l. 317 Please replace “while” with “although”, or similar

l. 337-341 This might fit better in the Discussion section

l. 367 “varying”

l. 373 “partially relatively inaccurate” → This is a little vague. Maybe expand upon it a

little.

l. 387-389 Please remove one instance of “plots with more species (two or four)”

l. 393 “higher value” than what?

l. 442 Maybe you can find a better phrasing than “diversity of human behaviour”

l. 457 “often costly”

l. 484 “large” instead of “excessive” (which means unreasonably much)

l. 485 “good transferability”

Figure 2: The text font is very small. It would also be better if the labels match the ones used in the text: “Ortho_July” and “Ortho_September” instead of “Ortho 1” and “Ortho 2”

Figure 4: The text font here is also very small.

Figure 6: The height of the transects seems to be different between plots (eg. plot 29 and plot 33). If they are all the same (2 m), please show them with the same extents in the figure as well.

I enjoyed reading the manuscript and its rigorous approach to image segmentation and have no additional comments in addition to those of Reviewer #1.