Summary

Freudenberg, Schnell and Magdon provide a new dataset of standardized measurements of tree species, tree attributes and their Sentinel-2 spectral time series for Germany from July 2015 to October 2022. The dataset contains ~388 thousand individual trees from the German National Forest Inventory (NFI) that are evenly spread out over the country. Due to confidentiality constraints of the NFI, the authors are not able to publish the exact coordinates of the measured trees but alleviate this limitation by spatially and temporally matching the measured trees to Sentinel-2 pixels and publishing their Sentinel-2 spectral data (bottom of atmosphere reflectance). I believe this dataset is unique in its size and spatial extent and therefore potentially very useful for scientists, forest managers and policy makers alike.     

Major comments

This data paper is well written and the dataset in sufficiently described to be useful to the wider community. The Jupyter notebook that is included in the download is a very useful way for potential users to become familiar with the dataset. Nonetheless I have a few major comments on clarity and on the figures.

First, the clarity of the general text can be improved. It was not entirely clear to me how the GNNS locations were used, how the exact locations of the individual trees were determined inside the plots and how their crown area was calculated. I was also confused about the coordinates provided in the dataset, called the “Inspire-grid” coordinates, are these the plot centers? I think the clarity could be improved if the authors are more clearly stating what exactly the dataset represents. It is my understanding that the dataset represents plot data of pure stands of tree species and their spectral data from Sentinel-2 and not individual trees, but please correct me if I am wrong!

Second, I noticed that throughout the text and in the figures and tables, the authors are not using the standard species naming guidelines. It is my understanding that scientific species names should be written in italics with genus name capitalized and the species name not.

Finally, the figures generally look great but the time series- and spectral signature plots (Figure 8 to 11) could be enlarged, there is enough space to enlarge the plots and it would make it easier for the reader to interpret the results. Furthermore, since Figure 8 and 9 depict averages of species, it could be a nice to show some variability around these averages in the form of shading or otherwise.   

Minor comments

L16 “disturbances” instead of “factors”?

L33 please shortly explain or define the F₁ score.

L61-L64 please explain how data on 387,775 individual trees and 70,242 non-tree locations result in “75.3 million data points for trees and 13.8 million observations for non-tree background”. Do these numbers refer to the number of images in the time series multiplied by the number of locations? This is presently unclear.

L64 “51 tree species and species groups” please clarify how many species and how many groups exactly, it could be 2 species and 49 groups, right?

L68 “it contains 24 925 of the 25 382 cluster plots” what happened to the 467 plots not included? Please explain why these were not included in the analysis.

L70 I assume scientific species names should be in italics “Pinus sylvestris”  also please use the full English name of the species “Scots pine” and “Norway spruce” or synonyms.

L73 Could give some references about forest disturbances in Germany since 2018 after “forest has likely decreased” such as: https://doi.org/10.1093/forestry/cpae038 and https://doi.org/10.3390/rs15174234

L92 “we can remove trees that are probably not visible from above by a heuristic.” A heuristic what? Function? Argument”?

L94 “the biggest (area-wise)” what is meant by area-wise? Crown area or basal area or something else?

Figure 5 shows polygon circles representing “modelled tree crowns” but I cannot find in the text how these were modelled. Please explain in detail how this was done because it is a critical part of the analysis. Was crown area measured in the field for each tree? If yes, how was this measured?

L127 “Every date was randomly shifted by up to three days.” Why was this done?

L146 Please specify the brand and model of the ultrasonic device.

L162-163 Please adapt to the species naming guidelines of the journal (I assume species scientific names should be in italics)

Figure 7 might be a nice addition to show two panels: a) species distribution of all trees in the NFI and b) species distribution of trees extracted from the NFI. Also in this figure, please use italics for scientific species names.

L166 “coniferous vs deciduous” is not a useful distinction, you can have coniferous trees that are also deciduous (e.g. Larix). Deciduous says something about the leaf phenology while coniferous says something about the phylogeny (conifers being a subset of gymnosperms) which are not mutual exclusive or useful groups to make in this study. It would be much more useful to refer to “broadleaved deciduous” and “evergreen needleleaf” or in the case of larch to “deciduous needleleaf”. Furthermore, the common holly (Ilex aquifolium) can also grow to tree size but is an “broadleaved evergreen” tree/shrub. Could be nice to add some columns to the dataset providing data on the leaf phenology (deciduous/evergreen), leaf shape (needle/broadleaf) and phylogeny (e.g. plant family).

L179 “Figure 10 shows the total observation count over time.” Make clear what observations these are, images, pixels, trees?

L190-L195 the text and the panels in Figure 12 are not in the same order, please change the order in either the text or the figure and label the panels in Figure 12 (a,b,c,d) and refer to the panels in the text.

Table A3 & A4 Please adapt to the species naming guidelines of the journal (I assume species scientific names should be in italics)

I've read the contribution by Freudenberg et al. with great interest. The lack of reference data is a big obstacle faced by studies mapping tree species across large areas. The authors seek to overcome that obstacle by developing a dataset that can be used by others to train tree species classification models. The dataset combines Sentinel-2 data and the entire National Forest Inventory from Germany. I am sure this contribution is welcomed by the community and I hope other countries will follow in their footstep. Unfortunately, there are a few major issue with the dataset and its presentation:

1) The dataset does not adhere to current practices for pixel-level training data. Typically, individual pixels or pixel blocks are selected and labeled based on the dominant tree species, forest type, or proportions of tree species (based on basal area). This method has been used in previous studies mapping tree species (as cited by the authors) and it is common practice when mapping land cover as well. The authors do not follow this approach. Instead, they select trees from the forest inventory and then extract Sentinel-2 pixels corresponding to each tree. Since trees are much smaller than the 10x10 and 20x20 meter Sentinel-2 pixels, this results in many duplicate pixels in the sample. In homogeneous field plots, pixel values and labels are replicated because many trees occupy the same pixel. In mixed species plots, pixels are replicated, and the same value is associated with different tree species. I have not encountered a study using such a sample for tree species classification, and the authors do not demonstrate the utility of this dataset. I believe pixel replication can potentially bias model training and error estimation. Therefore, I strongly suggest the authors follow current practices. For example, the authors could provide for each subplot: average reflectance, basal area proportions by tree species, crown area proportions by tree species, and/or other tree statistics.

2) The authors add random noise to the pixel reflectance to make the field plot locations untraceable. I understand that. However, the authors do not show a sensitivity analysis of the effect of adding noise. I would encourage the authors to test this effect on mapping accuracy. More generally, the authors could demonstrate the utility of their dataset for mapping tree species. The separability analysis is not really useful for map developers.

3) The writing could be improved with careful editing.

Detailed comments:

L12: Avoid single-sentence paragraphs. Remove. Readers already got that information from the abstract.

L19: traditional? Do you mean field information? I wouldn't term field inventories traditional. Both are needed.

L23: Extensive use?

L62: It is not possible to measure the reflectance of individual trees with Sentinel-2 data due to the spatial resolution. Sentinel-2 pixels represent a mixture of multiple tree canopies and background reflectance.

L73: Area of stocked forest -> forested area (as opposed to forest area) 

L81: Please also specify the basal area factors associated with these radii.

Fig 1. A figure of the sampling density by federal state would be more helpful. It is not possible to see the grid anyway

L85: Please clarify what you mean with "subset of tree species labels". Is a label referring to pixels overlaying the angle count plot with BAF 4? Also, be specific about what you did and what can be done, e.g., "This information was used to label..." rather than "This... allows..."

L90: stand area: you mean crown area?

L90: It is unclear (at this point) why you remove trees. In the previous section, you describe that you identify single-species stands. If you are removing trees in mixed stands, I wonder if your method tends to underestimate conifer trees, since their crown area is usually smaller and the crowns of the surrounding broadleaf trees are more flexible.

L92: same here. Write "..we removed trees..." rather than "..we can remove trees.."

L93: Use past tense consistently to describe what you did.

L96: Doesn't the forest inventory sampling design include non-forest observations? There are advantages with staying within a single sampling design.

L96: comma in front of ", we added.."

L114: I am still confused by your wording. It sounds like you are extracting pixels associated with individual trees. You probably mean "when the projected tree crowns from the plot covered more than a single Sentinel-2 pixel,..."? Please clarify what your spatial unit in the field is. It is probably the subplot and not individual trees, i.e., you obtain a single reflectance measure for each subplot and S2 image.

L115: Fair enough, but you are trying to be more precise than the data, considering that Sentinel-2 also has 20-m bands and the geolocation accuracy of the NFI and Sentinel-2.

L124: Ok. You do seem to extract S2 reflectance values for each tree. What is the rational and use case behind it? An S2 pixel is a mixture of different surface types including different trees, understory and other background. For tree species mapping, we are usually interested in modeling the relationship between such mixed signal either with the dominant tree species, a forest type category, or fractions of trees species. Attempting to link individual trees with Sentinel-2 is uncommon. As a result, you will produce multiple replicates of the same pixel or 3x3 pixel group and associate them with multiple trees of the same or multiple species. 

L125: I recommend to document your data in a table format. A table could show the field name of in your dataset along with a description. 

L142: Because you replicate pixels of subplot, your data still contains auto-correlated samples. There is also spatial autocorrelation within cluster plots. Do you have a recommendation how users can deal with the autocorrelation associated with cluster plots?

Figure 7: Why does the x-axis (count) only go up to 110? How does this correspond to your 350,000 trees and/or 25,000 cluster plots?

Figure 12: Please replace "Samples per 100km2" with a more meaningful unit, e.g. trees per ha or tree species proportion.

4.5 This section reports new results but is presented in the Discussion section.