Dear editor, Thank you very much for your time and effort. The review we received is included below in regular text and the changes that are included in the manuscript and the responses are written below each point in bold and cursive starting with the word response.

I, Femke van Geffen would like to, also on behalf of my team, the reviewers for their time and detailed comments on the manuscript and dataset. The comments greatly improve the work and are well appreciated.Kind regards,  Femke van Geffen

1. General comments:

These look like valuable datasets to make publicly available, from a vast region where there is a real paucity of data by any measure. These data applied to machine learning and other automated image analyses and interpretations should have value for remote sensing studies in the circumboreal region. I encourage publication, and could only note a few minor issues detailed below.

1. Significance
1. Uniqueness

These data are definitely unique in that they are from Siberia, a larger expanse of tundra–taiga than in N. America, and yet almost all available ABOVE datasets are based in N. America, especially Alaska.

1. Usefulness.

Is it useful or necessary to display images that were excluded from the dataset?

Response: This is included for informing on the quality of the data sets and to show details on our data processing standards: It shows the steps that were taken to make the dataset useful. Taking the images out that include people eliminates noise from the data such as the image shown in Fig. 8.

We also added the sentence:

L317-319:“The first fifteen images were excluded because they often contain a visual representation of the research team (for example, Fig. 8). Excluding these images reduces noise in the dataset as we aimed to include only forest and natural terrain in the images.” 

Figure 1. The study region names Chukotka and Yakutia are missing from the map or caption.

Response: Thank you for pointing this out, the regions are now included in the figure caption below figure 1.

Table 1. If my math is correct, it looks like Lake Ilirney should have 27 plots not n=25 as stated.

Response: This is correct there are 27 plots. The mistake was made in the Bilibino plots where plot EN18033 and EN18035 are actually not included due to lack of usable data for these plots. That led to a mistake in the total which led to a mistake for the plots of Lake Ilirney, which in fact does have 27 plots. It is now corrected in Table 1.

1. Data quality

L243. Were the cardinal directions adjusted for magnetic declination in the field, or prior to preparing the data product? This will affect how the plot sample area intersects with any georectified remotely sensed data.

Response: The magnetic declination in central Yakutia and Chukotka is very low  (around or less than 10 °). We oriented the vegetation plot by iteratively moving with GPS with the set up for the true North, not the magnetic North, and did not use a compass for the North orientation of the vegetation plot. Like this, despite also the challenge of accurate GPS measurements in the field, we consider the North-orientation of the vegetation plot as robust and reliable.

Presentation quality

Fig. 6. The background RGB image is shifted between the images on the left and right, making it unnecessarily difficult to visually compare the overlaid features and labels.

Response: Yes thank you, The figure was redone and can be seen in the revised manuscript. Two figures were combined into one new Fig. 6, hopefully a clearer image.

1. Specific comments:

L244-245. Could tree density per plot be calculated from these data? It sounds like the answer is no, if the only requirement was to sample 10 trees. How were the 10 trees selected from the size and spatial distribution of available trees in the plot?

Response: Yes, tree density per plot can be calculated from the data. The 10 trees were only selected to cover the whole size range of each species and investigate in-depth. All trees on the plots, with a height of >40 cm, were recorded and their height, vitality status, and species noted. This information per plot can be easily used to calculate the density.

In addition, the point clouds contain all trees on the plots. This information can also be extracted.

L556. Were the estimations made in the field or in the office based on visual interpretation of imagery?

Response: Thank you for your feedback. We have clarified it in the text.The recorded tree crowns refers to the trees measured in the field that were then linked manually to the orthoimages based on their locations and visual appearance. L410-415:

“In order to make assumptions and predictions about the content of the vegetation plots it is important to link the labelled individual trees from the fieldwork to the processed orthoimages. We located 872 trees and large shrubs in the orthoimages that were surveyed in Siberia during the two-month fieldwork expedition in 2018 (Kruse et al., 2019) (Fig. 16). “

The automatically detected tree crowns were estimated with detection software in R on the UAV orthoimages. Explained in the text at L273-279: “  The SiDroForest data collection also contains 19 342 automatically detected tree-crown polygons (Kruse et al. 2021b). The tree crowns were captured in the CHM by watershed segmentation analysis using the R package ForestTools (Plowright, 2019) and successive automatic generation of a polygon around them following Brieger et al. (2019). This automated tree-crown detection algorithm was run for all plots and the resulting shapefiles are provided for each plot that contained trees. Quality assurance was performed for each plot by carefully examining each plot based on expert knowledge and assigning a quality score of Q1 (good quality), Q2 (medium quality), or Q3 (poor quality) to the shapefile products.”

1. Technical corrections:

L82. Not just a C sink, but also potentially a C source.

L93-94. This sentence is incomplete.

L103. The word “Likeness” actually isn’t and should be deleted.

L128. Change “was” to “were”.

L129. Change “was” to “were”.

L153. The sentence that ends here needs a period.

L203-204. State why datasets one and two were prepared before doing this for datasets three and four.

L240. Change “projective” to “projected”.

L253. Change “are” to “is”.

L618. Should not the citation be Fig. 22 (instead of Fig. 23)?

L672. Change “as” to “a”.

Response: All changes were made based on the technical corrections suggested.

Dear Reviewer, Thank you very much for your time and effort. The review we received is included below in regular text and the changes that are included in the manuscript and the responses are written below each point in bold and cursive starting with the word response.I, Femke van Geffen would like to, also on behalf of my team, thank the reviewers for their time and detailed comments on the manuscript and dataset. The comments greatly improve the work and are well appreciated. Kind regards, Femke van Geffen

Review:

The datasets presented seem to be quite useful for an important region where data is scarce for research. I believe these field survey data would be helpful for both small-scale and large-scale studies in understanding the circumboreal region. A few comments listed below for consideration.

Q1: Is the data set significant – unique, useful, and complete?

I think the data is unique and complete, though the authors can mention more about the usefulness of these datasets. For example, some scientific questions that these datasets can help address? What specific applications can be expected from these datasets? Or how these datasets may help change the research landscape over time?

Response: We highlight the usefulness of the data set with first information in the introduction and describe its uniqueness and suggestions to how to use the data in the discussion. The purpose of each data type is also mentioned in the product data sections. I have clarified the usefulness more in the current version of the paper, please see lines below:

1-The satellite data is processed to a high degree and prepared for machine learning tasks with labeled patches.

L56: “The fourth dataset contains Sentinel-2 Level-2 bottom of atmosphere processed labelled image patches with seasonal information and annotated vegetation categories covering the vegetation plots (van Geffen et al., 2021b, https://doi.pangaea.de/10.1594/PANGAEA.933268). The dataset is created with the aim of providing a small ready-to use validation and training data set to be used in various vegetation-related machine-learning tasks. It enhances the data collection as it allows classification of a larger area with the provided vegetation classes. “

2- The orthoimages provide geographical information on individual trees and shrubs that were recorded on each plot. These individuals have information such as tree height and species and can be used to calculate the projected cover of each species and give general insights into the vegetation dynamics.

L153: ​​“Individual labelled trees surveyed during the fieldwork, including information on height, tree crown, and species. These tree-individual labelled point and polygon shape files (light green symbols) were generated and are linked to the UAV RGB orthoimages of the expedition vegetation plots.”

3- The point clouds provide information on the 3D structure of the forest at each location. The point cloud products such as Canopy Height Model can be used to extract the height of all the trees on the plots, not just the recorded ones from fieldwork.

L36-41 “The first dataset provides Unmanned Arial Vehicle (UAV)-borne data products covering the vegetation plots surveyed during fieldwork (Kruse et al., 2021, https://doi.pangaea.de/10.1594/PANGAEA.933263). The dataset includes structure from motion (SfM) point clouds and Red Green Blue (RGB) and Red Green Near Infrared (RGN) orthomosaics. From the orthomosaics, point-cloud products were created such as the Digital Elevation Model (DEM), Canopy Height Model (CHM), Digital Surface Model (DSM) and the Digital Terrain Model (DTM). The point cloud products provide information on the three-dimensional (3D) structure of the forest at each plot. ”

4- The synthetic dataset provides 10000 images that were generated using an algorithm. The purpose for this is that they can now be used as input for a neural network which needs many images as input for training.

L51-55: “The third dataset contains a synthesis of 10 000 generated images and masks that have the tree crowns of two species of larch (Larix gmelinii and Larix cajanderi) automatically extracted from the RGB UAV images in the common objects in context (COCO) format (van Geffen et al., 2021a, https://doi.pangaea.de/10.1594/PANGAEA.932795). As machine learning algorithms need a large dataset to train on, the synthetic dataset was specifically created to be used for machine learning algorithms to detect Siberian larch species.”

Q2: Is the data set itself of high quality?

Yes, the data is of high quality.

Presentation quality

I do not think the authors pointed out specific/suitable software for simple visualization and analysis as I typically need to install some software to view the data.

Response: Yes we did not specifically name suitable software. The Sentinel-2 and UAV orthoimages data can also be opened and viewed in any open source and commercial GIS and Remote Sensing software and for the point clouds AgiSoft or Cloudcompare can be used.  All data can be loaded in R. Suggestions to software have been added to chapter 2 data and methods and to specific sections.

L165-170: “ The SiDroForest products are in common software formats: there are point and polygonal shape files (shp), raster files are in the georeferenced tagged image format (tif), Geotiff, shapefile formats and JavaScript Object Notation (JSON) can be read and visualized in any open source and commercial GIS and Remote Sensing software tools and a wide range of libraries in R, python and other programming languages. The point clouds are provided in the standard LASer (LAS) binary file format that can be handled in any software that supports this format such as CloudCompare (CloudCompare, 2021) or R (R, 2020) or Python libraries specifically developed for this datatype.”

Specific comments

(1) The abstract can be improved by why these datasets are needed.

Response: We enhanced the abstract text in making our statements clearer. Limited data on boreal forest structure are available, especially for this region (Eastern Siberia) and in quality with labels that can be used for machine learning purposes. Please find all the changes in the revised manuscript with tracked changes.

(2)  The abstract seems too long and I would suggest further summarizing it and highlighting the datasets in brief.

Response: We have taken this suggestion into consideration. Please see new document with the shortened, edited abstract.

Abstract

The SiDroForest data collection is an attempt to remedy the scarcity of forest structure data in the circumboreal region by providing adjusted and labelled tree level and vegetation plot level data for machine learning and upscaling purposes. We present datasets of vegetation composition and tree and plot level forest structure for two important vegetation transition zones in Siberia, Russia; the summergreen–evergreen transition zone in Central Yakutia and the tundra–taiga transition zone in Chukotka (NE Siberia). The SiDroForest data collection consists of four datasets that contain different complementary data types that together support in-depth analyses from different perspectives of Siberian Forest plot data for multi-purpose applications.

i) The first dataset provides Unmanned Arial Vehicle (UAV)-borne data products covering the vegetation plots surveyed during fieldwork (Kruse et al., 2021, https://doi.pangaea.de/10.1594/PANGAEA.933263). The dataset includes structure from motion (SfM) point clouds and Red Green Blue (RGB) and Red Green Near Infrared (RGN) orthomosaics. From the orthomosaics, point-cloud products were created such as the Digital Elevation Model (DEM), Canopy Height Model (CHM), Digital Surface Model (DSM) and the Digital Terrain Model (DTM). The point cloud products provide information on the three-dimensional (3D) structure of the forest at each plot.

ii) The second dataset contains spatial data in the form of point and polygon shape files of 872 labelled individual trees and shrubs that were recorded during fieldwork at the same vegetation plots (van Geffen et al., 2021c, https://doi.pangaea.de/10.1594/PANGAEA.932821). The dataset contains information on tree height, crown diameter, and species type. These tree- and shrub-individual labelled point and polygon shape files were generated on top of the UAV RGB orthoimages. The individual tree information collected during the expedition such as tree height, crown diameter and vitality are provided in table format. This dataset can be used to link individual information on trees to the location of the specific tree in the SfM point clouds, providing for example, opportunity to validate the extracted tree height from the first dataset. The dataset provides unique insights into the current state of individual trees and shrubs and allows for monitoring the effects of climate change on these individuals in the future.

iii) The third dataset contains a synthesis of 10 000 generated images and masks that have the tree crowns of two species of larch (Larix gmelinii and Larix cajanderi) automatically extracted from the RGB UAV images in the common objects in context (COCO) format (van Geffen et al., 2021a, https://doi.pangaea.de/10.1594/PANGAEA.932795). As machine learning algorithms need a large dataset to train on, the synthetic dataset was specifically created to be used for machine learning algorithms to detect Siberian larch species.

iv) The fourth dataset contains Sentinel-2 Level-2 bottom of atmosphere processed labelled image patches with seasonal information and annotated vegetation categories covering the vegetation plots (van Geffen et al., 2021b, https://doi.pangaea.de/10.1594/PANGAEA.933268). The dataset is created with the aim of providing a small ready-to use validation and training data set to be used in various vegetation-related machine-learning tasks. It enhances the data collection as it allows classification of a larger area with the provided vegetation classes.

The SiDroForest data collection serves a variety of user communities. The detailed vegetation cover and structure information in the first two data sets are of use for ecological applications, on one hand for summergreen and evergreen needle-leaf forests and also for tundra–taiga ecotones. The first two data sets further support the generation and validation of land cover remote sensing products in radar and optical remote sensing. In addition to providing information on forest structure and vegetation composition of the vegetation plots, the third and fourth datasets are prepared as training and validation data for machine learning purposes. For example, the synthetic tree crown dataset is generated from the raw UAV images and optimized to be used in neural networks. Furthermore, the fourth SiDroForest data set contains Sentinel-2 labelled image patches processed to a high standard that provide training data on vegetation class categories for machine learning classification with JavaScript Object Notation (JSON) labels provided. The SiDroForest data collection adds unique insights into remote hard to reach circumboreal forest regions.

(3) I am not convinced how this field level data can help studies in this region. Only for optimized data containing annotated vegetation categories. What is the motivation for doing these machine learning applications? DO we need to look at data from a temporal perspective?

Response: The idea is that more detailed vegetation information can be extracted from this dataset per plot. Also, there is information on the vegetation structure of each plot in the SfM point clouds.

The machine learning applications for tree crowns created a set with n = 10 0000  crowns that can be used to automatically detect larches in UAV RGB images of forest plots.

It depends on the application if we should include temporal perspectives. The temporal information in the current dataset is the three seasons of satellite data included in the Sentinel-2 part of the dataset. At the current state this data collection is now an important and unique openly published data collection, a snapshot for the time window around 2018 for the taiga-tundra transition zone in Chukotka and the evergreen-summergreen transition zone in Central Yakutia. if in the future again plots of the same region will be assessed, change can be detected. This time stamp in 2018 can also support the assessment of land surface satellite products of the second late decade of this century compared to later satellite acquisitions in future and earlier satellite acquisitions in the past.

(4) The introduction should be shortened as well to highlight the contribution of datasets to the scientific literature or specific questions/challenges to be addressed

Response: Thank you for your feedback. We have rewritten the introduction and shortened it considerably (please see the track change file).

(5) Figure 1 suggested sampled sites are very limited. Can the authors justify why these sites were selected rather than other sites?

Response: The sites were selected based on the two important transition zones in these areas. The summergreen-evergreen and the Tundra-Taiga transition zone. The sites included in this dataset are from one very extensive, 2 month long expedition in 2018 where the team traveled from the City of Yakutsk following the only possible summer road in the direction of Mirny and then Lensk. We then followed the bioclimatic gradient leading to the easternmost larch dominated forests and westernmost mixed-species forests in the Lake Khamra region, this expedition part that assessed xxx forest plots took 4 weeks and was very efficient. From the Tundra-Taiga transition zone there is only summer road close to River Kolyma to Bilibino and then the expedition team needed to take helicopters to the mountainous tundra and forest tundra regions of Lake Rauchagytgyn and Lake Ilirney. This part of the expedition also took 4 weeks and was very time intensive with a lot of plots assessed (n= x). The sites were selected beforehand based on satellite imagery using the NDVI maximum summer values and change as well as disturbance products like Hansen et al. forest loss/gain for selecting various accessible sites from the road. This information has now also been added to the paper to make it more clear why we selected these sites. The 2018 two month-long expedition was a very extensive expedition covering a variety of environments with an unusual high number of field plots with successful UAV acquisitions.

(6) I am not sure whether the codes for generating these datasets are available as it is not mentioned explicitly in the manuscript.

Response: The code to generate the Synthetic tree crowns is referred to in the text, Kelley (2019) and the link to the github with the code is in the reference list: GitHub repository: https://github. com/akTwelve/cocosynth, 2019. This code can be downloaded and used to make the synthetic dataset. All code is in Python.

The individual trees and the Sentinel-2 were generated by hand and with free remote sensing and GIS (QGIS) software now put in in chapter 2. The SfM point clouds were made with Agisoft software, described in chapter 2. The tree crown detection was implemented in R which we describe and refer to Brieger et al. (2019) where it is described in much detail.

Dear Reviewer, Thank you very much for your time and effort. The review we received is included below in regular text and the changes that are included in the manuscript and the responses are written below each point in bold and cursive starting with the word response.I, Femke van Geffen would like to, also on behalf of my team, thank the reviewers for their time and detailed comments on the manuscript and dataset. The comments greatly improve the work and are well appreciated. Kind regards, Femke van Geffen.

Review 3:

General comments

The author collected tree - and plot-level forest structure data based on unmanned aerial vehicle and field investigation in two vegetation transition zones of Siberia, Russia. The datasets, including field plot level individual tree and shrub records (tree height, crown diameter, and species) and UAV products (e.g., Canopy Height) can be used for calibration and verification of model output, experiments, or observations. They are useful and important for future carbon dynamic studies and help to inform forest management, especially for the area where historical records and monitoring tend to be scarce. However, the manuscript is poorly organized and difficult to read. I encourage publication after addressing the following issues.

Significance:

The data are useful, complete, and fill in the region field data gap for Siberia boreal forests.

Data quality:

The data are easy to understand and presented readily and accessible to be used in other studies.

Presentation quality:

The manuscript was not logically articulated and was poorly written.

I recommend reframing the Introduction section. It should be more concise. I suggest shortening the detailed SiDroForest dataset and collection method description, but introducing the necessary or implication of each dataset (such as tree-level or plot–level records or canopy Height) and discussing the importance and challenges to collect these data.

Response: We have taken this suggestion into consideration. Please see new structured manuscript with the tracked changes that has been submitted.

The results seem like a duplicate of the method. I suggest including some further information and analysis, not only what the data were included or collected in the four datasets. For example, the frequency or distribution of tree species in the area for 3.2 Dataset 2, and the dominant vegetation classes and their distribution for Dataset 4 can be described.

Response: Thank you for your feedback. In fact, we detected some repetitions and optimized the data and methods and result section. We removed repetitive information and put highlights into the result chapter. Several of the results are visualized in the Appendix, e.g. the distribution of vegetation species in A2, and Percentage vegetation cover per plot in Yakutia for only large shrubs and trees (>1.3m) in A3. The dominant vegetation class of the Sentinel-2 patches are shown in Table 3. Please also see the track change file as an overview on our edits in chapter 2 ‘data and methods and chapter 3 ‘results’.

Specific comments:

Line 68: change “are” to “is”

Lines 106-109: The sentence is difficult to read.

Lines 119-121: delete one of the “use” in the sentence

Line 124: add “which was” before “derived from”

Line 129: change “was” to “were”

Line 150: add”,” before “and”

Response: Above specific comments were all taken into account and corrected for.

Lines 243-244 and Figure A4: Two 30-m-long tape or 15m?

Response: We investigated a circular plot area and used two 30-m long tape measures to segment the circle with a 15 m radius into four quadrants.

L195-200 “In the field, two 30-m-long tape measures were laid out along the main cardinal directions, intersecting in the plot centre, marking the main axes of a circular area with a radius of 15 m. A minimum of ten individuals of each tree and shrub species present were selected per plot. For each individual tree we measured the stem diameter at breast height and at the base. The tree crown diameter, tree height, and vitality were estimated as described in Brieger et al. (2019). There were three deviations from the standard method of vegetation inventory. On plot EN1814 and EN1865, all trees were recorded, and plot EN18070 was recorded by a transect with three segments: edge, transition, and centre.”

Lines 243-248: why were a minimum of ten individuals selected per plot? Why were plots recorded differently? If only part of the trees and shrubs were recorded, the data may not be able to represent the real forest information.

Response: All plots were recorded with the same method. All trees present on the plot area were recorded in the field by its height, species and vitality information. Additionally, a minimum of 10 individuals per species were selected for in-depth analyses and can be found in the Species polygons. Shrub cover per taxa was recorded in the field as well but only a minimum of three individuals selected for in-depth analyses. The UAV-derived Species polygon is supplied for classification tasks and marks a clear tree or shrub from a certain species. It includes a minimum of 10 recorded trees or shrubs per species, where possible the same as recorded in fieldwork.

Line 252 please clarify the 11 vegetation classes here.

Response:

A clarifying sentence is added to L201-204: “We post-fieldwork assigned 11 vegetation classes to the 64 plots (table A1). The class assignment was based on the previous classes determined by Shevtsova et al. (2020a) for Chukotka. For plots in Central Yakutia, we applied a similar method incorporating principal component analysis (PCA), tree density information from the UAV data, and recorded tree species information per plot (Fig. A2, A3 show the field data information).”

Lines 256-259 and Figure A4: The vegetation plot looks smaller than 30m ×30m (the double grid) if the red line is 15m long.

Response: Yes, the double grid in figure A4 is not the 30 m x 30 m vegetation plot, but the UAV double grid flight line set up around the 15 m radius vegetation plot, the blue grid covers around 50x50 m square overflight grid and the orange circular flight mission with a radius of ca. 25 m around the center was carried out with the camera pointing towards the center.

Figure A4: SiDroForest unmanned aerial vehicle (UAV) data acquisition and flight pattern consisting of a double grid (blue) and a circular mission (orange) around the vegetation plot. (Both images are attached with this message)

Figure A1 The two 15 m long grid lines (red) divide the plot area into four quadrants of similar size (yellow). From Brieger et al. (2019).

Line 274: with very low vegetation?

Response: The vegetation that is near to the ground i.e., low structure vegetation. Often this is small shrubs and small trees that are hard to segment because they are close together. We changed the text accordingly to ‘low structure vegetation’

Line 295: add “that was” between “area” and “not”

Lines 301-304: Please break the long sentence into several simple sentences.

Line 326: add “that were” before “corrected”

Response: All three suggestions above were followed.

Lines 338 -340: the sentence is difficult to read. Does that mean the tree crowns were captured by two methods: 1) watershed segmentation analysis and 2) successive automatic generation of a polygon around them?

Response: Yes. I corrected the sentence to be clearer.

Line 542: Link or Linking?

Response: Link.

Lines 553-557: what’s the meaning of this paragraph? Do you compare your field-measured crown diameters and detected crown polygons? If so, what’s the difference between your results and the results of Brieger et al.?

Response: The paragraph introduces that the automatically detected tree crowns are likely better fitting than the field estimations for each tree. We present here the detected tree crowns and did not run an analysis such as in Brieger et al. 2019.