Forests are three-dimensional, but a view from above shows little of what lies below the canopy. LiDAR (Light Detection and Ranging) scans a forest with laser pulses and records where they hit vegetation and ground. The result is a point cloud: millions of points, each with a position and an elevation. To study single trees, for example their height or crown size, we first have to find out which points belong to which tree. This step is called individual-tree segmentation.

In this tutorial, you will use the 3Dtrees tools (Gerberding and others 2026) in Galaxy to prepare one drone LiDAR point cloud, predict individual trees with SegmentAnyTree (SAT) (Wielgosz et al. 2024), and inspect the result in an interactive 3D viewer. The workflow was built by the 3Dtrees community, and the data are part of the THEIA data hub.

Tip: New to Galaxy?

If Galaxy is new to you, first try Galaxy Basics for everyone.

Comment: Which Galaxy server?

This tutorial was prepared on Galaxy Europe, where all 3Dtrees tools are installed. The screenshots come from its Earth System view.

Agenda

In this tutorial, we will cover:

1. Get the data
2. Discover the 3Dtrees workflow
   1. Standardize the point cloud
   2. Get a first impression with overview images
   3. Prepare the data for segmentation
   4. Identify individual trees with SegmentAnyTree
   5. Combine the predictions and transfer them to the original points
   6. Inspect the result in Potree
3. Next steps
4. Conclusion
5. Acknowledgements

Get the data

The data were collected by drone over a Mediterranean oak forest near Pic Saint-Loup in southern France (Alleaume and Jolivot 2021). The dataset has four files, one per flight area. We use only the smallest, Strip-20210804-150144_D.las (476,9 MB). LAS is a common file format for LiDAR points; LAZ is its compressed version.

Details: About the Pic Saint-Loup dataset

The flights took place in June 2021 with a YellowScan Surveyor scanner on a DJI Matrice 600 Pro drone, as part of the SENTHYMED/MEDOAK experiment (Adeline et al. 2024). The dates in the file names record when post-processing finished. The data have about 200–250 points per square metre, use the RGF93 / Lambert-93 coordinate reference system, and are published under the Etalab Open License 2.0.

The four files are separate flight areas, not tiles of one survey: A and B do not touch, while C and D overlap. Processing them together would need extra alignment and overlap checks, which is why this tutorial uses only file D.

Details: What is THEIA?

The dataset is part of THEIA, the data and services hub for continental surfaces of the Data Terra research infrastructure. The THEIA catalogue collects metadata from thematic and institutional repositories, including the INRAE repository that hosts this dataset.

Open image in new tab

Hands-on: Choose Your Own Tutorial

This is a 'Choose Your Own Tutorial' (CYOT) section (also known as 'Choose Your Own Analysis' (CYOA)), where you can select between multiple paths. Click one of the buttons below to select how you want to follow the tutorial

Import the data yourself, or start from a prepared history if the import does not work.

Get data History

Discover the 3Dtrees workflow

The following sections go through the 3Dtrees workflow step by step, from a raw point cloud to individual trees.

Details: What is 3Dtrees.earth?

3Dtrees.earth is a community building a global forest reference: researchers share high-quality 3D forest data to advance science, AI and sustainability. It is part of NFDI4Earth.

Open image in new tab

Standardize the point cloud

Scanners and processing software store point clouds in slightly different ways. The Standardization tool checks the file header, spatial bounds, point density and coordinate reference system, and writes a clean LAZ file that the following tools can rely on. It does not detect trees.

Hands On: Standardize file D

1. 3DTrees: LAS/LAZ Standardization ( Galaxy version 1.2.1+galaxy1) with the following parameters:
   • “Mode”: Single File Standardization
     • param-file “Point Cloud File”: Strip-20210804-150144_D.las

   Open image in new tab
   
   

   The tool returns standardized (the point cloud used from now on), metadata (a JSON summary) and convex_hull (the survey footprint as GeoJSON).

Quiz: Why standardize a point cloud?

Question 1: Which statements about the Standardization tool are correct? Select all that apply.

• It checks file information and prepares a consistent format for the next tools.
• It creates a summary and a footprint of the input.
• It assigns every tree a unique identifier.
• It proves that later segmentation results are accurate.
• It turns unrelated flight areas into one continuous survey.

Check Answers

Solution: Why standardize a point cloud?

For file D, the tool checks for example that the bounding box and point count in the header match the points, that the coordinate reference system (RGF93 / Lambert-93) is valid and in metres, that the points lie on land, and that the file follows the LAS/LAZ format conventions. It stores the results in metadata (point count, bounding box, coordinate reference system and attributes, before and after standardization) and the survey outline in convex_hull. Tree identifiers are not part of this step; they come from the segmentation model.

Get a first impression with overview images

Before predicting trees, look at the data. The Overviews tool draws images that open directly in Galaxy: top views show the extent of the survey and gaps in the canopy, section views show the ground and the vegetation layers above it.

Hands On: Create and inspect overview images

1. 3Dtrees: Overviews ( Galaxy version 1.2.0+galaxy1) with the following parameters:
   • param-file “Point Cloud Collection”: standardized (output of 3DTrees: LAS/LAZ Standardization tool)
2. Open the images in the Top View Images and Section View Images collections (two each).

Open image in new tab

The forest is less than 15 m tall but the survey is about 270 m long, so in the preview the section is a thin strip. Open the image at full size and zoom in. The images below show what you can see: the whole north–south section at true scale, and two close-ups.

Open image in new tab

Open image in new tab

Open image in new tab

Details: How these images were made

They were drawn outside Galaxy from file D. The full section is 10 m wide, like the default Section Width of the Overviews tool; the close-ups use a 4 m slice so that fewer crowns overlap. Colours show height above an estimated ground level.

Quiz: What can the overview tell you?

Question 1: Which overview image is best for spotting gaps in the canopy?

• The top view
• The section (side) view
• Neither of them

Question 2: Which overview image shows how the vegetation is layered above the ground?

• The top view
• The section (side) view
• Neither of them

Question 3: In which close-up do you expect individual trees to be harder to separate?

• Close-up A: taller crowns with gaps between them
• Close-up B: lower crowns that touch each other

Check Answers

Solution: What can the overview tell you?

Gaps are easiest to see from above, layers from the side. Where crowns touch, as in close-up B, there is no gap to mark where one tree ends, so separating trees is harder. Look at both areas again in the final viewer.

Prepare the data for segmentation

SmartTile prepares the point cloud for the model in two ways:

• Tiling: large surveys are cut into tiles, each with an overlapping buffer that gives the model context around trees at the edge. File D is below the default Tiling threshold (10000 MB), so it stays one tile.
• Subsampling: the points are reduced to one per 10 cm voxel (a small cube). SegmentAnyTree works at this resolution; the predictions are transferred back to every original point later.

Hands On: Tile and subsample

1. 3DTrees: SmartTile ( Galaxy version 2.4.1+galaxy0) with the following parameters:
   • “Task”: Tile and subsample
     • param-file “Input LAZ/LAS files”: standardized (output of 3DTrees: LAS/LAZ Standardization tool)
     • “Tiling mode”: Buffered spatial tiles
     • In “Output”:
       • param-repeat “1: Output”
         • “Resolution”: Subsampled
           • “Voxel size (cm)”: 10
         • “Format”: LAZ

   Open image in new tab
   
   

   The outputs are Tiles, output 1 (the subsampled tiles, input for SegmentAnyTree), a Tiling preview map, and Tile bounds JSON (the tile layout, needed for the merge step).

Quiz: Why use overlapping processing tiles?

Question 1: Why is a buffer useful when a tree crosses a processing boundary?

• It gives the model points from beyond the tile edge, so trees near the edge keep their context.
• It removes duplicate trees from the result.
• It reduces the number of points to process.
• It guarantees that every tree is completely inside one tile.

Question 2: What happens when you make the buffer wider? Select all that apply.

• The model gets more context around trees near the tile edges.
• More points are repeated in neighbouring tiles, so processing takes longer.
• The same tree is more likely to be predicted in neighbouring tiles.
• The predictions become correct everywhere.

Check Answers

Solution: Why use overlapping processing tiles?

The buffer gives trees at the edge of a tile their surroundings, but it repeats points: a wider buffer costs more time and makes it more likely that a tree is predicted twice. The merge step later keeps each tree only once.

Identify individual trees with SegmentAnyTree

SegmentAnyTree is a deep-learning model trained on laser scans of many forest types (Wielgosz et al. 2024). It adds two values to each point:

• PredSemantic_SAT, the semantic class: vegetation or non-vegetation (for example ground).
• PredInstance_SAT, the instance: an identifier shared by all points of one predicted tree.

Two neighbouring trees therefore have the same class but different instance identifiers.

Hands On: Predict trees

1. 3Dtrees: SegmentAnyTree ( Galaxy version 1.2.3+galaxy0) with the following parameters:
   • param-collection “Input Dataset”: Tiles, output 1 (output of 3DTrees: SmartTile tool)
   • “Predicted instance dimension name”: PredInstance_SAT
   • “Predicted semantic dimension name”: PredSemantic_SAT

   Select the input with the param-collection Dataset collection button, so that Galaxy runs the model on every tile. The dimension names are the defaults.

   Open image in new tab
   
   

   The job needs a GPU. In our run it waited about an hour in the queue and then ran for about 20 minutes. The output collection is called Processed Files.

Quiz: Instance and semantic predictions

Question 1: Two neighbouring crowns are both predicted as vegetation but have different PredInstance_SAT values. What does that mean?

• They have the same class (vegetation) but are two different individual trees.
• They are the same tree, split into two classes.
• One of them is ground and the other is vegetation.
• They belong to different tree species.

Question 2: Which analyses need the individual-tree instances (PredInstance_SAT)? Select all that apply.

• Predicting the species of each tree
• Measuring the height of each tree
• Counting the trees in the survey area
• Mapping where there is vegetation at all

Check Answers

Solution: Instance and semantic predictions

In this tutorial, the semantic class separates vegetation from non-vegetation, and the instance identifies the individual tree. The instances are what make tree-based studies possible: predicting the species of each tree, measuring tree height or crown size, or counting trees.

Combine the predictions and transfer them to the original points

SmartTile now puts the tiles back together. Using the tile layout, it keeps each tree only in the tile that owns it and joins trees cut by tile borders. With Remap target set to the original file, it then copies the predictions from the 10 cm points to every point of the full-resolution file, which keeps all its original attributes.

Hands On: Merge the predictions

1. 3DTrees: SmartTile ( Galaxy version 2.4.1+galaxy0) with the following parameters:
   • “Task”: Filter / merge segmented tiles
     • param-collection “Segmented tiles”: Processed Files (output of 3Dtrees: SegmentAnyTree tool)
     • “Voxel size of the segmented tiles (cm)”: 10
     • “Model”: SegmentAnyTree, ForestMamba or other point predictions
       • “Instance dimension”: PredInstance_SAT
     • param-file “Tile bounds JSON”: Tile bounds JSON (output of 3DTrees: SmartTile tool)
     • “Remap target”: Original files
       • param-file “Original LAZ/LAS files”: standardized (output of 3DTrees: LAS/LAZ Standardization tool)

   Open image in new tab
   
   

   The result, Originals with predictions, is file D at full resolution with both predictions added. In our run this step took about 50 minutes.

Inspect the result in Potree

Potree shows the point cloud in 3D in your browser, so you can rotate and zoom to check the predicted trees: are there missed trees, trees joined together, or trees split into pieces?

Hands On: Create and open the interactive view

1. 3Dtrees: Potree Converter ( Galaxy version 1.2.1+galaxy0) with the following parameters:
   • param-file “Point Cloud Files”: Originals with predictions (output of 3DTrees: SmartTile tool)
   • “Generate interactive Viewer”: Yes
   • “Special Coloring”: Enabled
     • “Special Coloring Instance Attributes”: PredInstance_SAT
     • “Write Special Coloring JSON”: Yes

   Tree identifiers are just numbers, so neighbouring trees would get almost the same colour. Special Coloring adds coloring_id_sat, which gives nearby trees clearly different colours.

2. Open the Potree Visualization output with its galaxy-eye eye icon. Loading can take a minute. Each predicted tree has its own colour; points that belong to no tree are grey.

   Open image in new tab
   
   

3. Look at a few neighbouring trees from above and from the side, for example around close-ups A and B. To see the vegetation/non-vegetation classes instead, open Scene in the menu on the left, click the point cloud potree-vis, and set Attribute to PredSemantic_SAT.

   Open image in new tab
   
   

Quiz: Does the segmentation look plausible?

Question 1: Which observations support a plausible segmentation? Select all that apply.

• Most instance colours follow recognisable tree crowns rather than scattered patches.
• Neighbouring, visibly separate crowns usually have different colours.
• Visible ground points are mostly in the non-vegetation class of PredSemantic_SAT.
• Many crowns are cut exactly along straight processing boundaries.
• One colour spans many unrelated crowns.

Question 2: Why is this only a visual check and not an accuracy assessment? Select all that apply.

• A result can look plausible and still contain missed, merged or split trees.
• We did not compare the predictions with independently labelled reference trees.
• Measuring accuracy needs reference labels and a method to compare them with predictions.
• Potree proves that every displayed tree is correct.

Check Answers

Solution: Does the segmentation look plausible?

Colours that follow single crowns and differ between neighbours, and ground in the non-vegetation class, suggest a good result. Straight cuts along tile borders or one colour across several crowns suggest errors. A plausible picture is not proof, though: measuring accuracy needs trees that were labelled independently, for example in the field.

Next steps

The tree labels are a starting point for measuring tree height, crown size or tree density. Other 3Dtrees tools in Galaxy go further:

• 3Dtrees: ForestMamba ( Galaxy version 1.0.0+galaxy0) is an alternative tree segmentation model that also separates leaves from wood.
• 3Dtrees: DetailView ( Galaxy version 1.1.1+galaxy0) predicts the species of segmented trees, within the classes it was trained on.

Conclusion

You turned one drone LiDAR file into a point cloud in which every point carries a predicted class and, for tree points, a tree identifier, and you checked the result in 3D. The predictions depend on forest type, point density and survey method, so always inspect them before using them for ecological analysis.

Acknowledgements

This work is part of the collaboration between the EOSC node Data Terra and NFDI4Earth.

Open image in new tab

Open image in new tab

Open image in new tab

You've Finished the Tutorial

Please also consider filling out the Feedback Form as well!

Key points

• LiDAR records the 3D structure of vegetation and terrain

• SegmentAnyTree separates vegetation from non-vegetation points and assigns an ID to each predicted tree

• Visual inspection can reveal obvious problems but is not a quantitative accuracy assessment

Frequently Asked Questions

• Tutorial FAQs
• Topic FAQs
• Galaxy FAQs
• GTN Matrix Chat
• Galaxy Help Forum

References

1. Alleaume, S., and A. Jolivot, 2021 Données lidar acquises par drone au pied du Pic St Loup (34). Etalab Open License 2.0. 10.15454/DMYWPB
2. Adeline, K., J. B. Féret, H. Clenet, J. M. Limousin, J. M. Ourcival et al., 2024 Multi-scale datasets for monitoring Mediterranean oak forests from optical remote sensing during the SENTHYMED/MEDOAK experiment in the north of Montpellier (France). Data in Brief 53: 110185. 10.1016/j.dib.2024.110185
3. Wielgosz, M., S. Puliti, B. Xiang, K. Schindler, and R. Astrup, 2024 SegmentAnyTree: A sensor and platform agnostic deep learning model for tree segmentation using laser scanning data. Remote Sensing of Environment 313: 114367. 10.1016/j.rse.2024.114367
4. Gerberding, K., and others, 2026 3Dtrees: title to be added on publication. Manuscript in preparation. Placeholder: replace with the published reference.. https://3dtrees.earth

Feedback

Did you use this material as an instructor? Feel free to give us feedback on how it went.
Did you use this material as a learner or student? Click the form below to leave feedback.

Citing this Tutorial

1. Kilian Gerberding, Marie Josse, 3Dtrees: From LiDAR point clouds to individual trees (Galaxy Training Materials). https://training.galaxyproject.org/training-material/topics/climate/tutorials/3dtrees/tutorial.html Online; accessed TODAY
2. Hiltemann, Saskia, Rasche, Helena et al., 2023 Galaxy Training: A Powerful Framework for Teaching! PLOS Computational Biology 10.1371/journal.pcbi.1010752
3. Batut et al., 2018 Community-Driven Data Analysis Training for Biology Cell Systems 10.1016/j.cels.2018.05.012

BibTeX

@misc{climate-3dtrees,
author = "Kilian Gerberding and Marie Josse",
	title = "3Dtrees: From LiDAR point clouds to individual trees (Galaxy Training Materials)",
	year = "",
	month = "",
	day = "",
	url = "\url{https://training.galaxyproject.org/training-material/topics/climate/tutorials/3dtrees/tutorial.html}",
	note = "[Online; accessed TODAY]"
}
@article{Hiltemann_2023,
	doi = {10.1371/journal.pcbi.1010752},
	url = {https://doi.org/10.1371%2Fjournal.pcbi.1010752},
	year = 2023,
	month = {jan},
	publisher = {Public Library of Science ({PLoS})},
	volume = {19},
	number = {1},
	pages = {e1010752},
	author = {Saskia Hiltemann and Helena Rasche and Simon Gladman and Hans-Rudolf Hotz and Delphine Larivi{\`{e}}re and Daniel Blankenberg and Pratik D. Jagtap and Thomas Wollmann and Anthony Bretaudeau and Nadia Gou{\'{e}} and Timothy J. Griffin and Coline Royaux and Yvan Le Bras and Subina Mehta and Anna Syme and Frederik Coppens and Bert Droesbeke and Nicola Soranzo and Wendi Bacon and Fotis Psomopoulos and Crist{\'{o}}bal Gallardo-Alba and John Davis and Melanie Christine Föll and Matthias Fahrner and Maria A. Doyle and Beatriz Serrano-Solano and Anne Claire Fouilloux and Peter van Heusden and Wolfgang Maier and Dave Clements and Florian Heyl and Björn Grüning and B{\'{e}}r{\'{e}}nice Batut and},
	editor = {Francis Ouellette},
	title = {Galaxy Training: A powerful framework for teaching!},
	journal = {PLoS Comput Biol}
}

                   

Funding

These individuals or organisations provided funding support for the development of this resource

DATA TERRA

Galaxy Administrators: Install the missing tools

You can use Ephemeris's shed-tools install command to install the tools used in this tutorial.

shed-tools install [-g GALAXY] [-a API_KEY] -t <(curl https://training.galaxyproject.org/training-material/api/topics/climate/tutorials/3dtrees/tutorial.json | jq .admin_install_yaml -r)

Alternatively you can copy and paste the following YAML

---
install_tool_dependencies: true
install_repository_dependencies: true
install_resolver_dependencies: true
tools:
- name: 3dtrees_detailview
  owner: bgruening
  revisions: 42e755abe831
  tool_panel_section_label: GIS Data Handling
  tool_shed_url: https://toolshed.g2.bx.psu.edu/
- name: 3dtrees_forestmamba
  owner: bgruening
  revisions: ab9b0561b285
  tool_panel_section_label: GIS Data Handling
  tool_shed_url: https://toolshed.g2.bx.psu.edu/
- name: 3dtrees_overviews
  owner: bgruening
  revisions: '049c348797e9'
  tool_panel_section_label: GIS Data Handling
  tool_shed_url: https://toolshed.g2.bx.psu.edu/
- name: 3dtrees_potree
  owner: bgruening
  revisions: 0c99fafe841d
  tool_panel_section_label: GIS Data Handling
  tool_shed_url: https://toolshed.g2.bx.psu.edu/
- name: 3dtrees_segmentanytree
  owner: bgruening
  revisions: 2428651639f8
  tool_panel_section_label: GIS Data Handling
  tool_shed_url: https://toolshed.g2.bx.psu.edu/
- name: 3dtrees_smart_tile
  owner: bgruening
  revisions: b3279a32a87b
  tool_panel_section_label: GIS Data Handling
  tool_shed_url: https://toolshed.g2.bx.psu.edu/
- name: 3dtrees_standardization
  owner: bgruening
  revisions: e36c4b5d5cd9
  tool_panel_section_label: GIS Data Handling
  tool_shed_url: https://toolshed.g2.bx.psu.edu/
- name: 3dtrees_standardization
  owner: bgruening
  revisions: a3f2cd19d45e
  tool_panel_section_label: GIS Data Handling
  tool_shed_url: https://toolshed.g2.bx.psu.edu/

Feedback

No feedback has been recieved yet for this training. Be the first one by filling in the feedback form.