Datasets¶

There are three datasets within the MillionTrees package: TreeBoxes, TreePoints, and TreePolygons. These datasets contain many source datasets from dozens of papers and research efforts. Below, each source is briefly described. The images for each dataset are generated directly from the dataloaders to allow rapid verification of the annotation status and are regenerated automatically when a new dataset is released or updated. Sample images within a dataset are randomly selected to ensure transparency. This does mean that some images are sparsely annotated compared to the rest of the images in a source.

Note: The datasets below are processed and will be part of the final release. The current release is pre-release and not final. Only publicly available datasets are included at this time.

Dataset Filtering and Management¶

MillionTrees datasets can contain millions of annotations. Use filtering capabilities to manage dataset size and preview data before downloading.

Source Filtering¶

List and filter available sources:

from milliontrees.datasets.TreePoints import TreePointsDataset
dataset = TreePointsDataset()
sources = dataset.sources
print("Available sources:", sources)
# Available sources: ['Kattenborn_NewZealand', 'NeonTreeEvaluation', 'OFO_unsupervised', 'NEON_unsupervised'...]

Include only specific sources:

dataset = TreePointsDataset(
  include_sources=['Amirkolaee et al. 2023']
)

Exclude specific sources (exact names or glob patterns supported):

# Exclude a single source by name
dataset = TreePointsDataset(exclude_sources=['NEON_unsupervised'])

# Exclude by pattern (wildcards)
dataset = TreePointsDataset(exclude_sources=['*_unsupervised'])

Release sizes (mini, small, full)¶

Each geometry dataset is published in three sizes:

mini (mini=True): up to 3 images per source; random split only — fastest smoke tests.
small (small=True): up to 50 images per source; all split schemes (random, zeroshot, crossgeometry).
full (default): complete packaged release.

dataset = TreePointsDataset(mini=True)
dataset = TreeBoxesDataset(small=True, split_scheme="zeroshot")

Do not pass mini=True and small=True together.

Boxes¶

Dumortier 2025¶

sample_image

Citation: Dumortier, J. (2025). Annotated tree crown bounding boxes in urban/rural environment [Data set]. Zenodo. https://doi.org/10.5281/zenodo.15155081

Link: https://zenodo.org/records/15155081

Kaggle Palm Counting¶

Source Name: “Kaggle_Palm_Counting”¶

https://www.kaggle.com/datasets/praneethsaikolla/palm-tree-detection-dataset

sample_image

Kwon et al. 2023¶

sample_image

Citation: Ryoungseob Kwon, Youngryel Ryu, Tackang Yang, Zilong Zhong, Jungho Im,
Merging multiple sensing platforms and deep learning empowers individual tree mapping and species detection at the city scale,
ISPRS Journal of Photogrammetry and Remote Sensing, Volume 206, 2023

Location: Suwon, South Korea

Radogoshi et al. 2021¶

sample_image

Link: https://lila.science/datasets/forest-damages-larch-casebearer/

Dataset citation: Swedish Forest Agency (2021): Forest Damages – Larch Casebearer 1.0. National Forest Data Lab. Dataset.

Location: Sweden

Sun et al. 2022¶

sample_image

Link: https://www.sciencedirect.com/science/article/pii/S030324342100369X

Santos et al. 2019¶

sample_image

Link: Dataset Ninja

Citation

@Article{s19163595, AUTHOR = {Santos, Anderson Aparecido dos and Marcato Junior, José and Araújo, Márcio Santos and Di Martini, David Robledo and Tetila, Everton Castelão and Siqueira, Henrique Lopes and Aoki, Camila and Eltner, Anette and Matsubara, Edson Takashi and Pistori, Hemerson and Feitosa, Raul Queiroz and Liesenberg, Veraldo and Gonçalves, Wesley Nunes}, TITLE = {Assessment of CNN-Based Methods for Individual Tree Detection on Images Captured by RGB Cameras Attached to UAVs}, JOURNAL = {Sensors}, VOLUME = {19}, YEAR = {2019}, NUMBER = {16}, ARTICLE-NUMBER = {3595}, URL = {https://www.mdpi.com/1424-8220/19/16/3595}, DOI = {10.3390/s19163595} }

Location: Barro Colorado Island, Panama

Velasquez-Camacho et al. 2023¶

sample_image

Link: https://zenodo.org/records/10246449

Location: Spain

Weinstein et al. 2021¶

sample_image

Link: https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1009180

Location: NEON sites within the United States

An extension of this published resource was made by the Weecology Lab at the University of Florida

sample_image

World Resources Institute¶

NAIP Imagery from across the United States

sample_image

SelvaBox (CanopyRS)¶

sample_image

Citation: SelvaBox: A high-resolution dataset for tropical tree crown detection in dense canopies
Link: https://huggingface.co/datasets/CanopyRS/SelvaBox
Location: Brazil, Ecuador, and Panama
Description: High-resolution dataset (4.5 cm GSD) with over 83,000 human bounding box annotations for tropical tree crowns in dense canopies. The dataset comprises 14 rasters from three different countries with varying spatial extents.

Veitch-Michaelis et al. 2024.¶

sample_image

Link: HuggingFace Dataset

https://zenodo.org/records/11617167

Location: Global

Citation: Veitch-Michaelis, J., Cottam, A., Schweizer, D., Broadbent, E., Dao, D., Zhang, C., Almeyda Zambrano, A., & Max, S. (2024). OAM-TCD: A globally diverse dataset of high-resolution tree cover maps (1.0.0) [Data set]. Zenodo.

For more information about the dataset collation, see: Veitch-Michaelis, J. et al. “OAM-TCD: A globally diverse dataset of high-resolution tree cover maps.” Advances in neural information processing systems 37 (2024): 49749-49767.

Zamboni et al. 2021¶

sample_image

Link: https://github.com/pedrozamboni/individual_urban_tree_crown_detection

Location: Mato Grosso do Sul, Brazil

Puliti and Astrup 2022¶

Source Name: “Puliti and Astrup 2022”¶

sample_image

Link: https://zenodo.org/records/14711562

Location: Norway

NIBIO UAV tree damage dataset. YOLO-format bounding boxes for individual trees over Norwegian forest sites; all classes are combined as Tree for the MillionTrees benchmark.

Reiersen et al. 2022¶

Source Name: “Reiersen et al. 2022”¶

sample_image

Location: Ecuador

ReForestTree dataset. Bounding boxes were generated by running DeepForest and then human-cleaned, so this source is used for training only and not for evaluation.

Šrollerů et al. 2025¶

Source Name: “Šrollerů et al. 2025”¶

sample_image

Link: https://zenodo.org/records/15591546

Location: Krkonoše Mountains (Bílé Labe Valley), Czechia

UAV, aerial, and terrestrial 3D point clouds from a treeline ecotone with tree-level reference ground-based measurements across three plots (low/mid/high upper-forest limit). LiDAR-derived crown bounding boxes are projected into the matching orthoimagery for MillionTrees.

Allen et al. 2025¶

Source Name: “Allen et al. 2025”¶

sample_image

Citation: Allen, M.J., Owen, H.J.F., Grieve, S.W.D., & Lines, E.R. Manual Labelling Artificially Inflates Deep Learning-Based Segmentation Performance on RGB Images of Closed Canopy: Validation Using TLS. Remote Sensing of Environment (in press). https://arxiv.org/pdf/2503.14273

Location: Joensuu, Finland (boreal) and Alto Tajo, Spain (Mediterranean)

Axis-aligned bounding boxes derived from TLS crown footprints projected into drone orthomosaics. All rows are assigned to the validation split — held out for independent post-hoc evaluation, not hyperparameter tuning.

Points¶

Amirkolaee et al. 2023¶

sample_image

Citation: Amirkolaee, Hamed Amini, Miaojing Shi, and Mark Mulligan. “TreeFormer: A semi-supervised transformer-based framework for tree counting from a single high-resolution image.” IEEE Transactions on Geoscience and Remote Sensing 61 (2023): 1–15.

Link: https://github.com/HAAClassic/TreeFormer

Location: London, England

GSD: 0.2m

Dubrovin et al. 2024¶

sample_image

Link: Kaggle Dataset

UAV orthomosaics with accompanying field-survey tree locations. We include the field survey points as per-image point annotations for individual trees.

Location: Perm Krai, Russia

Citation: Dubrovin, Ivan, Clement Fortin, and Alexander Kedrov. “An open dataset for individual tree detection in UAV LiDAR point clouds and RGB orthophotos in dense mixed forests.” Scientific Reports 14.1 (2024): 21938.

Link (article): https://www.nature.com/articles/s41598-024-72669-5

DOI: https://doi.org/10.1038/s41598-024-72669-5

Ventura et al. 2022¶

sample_image

Citation: J. Ventura, C. Pawlak, M. Honsberger, C. Gonsalves, J. Rice, N.L.R. Love, S. Han, V. Nguyen, K. Sugano, J. Doremus, G.A. Fricker, J. Yost, and M. Ritter.
Individual Tree Detection in Large-Scale Urban Environments using High-Resolution Multispectral Imagery.
International Journal of Applied Earth Observation and Geoinformation, 130, 103848 (2024)

Link: https://github.com/jonathanventura/urban-tree-detection-data

Location: Southern California, United States

GSD: 0.6m

National Ecological Observatory Network¶

sample_image

Location: Multiple sites across the United States, see NEON Field Sites

Link: https://data.neonscience.org/data-products/DP1.10098.001

GSD: 0.1m

Auto-arborist¶

The auto-arborist dataset is a compilation of street-level surveys performed by local cities. Street trees were labeled with points. Annotations are limited to street trees, and do not include non-street trees. MillionTrees datasets are associated with state orthophoto programs that vary in ground resolution, from 20cm to 60cm.

https://google.github.io/auto-arborist/

Citation: Sara Beery, Guanhang Wu, Trevor Edwards, Filip Pavetic, Bo Majewski, Shreyasee Mukherjee, Stanley Chan, John Morgan, Vivek Rathod, Jonathan Huang; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 21294-21307

GSD: 0.2m - 0.6m

Chen and Shang 2022¶

Source Name: “Chen & Shang (2022)”¶

sample_image

Location: Yosemite National Park, California, USA

Density-map annotations from the Yosemite tree counting dataset, converted to per-tree points via morphological centroid extraction.

OSBS megaplot 2025¶

Source Name: “OSBS megaplot 2025”¶

sample_image

Location: Ordway-Swisher Biological Station (NEON OSBS site), Florida, USA

Field-mapped visible-tree stems from the 2025 OSBS megaplot survey, paired with the OSBS 2025 RGB mosaic and split into uniform image tiles.

Young et al. 2025¶

Source Name: “OFO field 2025”¶

sample_image

Link: Open Forest Observatory

Field-validated stem maps from the Open Forest Observatory ground reference catalog. David Young (UC Davis) shared a concatenated set of 330 plot-drone pairs in which each tree has been manually mapped on the ground and the plot has been algorithmically registered to a specific drone mission’s photogrammetry products. The field trees are duplicated per overlapping mission with a mission_id attribute denoting which orthomosaic each row should be paired with, and a withhold_from_training flag that routes those trees to the test split.

The dataset is built by data_prep/process_ofo_field.py, which downloads the matching missions_03 orthomosaic for each mission_id from https://js2.jetstream-cloud.org:8001/swift/v1/ofo-public/drone/missions_03/{mission_id}/photogrammetry_03/full/{mission_id}_ortho-dsm-ptcloud.tif, tiles it into uniform patches, and projects field-tree points into image coordinates. Trees that are not flagged predicted_overstory (overhead visible) are dropped by default since field stems under the canopy cannot be observed from nadir drone imagery.

Location Forests across California (Sierra Nevada and Lake Tahoe basin), USA

Allen et al. 2025¶

Source Name: “Allen et al. 2025”¶

sample_image

Location: Joensuu, Finland (boreal) and Alto Tajo, Spain (Mediterranean)

Tree-top points at polygon centroids from TLS-derived crown footprints. All rows use the validation split.

Polygons¶

Araujo et al. 2020¶

sample_image

Link: https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0243079

Location: Manuas, Brazil

Ball et al. 2023¶

sample_image

Link: https://zenodo.org/records/8136161

Location: Danum, Malaysia

Bohlman 2008¶

sample_image

Unpublished data from Barro Colorado Island, field verified after photo-interpretation.

Both crowns and points are available.

Location: BCI, Panama

Cloutier et al. 2023¶

Source Name: “Cloutier et al. 2023”¶

sample_image

Link: https://zenodo.org/records/8148479

Location: Quebec, Canada

Feng et al. 2025¶

Source Name: ‘Feng et al. 2025’¶

sample_image

Citation: Feng, Zhengpeng, Yihang She, and Srinivasan Keshav. “SPREAD: A large-scale, high-fidelity synthetic dataset for multiple forest vision tasks.” Ecological Informatics 87 (2025): 103085.

Link: Zenodo

Location: Simulation

Note this dataset is simulated and therefore somewhat more controvertial in its inclusion. We believe the size, diversity and realism of the images merit inclusion.

Firoze et al. 2023¶

Source Name: “Firoze et al. 2023”¶

sample_image

Link: https://openaccess.thecvf.com/content/CVPR2023/papers/Firoze_Tree_Instance_Segmentation_With_Temporal_Contour_Graph_CVPR_2023_paper.pdf

Location: Indiana, United States

Hickman et al. 2021¶

sample_image

Link: https://zenodo.org/records/5515408

Location: Sabah, Malaysia

Jansen et al. 2023¶

sample_image

Link: https://zenodo.org/records/7094916

Location: Northern Australia

JustDiggit¶

Source Name: “Justdiggit 2023”¶

sample_image

Link: JustDigIt

Location: Tanzania

Citation status uncertain, contact Tjomme Dooper fruit punch AI.

Li et al. 2023¶

sample_image

link Manuscript github

Location: Denmark

Lefebvre et al. 2024¶

Source Name: “Lefebvre et al. 2024”¶

sample_image

Dataset Link

Lucas et al. 2024¶

Source Name: “Lucas et al. 2024”¶

sample_image

Link: Manuscript

Citation Lucas, Moritz, et al. “Individual tree detection and crown delineation in the Harz National Park from 2009 to 2022 using mask R–CNN and aerial imagery.” ISPRS Open Journal of Photogrammetry and Remote Sensing 13 (2024): 100071.

Kattenborn et al. 2023¶

sample_image

Link: https://zslpublications.onlinelibrary.wiley.com/doi/full/10.1002/rse2.146

Location: New Zealand

Miranda et al. 2024¶

sample_image

Link: Courtesy of Alejandro Miranda

Safonova et al. 2021¶

sample_image

Link: https://www.mdpi.com/1424-8220/21/5/1617

Location: Spain

Takeshige et al. 2025¶

Source Name: “Takeshige et al. 2025”¶

sample_image

Link: https://esj-journals.onlinelibrary.wiley.com/doi/10.1111/1440-1703.12555

Location: Japan

Troles et al. 2024¶

Source Name: “Troles et al. 2024”¶

sample_image

Citation: Troles, J.; Schmid, U.; Fan, W.; Tian, J. BAMFORESTS: Bamberg Benchmark Forest Dataset of Individual Tree Crowns in Very-High-Resolution UAV Images. Remote Sens. 2024, 16, 1935. https://doi.org/10.3390/rs16111935

Link: https://www.mdpi.com/2072-4292/16/11/1935

Location: Bamberg, Germany

Wagner et al. 2023¶

sample_image

Link: https://www.mdpi.com/2504-446X/7/3/155
https://www.mdpi.com/2072-4292/16/11/1935

Location: Australia

Vasquez et al. 2023¶

sample_image

Link: Figshare

Location: Barro Colorado Island, Panama

There is also a training-only portion of this that was used in conjuction with a model to predict labels that were then verified.

Allen et al. 2025¶

Source Name: “Allen et al. 2025”¶

sample_image

Location: Joensuu, Finland (boreal) and Alto Tajo, Spain (Mediterranean)

Crown polygons delineated from co-located terrestrial LiDAR (TLS) following the pipeline in Allen et al. (2025). Orthomosaics are tiled for MillionTrees packaging. This source is validation-only: use after training and standard test evaluation for independent TLS ground-truth assessment. Do not use for hyperparameter tuning.

Khan et al. 2026¶

Source Name: “Khan et al. 2026”¶

sample_image

Link: https://zenodo.org/records/19695972

DOI: https://doi.org/10.5281/zenodo.19695972

Location: Halle (Saale), Germany

DeepTrees-Halle: per-tile polygon shapefiles over the city of Halle. Classes are 0 (tree), 1 (cluster of trees), and 2 (unsure); MillionTrees retains only individual trees (class 0).

NEON combined crowns¶

Source Name: “NEON combined crowns”¶

sample_image

Location: Harvard Forest (NEON HARV site), Massachusetts, USA

Field-mapped crown polygons (combined from multiple campaigns at Harvard Forest) paired with NEON AOP high-resolution camera imagery (DP3.30010.001). The crown polygons are merged into a single GeoPackage, draped over the AOP mosaic, and tiled for MillionTrees.

NEON MultiTemporal¶

Source Name: “NEON MultiTemporal”¶

sample_image

Location: Multiple NEON sites across the United States, multiple flight years

Per-plot bounding box, point, and polygon annotations over NEON AOP camera imagery ({site}_{plot}_{year}.tif) for multi-year evaluation. All rows are assigned existing_split="test" and are used only for evaluation across geometry types.

Schütte et al. 2025¶

Source Name: “Schütte et al. 2025”¶

sample_image

Location: Berlin and Osnabrück, Germany

Urban tree crown polygons from the ITCD Urban Berlin/Osnabrück dataset, drawn over 20 cm DOP false-color orthoimagery and tiled for MillionTrees.

Zuniga-Gonzalez et al. 2023¶

Source Name: “Zuniga-Gonzalez et al. 2023”¶

sample_image

Location: London / Cambridge, United Kingdom

UrbanLondon urban tree crown polygon dataset at 0.25 m GSD with an existing train/test split preserved in MillionTrees.

Unsupervised¶

Weinstein et al. 2018¶

Coregistered LIDAR and RGB were acquired over 27 sites in the National Ecological Observation Network, USA. These sites cover a range of forest habitats. A weakly supervised LiDAR tree detection algorithm was used to predict tree locations. These locatons were draped over the RGB data to create a very large weakly supervised dataset. There is currently over 40 million tree locations from the original dataset and more can be generated with ongoing data collection.

Citation: Weinstein, B.G.; Marconi, S.; Bohlman, S.; Zare, A.; White, E. Individual Tree-Crown Detection in RGB Imagery Using Semi-Supervised Deep Learning Neural Networks. Remote Sens. 2019, 11, 1309. https://doi.org/10.3390/rs11111309

Location: Forest across the United States (NEON)

sample_image

Open Forest Observatory¶

High resolution drone imagery used to create photogrametry-derived predictions of tree crowns.

https://openforestobservatory.org/

https://besjournals.onlinelibrary.wiley.com/doi/10.1111/2041-210X.13860

Source Name: “Young et al. 2025 unsupervised”¶

Location Forest across the United States

sample_image