Global coverage
Crowd-sourced orthoimagery from the deadtrees.earth platform, with joint labels for living tree cover and tree mortality.
Benchmark sites by biome group
Training and validation sites by biome group
Get the data
| Dataset | Content | Size | Licence | |
|---|---|---|---|---|
| DTE-aerial-bench | 525 patches · 25 sites · expert labels · 5/10/20 cm | 2 GB | CC BY 4.0 | Hugging Face |
| DTE-aerial-train (CC BY) | 308,651 training + 34,885 validation patches · by biome and resolution | 0.98 TB | CC BY 4.0 (+ 353 MIT patches) | Hugging Face |
| DTE-aerial-train (CC BY-NC-SA) | 2,090 training patches · 5 sites · non-commercial | 6 GB | CC BY-NC-SA 4.0 | Hugging Face |
| Model | MiT-B3 trained on DTE-aerial-train | 0.2 GB | CC BY 4.0 | Hugging Face |
The training set used in the paper is the union of both DTE-aerial-train repositories.
from huggingface_hub import snapshot_download
# metadata for all patches, plus all boreal training shards (WebDataset: <key>.tif, <key>.mask.tif, <key>.json)
snapshot_download("ayushi3536/deadtree.earth-aerial-train_cc_by", repo_type="dataset",
allow_patterns=["metadata/*", "data/train/boreal/*"])
# mask values: 0 background, 1 tree cover, 2 mortality, 255 outside the area of interest
DTE-aerial-bench
Experts label each DTE-aerial-bench site at 5 cm as 16 tiles of 1024 × 1024 pixels. Labels at 10 cm (4 tiles) and 20 cm (1 tile) are derived from the 5 cm labels by downsampling. Every resolution therefore covers exactly the same area with consistent labels, and resolution is the only factor that changes.
Image
Labels
Benchmark site 4181 (Temperate, Chile): the same 204.8 m × 204.8 m area at each resolution. Every tile has 1024 × 1024 pixels, so the site is 16 tiles at 5 cm, 4 at 10 cm and 1 at 20 cm. Groups of four 5 cm tiles are stitched and downsampled into one 10 cm tile, and the four 10 cm tiles into the single 20 cm tile. Click a step to jump to it. Labels: tree cover, mortality, background.
25 sites × (16 + 4 + 1) tiles = 525 benchmark patches.
DTE-aerial-bench
Experts label each benchmark patch completely: living tree cover, and mortality from partial crown dieback to fully dead crowns. Mortality can therefore be measured as a share of the forested area. Choose a biome group and drag the slider to compare image and labels.
DTE-aerial-bench
MiT-B3 trained on DTE-aerial-train (mean of three seeds) against the previous deadtrees.earth model DT-V1 for mortality and against OAM-TCD for tree cover. Scores are averaged over the benchmark sites.
Acknowledgements
DTE-aerial is built from imagery shared on
deadtrees.earthan open platform for sharing aerial imagery of forests and mapping tree mortality worldwide.
Visit deadtrees.earth →Thank you to all data contributors, credited per site in the dataset metadata.
Supported by the European Space Agency (DeepFeatures, FORTRACK) and by the German Aerospace Center on behalf of BMWK (ML4Earth).
Citation
If you use DTE-aerial-bench, DTE-aerial-train or the model, please cite:
@misc{sharma2026dteaerial,
title = {deadtrees.earth-aerial: A Multi-Resolution Aerial Image Dataset
for Tree Cover and Mortality Detection},
author = {Sharma, Ayushi and Mosig, Clemens and Drees, Lukas and Soltani, Salim
and Vajna-Jehle, Janusch and Sheppard, Aaron and Ahmadi, Belqis
and Schmid, Jonathan and Neumeier, Paul and Jacobs, Nathan
and Wegner, Jan Dirk and Kattenborn, Teja},
year = {2026},
eprint = {2605.19605},
archivePrefix = {arXiv},
doi = {10.48550/arXiv.2605.19605},
url = {https://arxiv.org/abs/2605.19605}
}