ayushi3536/deadtree.earth-aerial-bench

deadtrees.earth-aerial: DTE-aerial-bench

A Multi-Resolution Aerial Image Dataset for Tree Cover and Mortality Detection

📄 Paper: arXiv:2605.19605 · 💻 Code: GeoSense-Freiburg/DTE-aerial · 🌍 Platform: deadtrees.earth · 🔎 Interactive viewer: deadtrees.earth/releases/dte-aerial-bench

This dataset is part of the paper deadtrees.earth-aerial: A Multi-Resolution Aerial Image Dataset for Tree Cover and Mortality Detection (Sharma et al., 2026). If you use it, please cite the paper (see Citation).

Overview

DTE-aerial-bench is an expert-annotated benchmark for segmenting tree cover and tree mortality (visible dieback and dead crowns) in aerial RGB imagery, built for evaluation across biomes and spatial resolutions. It contains 525 patches of 1024 × 1024 pixels from 25 sites (orthophotos) shared on the crowd-sourced deadtrees.earth platform: 11 temperate, 6 dryland (Mediterranean), 4 (sub)tropical, and 4 boreal and montane.

Experts labelled every site exhaustively at 5 cm, with iterative multi-reviewer consensus (up to five rounds per site). The 10 cm and 20 cm labels are derived from the 5 cm labels by downsampling, so all three resolutions cover exactly the same area:

Resolution	Patches per site	Patches in total
5 cm	16	400
10 cm	4	100
20 cm	1	25

These are exactly the data used for all benchmark results in the paper.

Relation to other repositories

DTE-aerial-bench is the test split of deadtrees.earth-aerial. None of its 25 orthophotos occurs in the training or validation data, which are published separately:

• ayushi3536/deadtree.earth-aerial-train_cc_by (CC BY 4.0, plus a few MIT-licensed orthophotos)
• ayushi3536/deadtree.earth-aerial-train_cc_by_nc-sa (CC BY-NC-SA 4.0)

The model trained on DTE-aerial-train (MiT-B3) is available at ayushi3536/DTE-aerial-model.

Labels

Value	Class
0	background
1	tree cover
2	tree mortality (visible mortality symptoms / crown decline)

The benchmark masks contain only these three values (no void pixels). The mortality class describes visible symptoms in RGB imagery as interpreted by expert annotators. It does not measure physiological tree death.

Dataset structure

File	Size	Content
DTE-aerial-bench.parquet	1.0 GB	the 525 patches and masks with their metadata in one table (shown in the dataset viewer)
DTE-aerial-bench-meta.csv	0.1 MB	one row per patch (metadata)
DTE-aerial-bench-tiles.tar	1.0 GB	RGB patches, tiles/<id>/<name>.png
DTE-aerial-bench-masks.tar	9 MB	label masks, masks/<id>/<name>_mask.png
SHA256SUMS		checksums
croissant.json		Croissant metadata with Responsible AI fields

There are two formats with identical content:

• The CSV and the two tar archives are the format read by the evaluation code. Use them to reproduce the paper.
• The Parquet file holds the same patches and masks, with the metadata, for quick loading in Python and the dataset viewer.

Patches are 1024 × 1024 RGB PNGs (uint8); masks are 1024 × 1024 single-band uint8 PNGs aligned pixel by pixel with their patch. File names end in the resolution, e.g. 1371_0_0_5cm.png, 1371_1762700398977_0_10cm.png and 1371_20_1762700398977_20cm.png.

Metadata fields

Field	Description
id	orthophoto (site) id on deadtrees.earth
tile_path, mask_path	relative paths of the patch and mask (member paths in the tar archives)
resolution	5cm, 10cm or 20cm
lat, lon	orthophoto centroid (WGS84)
biome	Olson et al. (2001) biome of the orthophoto centroid
license, contributors, citation_doi	licence, data contributors and source DOI or link (where available)
orthophoto_res	native resolution of the source orthophoto (m)
acquisition_date_year, _month, _day	acquisition date

The Parquet file additionally has key (patch file name without extension), image and mask ({bytes, path}), and frac_background, frac_tree_cover, frac_mortality (class fractions of the mask; frac_ignored is always 0).

Usage

Load with 🤗 Datasets:

from datasets import load_dataset

ds = load_dataset("ayushi3536/deadtree.earth-aerial-bench", split="test")
sample = ds[0]   # sample["image"], sample["mask"] (PIL images), sample["resolution"], sample["biome"], ...

Download the raw files and reproduce the paper's evaluation:

huggingface-cli download ayushi3536/deadtree.earth-aerial-bench --repo-type dataset \
    --include "DTE-aerial-bench-*" SHA256SUMS --local-dir DTE-aerial-bench
cd DTE-aerial-bench
sha256sum -c SHA256SUMS --ignore-missing
tar -xf DTE-aerial-bench-tiles.tar
tar -xf DTE-aerial-bench-masks.tar

This gives the layout expected by the evaluation code:

DTE-aerial-bench/
├── DTE-aerial-bench-meta.csv
├── tiles/<id>/<name>.png
└── masks/<id>/<name>_mask.png

Then set data.input_dir in config/evaluation.yml of the code repository to this folder and run python eval.py --cfg config/evaluation.yml --checkpoint DTE-aerial-model/DTE_aerial_model.safetensors.

Licensing

All 525 patches are licensed CC BY 4.0. You may share and adapt them for any purpose, including commercially, with attribution. When you use the data, please cite the paper and also attribute the original contributors: contributors and, where available, citation_doi give the provenance of every site.

Intended use and limitations

• Intended use: standardised evaluation of tree-cover and tree-mortality segmentation models across biomes and spatial resolutions (5, 10 and 20 cm), including site-level quantities such as the mortality fraction.
• Out of scope: surveillance, property-related or other applications targeting people or private property.
• Limitations:
  • With 25 sites, the benchmark is temperate-heavy and per-biome results (in particular for boreal and montane forests, 4 sites) are subject to site-level uncertainty.
  • The mortality class is based on visible symptoms and can be confused with phenology, illumination, shadow or species-specific crown colour; labels are not validated against field observations.
  • The paper describes these limitations in detail.

Citation

If you use this dataset, please cite:

@misc{sharma2026deadtreesearthaerialmultiresolutionaerialimage,
      title={deadtrees.earth-aerial: A Multi-Resolution Aerial Image Dataset for Tree Cover and Mortality Detection},
      author={Ayushi Sharma and Clemens Mosig and Lukas Drees and Salim Soltani and Janusch Vajna-Jehle and Aaron Sheppard and Belqis Ahmadi and Jonathan Schmid and Paul Neumeier and Nathan Jacobs and Jan Dirk Wegner and Teja Kattenborn},
      year={2026},
      eprint={2605.19605},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      doi={10.48550/arXiv.2605.19605},
      url={https://arxiv.org/abs/2605.19605},
}

Please also credit the data contributors listed in the metadata.