DTE-aerial-model

Semantic segmentation model for tree cover and tree mortality in aerial RGB imagery (5–20 cm), from the paper deadtrees.earth-aerial: A Multi-Resolution Aerial Image Dataset for Tree Cover and Mortality Detection (Sharma et al., 2026).

📄 Paper: arXiv:2605.19605 · 💻 Code: GeoSense-Freiburg/DTE-aerial · 🌍 Project page: DTE-aerial

Model

• Architecture: SegFormer with a MiT-B3 encoder.

• Input: RGB patches of 1024 × 1024 pixels.

• Output: 3 classes:

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

• Training data: DTE-aerial-train, 311K patches from 1,959 orthophotos at 2.5, 5, 10 and 20 cm. It is the union of the CC BY and CC BY-NC-SA repositories. The validation split (35K patches) was used for checkpoint selection.

Results on DTE-aerial-bench

F1 per class, averaged over the 25 benchmark sites:

Tree cover	Mortality
0.89	0.60

The paper reports the mean over three seeds, with per-biome and per-resolution results.

Usage

from huggingface_hub import hf_hub_download

ckpt = hf_hub_download("ayushi3536/DTE-aerial-model", "DTE_aerial_model.safetensors")

Evaluate on DTE-aerial-bench with the code repository:

python eval.py --cfg config/evaluation.yml --checkpoint <path to DTE_aerial_model.safetensors>

Intended use and limitations

• Intended use: research on tree-cover and tree-mortality mapping from drone and aircraft imagery.
• Limitations:
  • The mortality class reflects visible symptoms in RGB imagery and can be confused with phenology, illumination, shadow or species-specific crown colour.
  • Validate locally before any operational use.

Citation

If you use this model, 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},
}