isp-uv-es/treesatai-taco

TreeSatAI Benchmark Archive

This is a repackaging, not a new dataset. It is TreeSatAI Benchmark Archive by Ahlswede et al., TU Berlin / Northwest German Forest Research Institute, converted to TACO. Pixel values and labels are kept as released except where the description below says otherwise. All credit belongs to the original authors: if you use it, please cite them and follow their licence.

original dataset · paper · licence: CC-BY-4.0

Repackaged into TACO by the Image and Signal Processing Group (ISP), Universitat de València, within the ELLIOT project.

Citation

Please cite the original work:

@article{ahlswede2023treesatai,
  title   = {TreeSatAI Benchmark Archive: a multi-sensor, multi-label dataset for tree species classification in remote sensing},
  author  = {Ahlswede, Steve and Schulz, Christian and Gava, Christiano and Helber, Patrick and Bischke, Benjamin and F{\"o}rster, Michael and Arias, Florencia and Hees, J{\"o}rn and Demir, Beg{\"u}m and Kleinschmit, Birgit},
  journal = {Earth System Science Data},
  volume  = {15},
  number  = {2},
  pages   = {681--695},
  year    = {2023},
  doi     = {10.5194/essd-15-681-2023}
}

About the data

50381 co-located triplets over Lower Saxony, Germany, built from the torchgeo mirror:

• A 20 cm colour-infrared aerial patch, band order NIR, G, B, R.
• A Sentinel-1 patch: VV, VH, and VV/VH as a quotient of the dB values.
• A 12-band Sentinel-2 L2A patch.

50,381 samples · splits: test 5,044 · train 40,796 · validation 4,541 · tasks: classification

Packaged as TACO v3.

Full description

Each sensor appears at both the labelled 60 m footprint and a 200 m context footprint; the 200 m footprint is Sentinel only, as there is no 200 m aerial.

Target. The area fraction of each of 15 tree genera, stored verbatim as a fixed-width vector. The fractions do not sum to 1 in the release (0.005 to 2.0 per sample) and are not renormalised. The multi-hot target is fraction > 0, and the benchmark's usual 0.07 threshold is recoverable from the fractions.

Projections. The aerial patch is EPSG:25832 and the Sentinel patches EPSG:32632, as delivered.

Splits. The official partition: 45337 train, 5044 test, no val. validation (4,541) is carved out of the release's train, and split_original keeps the release's own value.

Getting started

git clone https://github.com/OscarPellicer/taco
pip install -e "taco/python[ml]"     # builds the native reader: needs CMake and Ninja

Read it straight from the Hub:

import os
from huggingface_hub import hf_hub_download, snapshot_download
from taco.ml import Dataset, plot_sample

path = hf_hub_download("isp-uv-es/treesatai-taco", "treesatai.zip", repo_type="dataset")
ds = Dataset(path)
plot_sample(ds[0])

or from a local copy:

ds = Dataset("treesatai.zip")
sample = ds[0]                      # {slot name: SlotValue}, arrays decoded
sample["aerial"].array.shape

Metadata without decoding anything:

import taco
taco.read("treesatai.zip")              # one Arrow table, levels joined

Samples

What a sample contains

role	slot	holds	modality	detail
input	aerial	raster	optical	4 band(s), requantised
input	s1	raster	sar	3 band(s), unit dB, physical
input	s2	raster	multispectral	12 band(s), unit 1, scaled
input	s1_200m	raster	sar	3 band(s), unit dB, physical
input	s2_200m	raster	multispectral	12 band(s), unit 1, scaled
target	genera	class_multihot		15 classes

Licence

CC-BY-4.0

Providers: Ahlswede et al., TU Berlin / Northwest German Forest Research Institute

Acknowledgements

TACO was designed by César Aybar and is specified at https://asterisk.coop/taco/spec/.

Built by Oscar Pellicer within the ELLIOT project at the Image and Signal Processing Group (ISP), Universitat de València.