isp-uv-es/bigearthnet-v2-txt-taco

BigEarthNet v2 with BigEarthNet.txt

This is a repackaging, not a new dataset. It is BigEarthNet v2 with BigEarthNet.txt by BigEarthNet v2 / TU Berlin and partners, BigEarthNet.txt / BIFOLD and TU Berlin, ESA Copernicus, 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: CDLA-Permissive-1.0

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{Herzog2026BigEarthNetTXT,
  title   = {BigEarthNet.txt: A Large-Scale Multi-Sensor Image-Text Dataset and Benchmark for Earth Observation},
  author  = {Herzog, Johann-Ludwig and Adler, Mathis J{\"u}rgen and Hackel, Leonard and Shu, Yan and Zavras, Angelos and Papoutsis, Ioannis and Rota, Paolo and Demir, Beg{\"u}m},
  journal = {arXiv preprint arXiv:2603.29630},
  year    = {2026}
}

About the data

464,044 co-registered Sentinel-1/Sentinel-2 image pairs from BigEarthNet v2 with 9,553,962 BigEarthNet.txt textual annotations.

464,044 samples · splits: test 116,835 · train 229,114 · validation 118,095 · tasks: image-captioning, visual-question-answering, visual-grounding

Packaged as TACO v3.

Full description

Layout. Each image pair occurs exactly once; annotation ids, prompts, responses, types and categories are aligned parallel sequences, which avoids the approximately 21-fold pixel duplication of an instruction-per-sample layout. BigEarthNet's 19-class land-cover labels are retained as a multi-hot target.

Imagery. S1 is a 10 m VV/VH dB COG. S2 L2A is kept on its native 10 m, 20 m and 60 m grids in three COGs, without conversion-time resampling.

Splits. BigEarthNet.txt train, validation and test splits are preserved; its manually verified bench split (1,082 pairs) is written as test, and split_original keeps the release's own value.

Getting started

git clone --recursive https://github.com/OscarPellicer/taco
pip install -e "taco/python[ml]"     # builds the reader: C++23, CMake, Ninja, pkg-config, libcurl >= 7.83, OpenSSL >= 3

Read it straight from the Hub:

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

path = snapshot_download("isp-uv-es/bigearthnet-v2-txt-taco", repo_type="dataset")
ds = Dataset(path)                  # the repository is the container: parts + .tacocat
plot_sample(ds[0])

or from a local copy:

ds = Dataset("bigearthnet-v2-txt")
sample = ds[0]                      # {slot name: SlotValue}, arrays decoded
sample["s1"].array.shape

Metadata without decoding anything:

import taco
taco.read("bigearthnet-v2-txt/.tacocat")              # one Arrow table, levels joined

Samples

What a sample contains

role	slot	holds	modality	detail
input	s1	raster	sar	2 band(s), unit dB, physical
input	s2_10m	raster	optical	4 band(s), unit 1, scaled
input	s2_20m	raster	multispectral	6 band(s), unit 1, scaled
input	s2_60m	raster	optical	2 band(s), unit 1, scaled
input	prompts	text_sequence	text
target	responses	text_sequence	text
target	land_cover	class_multihot		19 classes

Licence

CDLA-Permissive-1.0

Terms of use:

• The CDLA-Permissive-1.0 licence text must accompany redistributed copies, and recipients are bound by it.
• Existing attribution notices must be kept, and modified files must be marked as modified (CDLA-Permissive-1.0).
• Credit the original authors and the data sources listed below.

Required credits:

• BigEarthNet v2.0 (reBEN), Remote Sensing Image Analysis Group, Technische Universität Berlin
• BigEarthNet.txt, BIFOLD and Technische Universität Berlin
• Contains modified Copernicus Sentinel data [2017-2018]

Providers: BigEarthNet v2 / TU Berlin and partners, BigEarthNet.txt / BIFOLD and TU Berlin, ESA Copernicus

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.